Sensing Direction of Human Motion using Single-Input-Single-Output (SISO) Channel Model and Neural Networks

Object detection Through-the-Walls enables localization and identification of hidden objects behind the walls. While numerous studies have exploited Channel State Information of Multiple Input Multiple Output (MIMO) WiFi and radar devices in association with Artificial Intelligence based algorithms (AI) to detect and localize objects behind walls, this study proposes a novel non-invasive Through-the-Walls human motion direction prediction system based on a Single-Input-Single-Output (SISO) communication channel model and Shallow Neural Network (SNN). The motion direction prediction accuracy of SNN is highlighted against the other types of Machine Learning (ML) models. The comparative analysis of models in this study shows that unique human movement patterns, superimposed on received pilot radio signal, can be classified precisely by SNN, with an accuracy of approximately 89.13% compared to the other ML based models. The results of this study would guide scholars, active in developing human motion recognition systems, intrusion detection systems, or Well-being and healthcare systems, and in processes that innovate and improve processing techniques for monitoring and control.

R ECENT advancements in wireless and communication technology are driving researchers towards developing systems that exploit Channel State Information (CSI) of radio communication devices for localization and pattern recognition based applications. Several research studies focussed on areas of human motion detection, gesture detection, and facial recognition systems have been proposed [1]- [5]. Wireless communication devices available in the markets at present are grounded on Software Defined Radio (SDR) sub-units that enable support to multi radio subcarrier transmission, thereby enabling communication of different types of data. These SDR sub-units facilitate the acquisition of CSI in wireless communication devices, and the collected CSI data is then explored to uncover hidden human motion patterns. Some Commercially Available Off-The-Shelf (COTS) devices even output CSI directly via their inbuilt radio subsystems [6]. Likewise, WiFi devices available today, also output CSI and outperform against the other human motion sensing systems. However, Human Activity Recognition (HAR) processes using WiFi based solutions prerequisite acquisition of CSI for accurate estimation and prediction of human activities [7], [8].
A review of recent studies unveils that CSI signals show randomized behavioral patterns. This randomized behavior prevails due to signal reflections in the multipath signal propagation and mainly serves as a source of distortion in the received signals [9]- [11]. The statistical random variations observed in the CSI data primarily exist due to the Non Line of sight (NLOS) propagation characteristics. Any errors introduced by the reflections, scattering, and diffractions of signals are mainly considered as obstacles in signal propagation [12]. Random patterns in CSI exist due to the dynamic motion of objects in the near vicinity of the transmission source that inevitably influence the elementary behaviour of received signals. Although CSI employs physical layer properties to differentiate between the amplitude and phase characteristics whilst being sensitive to the dynamics of the environment, the multipath effect is observed to be constructive, when it comes to its applications in localization systems that are based on fingerprinting [13]. While the amplitude of CSI could be a stable metric, the added external noise to the signal from the system and the environment distorts the signal considerably [14]. The distance between a transmitter and receiver, and radio reflection, refraction, diffraction, absorption, polarization, and scattering, termed as multipath effects, considerably attenuate the amplitude and introduce phase shift in CSI signals [15]. Apparently, the complex nature of acquired CSI data revealing random probability distribution, requires efficient processing mechanisms of noise removal and signal decomposition, before some meaningful information can be extracted from it, or in particular signal features that are imperative in the pattern recognition processes and applications. To identify the behavioral characteristics or patterns in any given CSI dataset, successive remarkable attempts have been made so far by the research communities, uncovering the potential of CSI for human motion recognition. Various approaches of signal processing have been used so far, for instance, Discrete Wavelet Transform (DWT) is used for noise removal and Linear Discriminant Analysis (LDA) for feature engineering by [16], feature extraction using Principal Component Analysis (PCA) by [9] [10] [17] , the CLEAN algorithm by [18], Doppler spectrum by [19] and even scale and time shift projections by [16]. Signal processing methods as such prerequisite high speed processing elements, and in addition to being process intensive, running these algorithms on tiny or resource constrained hardware devices may not be an optimal solution. Also, the primitives -rising edges, falling edges, and pauses defined at the primitive layer in [16] attempt to address issues as mentioned, however, in a generic sense. Therefore, the need to reduce the complexity of CSI processing and application of innovative filtering techniques merely result in an increase in the number of processing elements as well as leads to additional system costs.
Whereas numerous research studies based on WiFi have been used for developing applications for localization and pattern recognition, the resulting systems are highly accurate in recognizing human activities, in particular human motion, and mostly employ either machine learning or deep learning algorithms for pattern recognition or classification. CSI can reveal patterns in the data, which come from different domains, for instance, motion sensing systems, object recognition systems, and pose estimation. However, it requires processing, modelling, and training specific to the each domain. WiFi sensing performance relies mostly on noises and outliers that exist in the collected raw CSI data [9], and the CSI data acquisition with off-the-shelf WiFi cards, may not always be a readily available option, since CSI data acquisition features are unavailable in some WiFI cards. In addition, there are concerns relating to processing complexity of CSI signals, portability of WiFi devices, adaptability to current networks, unreliability to due fading channel propagation, lower precision due to random nature of CSI, and inefficient system designs due to complex signal processing architectures. The radio subunits provide CSI that often requires hardware and software signal processing mechanisms, and subsequently the developed systems adumbrate need of complex signal processing circuitry.
Although the motion detection accuracy of Wifi is high compared to the other radio devices, CSI data is dependent on the availability of multiple WiFi channels for consistent monitoring of activities under consideration. If one of the channel experiences interruption, the detected activity, for example human motion prediction, accuracy may decrease significantly. This leads us to propose a single communication channel only for activity recognition purposes instead of employing multiple radio communication channels. The proposed method alleviates the level of processing, when compared to the processing used for multiple CSI subcarriers, and implements a Single Input Single Output (SISO) channel model for activity recognition, in this study human motion direction prediction. Using SISO, there exists a single transceiver pair, modelling the system to enable detection of human motion. This study presents a testbed for modelling SISO communication system model, and an experimental setup to evaluate the prediction accuracy of the single channel based human motion direction prediction system. A human surveillance system is presented that predicts the direction of human motion under the realm of detection through the walls. For human motion classification, this study employs machine learning models to identify different human motion patterns of four different humans employed in this study. The motion patterns were collected in a controlled lab environment. Therefore, the study extends scope of the current research in human activity or motion recognition systems, and proposes the development of an efficient human motion direction identification system. In addition to demonstrating the SISO model based human motion direction prediction, the study also highlights the practical implementation of SDRs for human motion surveillance and activity monitoring.
In this paper, a Shallow Neural network [20] based system model is proposed that acquires and processes received Radio Signal Strength (RSS) data from a SISO communication channel to identify the direction of human motion, through the walls. A testbed is setup and configured in a real time scenario to validate the proposed model, and estimate the detection accuracy of the proposed system model. The proposed system processes unique human motion signatures, superimposed on the RSS, and then attempts to unveil the direction of human motion from RSS using various machine learning models. In abreast to showcasing predictions on human motion direction, which CSI-based and other HAR studies have never focused before, the study shows direction prediction accuracies resulted by different machine learning models for comparative analysis, enabling us to determine the best prediction model. The study hypothesis bases on the assumption that Shallow Neural Networks (SNNs) can provide the optimal human motion direction prediction accuracy (SNNs) by classifying the human motion patterns, acquired from a continuous time pilot carrier signal transmission. Therefore, SNN prediction model is implemented in LabVIEW to validate the prediction accuracy of the devised human motion direction recognition model, in addition to training various other machine learning models for comparison. The study results show that SNNs show higher prediction accuracies than other machine learning models, whilst presenting a light network model compared to the deep learning frameworks with multiple hidden layers, typically implemented in human motion recognition systems using WiFi.

A. CONTRIBUTION
This study proposes a system model for human motion detection through the walls, and employs just a single carrier frequency of 2.4GHz, enabling detection human motion and to trace the direction of human motion using a SISO based channel model. The main contributions of this study are listed here under: • The study proposes a testbed to validate the proofof-concept i.e., to identify the direction of human motion in indoor environments using SNN. The proposed testbed implements a Single Input Single Output (SISO) topology. • The study provides a system model that enables human motion direction recognition using a single communication channel. SNN is used to predict the six different directional motion patterns of four participants. • The study provides a novel method to identify the direction of human motion based on Software Defined Radios (SDRs). A case study of human motion surveillance is described to highlight the applications of SDRs in Through the walls human motion sensing and detection. • The study evaluates accuracy of various machine learning models on the collected human motion dataset. A comparative analysis of SNN with other machine learning models is also given to determine the best prediction model. • The study provides statistical validation of the two proposed models to determine the efficient model for predicting human motion direction.

B. PROBLEM STATEMENT
From the literature review, we conclude that most of the studies have addressed challenges pertaining to human motion recognition. WiFi and radar sensing technology in association with machine learning and Deep Neural Networks (DNNs) have been widely investigated and applied in numerous applications. The systematic literature review reveals that prolific attempts have been made to address challenges pertaining to human motion recognition. However, there is no research study till date that addresses the challenge of detecting the direction of human motion from the RSS of a pilot signal, more specifically with Universal Software Radio Peripherals (USRPs) [21] using omni-directional antennas. Using our proposed method, we the authors of this study claim to be the first to address this challenge, and present a novel method that employs a single Continuous Wave (CW) -a pilot radio subcarrier, and the Shallow neural network for predicting the direction of human motion in a controlled environment.

C. STUDY OUTLINE
The study is organized according to the following sections. Section-II provide a detailed literature review of related studies. Section III discusses a theoretical perspective of VOLUME 12, 2021 the study, while the Section-IV provides the experimental design method. Section-V, method of data collection acquisition using SDRs. Section-VI provides the SNN based machine learning solution to motion pattern analysis and classification, whereas Section-VII discusses study results and compares resulting ML models. Finally in the Section-X, conclusions are presented and suggestions are made for future research in Section

II. RELATED WORK
With the current advancements in wireless communication and sensing technologies, their applications to HAR systems are gaining wider acceptance at global level. In recent years, wireless technologies have played a vital role in the development of HAR systems. HAR is intuitively the main driving technology in human-computer interaction and activity analysis, and is nowadays the promising area of research in the field of artificial intelligence. A number of research studies have focused on issues concerning HAR [16]. HAR methods used in the current ongoing research are broadly divided into three main categories -image processing, wearable sensing devices, and wireless radio signal processing. In computer vision, using image processing techniques on modelled human motion and analyzing effects in real life settings, are currently driving the development of many industrial and commercial applications [17] [22]. For example, real-time Human Motion Recognition (HMR) [18], systems using radar images responsible for distinguishing human motions [19], intelligent surveillance platforms [23] and many others. In addition to the applied image processing techniques, HMR systems base on various types of sensors such as Micro-electromechanical System (MEMS) Sensors [19], motion sensor using U-Net [24], smartphone-based inertial sensors [25], magnetic field sensors [26], Continuous-Wave Radar Sensors [27] and velocity sensor array [28]. Currently, prolific sensor based HMR systems are widely available, however, research has shown that signals acquired from WiFi devices can help in recognizing human motions and positions [29]. Several WiFi based solutions such as WiAct [30], WiRoI [31], WiDrive [32], WiFi Motion Detection [33], WiTT [23] and many others have be already proposed and implemented.
These WiFi based studies devise solutions based on CSI, which the SDR devices acquire by analyzing the multiple radio sub-carriers. However, preprocessing the CSI is a difficult task on hand and poses several critical challenges whilst deriving human activity information from it. Typically, the quantitative measure of CSI phase reveals that the sampling time mismatch leads to random phase shifts in the signals [34]. Consequently, the devices having same hardware and software architectures, including physical properties may show varying results, even if the devices receive the same type of activity data, as input. The lower sampling rate of processing units within the SDRs results in processing delays, eventually limiting their application in real time systems. While CSI subcarriers are influenced by the dynamic nature of a communication channel and intensity level of noise, extracted CSI data are input to the signal processing units for eliminating the unwanted noise, applying signal transforms, and/or extracting actual motion signals. Next, pre-processed CSI data are input to algorithms that employ modelling, learning, or hybrid algorithmic techniques to deliver the output, targeted at different WiFi based sensing applications. This requires cleansing, processing and analysis of each CSI subcarrier, which in turn demand efficient hardware resources. Moreover, the sensing environments may elevate additional processing demands due to dynamic background of the monitoring environments. The background object elimination algorithms, for instance by [34], may eliminate such effects and estimate likelihood criteria to isolate actual targets from the other environmental entities, however the present estimation criteria demand precise estimations of the noisy descriptor variables that mostly rely on ambient parameters [35].
As far as Hardware devices are concerned, WiFi network cards, for instance Intel 5300 network card, are installed and fixed inside notebooks or desktop computers. The fixed orientation limits their application in testbeds, where user defined orientation is desired to acquire CSI statistics. Similarly, Multiple-Input-Multiple-Output (MIMO) systems [17], [18] employ multiple transmitters and receivers, include multiple antennas for reception and transmission [2], [36]. Increase in hardware and advanced software requirements, extends the level of costs of MIMO devices compared to SISO devices that rely on simple hardware, software and antenna types. MIMO systems adumbrate higher resource requirements including hardware complexity when compared to SISO systems. Individual RF units are prerequisite to process radio signals of each antenna, and MIMO topology employs advanced mathematical or signal processing algorithms that necessitate powerful Digital Signal Processors (DSPs) to execute the developed algorithms. Apparently, increase in the requirement of hardware resources leads to increase in the power requirements. Eventually, complex and computationally intensive signal processing algorithms lead to high battery consumption, thereby draining out batteries faster and decreasing life time of batteries used to drive MIMO devices.
The proliferation of hardware elements as such increases infrastructure and deployment costs, and also limits their applicability in the processes leading to the development of low cost solutions that enable HAR. Some studies also affirm that adding additional subcarrier frequencies in the WiFi recognition process does not help in improving the human recognition accuracy. The study by [10] and [17] reports that CSI acquired with WiFi operating at 5GHz of frequency with 114 and 132 subcarriers, shows 94.0% and 94.1% of True Positive Rate (TPR) only in detection accuracies, respectively. This implies that recognition accuracy or in other words the TPR is independent of the number of CSI subcarriers, and also proves that employing additional radio subcarriers unlikely shows an increase in the level of human motion recognition accuracy. Hence, HAR systems that implement a number of subcarrier frequencies, demonstrate a limit in the level of accuracy of human motion detection with multi subcarrier channel based systems.
On the other hand, HMR systems grounded on radar sensing technology, immensely eliminate the need of deploying numerous sensors, and are mostly preferred over other sensor based systems. HMR systems based on radar sensor systems have been extensively used in various applications, including surveillance [37] [38], activities concerning search [39] [40] and rescue [41], smart environments [42] [43], and in ambient assisted living [44] [45] [46]. Research studies, for instance motion capturing system [47], contactless UWB Radar [48], Magnetic Induction [49] exclusively demonstrate the featured capabilities of non-intrusive motion recognition systems. Among many wireless sensing technologies, radar sensing has emerged as an efficient solution that relatively focuses on areas such as commerce, defense, and security. Although widely used in military and defense services, radar sensing systems are also implemented in Human-Computer Interaction (HCI) in medical applications, in automotive fields and smart sensing systems [50]. Radar technology employing frequencies that range from few GHz to hundreds of GHz (mmWave) is used for medium or very close range sensing, and presents an environmentally robust sensing systems. Recent applications of radar include gesture recognition [51] [52], monitoring patients [53], monitoring movement of Ataxia Patients [54], detection and tracking of the human body [55] and various other applications. Amalgamation of radar sensing devices and Machine learning, particularly ones that work in association with Deep Neural Networks (DNNs), have shown promising results in applications, such as counting number people in a room [56], human identification [57], and even in unknown human classifications [58].
Typical experimentation with radar sensing systems, prerequisite distinguishing factors to be taken into account. The characteristics, such as technology, modulation technique, frequency, communication bandwidth, antenna type, including the processing capability of algorithms, serve as major factors that influence the overall application area of radar based sensing and tracking systems. From our literature review of recent studies, we divide radar sensing technology into four dimensions, a) technology b) frequency c) antenna types d) implemented modulation techniques and e) Data representation. We consider that these five dimensions are the key factors that affect the overall performance of a human radar based sensing and tracking system, as numerous studies have relied on these parameters so as to achieve higher performance and level of sensing accuracy. The Table-1 lists these five main parameters with corresponding definitions.
Most of the human motion recognition applications implement artificial intelligence based solutions to unveil the hidden characteristics in the data, whereby the data acquired from radar sensors is typically processed off-line than in actual real time [30]- [34]. Real time human motion recognition applications, in general, exclusively demand lower complexity of signal processing hardware, so as to enable fast and efficient human motion data processing, also require hardware circuitry to ensure active real time response. Nonetheless, to implement the DNN algorithms on a low level hardware architectures is a challenge in itself because DNNs exhibit highly complex architectures that demand high end processing elements employing complex signal processing techniques.

A. RESEARCH MOTIVATION
In recent years, Object Detection Through-The-Walls (ODTTW), has been a major source of attraction to researchers. Due to its wide range of capabilities, ODTTW enables support for security and military defence applications. In a typical ODTTW application, a radar transmitter sends out electromagnetic pulses through the wall to sense the presence of an object, located on the other side of the wall. And the returned signal reflections from a wall and target, are processed to detect the presence of an object. ODTTW does not need to establish any physical contact with objects that lie behind walls, so it allows passive detection of hidden objects. Sensing systems, as such, are categorized as passive sensing systems, exhibit features such as context awareness in close proximity, or monitoring an object that is located on either side of a wall. The method of transmission of radar signals, to detect objects that are located behind walls, finds application in tasks such as surveillance, search and rescue operation, enforcement of law, and other operations that save people's lives [59], [60]. There are other methods as well that help in sensing and localization of objects covered by walls. Sensing methods, as such, employ 60-70 GHz band of spectrum [61]- [64], to passively sense the presence of objects hidden behind the walls. On the other hand, Through Wall Radar Imaging (TWRI) methods are used to reconstruct static targets, as well as to present the information on the monitored scene of a hidden object. Nevertheless, the parameters defining the characteristics of a wall, and others that define the position there, are necessary, VOLUME 12, 2021 prior conditions to the process rebuilding a scene, although wall parameters are usually unknown in real scenarios [65].
One of the major problems, underlying in environments as such, is the unknown width of a wall or barrier; an essential parameter, which is often neglected in the process of imaging or modelling the background space of a wall. These types of barriers can completely block the transmission of radar signals, if walls are highly dense in their physical structure. However, if the frequencies used for transmission are low, then walls can allow penetration of signals with low frequencies. TWRI studies have mostly focused on stationary objects, but some studies have attempted to model, and investigate scenarios, wherein attempts have been made to localize moving objects [66]. To sense non-moving objects that are covered by walls, two preconditions need to be satisfied. The first condition is to remove clutter from walls, which result in multiple reflections of a radar signal due to the irregular nature of objects on the wall, and the second is to build an image which models the reconstructed scene of the background [62], [67]. Nonetheless, if the strengths of reflected radar signals, are stronger, due to high density of walls, than the reflected radar signals that emerge, due to a stationary object, lying behind a wall, then the process of detecting an object, becomes very difficult because of clutters that lie in between radar, and a target object [68]. On the other hand, the frequencies that range below the radar frequency band can easily penetrate through walls, and thus enable the detection of objects located behind walls, possibly. Therefore, the frequencies that lie in the range of 2GHz to 5GHz are most suitable for human motion detection.

III. SYSTEM MODEL
Our study design bases on the system model that makes use of SDRs that allow customization at both the hardware and software level. The proposed system model comprises of two SDRs, with one acting as a transmitter and other as receiver. SDRs allow instant software update to be made to the radio hardware, and this level of customization is achieved through an application programming interface, such as LabVIEW. The LabVIEW application runs on a computer, which sends out control and data commands to the SDR via a USB cable. The software updates made to reconfigure the SDR hardware, make it more easier readjust the system design of SDRs, and SDRs can be programmed, or reprogrammed instantly to meet the local and international commercial guidelines of RF transmission. SDRs are portable devices, allow reconfiguration of hardware parameters such as controlling the IQ rate, bandwidth, data sampling, filtering, and radio internal block circuitry, which include implementation of coding and decoding techniques. The two programmable radio devices are placed 6 meters apart in two lab rooms. SDRs are connected to two desktop computers, with one computer connected to the projector that displays the application interface on the sidewall to indicate to the participant, when to start and stop human walking activity.
To determine the necessary settings of proposed testbed system and their control configurations, this study employed an experimental setup that determines these necessary parameters using the Randomized Control Trial (RCT) method. A topology based on NI-USRPs was designed to collect data of human movements in various directions. The sections next provide the theoretical background of the proposed model, procedure of the experiment design and setup, data collection method, and application of proposed machine learning model to the collected data.

A. BACKGROUND OF SIGNAL PROPAGATION
A transmitted signal propagates from the transmitter to the receiver at a particular frequency of interest. Wireless signals travelling the through space may arrive at the receiver via various paths, and each path typically delays and attenuates the signal. The signal transmission by wireless devices, may encounter many obstacles in communication medium, and the CSI acquired from a radio subsystem provides the Channel Frequency Response (CFR) that describes the propagation behavior of signal through the communication medium. The CFR outsourced by CSI, characterizes the signal transmission from transmitter to the receiver and is expressed by eq (1): where b q (t) and φ q represents the amplitude attenuation and the propagation delay parameter, respectively. The parameter k represents the frequency of the transmitted carrier or continuous wave. It should be observed that the relocation and movement of the transmitter, receiver, and the ambient entities and humans influence the amplitude H and phase ∠H of the CSI. This implies that the CSI embeds the effects of objects in the nearby vicinity of the transmitter and receiver, and the observed CSI basically reveals or characterizes the sensed environment, using mathematical modelling. The same rationale governs the implementation of CSI based WiFi sensing systems. Signal propagation in indoor environments usually follows a similar behaviour showed in the external environments, and the transmitted signal may travel through various path to arrive at the receiver. While the receiver receives the actual transmitted signal, some alias forms arising out of multi path propagation of the actual transmitted signal may also arrive at the receiver [10]. The voltage level (S) of pilot carrier signal available at the receiver is expressed by eq(2): where S p represents the amplitude of the p th multipath component, θ p denotes the phase of the p th multipath component and N is the total number of multipath components. S represents the overall signal strength received at the receiver. The Radio Signal Strength (RSS) [10] of pilot measured in terms of decibels scale (dB) is expressed by eq (3): The level of received signal strength depends on the signal propagation distance, wavelength, and time. Due to dynamics of the environment, the strength of pilot signal varies considerably,in fact even with minor multipath components dynamics. A study by [23] reports that 5 dB of change in the RSSI has been observed over a minute, on an immobile receiver, placed in a controlled lab environment. This implies that the having fixed location for transmitter and receiver is more stable compared to having a mobile setup. Hence, this study makes use of an immobile setup for both the transmitter and receiver, so as to ensure RSSI variations are reduced to minimum possible values.

1) SISO Channel Model and Feature Extraction
Using a conventional channel estimation method based on standard-defined pilot sub-carrier arrangement [69], the MIMO channel response is monitored in the frequency domain over k pilot sub-carriers λ ∈ F = (λ 1 , λ 2 , · · ·, λ k ). This study develops features based on SISO channel data to identify directional motion patterns. A moving persons' body movements modify the statistical properties of the RSS in the frequency domain, such as the strength moments can evaluate as an average over the time in a given frequency sample. For example, attenuation and fading of RSS due to the body of a person are referenced by the statistical properties, like mean and variance. In the following, the study considers statistics of the overall SISO power frequency profile. The frequency-domain CSI strength vector for link u at time t is defined by eq (27): Where CSI S u,t is the strength observed at time t ∈ T on sub-carrier λ ∈ F and link u ∈ U. VOLUME 12, 2021 The complex baseband channel response observed on frequency λ ∈ F, over the link λ = λ (β t , β r ) between the transmitting antenna β t = (1, · · ·, Z t ) and receiving antenna β r = (1, · · ·, Z r ) at time t ∈ T is denoted as H λ,u,t .
While employing SISO channel model, this study employs and extracts a single channel strength sample only S λ1,1,t from the frequency vector S u,t = S λ1,u,t · · · S λ k ,u,t T . Where λ 1 represents a single channel pilot radio carrier, and u = 1 represents a single communication link.
Six distinctive statistical indicators model the features of corresponding probability function S (λ,1,t) ∼ P r[S f,1,t ] , and are defined in the Section-(VI-C).

IV. EXPERIMENTAL SETUP
The experimental method applied to the devised system model (see Fig.1) employs a SISO channel that comprises of one Transmitter (TX) and Receiver (RX) as a pair for the initial testing and setup. The transceiver pair, both have small antennas with omnidirectional radiation pattern and operate in commercial ISM 2.4 GHz band. Antenna specifications are -VERT2450 Dual Band 2.4 to 2.48 GHz and 4.9 to 5.9 GHz omni-directional vertical antenna, at 3dBi Gain. To acquire data, the study sets up six directions for observing six directional human motion patterns and selects four participants for data collection. The environment for the experiment included two adjacent rooms, separated by a common path of a corridor between the two. A desktop computer along with the SDR was setup in each room to devise the proposed testbed for the experiment. The corridor environment was set out to have no obstacles in it's path and provided participants a free space to walk through. The motion patterns of each participant were monitored inside the receiver room, however participants were apprised when to start and stop the motion, i.e. walking, using two indicators mounted on the left wall of the corridor.

1) Instrument Setup
Two NI-USRP 2921 [25] SDRs were employed with TX and RX applications, which were developed in LabVIEW [24]. The TX application allowed controlled pilot signal transmission at a frequency of 50KHz, and the RX application was used to estimate the Power Spectral Density (PSD) of pilot signal. TX and RX devices were placed at a distance of 6 meters apart from each other, and the gain of both the devices was set to default i.e. to zero, for the purpose of initial testbed setup. different reference positions in this scenario, while placing the NI-USRPs 6m apart from the corridor path (see Fig.1). In each reference direction, we collected three datasets consisting of RSS values, for each move of each participant.
• Initial Testbed Setup and RSS Stabilization: To test the validity of proposed testbed, first orientation tests were performed to ensure there is minimum level of variation in RSS. Table 2 provides measured values of the power level of RSS on a dB scale, while selection is made to the change the orientation of the NI-USRP TX and RX. To ensure both devices were placed in a optimal orientation, the PSD of RSS enabled pilot signal power measurement, whilst having no objects in between the Line Of Sight (LOS) path. Apparently, a controlled environment was desired and consequently for the RCT an empty corridor space with the presence of any human, was chosen for the study. After the selection of a place with minimal environmental variations, the nominal variation of the pilot signal stabilized. The (see Fig.2) depicts the variations in RSS, which were observed in the empty space at a room temperature of 24 o c, and the Fig. 3 and Fig. 4 present the observations made to TX and RX alignment patterns, and individual gain of TX and RX used during the initial testbed setup. The optimal configuration setup ended in the setup with TX and RX gain set to 50, whilst placing both the devices at 0 o with LOS as reference axis. The TX and RX are placed 2.2 meters above the ground level so that the chosen average height matches with the average height of the participants. Whereas placing the NI-USRPs at 0 o with reference to the LOS showed maximum strength of pilot signal at the receiver, RSS of the pilot decreased significantly to low levels at other orientation values. For the proposed system model, a direct LOS communication link is established between the TX and RX, ensuring high RSS power level at the receiver (see Fig. 3).

2) RF Pilot Signal Processor Application
We build an automated RF Pilot Signal Processor (RFPSP) application in LabVIEW to record and collect RSS data.

V. DATA COLLECTION AND SYNTHESIS
Human motion data was collected in phased manner, and we employed an automated approach to our data collection process. Firstly, the NI-USRPs were setup to provide the motion data in the form of files, which were next subject to processing via the developed LabVIEW application. After processing, the LabVIEW generated prediction results, and for comparative analysis we employed similar datasets to test various other machine learning models in MATLAB.
The different phases employed in the data collection process are:

A. DATA COLLECTION AND DISTRIBUTION
To enabled automated human motion data collection, we employed our RFPSP application that enabled us to collect three individual datasets for each participant. A total of 1780 samples, collected within 15 seconds duration, defined each move and for 6 directional motion patterns each dataset contained 6 × 1780 = 10680 RSS samples. Thus, three similar signatures (moves) of each move were prepared. Among the three, two of the datasets were used for training and third was reserved for testing the used SNN model. Each dataset in the set of three was further partitioned into two datasets, with one designed as Training and the other as Test dataset. For the Training dataset contained 21360 data samples, including manually assigned labels to each class of motion pattern. The Test dataset containing 10680 samples only, enabled evaluation of the trained SNN model and to figure out the prediction accuracy. Since two of the three datasets were used to train the model, and only one dataset was reserved used to test, the test set contained 25% of the overall data, whereas train dataset contained 75% of the overall data.

B. DATA PRE-PROCESSING AND INCONSISTENCY CHECK
The observed Raw RSSI measurements contained noises and outliers that could significantly reduced the quality of signal describing motion signatures. Among many methods for filtering out unwanted noise, Moving Average and Median Filters offer simple solutions and are commonly adopted methods for eliminating high frequency noise components. Moving Average filter (window size of 60) is applied on the collected raw datasets of RSS, filtered out the unwanted noises and random variations. In the Moving Average and Median filters, an average or median data sample replaces a range of data points. A sliding window and multiplying factors, used provide individual weights to the averaged data samples. All datasets containing RSS samples contained random noise components. As an example, the Fig.5 and shows the raw plot signal before Moving Average filter [70] is applied and Fig.6 shows the averaged and cleansed signal after applying the filter. The collected human motion data showed data unusual variations as a result of movements, such movement of head, leaning backwards on the wall, raising hand, raising arm. These unusual movements were observed just before the data collection ended. Abreast to unwanted motion signatures,wherein redundant data as such was removed manually by padding normal RSS values in the dataset, some datasets showed missing RSS samples. As the USB type communication was used between RX and the Desktop computer, the RX setup incurred communication delays that resulted in the loss of data samples over the USB communication interface. The missing data samples were manually added using interpolation method to ensure consistency in the collected data.
• Moving Average Filter: The acquired human motion patterns represent both the linear and non-linear data components. We decompose the acquired time series data sequence using the Moving Average Filter (MAF), a classical method of time series decomposition [70]. An typical MAF of order D is expressed by the  mathematical expression, eq (5): Where x[i] and y[i] represent the input and output set of samples, respectively. D represents the number of samples selected for averaging. Also called as recursive form of MAF, the eq(5) can be described as a difference equation represented by eq (6): where M = 2p + 1. The trend cycle is estimated by averaging the time series with k periods of time t. The filtered data contains data samples, have data value close to each other since the averaging removes randomness in the given data. In the expression given in eq(10), the detrend sequence T 2 is the difference between the trend part of time series T 1 and the original time series T .
where T is the sequence of original human motion samples; T 1 the sequence of samples representing a trend; T 2 is the sequence of sample representing detrend in the sequence.

VI. PATTERN DETECTION AND CLASSIFICATION
To identify human motion direction, patterns of which are hidden in the collected raw RSS data, a machine learning based solution is implemented to trace the motion signatures out of the relative datasets of each participant. Learning based algorithms are implemented as predictors in most of the recognition based applications, and typically implement Support Vector Machine (SVM) [72] for categorization. SVMs have been mostly implemented as predictors in many studies, and these uncover embedded patterns in linear and non-linear datasets, including classification of data points that are left over from the training dataset [73]. In addition to SVM, there also exist other types of classifiers, such as K-Nearest Neighbour (KNN), Dynamic Time Wrapping (DTW), Hidden Markov Model (HMM) or Convolutional Neural Networks (CNN). While SVMs prerequisite large dataset containing support vectors that apparently determine the function of the output, SVMs are observed to be computationally expensive; hence are not reliable and may not be applicable in real-time based solutions. Furthermore, despite knowing the fact that deep learning algorithms offer high prediction accuracy, this study leaves deep structures since structures as such require additional parameters definitions, for instance kernel size, quantity of neurons etc, whilst having many layers of hidden neuron that complicate the overall architecture of the algorithm. However, SNNs can be used on mobile devices and offer more light weight architecture than other types of classifiers. Therefore, this study implements Shallow Artificial Neural Networks as a machine learning based solution since ANNs outperform SVMs in terms of computational efficiency [58], [74], [75].

A. SHALLOW NEURAL NETWORK (SNN)
ANN algorithms mimic the neural system functions of living organisms, and typically implement fully connected computational units called neurons. ANN organizes these neurons in different layers, wherein the first layer is called as the input layer, the second containing few hidden layers, and the third layer called as is the output layer. Among the three layers, output of each hidden layer feeds the input of the next layer. This study employs a single hidden layer SNN as the machine learning solution to recognize the directional patterns in human motion data. Fig.7 shows a SNN that includes n inputs, d hidden neurons placed in a single hidden layer, and a only one output layer y. SNN model defines the input-to-output relation as a function, f : R n → R that associates the input vector x ∈ R n to a scalar based on the following relationship, eq(11): Where w l ∈ R represent vector with edge weights that associate input to hidden node lth and v l ∈ R represents edge weight connecting the hidden node lth to the output layer. The expression ξ : R → R is the activation function used for the each hidden node in the hidden layer. The ma-  trices w ∈ R d×x and v ∈ R d represent a compact notation of the weights w l /v l , eq(12): The cost function of a given Fully Connected Neural Network (FCNN) [76]. The network has d units organized in a single hidden layer, and includes the linear output unit that computes sum-of-squares error for the given set of input data expressed by eq(13):. where In the eq (14):, {ω k } Q p=1 denotes the weights of the hidden layer and {γ k } Q p=1 , the weights of output layer. The expression ψ(c, ω, x) represents the sigmoidal for Multilayer Perceptrons (MLPs) as the activation function for the hidden units.

B. FEATURE EXTRACTION
With current testbed, only a single vector containing a stream of RSS samples corresponding to each move is collected. SNN algorithm requires more features vectors for VOLUME 12, 2021 . . . . . . effectively classifying the direction of moves. Thus, additional feature vectors based on the dynamic non-periodic signal properties were formulated. The extracted feature vector space consisted of RSS variations and associated radio signal properties for each participant observed in six directions. The properties included various signal properties such as meanx, standard deviation (sd), difference d, variance σ, and signal energy over the sampling interval. The directional motion datasets (see eq (15)) for each of the four participants were collected and class labelled each move of the six associated moves as 1,2,3,4,5 and 6 respectively. Here the superscripts are indexes representing each participant, and subscripts are indexes that denote the collected data points. The input to the SNN algorithm is a matrix of the form shown expression eq (16): The subset s z 1,1 × s z u,n is a partial-matrix representing manually extracted radio signal properties from RSS for the participant s 1 with assigned label vector q 1 1 to q 1 6 , categorize every move of comprising 1500 RSS samples. With each move of s 1 1,1→n to s 1 1500,1→n is associated a move label q 1 r .
The other three participants' feature vectors s 2 1,1 to s 2 u,n , s 3 1,1 to s 3 u,n and s 4 1,1 to s 4 u,n are assigned class labels q 2 r , q 3 r and q 4 r respectively . The superscript z : 1 → 4 denotes the participant index. The subscript u is an index of the samples that range from u : 1 → 9000 for each of the subset matrix for each participant, while r : 1 → 6 represents index of the six directional movements undertaken by the subjects.

C. DEFINING FEATURE VECTORS
The number of features to be defined for the motion classification is determined by signal itself. Since this study employs a single carrier component, the directional motion signature pattern is determined by signal amplitude, embedding motion characteristics in the form of signal variations over time. With the inspiration from [77] to uniquely identify each motion pattern, we construct a set of manually crafted features from the RSS of pilot signal. The manually crafted feature vector set considers six different features of the received signal pattern, so as to address signal characterization properties that include mean, standard deviation, skewness, Real-Signal Kurtosis (RSK), average power, and energy of motion signalling period, wherein each received signal, with a signal sample z of the RSS signal, consists of 1 by η data samples. Following are the expressions, eq (17,18,19,20,21,22,23), used in this study to manually craft the six possible feature vectors that distinguish each motion pattern uniquely.
• Mean (µ) -Mean [78] is an average value of time varying signal, and it is computed over a period of time. µ is expressed by the eq(17), is an "equivalent" constant level, represents the same area as the signal over the same time: • Standard deviation (σ) -Standard deviation [78] provides a measure of spread of the amplitude values and indicates the width of the bell curve for a Gaussian noise signal or a given signal and is expressed by eq (18): • Skewness (γ) -skewness (I/Q) [79] is a higher-order statistical attributes of a time series. Skewness reveals the symmetry in the estimated probability density function (PDF) of the signal amplitude. γ is expressed by eq (19): • Real-Signal Kurtosis (RSK) -is proposed by [78] and provides estimate of kurtosis of In phase and Quadratic (I/Q), components of the signal. Kurtosis is given by eq (20).
Averaging the estimated kurtosis components (I/Q), determines the RSK feature, which is expressed by eq (21): • Average Power -average power [80] of a discrete time periodic signal over its length is given by the relation. Average Power is expressed by eq (22): • Energy -energy [80] of each signal over its length is given by the relation:

D. TRAINING SNN MODEL
SNN model performance measurements were carried out with the datasets, split into training and test dataset. The SNN model was tuned with the random optimization program, which is available in LabVIEW application suite, to precisely estimate the accuracy levels of available SNN models, whilst aiming to achieve high performance. Multiple iterations on 12 different neuron counts concluded in the selection of the optimized SNN model (see Fig.19). The SNN model was created for each participant to individually check model accuracy for each of the six movies, and also created was an overall model that trained on all 24 moves of the four different subjects, so as validate the overall accuracy of the SNN model on a aggregated dataset. Figures  8 to 11 show the individual unique motion patterns of each participant, whereas Fig. 12 shows an outcome of the combined move set. An SNN type classifier was configured and implemented using evaluated optimized settings for each of the participants. The SNN type classifier was input a dataset in the CSV file format that contained identified feature vectors and the associated class labels. After the SNN tuning parameters were set to the optimal settings, the classification accuracies observed during the tuning parameter optimization and actual training, remained almost the same. For the actual training With optimized tuning parameter selections, the model prediction accuracies remained exactly the same as observed during the search for optimized model parameters.

E. TESTING SNN MODEL
The preprocessed test data set was split up into individual datasets corresponding to each of the four participants,  In light of an overall data set and individual datasets, we developed individual SNN models that were trained on each participant individual data set, and an overall SNN model that was trained on the combined dataset of all the four participants. Next, the each individual test data set, having sequentially arranged data of six different moves of each participant, was input to the SNN classifier to evaluate the models' accuracy. The individual test set contained 9000 samples, including associated class labels for each directional move. Furthermore, we input the overall test dataset that contains all 24 directional move patterns of four participants to the SNN classifier, which we name as an "overall SNN model". Figures 14 to 17 show the actual and predicted outcomes of individual models, whereas Fig. 13 shows an outcome of the overall SNN model, used to find direction of motion of all the four participants.

VII. RESULTS AND ANALYSIS
The two NI-USRP 2921 TX/RX devices,which were installed in two adjacent rooms, separated by a distance of 6 meter, established a communication transmitter-receiver pair. During the experiment, our proposed testbed setup showed instability in the received signal strength because of unwanted activities by human in the corridor. The pilot signal's dynamic nature showed little instability due to mobile devices that were carried by the humans in the room with RX. The NI-USRP RX device enabled us to observe the normal transmission of pilot in the corridor area, and as the RF pilot signal presents a highly dynamic and sensitive nature, we observed considerable variations in the received signal strength with little human body movements, which include turning around, leaning on the wall, making gestures with hand, etc. To ensure that our data collection excluded such variations, the testbed environment was limited to the participant only, and with absence of other human movements, except the motion executed by the subject(s) under investigation, enabled least variations in RSS of pilot signal. Fig. 2 shows the six recorded pilot signal strength variations under no human movement. Fig. 2 depicts dynamic variations in pilot RSSI at room temperature. The random variations of pilot having -35 dBm to -38.5 dBm approximate strength was observed even in absence of human movements, and this clearly shows the variability of the RSS. However, when each of the participant performed the activity of walking though the corridor path, the dynamics of RSS showed huge variations that eventually resulted in a pilot signal with embedded unique motion patterns. The patterns embedded in the RSSI of the pilot hold the unique motion patternstyle of the participant(s). Thus, the unique sensitive pattern detection system can be categorized as a highly sensitive motion recognition system since even a micro moment gets embedded in the pilot. When the motion patterns are subject to analysis, the study results reveal the sensitive critical information contained in RSSI dataset, could be systematically explore to identify unique direction motion patternssignatures of any person.

A. PERFORMANCE METRICS
We use measurement parameters -precision, recall, and F-measure metrics to evaluate the proposed models of human motion direction prediction [81]- [83]. The precision parameter is a positive value that quantifies the prediction results of a classification process. It is computed by the following expression given in eq (24): where T P and F P denote the true positive and false positive values, respectively.
The Recall evaluation parameter quantifies sensitivity, and is given by the eq (25): where F N denotes the false negative value. The weighted average of recall and precision is known as F-measure, and it is quantifies the weighted average using following expression given in eq(26): where F 1 denotes the false negative value.

VIII. PERFORMANCE EVALUATION -SINGLE VERSUS MULTIPLE SNN MODELS
We evaluated the training and test datasets using SNN algorithm, the accuracies observed during SNN training and testing varied drastically. Testing SNN on individual datasets  resulted in prediction accuracies ranging between 80% to 83%, whereas testing SNN on the overall data set results in just 75%-78%.

A. TRAINING AND TESTING INDIVIDUAL SNN MODELS
LabVIEW application was used to build, train and test the SNN models. With individual models, trained to predict   accuracies of more than 85% resulted in overfitting, and as a result of computing the prediction accuracy on the test datasets, reduced nearly to about 30% to 40%. After several attempts, we trained our individual SNN models to predict training accuracies ranging between 80%-85% that finally resulted in best fit. These best fit models actually showed an increase in the actual prediction accuracy of each model, with accuracy of nearly 83%. Moreover, to ensure that the four models could identify each participant uniquely, we conducted as cross validation test. Figure 18 shows the cross validation of models along with the observed motion direction prediction accuracies. Each model is tested against all the four individual datasets. In each individual test that resulted in a motion direction predicted outcome, each SNN model showed wrong predictions of moves 3 and 4 that correspond to the moves -from West to East, and viceversa, while rest of the directional motion patterns were detected with almost 100% accuracy.

B. TRAINING AND TESTING AN OVERALL SNN
For the training of SNN on an overall dataset resulted in higher prediction accuracies reaching almost 82%. With the accuracies observed during training sequence, the test set accuracies varied for neurons less than x − inputs and greater than 96. However, when the overall model was trained to an accuracy level ranging between 75%-78% using neuron count greater than x − inputs and less or equal to 96 , the test accuracies remained almost constant at 70% (See Figure 19).

IX. COMPARATIVE ANALYSIS
To compare and trace out the best prediction model that detects the direction of human motion precisely, we trained and tested various other ML models available in MATLAB. We employed Decision tree algorithms, Linear Discriminant Analysis (LDA), Naïve Bayes, SVMs, and KNN algorithms to test our proposed model, which resulted in contrasting prediction accuracies. Tables -5, 6 and 7 show the observed results of these various machine learning models. Our observation of the collected results reveal that KNN outperformed the rest of the models. However, the Shallow Neural Network (SNN) based prediction model showed the highest prediction results compared to the rest of the models.
The KNN seems to compete equally well with the SNN, whilst showing a consistent motion direction prediction accuracy of approximately 83% on the individual test set, and 71% accuracy on the overall data set of the three male participants. However, it should be noted that the KNN just showed a motion direction prediction accuracy of 67% only on the individual test data set of the female participant (P3). Although this fact is not investigated yet, in the future studies this unusual behaviour needs further research. Therefore, a study on the gender based motion detection system is imperative.
For the rest of models, showed varying results and no consistency in human motion direction classification. The Decision Tree based models revealed the varying prediction accuracies on the individual datasets and 50%-60% prediction accuracies on overall datasets. Similar to Decision tree type models, Naive Bayes type models showed a consistent accuracy of 50-60% on overall data set, whilst failing to maintain the consistency on individual datasets. On the other hand, LDA showed a minimal improvement in the accuracy that ranged between 60% to 70% on average. Although SVMs are best suited for the classification of non linear data entities, SVM models employed in the study revealed 67% average accuracy on individual datasets and 71% on overall datasets only.
Overall, the models based on KNN show better prediction accuracies than others, however fails to compete with a single hidden layer neural network.

A. CROSS-VALIDATION OF SHALL NEURAL NETWORK MODELS
Cross-validation technique assesses and evaluates machine learning algorithm's performance. Generated datasets are input to machine learning algorithm to assess the model's performance, and with unknown datasets as input to an algorithm, allows assessment of prediction outcomes. Assessment techniques require random partitioning of a given dataset into a number of training and testing datasets. During several iterations, training sets train a supervised learning algorithm, while a single testing dataset is reserved during each iteration to test accuracy of model.
A commonly used Common Cross-Validation Technique called k-fold method, partitions a given dataset into k folds (random subsets) of nearly equal sizes. k-folds consist of k-1 subsets reserved to train a machine learning algorithm and one test subset that evaluates the trained machine learning model's performance. To assess the model's performance, the process iterates ω times to generate a new single subset (test set) for validation. Average estimation error across all k partitions is typically expressed by .
This study employs 5-fold Cross-Validation method to partition dataset into 5 equal folds of nearly equal sizes. The Cross-validation is configured at an iteration count of ω = 1000. To perform comparative analysis and cross validate the developed models, comparative analysis of two model is carried out using MATLAB [84]. Narrow Neural Network (NNN) is a form of SNN available in MATLAB , and the NNN model can be trained and tested using customized parameters. Since the input layer of the individual training models see 6 features, each NNN model's first layer is set with 6 neurons, and four such individual models are trained and tested for four different participants.
Although each model is trained and tested individually, the ensemble model containing four models with a common input, results in class probability as a common output of all the four. Each individual classifier has T anh as activation function at its output, and the probabilities of each model are fed to objective function max(P 1 , P 2 , P 3 , P 4 ). The objective function outputs the class of identified direction having maximum estimated probability by the objective function (see Fig. 21). Compared to ensemble model, the overall model also sees the same directional moves of four different participants at input. Thus, the input needs to have 24 neurons to classify 6 × 4 = 24 moves. NNN is configured with an input layer of 24 neurons and an iteration count of ω = 1000. 5fold Cross-Validation again evaluates the performance of the overall model for comparison to the Ensemble model. Fig. 22) shows the block diagram of overall model with 6 input feature vector X, and 24 directional classifications at output Y .

B. HYPOTHESIS TEST
To test the statistical significance of our proposed models, the study proposes the following two hypothesis, and applies Paired t test to compute statistical parameter t and probability P .
H 0 : the population mean of differences between observed ML model accuracies of Ensemble (NNN model) and overall model is zero, H 0 : µ d = 0.  1) Paired t test t test [85], a statistical analysis technique, is also known as Student's t test. t test is applicable to one sample only or two samples with the difference in between their two independent means. t test reveals the statistical difference that may exist between two independent sample means or within a sample mean. However, difference between two samples is close to zero if the sample means are same, and additional statistical tests may be applied on data to find out the statistical significance between the variables.
Paired t test statistically tests the difference between two paired results of a sample. The resulting difference between two paired samples is close to zero, if applied treatments on the two samples are same, and computed differences between the sample means used during the test result in a difference of 0.
The variance of the difference of two variables, for example, A and B, V ar(A − B) , is expressed by the equation eq (27) [85], where σ 2 1 and σ 2 2 represent the variance of variables A and B, respectively. ρ represents the correlation coefficient for the two variables. Correlation coefficient ρ equals to 0 for two independent groups used in an independent t test, VOLUME 12, 2021 and the sum of the two variances represents the variance of the difference between the two variables.
However, ρ may or may not be equal to 0 for paired variables. Apparently, t statistic is different for the two dependent samples and the expression equation eq(28) is represents a modified form of equation eq (27) for the paired number of samples.
whereχ 1 andχ 2 represent the sample means of two variables A and B. Variable s represents sample variance, expressed by equation eq (29). To determine the sampling distribution of the mean, sample variance is used instead of the population variance [85]: Here x i represents data sample at index i;x represents sample mean, and n 1 = n 2 = n.. The expression s 2 1 + s 2 1 − 2ρs 1 s 2 , computes statical variance between the two variables and takes into account the correlation coefficient ρ. As a result, t statistic in a paired t test is given by the equation eq(30): In equation eq (30), for the correlation coefficient ρ > 0, t statistic increases as the denominator becomes smaller. Subsequently, statistical power of paired t test increases compared to an independent t test. On the other hand, for the correlation coefficient ρ < 0, statistical power decreases and lowers compared to the independent t test.
This study applies t test, and the observations of Ensemble NNN and overall model substitute the samples A and B introduced in the prior section.
The test statistic is t = 112.987, with 999 degrees of freedom, and P < 0.0000 (reported up to four decimals only). Since observed p-value is less than α = 0.05, the study rejects the null hypothesis and concludes that there is a significant difference between the proposed ML models having mean validation accuracies of 74.02% (for overall model) and 89.13% (for Ensemble NNN model) approx. for prediction of human motion direction, respectively. The test clearly shows that Ensemble model can offer high prediction accuracy to determine direction of human motion using RSS.

C. OVERALL ANALYSIS AND RESEARCH DIRECTIONS
Our analysis on observed results shows that individually trained models for recognizing unique signatures (directional moves) of each participant provides higher prediction accuracies than the overall trained model for all four participants. This leads us to conclude that in a scenario with n-participants, n-models can be built to ensure that the proposed system delivers higher accuracy in predicting the individual signatures (directional moves). However, this demands distributed processing platforms, wherein at nnode located at m locations can each hold a pre-trained model. Upon receiving a common input (a sample set of moves of a person) over a communication network, these individual nodes can output predictions. The node with the highest predicted score will enable the identification of the move along with the identity of the person. Distributed processing systems as such would require nodes having embedded devices, like FPGA based platforms, supporting and running machine learning algorithms for higher efficiency. In continuation of this study, implementation of a newly proposed distributed processing system model is aimed in the next study, wherein delivery of results showing real time performance of distributed embedded systems employing machine learning algorithms is expected, including network performance measurements.

D. ADDITIONAL TESTBED REQUIREMENTS
While the human motion patterns were observed via an automated system application, USRPs posed problems that related to clock synchronization. The USRP radio device configured as RX, encountered latency, including drift in the received pilot signal. The system attempted to detect directional locomotion patterns within a microsecond time duration that required very high speed communication interface between the USRP and the Desktop Computer. Similarly, due to local oscillator frequency mismatch between the TX and RX, the pilot signal appearing on the RX, was offset by a frequency of 4KHz. The USRP sampling rate of RX device to manually calibrated during the testbed   optimization process. The frequency offsets have been often reported in the USRP devices [86], and to calibrate the TX and RX device properly, an external reference clocks signal generator is prerequisite to further studies. Hence frequency offsets in received pilot signal can be stabilized by using same reference clock signal for both the RX and TX. The process of finding optimized tuning parameters during the initial testing and calibration phase helped to establish an accurate and efficient testbed. The calibration parameters that results in the correct tuning of the testbed included power, sampling frequency, pilot signal frequency, distance and receiver orientation.

X. CONCLUSIONS
This study focused on the identification of direction of human motion prevailing behind the walls, using RSS of a pilot radio carrier and Shallow Neural Network. The study implemented proposed a system model and a testbed, employing state-of-the-art SDRs, to collect and analyze human motion patterns in indoor environments. The results of this study show that the use of RSS, embedding human motion patterns and application of SNN to classify such embedded motion patterns in RSS data, recognizes and predicts the direction of human motion with 89% approx. at 1000 iterations detection accuracy, however, using individual data models trained on each subjects motion datasets. The study also provides a comparative analysis of various machine learning models that were trained on the collected data, with prediction results showing that KNN can also help in prediction of human motion, however, the accuracy is lower than the SNN based prediction models. The proposed system model can be reproduced and validated on any commercially available SDRs for further research. A feature vector set has been proposed as input to the SNN model. The NI-USRP 2921 SDR model implemented was implemented, which enabled fine tuning of both the transmitter and receiver designs, so as to collect the directional human motion data.
In the future studies, RSS would be employed to precisely localize and detect human identity, as well as gender, walking behind the walls in given indoor or outdoor environment. A real time motion recognition system is anticipated based on a distributed processing model. This would enable processing of each subjects data on an individual processing unit connect to the distributed processing system. To design the new distributed processing testbed, it is envisioned to interface multiple SDRs, as individual radio units, interfacing high speed desktop computers. Each desktop computer would be running an application to implement SNN model trained on a specific subject. This would allow real time comparison of inputs received from multiple SDR units in real time. Further research in current study would guide researchers in developing real-time monitoring gadgets or apparatus and also to test and evaluate system performances in other available ISM bands. Moreover, it will lead researchers to develop more efficient, robust, and highly accurate HMRs if FPGA based processing units or cloud services are employed for high speed designs, since these offer more flexibility for modular design and prototyping.