A New Wide-Angle, Low-Resolution Infrared Array Sensor with Improved DL Algorithm for Activity Detection

A deep learning based method for activity detection that makes use of low-resolution, wide-angle infrared (IR) array sensors. In addition to the primary difficulty of this work—how to further enhance the performance of IR array sensor-based systems for activity detection also tackle the following issues: When comparing to a regular IR array sensor, a wide-angle sensor is used to get a better overall picture. As a result, it is difficult to learn how to adapt to the varying temperature distributions caused by activities performed in various locations. Additionally, in contrast to other efforts, aim to carry out the activity detection with the minimum amount of information feasible. This study intends to detect activity in a time frame of less than 1 second, whereas previous efforts have used a time window equivalent to 10 seconds. However, by using deep learning method outperform past work's accuracy while keeping the technique lightweight enough to operate on smartphones with limited processing capacity. Because of the neural network's capability to acquire the features, our hybrid deep learning model is particularly well adapted to the categorization of distorted photos. First, a Convolutional Neural Network (CNN) is used to differentiate between ceiling data and wall data, and then the combined data is also classified. Second, the CNN's output is fed into a Long Short-Term Memory (LSTM) with a window size equal to 5 frames so that the sequence of actions may be categorized. To detect the activity, our model uses a window size equal to 5 frames (i.e., less than a second). Our hybrid deep learning model achieves 99% accuracy in the total identification of activities while outperforming the state-of-the-art methods that use a longer time frame.