Target Classification Aided Variable-Structure Multiple-Model Algorithm

Because the varieties of general aviation aircraft and their performance differences become increasingly large, the traditional multiple model tracking algorithm needs to employ more motion models to describe the actual maneuver model of a moving target. This fact is easy to degrade the tracking accuracy and brings a large computational load. Thus, a target classification aided variable-structure multiple-model algorithm (TCA-VSMM) is proposed to solve the above problem. In the proposed TCA-VSMM algorithm, the target classification aided is introduced to improve the accuracy of state estimation. Then, the target classification aided in the screening of the motion model set is analyzed. Concretely, the target classification information and velocity information in automatic dependent surveillance-broadcast (ADS-B) measurements are incorporated into the variable-structure multiple-model filter. Consequently, the screening model set is more approximated to the real motion model of a moving target. Experiments show that the TCA-VSMM algorithm can obtain better performance with small computation load and high estimation accuracy than those of the model-group switching variable-structure multiple-model algorithm.


I. INTRODUCTION
Under the target tracking, the tracking filter is the mathematical calculation procedure. The contemporary measurement and former states were adopted to evaluate the current states of the objective [1]. The noise was filtered from the measurement data to perceive the states and trajectory of the target. When the target maneuvered, the motion mode change of the target cannot be corrected and predicted in time by each model filter, such as Kalman filter and extended Kalman filter, etc. Besides, it happened in a limited tracking performance [2]. When tracking maneuvering targets, using the multiple-model tracking algorithm is superior to the single-model one [3]. For example, the actual motion mode of the target cannot be accurately reflected by the motion model if one model is only used to track the target in the filtering algorithm. When the motion model is greatly different from the actual motion model, the performance of the tracking algorithm sharply drops, resulting in incorrect state estimation or loss of the target. When the target is in motion, there The associate editor coordinating the review of this manuscript and approving it for publication was Wenming Cao .
is just an actual motion model to meet every second. However, it is challenging to choose the right motion model while the target is maneuvering. Consequently, the multiple-model algorithm created a motion model that produced a restricted number model to reach the actual model set.
There are multiple classifications of general aviation aircraft, such as unmanned aerial vehicle, ultra-light aircraft, light aircraft, small aircraft, rotor aircraft, glider, paraglider, and lighter-than-air aircraft [4]. The turning performance of the unmanned aerial vehicle is good. The acceleration performance of ultra-light aircraft is different according to different models; the turning performance of Small aircraft is different according to different models; Rotor aircraft has vertical takeoff and landing capabilities, but the acceleration performance is not good; the turning and acceleration maneuver abilities of glider and paraglider are weak. These aircraft have large differences in maneuverability and diverse flight modes, causing an overlarge number of models in the model set that fully describes the different flight modes of these aircraft. Therefore, the delay of switching is increased to the model subset that best matches the target motion when using VSMM algorithms for tracking filtering. Meanwhile, the real-time performance of the algorithm is reduced. By introducing information of target classification, the model set can be screened to reduce the number of inappropriate models and to improve the tracking performance. With the increasing demand for accurate tracking of general aviation aircraft, it is imperative to make full use of the measurement obtained by sensors and introduce it into the algorithm for improving the tracking performance.
The application background of this paper is mainly aimed at the problem of target tracking in the complex environment of general aviation aircraft. There are many classifications of general aviation aircraft and their maneuverability greatly varies. Therefore, the influence of target classification is mainly considered.

II. RELATED WORK AND CONTRIBUTIONS A. RELATED WORK
Multiple possible maneuvering and non-motorized motion models are used by the multiple-model algorithm to work in parallel to overcome the uncertainty of the target motion model, which has more stable performance and better effects. Besides, the multiple-model algorithm is a mode of first estimating and later discriminating, which is, obtaining sufficient information before estimating all possible model sets of the target. Then, which model is the most suitable one is determined, resulting in improving the accuracy of the discrimination. Therefore, when the motion model set includes the real motion model of the target, the target maneuver can be accurately detected by the multiple-model tracking algorithm [3]. Based on the above advantages, the multiple-model target tracking algorithm using multiple filters working in parallel has been widely used [5]- [8].
The Interaction Multiple-Model (IMM) algorithm is based on a fixed model set, which requires a large number of models to ensure tracking accuracy [9]. The large model set leads to a huge amount of calculations; an overly detailed model space may destroy Bayesian reasoning's requirement for independent models, resulting in reducing tracking performance [10]. To solve this problem, Xiao Rong Li proposed the third-generation multiple-model algorithm in reference [11]: Variable-Structure Multiple-Model (VSMM). The algorithm uses a two-level hierarchical structure: the high-level structure is responsible for determining the model set; the low-level structure performs multiple-model filtering based on the determined model set. A method called Model-Set Adaptation VSMM is used in the high-level structure to obtain the best model set. Based on the concept of the VSMM algorithm, researchers also proposed an equivalent-model augmentation VSMM algorithm, and so on [12]- [14].
The setting of the model set is the key to the success of the multiple-model algorithm. A good set of motion models contains the maneuvers of the target in speed and regulation; the tracking error can be minimized by such a design. Typically, the model set covers the following models, including constant velocity (CV), constant acceleration (CA), and coordinated turn (CT), etc. to be arranged [15]. Besides, the size of the model set takes into account the computational complexity of the algorithm to avoid additional consideration. Then, the real-time production of the algorithm is decreased by numerous models. The problem of model set adjustment has been discussed in some review papers by Rong Li [16], [17]. This series of papers focuses on the setting and adjustment of the model set in the multiple-model algorithm. Its principles and methods are relatively mature, demonstrating its research progress and existing problems. However, with the development, a growing number of new-tracked targets such as general aviation aircraft have emerged. This classification of target uses a new Automatic Dependent Surveillance-Broadcast (ADS-B) sensor that can not only provide state information such as target position and velocity but also obtain information like target classification [18], [19]. This kind of information can be used for the aided adjustment of the motion model set in the multiple-model algorithm. The TCA-VSMM algorithm for target classification aided adjustment of the motion model set is generally divided into two parts: the acquisition and processing of information and the multiplemodel filtering. Firstly, the acquired measurements are processed; then, the target classification information is extracted; next, it is used to screen the motion model set; finally, the multiple-model filtering is carried out.

B. THE CONTRIBUTIONS
This paper proposed a TCA-VSMM maneuvering target tracking algorithm aided by target classification information which adapts to the new development of sensor technology and signal processing technology and makes full use of the knowledge acquired by the sensor. The TCA-VSMM algorithm suggested in the paper increases the utilization rate of knowledge information, reduces the uncertainty of the estimation results, improves the data processing capability, and enhances the performance of target tracking by using target classification information. This paper explains the theoretical framework of the TCA-VSMM algorithm in which using the target classification information assistance in the multiplemodel algorithm can exert restraining action on the target motion model set, eliminate the error model, and make the motion model set more coincide with the target real motion mode, thus improving tracking performance. This paper analyzes why employing the target classification information to filter the model set in the multiple-model algorithm can add the accuracy of the estimation and reduce the estimated Root Mean Square Error (RMSE). Based on this theory, this paper studies the specific application of the TCA-VSMM algorithm in ADS-B data filtering from the general aviation monitoring system. The theoretical framework mentioned in this paper can also be applied to filter motion model set using obstacle information, road information, attitude angle information, etc.
The remaining content of this paper is arranged as follows. The basic theory of TCA-VSMM is introduced in the third part; in the fourth part, a TCA-VSMM algorithm is proposed for the problem of overlarge motion model set caused by various classification of general aviation aircraft and maneuvering modes, also conducted experiments; the conclusion of this paper is drawn in the fifth part.

III. THEORETICAL FRAMEWORK OF TCA-VSMM ALGORITHM
This section mainly dwells on the principles of the MGS-VSMM algorithm and the principles of the VSMM algorithm to improve the tracking performance proposed in this paper. It is theoretically proved that the introduction of target classification screening can make the model set during the multiple-model filtering process closer to the optimal model set, thereby improving the performance of the multiple-model algorithm.

A. PRINCIPLE OF VSMM ALGORITHM
Maneuvering target tracking is a state estimation issue in which the motion model has an unexpected change. While the target maneuvered, the procedure of developing its motion pattern corresponds to various motion models. The multiple-model algorithm uses multiple filters (one filter for one model) to work in parallel to match the target maneuver. The set of models used by these filters is called a model set. Thus, the model set with r models can be written as The state equation and the measurement equation are revealed as: where x k and z k are the state and measurement vectors individually, F i k (·) and H i k (·) are the state transition function and the measurement function, respectively; i ∈ {1, 2, . . . ,r} , m i k is the i-th adopted model at time k. The w i k (·) and v i k (·) represent the process noise and the measurement noise, respectively. They are mutually independent zero-mean Gaussian white noise with covariance cov w i k = Q i k and cov v i k = R i k . The IMM algorithm utilizes cooperation to re-initialize the filters. The transition between distinct models is defined by the probability that the k−1 time m j k−1 transferred to m i k of probability to p ji k , which can manifest as where m i k is the model applied for filtering at time k; and m j k−1 is the model applied for filtering at time k−1; r is the number of models in the model set.
The VSMM algorithm is an extension of the IMM algorithm. The multiple-model algorithm with a fixed model set such as the IMM algorithm can be regarded as a special case of the VSMM algorithm. The core idea of the VSMM algorithm is that the current model set is switched or adaptively generated from all available information such as current measurements. The VSMM algorithm includes the following two parts: -Model set adaptation. The obtained measurements and prior knowledge are used to obtain the model set required at each moment. The adaptive methods of existing model sets can be divided into model set switching, cropping, and supplementation. These methods are often used in combination in practical applications.
-Multiple-model filtering. The model set obtained in the first part is used to perform corresponding multiple-model filtering.
The specific method is similar to the IMM algorithm with a fixed model set. The structure of the model-group switching variable-structure multiple-model (MGS-VSMM) algorithm shown in Figure 1. The core of the VSMM algorithm is model-set adaptation (MSA) [16]. The MSA algorithm can be divided into two types according to the generation method of the model set: the activated model set and the generated model set [20]. In the class of the activated model set, first, a general model set is determined; then, several candidate model subsets and switching strategies are determined; finally, the model subset is adaptively selected according to the strategy. Model-group switching (MGS) [17] and likely model set (LMS) algorithms [21] are the representatives of such methods. In the class of the generated model set, new models are continuously generated according to the strategy, thereby obtaining new model sets. Expected mode augmentation (EMA) [14] and adaptive grid structure (AGS) [12] are the representatives of this type of algorithm. The VSMM algorithm is described using the activation model set class as an example. Figure 1 is a basic structure diagram of the activation-type VSMM algorithm. It illustrates the basic principles and structure of this type of the MGS-VSMM algorithm.

B. PROPOSED THE TCA-VSMM ALGORITHM
After screening the model set using the target classification, the TCA-VSMM algorithm adjusts the functional elements then performs multiple-model filtering. The functional elements refer to the motion model set. This classification of algorithm introduces screening to adjust the motion model set in the VSMM algorithm, thus improving the tracking performance. The iterative principle flow of the TCA-VSMM algorithm is shown in Figure 2. At k time, in this figure, after processing target classification, the functional elements are adjusted, followed by the implementation of multiple-model filtering.
The total model set in Figure 2 is the union set of the target model set in the domain. For example, the target classification has K classes, then the target classification space is marked as T = {C 1 , C 2 , . . . ,C K }, and set the motion model set M i is set to correspond to the target classification C j , j = 1, 2, . . . K .
The screening process in Figure 3 uses the acquired target classification to screen the total model set. In the screening, S is the optimal model set, and A is the model set used in the current filtering, which is obtained by screening the total model set B. The total set of motion models is currently known by B is summarized by a large amount of knowledge, covering various objectives, applications, and environmental conditions, etc. Since target classification is sufficient, it is impossible to omit any model. Rather, there may be extra models, so the screened A contains S.
The model set B is initially set large, i.e. S ⊂ A ⊂ B. With the target classification aided, the result, after the model set A is deleted from B, the result is supposed as C. The distribution of the four sets, A, B, C, and S are demonstrated in Figure 3.
The outer ellipse stands for set B. The line shadow stands for set C. The spot shadow stands for the optimal model set S. The inner ellipse stands for the model set A is aided with the target classification. When C ∩ S = ∅, the state estimation of the model set A is better than that of the model set B and closer to that of the optimal model set S, referring to the conclusion in Literature [11] by X. Rong Li. Suppose the model set with target classification aided as A and the state estimation obtained by the model set A, S, and C isx A ,x S , andx C , respectively. The intersection of S and A is J, which is the optimal model part in A. Lis a set of missing and unnecessary models, i.e. the model sets in A that do not match S; the state estimation isx L .
According to the theorem 4.1 in Literature [11], suppose b = p{m j | z,A} p{m j | z,S} , ∀m j ∈ J, sô If b = 1, thenx S =x A , indicating that the state estimation of A equals to that of S. If b = 0, then A is completely wrong. In the condition that Sis contained in A, when J = S, 0 ≤ b ≤1; the larger the value of b is, the better the performance is. And the maximum is 1, so the performance is optimal. In the condition that A is contained in S, when J = A and b ≥ 1, the smaller the value of b is, the better the performance is. And the minimum is 1, so the performance is optimal. The state estimationx A is demonstrated aŝ The state estimationx B is expressed aŝ VOLUME 8, 2020 The posterior probability of m j ∈ B is Because of B = C +A, we further find that When m j ∈ A ⊆ B, we find that Therefore, When Therefore, We can find that Moreover, Suppose that C ∩ S = ∅, i.e. the model set C is excluded from the optimal model set S. The assumption should be further clarified later. Due to the uniqueness and independence of the setting of the model set, from C ∩ S = ∅, we get J = S ∩ B = S ∩ A. It can be then deduced that Here, b B ≤ b A , ∀m j ∈ J. From the relation formula of A, B, and S, 0 ≤ b B ≤ b A ≤ 1, and b B ≤ b A indicate that b A is closer to 1 than b B . Therefore, the performance of the model set A's state estimation may be better than that of B In other words, the overall set C is completely not in the optimal model set S. We conclude that applying information aided can promote the performance of the multiple-model algorithm.
This paper assumes C ∩ S = ∅, that is, the intersection between the deleted model set C and the optimal motion model set S is empty. If this hypothesis cannot be supported, the above derivation process is meaningless. This paper proposed a series of algorithms on the theoretical basis of that proof and takes the adjustment of the motion model set as an example to demonstrate that the hypothesis is valid. C ∩ S = ∅ is proved to be true in the algorithm for aided adjustment the model set of the target classification, the maneuvering ability of different types of targets varies, and each target has its maneuvering feasible range, which cannot exceed its allowable maximum. Therefore, C ∩ S = ∅ is established by the deleted model set C of the target classification. If C ∩ S = ∅ is not necessarily guaranteed. For example, in the obstacle-assisted model set screening algorithm, the deleted model set C and the optimal model set S may have an intersection. The problem can be analyzed from the perspective of probability. The intersection is based on empty probability, and the probability of deleting model error is extremely low, so the proposed algorithm can satisfy C ∩ S = ∅ based on the probability which can be verified in simulation experiments of specific algorithms.

IV. IMPLEMENTATION PROCESS OF TCA-VSMM ALGORITHM
There is a wide range of general aviation aircraft, whose maneuvering performance greatly varies with the structure, manufacturing process, power source, etc. In target tracking, it is necessary to establish a model set that covers all types of general aviation aircraft. There are many models in the set, which compete with one another and degrade the performance of the multiple-model filtering. The original setting method of the model set is no longer suitable for general aviation target tracking filtering. As a result, the TCA-VSMM algorithm is applied in the general aviation aircraft tracking.

A. TARGET CLASSIFICATION INFORMATION IN ADS-B DATA
ADS-B is a new sensor technology that redefines three elements in today's air traffic control, which are communication, general aviation, and surveillance. A stands for ''Automatic'', i.e., all-day operation without duty; D stands for ''Dependent'', i.e., to rely only on precise global satellite navigation and positioning data; S stands for ''Surveillance'', i.e., acquiring the aircraft position, altitude, velocity, heading, identification number, and other information; B stands for ''Broadcast'', i.e., broadcasting respective data information among the aircraft or between the aircraft and the ground station. ADS-B features long operational distance due to the use of broadcasting, which promotes the technology to be rapidly popularized and applied.
The airborne ADS-B equipment of the general aviation aircraft broadcasts the ADS-B messages to the remote station and other aircraft every other cycle. Generally, one data frame of a message only contains information of a coordinated flight target, which is encoded according to the order of data items in the User Application Profile. EUROCONTROL is a pan-European, civil-military organization dedicated to supporting  European aviation. EUROCONTROL uses the standard data item CAT021 to transmit ADS-B information according to uniform standards. The ADS-B measurements [22] are obtained by analyzing the CAT021 protocol, which mainly includes aircraft identification (data source, etc.), positioning (longitude, latitude, altitude), velocity, steering, etc. and other related additional data (air pressure, wind direction, wind velocity, air pressure height, temperature, etc.). In practical applications, this data does not need to be fully available. According to the analysis of the measurement data of the ADS-B, the data has a unique target identifier, as shown in Table 1. According to the prior knowledge, the target identifier has a one-to-one correspondence with the target classification, therefore, the target classification information of general aviation aircraft can be determined by the identifier (in Table 1, identifier 7865504 is the Cessna-172 aircraft).
In the tracking of general aviation aircraft, it is necessary to establish the corresponding motion model set for each classification of general aviation aircraft to make the motion model adopted in filtering better match the flight condition. The flight model database of different classifications of the general aviation aircraft is the total model set. The target tracking problem is a random estimation problem, and the model is a constraint in filtering estimation. The randomness of the maneuvering target is larger, so multiple models are used to match the filter to overcome the uncertainty. Researchers have developed many methods to build model sets, where the most common one is using the technical performance indicators of different classifications of aircraft. The ADS-B error is from Global Positioning System (GPS) positioning. Most existing methods rely on representing the actual measurement error distribution with the Gaussian model.

B. AN ITERATIVE IMPLEMENTATION PROCESS
This paper studies the matching degree between the model set and the actual navigable flight mode and proposes a TCA-VSMM algorithm to satisfy the requirement for tracking and monitoring general aviation aircraft, so it obtains a better tracking performance. The algorithm roughly tracks the target before classifying using the total set of the motion model; then, it screens the model set according to the target classification; finally, it performs multiple-model filtering. Based on the multiple-model algorithm, the TCA-VSMM algorithm adds the building part of the total set of motion models and the filter part of the model set. The flow of the algorithm is shown in Table 2. Let the total model set be M, which is the union of all classifications of target's motion model sets, where the number of models is n; the model set in the current algorithm is M k ; the model set is M i when the target classification is C i , where the number of models is s, VOLUME 8, 2020 which is less than n. Assuming that the target classification of the general aviation aircraft has been determined to be C j , then the model set M j of the corresponding s models is used in multiple-model filtering. In the tracking of general aviation aircraft, the target classification information comes from the data identifier entries in ADS-B data. The calculation process of the filtering algorithm in period [k−1, k] is shown in Table 2. In Table 2, is the state estimation under different models at time k-1, and the mixed state estimation is obtained by integrating model probability µ i k as the input. The s filters working in parallel use different models withx i k−1 and z k as inputs for filtering, and the output of state estimation isx i (k|k) . In the updating part of the model probability, the likelihood function i k and model probability µ j k−1 of the previous moment are used as inputs, so we obtain the updated model probability µ i k , state estimationx i (k|k) , and covariance estimation P (k|k) . In the iterative process of the algorithm, the target classification information is used to screen the model set to reduce invalid models and improve the filtering speed and precision.

V. SIMULATION EXPERIMENT AND DISCUSSION
This section verifies the tracking accuracy and computing speed of the TCA-VSMM algorithm through simulation experiments and real experiments. The computer configuration is as follows: a CPU of the Intel i7 3.0 GHz, the memory is 4 GB, and the software is MATLAB R2014a. In the ADS-B measurements, an Estimated Position Uncertainty (EPU) is defined as the estimation accuracy error range of the target in the horizontal position, and the actual position falls within the circle with a 95% probability and a 5% probability outside the circle that is centered on the aircraft reporting position. This error representation is known as Circular Error Probability (CEP), which is widely used in the field of navigation positioning [23]. When the probability is 95%, the circular probability error can be expressed as CEP 95 . For example, when the Navigational Accuracy Category for Position (NACp) level is 9, the theoretical containment radius is (10m, 30m]. Therefore, after the NACp in the ADS-B measurements is obtained, the CEP range of the measurement can be ascertained. It can be seen from the range of the EPU that the measurement noise range of the ADS-B measurement is fairly wide.
The measurement noise variance is adopted in filtering, for which the EPU needs to be converted into the corresponding RMSE. The corresponding relationship with RMSE commonly used in target tracking is as follows, According to the corresponding formula in reference [23], it will be obtained as follows: RMSE = 1.1×CEP 95 .

A. SIMULATION EXPERIMENT
The motion trajectory of the simulation target is shown in Figure 4. The velocity values are assumed to be derived from ADS-B measurements. Let the initial state of the target be [100 m, 40 m/s, 1000 m, 40 m/s] and let the sampling frequency be 1 Hz to collect a total of 100 points. This simulation assumes that the target is flying at a CV model and a CT model [15]. The target turns at CT +2 degrees/second in 15-30 seconds, CT -2 degrees/second in 50-60 seconds, CT +3 degrees/second in 80-90 seconds and CV for the remaining time. The measurement data generated by the simulation are obtained by adding Gaussian white noise with a mean square error of 33 meters to the real trajectory data, corresponding to the maximum value of the NACp level is 9. On the horizontal plane, let the motion state be x = [x,ẋ, y,ẏ], where x and y are the position coordinates of the target;ẋ andẏ are the velocity of the target along the X and Y axes, respectively. T is the sampling period of the measurement; ω k is the turning angular velocity, which is preset according to the actual situation. Angular velocity ω k in the CT model can be designed according to the design approach of the motion model or the known real-mode distribution function [22]. The model transition probability matrix is set by the fixed setting method. Let the maneuver model on the main diagonal  element be 0.7, the no maneuver model be 0.9, and the non-diagonal elements be evenly distributed.
The initial selection model set of the filter is M k = M, including thirteen models: CV, CT +2 degrees/sec, CT -2 degrees/sec, CT +3 degrees/sec, CT -3 degrees/sec, CT +5 degrees/sec, CT -5 degrees/sec, CT +7 degrees/sec, CT -7 degrees/sec, CT +10 degrees/sec, CT -10 degrees/sec, CT +13 degrees/sec and CT -13 degrees/sec. The model transition probability is initially selected as a 13 × 13 matrix, which corresponds to the initial model set M; thus, the model set of target classification information is C i , so M k = M i , which contains five models: CV, CT +2 degrees/sec, CT -2 degrees/sec, CT +3 degrees/sec, and CT -3 degrees/sec. The remaining nine models are discarded because they are not suitable for the target of this classification, and the model transition probability matrix is adjusted to a 5 × 5 matrix instead of M i ; this process realizes the screening of the model set.
In the simulation, it is assumed that the first 50 points of the target track do not obtain the target classification information, and the model set is M k = M. When the sensor collects the last 50 points, the target classification information C i is obtained, and the model set is adjusted to M k = M i . After 50 Monte Carlo simulations, the average calculation time of each point of the simulation track is shown in Figure 6. When the classification information of the target is unknown, the filtering calculation time of the first 50 points of the TCA-VSMM algorithm is basically identical to those of the MGS-VSMM algorithm [9]. However, much less calculation time is required for the TCA-VSMM algorithm  after introducing the aid of target classification from point 51.
The most well-known standard for filtering is the RMSE. The filtering results of the target motion trajectory in the simulation experiment are shown in Figure 7 and Figure 8, where the TCA-VSMM algorithm is used in the entire process. The TCA-VSMM algorithm has better tracking accuracy on the filter tracking of the target position and velocity than the MGS-VSMM algorithm. The RMSE calculated by the algorithm at the position and velocity is reduced compared to the MGS-VSMM algorithm and deleted model set C. The main reason is that the target classification aid is introduced in the process of multiple-model filtering, which discards impossible models and reduces the number of models in the model set from 13 to 5. The reduced competition among models improves the filtering accuracy and reduces the computation load of the algorithm, which makes the TCA-VSMM  algorithm more suitable for tracking target. The two figures also show that the filtering effect is more obvious when zero-mean Gaussian noises with standard deviations increase, the standard deviation increases from σ = 11m(NACp = 10) to σ = 33m(NACp = 9). However, some large errors remain in Figure 7 and Figure 8. The algorithm should consider the computational complexity and adopt the compromise processing method. At present, it can only ensure better performance on the whole, but the processing method must be further improved.
Section III concludes that b A is closer to 1 than b B , and the state estimation obtained by model set A is better than that by model set B. Since p m j | z, S is difficult to solve, Assuming that C ∩ S = ∅, the condition of using TCA-VSMM algorithm is satisfied. Therefore, the experiment can verify that b B b A ≤ 1, and the performance of the tracking algorithm is improved by introducing the target classification aid. Eight points are selected to calculate the ratio of the probability of the MGS-VSMM algorithm model to that of the TCA-VSMM algorithm model. As shown in Table 3 the TCA-VSMM algorithm can confirm that b B b A ≤ 1, which improves the tracking performance.

B. EXPERIMENTAL ANALYSIS OF MEASURED DATA
To test the effectiveness of the algorithm in this paper, an ADS-B receiver is used to obtain the observations of the aircraft. Measured data used in this study originated from ADS-B (116 sampling points) observations for a target. The program is run on a computer with 200 Monte Carlos runs. Figure 9 shows the real trajectory obtained using the ADS-B and this trajectory with Gaussian white noise (σ = 33 m) is added. The curve of the TCA-VSMM algorithm and Real noise coincide. Figure 10 shows the position RMSE statistics of MGS-VSMM algorithm TCA-VSMM algorithm and real noise. It shows that the performances of the TCA-VSMM algorithm are better than the performances of the MGS-VSMM algorithm in real data example, and we can see that the position RMSE of the MGS-VSMM algorithm is much larger than the position RMSE of the  TCA-VSMM algorithm. A factor responsible for these can be attributed to the same reason for the good performance of the proposed algorithms as mentioned in Section III.B. The runtime statistics of each algorithm are illustrated in Figure 11. In each case, the program is run on a computer with 50 Monte Carlos runs. As is clearly seen in Figure 11, the TCA-VSMM algorithm is the faster the MGS-VSMM algorithm for target tracking. One major reason for this is that the MGS-VSMM algorithm must calculate the more results of motion models, and lead to the increasing of the computational load as the number of motion models increases. In a word, compared with the MGS-VSMM algorithm, the TCA-VSMM algorithm has better tracking performance.

VI. CONCLUSION
This paper mainly studies the screening adjustment method of the motion model set in the multiple-model algorithm to track general aviation aircraft. The variety of general aviation aircraft and huge differences in performance results in a large set of motion models, which causes low tracking accuracy and large computational load for multiplemodel target tracking algorithms. Therefore, this paper proposed a TCA-VSMM algorithm to address these problems.
This algorithm introduces the target classification information in the ADS-B measurement into the aided filter of the multiple-model algorithm. Then, the models of the model set are screened to make them closer to the actual motion model of the target. Thus, the tracking accuracy is improved, and the computation load of the general aviation aircraft tracking is reduced. Simulation experiments and real experiments show that the TCA-VSMM algorithm has better estimation performance. Compared to the MGS-VSMM algorithm, the algorithm proposed in this paper is more suitable for tracking general aviation aircraft. In practical applications, there are many types of multiple-model filtering algorithms that use screening in adjusting the model set, which require further research.