A method for tracking maneuvering targets in cluttered environments based on interactive autoencoders

Abstract

This invention discloses a maneuvering target tracking method in clutter environments based on interactive autoencoders, applied to the field of radar detection technology. Addressing the problem that existing technologies only consider the inter-frame relationships of different historical trajectory points from the perspective of frame relationships, neglecting the clutter environment and resulting in low target tracking accuracy, this invention first introduces a pyramid-shaped dynamic mapping scheme, and then designs a maneuvering target tracking method in clutter environments based on an interactive Transformer-GAT autoencoder network. This invention can achieve high-precision and stable tracking of maneuvering targets in clutter environments.

Description

Technical Field

[0001] This invention belongs to the field of radar detection technology, and specifically relates to a tracking technology for moving targets in cluttered environments. Background Technology

[0002] Model-driven methods require accurate kinematic models for tracking maneuvering targets in cluttered environments. In fact, constructing these precisely matched models is challenging, and it's impossible to obtain this prior information accurately and in a timely manner before tracking. Furthermore, classic model-driven algorithms only establish relationships between adjacent frames, and data association algorithms such as PDA and IPDA only consider the latest set of measurements when updating state, neglecting historical tracking data. Therefore, these model-driven methods ignore the spatiotemporal relationships between intra-frame measurements and multi-frame historical trajectory information. Without fully and effectively utilizing the potential relationships existing within and between frames, the operability of these model-driven methods decreases in the presence of clutter.

[0003] In recent years, deep neural networks have been widely used to replace the filtering or data association parts in classical tracking algorithms, providing a good example for introducing deep learning techniques into the problem of maneuvering target tracking. The paper "L. Shen, H. Su, Z. Li, C. Jia, and R. Yang, Self-attention-based Transformer for nonlinear maneuvering target tracking, IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1-13, 2023" proposes two Transformer-based nonlinear trackers to improve the performance of nonlinear maneuvering target state estimation. The paper "S. Kar, S. Babu, D. Jain, Y. Shali, DSSaritha, RVBhat, and BNBharath, Deeplearning-based track prediction and correction for a radar target, IEEE Trans. on Radar Syst., vol. 1, pp. 395-400, Jul. 2023" implements radar target prediction and correction using a temporal convolutional network. However, these methods only focus on the inter-frame relationships between different historical trajectory points from the perspective of frame relationships, without considering the clutter environment, which limits their further development. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for tracking maneuvering targets in clutter environments based on an interactive Transformer-GAT autoencoder network, which can achieve high-precision and stable tracking of maneuvering targets in clutter environments.

[0005] The technical solution adopted in this invention is: a method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network, comprising:

[0006] S1. Extract and fuse multi-dimensional spatiotemporal features from historical track points and current frame measurement data;

[0007] S2. Concatenate the results from step S1 to obtain the node feature matrix V. This represents the matrix transpose operation. This represents the multidimensional spatiotemporal feature matrix of the measurement data in the current frame. The multidimensional spatiotemporal feature matrix of the historical track is represented; the edge relationship matrix E is obtained based on the fact that any two historical track points are connected by an edge, any historical track point is connected to any current frame measurement point by an edge, and any two current frame measurement points are connected by an edge; a pyramid-shaped dynamic graph G = (V, E) is constructed based on the node feature matrix V and the edge relationship matrix E.

[0008] S3. Construct an interactive Transformer-GAT autoencoder network;

[0009] S4. Input the pyramid-shaped dynamic graph G=(V,E) obtained in step S2 into the interactive Transformer-GAT autoencoder network constructed in step S3 for training.

[0010] S5. Perform target tracking based on the trained interactive Transformer-GAT autoencoder network.

[0011] The beneficial effects of this invention are as follows: This invention provides a method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network, enabling high-precision and stable tracking of maneuvering targets in cluttered environments. This invention employs a pyramid-shaped dynamic mapping scheme to construct the spatiotemporal relationship between historical tracks and current frame measurements; and uses an interactive Transformer-GAT autoencoder network to learn this spatiotemporal relationship. Therefore, this invention has the advantages of strong adaptability and high robustness, and can be applied to existing radar equipment. Attached Figure Description

[0012] Figure 1 This is a flowchart of the processing in this embodiment;

[0013] Figure 2 This is a schematic diagram of a pyramid-shaped composition scheme;

[0014] Figure 3 This is a simulated scenario of tracking a maneuvering target in a cluttered environment.

[0015] Figure 4 for Figure 3 The tracking results are shown in the simulation scenario. Detailed Implementation

[0016] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0017] The flowchart of this invention is as follows Figure 1 As shown, the specific steps for its implementation are as follows:

[0018] Step 1: Extract and fuse multidimensional spatiotemporal features from historical track points and current frame measurement data.

[0019] For from the first Frame to the The set of historical track points of a frame The matrix has a shape of 3 rows and 2 columns, and contains elements from the first row to the second row. Frame to the The x and y coordinates of the historical track points of the frame, the first The historical track points of a frame can be represented as For from the first The set of measurement data for the current frame, which contains the measurement data of the first frame. The x and y coordinates of the measurement data of the frame, where the first... The i-th measurement point of the frame can be expanded as follows: The extracted multidimensional spatiotemporal features include: the x-coordinate and y-coordinate from the original data, as well as the distance s and yaw angle. Curvature K, velocity v, acceleration a, and heading angle θ. Feature fusion involves concatenating the remaining six features from the original data, based on the x and y coordinates. and Their calculation formulas are as follows: in and It is by and Calculate distance s and yaw angle The curvature K, velocity v, acceleration a, and heading angle θ yield shapes of 3×6 and 3×6 respectively. eigenmatrix Indicates the first The number of measurement points per frame. After feature extraction and fusion, a multidimensional spatiotemporal feature matrix of the historical trajectory is obtained. and the multidimensional spatiotemporal feature matrix of the current frame measurement data The concatenation of these two elements forms the node feature matrix for the subsequent pyramid-based dynamic graphing scheme, and its calculation formula is as follows:

[0020] Step 2: Pyramid-based dynamic composition scheme

[0021] The graph construction consists of two parts: node selection and edge establishment. The three historical trajectory points closest to the current time frame and the measurement point in the current frame are selected as nodes for graph construction, and their node feature matrix is ​​obtained in step 1. Regarding edge establishment, any two historical track points are connected by an edge, any historical track point is connected to any current frame measurement point by an edge, and any two current frame measurement points are connected by an edge. Because the final graph resembles a pyramid, it is named the pyramid-shaped graphing scheme, as illustrated in the diagram below. Figure 2 As shown. The graph obtained through the pyramid-shaped graph construction scheme is denoted as G = (V, E), where E is the edge relation matrix connecting each node. The graph construction result G will serve as the raw input to the encoding part of the interactive Transformer-GAT autoencoder network.

[0022] Step 3: Construct the encoding part of the interactive Transformer-GAT autoencoder network

[0023] The encoding part of the interactive Transformer-GAT autoencoder network includes: a Transformer encoding layer, a GAT (Graph Attention Network) encoding layer, a linear layer, and a global max pooling layer. The Transformer and GAT encoding layers extract temporal and graph structure features respectively, the linear layer performs dimensionality upscaling and downscaling of the features, and the global max pooling layer outputs the final coordinate prediction result. The original input to the Transformer and GAT encoding layers is the graph construction result G = (V, E) from step 2. Specifically... Figure 1 As shown, it includes 5 Transformer coding layers and 4 GAT coding layers. The 5 Transformer coding layers are denoted as TE0, TE1, TE2, TE3, and TE4, and the 4 GAT coding layers are denoted as GE1, GE2, GE3, and GE4. The input of TE0 is V, and the output is TE. (0) The inputs to TE1, TE2, TE3, and TE4 are the results of fusing the output of the previous Transformer coding layer after passing through a linear layer with the output of the corresponding GAT coding layer; the input to GE1 is V and TE.(0) The fusion result, and E, the inputs of GE2, GE3, and GE4 are the fusion of the output of the previous GAT coding layer and the output of the corresponding Transformer coding layer, and E.

[0024] The formula for calculating the encoding part of the interactive Transformer-GAT autoencoder network is as follows:

[0025]

[0026]

[0027] LTE (l-1) =LL(TE (l-1) )

[0028]

[0029]

[0030] Among them, GE (l-1) TE represents the output of the (l-1)th GAT coding layer. (l-1) η represents the output of the (l-1)th Transformer coding layer. G This represents the weighting coefficients in the output of the GAT encoder. Let E represent the node feature input of the l-th GAT coding layer, E represent the edge relation matrix of the graph construction result, GEl(·) represent the l-th GAT coding layer, and LL(·) represent the linear layer. (l-1) η represents the output of the Transformer encoding result of layer l-1 after passing through a linear layer. T This represents the weighting coefficients in the output of the Transformer encoder. TEl(·) represents the node feature input of the l-th Transformer coding layer.

[0031] Step 4: Construct the decoding part of the interactive Transformer-GAT autoencoder network

[0032] The encoding part of the interactive Transformer-GAT autoencoder network constructed in step 3 is introduced into an autoencoder structure to obtain the decoding part of the interactive Transformer-GAT autoencoder network. The node features are then reconstructed using the Transformer decoder and the GAT decoder. Specifically, as follows... Figure 1As shown, it includes 5 Transformer decoding layers and 4 GAT decoding layers. The 5 Transformer decoding layers are denoted as TD0, TD1, TD2, TD3, and TD4, and the 4 GAT decoding layers are denoted as GD1, GD2, GD3, and GD4. The input of TD0 is the output of TE4, TE. (4) The output is TD (0) The inputs to TD1, TD2, TD3, and TD4 are the results of the previous Transformer decoding layer after passing through a linear layer; the input to GD1 is the output of GE4. (4) The inputs to E are the outputs of the previous GAT decoding layer and E.

[0033] The calculation formula for the decoding part of the interactive Transformer-GAT autoencoder network is as follows:

[0034]

[0035]

[0036] TD (4) =TD4(LTD) (3) )

[0037] Where GDl(·) represents the l-th GAT decoding layer, Let E represent the node feature input of the l-th GAT decoding layer, and let E represent the edge relation matrix of the graph construction result. (4) This represents the output of the 4th GAT coding layer, GD (l-1) This represents the output of the (l-1)th GAT decoding layer. TDl(·) represents the node feature input of the l-th Transformer decoding layer, LL(·) represents the linear layer, and TE represents the linear layer. (4) This represents the output of the 4th Transformer coding layer, LTD. (l-1) This represents the output of the (l-1)th GAT decoding layer.

[0038] Step 5: Output track tracking results

[0039] The output of the 4th Transformer coding layer (TE) (4) The output of the 4th GAT coding layer GE (4) The weighted fusion is calculated as follows:

[0040] TGE = η G ×GE (4) +η T ×TE(4)

[0041] Where, η G η represents the weighting coefficients of the GAT encoder output. T These represent the weight coefficients in the Transformer encoder output. The fused TGE result is passed through a global max-pooling layer and three linear layers to obtain the prediction result for the current frame.

[0042] Step 6: Design a hybrid supervised training loss function

[0043] The hybrid supervised training loss function includes two loss functions: a current frame waypoint prediction loss function and a graph node feature reconstruction loss function. The network is trained using both loss functions. The formula for the hybrid supervised training loss function is:

[0044]

[0045]

[0046] in, It is the first Frame coordinate prediction results It is the first The true coordinates of the frame, Loss CP V represents the current frame waypoint prediction loss function. (i,j) This represents the j-th feature of the i-th node. This represents the j-th feature of the i-th node in the node feature matrix reconstructed by the GAT decoder layer. This represents the j-th feature of the i-th node in the node feature matrix reconstructed by the Transformer decoder layer. Indicates the first Number of measurement points per frame The total number of nodes after adding the historical track points from the 3 frames.

[0047] The effectiveness of this invention is further illustrated by the following simulation verification:

[0048] Experimental results:

[0049] Simulate the motion of a maneuvering target in a cluttered environment. A radar at coordinates (0,0) observes a maneuvering target moving in the XY plane. The radar obtains the target's x and y positions. The target's starting point is (x1, y1) = (0.83 × 10⁻⁶). 4 m, 1.22 × 10 4 m), initial velocity is (v x1 ,v y1The target's trajectory consists of 22 maneuver segments, each lasting 10 seconds, with 10 track points per segment. To enhance the target's maneuverability, the trajectory includes five representative maneuver segments: a "reverse turn" maneuver (segments 2-6), a "figure-eight" maneuver (segments 7-9), a "spiral escape" maneuver (segments 13-16), an "S-shaped" maneuver (segments 17-18), and a "sudden turn" maneuver (segments 19-21). The simulated maneuvering target motion scenario in a cluttered environment is as follows: Figure 3 As shown. The clutter is added in the following way: Assume the first... Frame and the The radial Euclidean distance between the track points of the frame is The amount of clutter added in each frame is k. The position of the i-th clutter in the frame and the position of the i-th clutter in the frame The radial Euclidean distance between historical track points of a frame satisfies:

[0050]

[0051] in Representing a uniform distribution, ρ1 = 0, ρ2 = 2. The i-th clutter in the frame relative to the i-th The yaw angle of the historical track point of the frame satisfies:

[0052]

[0053] in,

[0054] Therefore, the location where clutter is added can be represented as

[0055]

[0056]

[0057] Tracking results in the simulation scenario are as follows Figure 4 As shown in the figure. Statistical analysis reveals that the average root mean square error (RMSE) of tracking in the x-direction is 87.40m, with a variance of 63.48m. The average RMSE of tracking in the y-direction is 95.14m, with a variance of 69.37m. The average RMSE of tracking for each frame is 142.43m, with a variance of 72.31m, and the average RMSE of tracking for each track segment is 157.63m, with a variance of 26.01m.

[0058] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network, characterized in that, include: S1. Extract and fuse multi-dimensional spatiotemporal features from historical track points and current frame measurement data; S2. Concatenate the results from step S1 to obtain the node feature matrix V. This represents the multidimensional spatiotemporal feature matrix of the measurement data in the current frame. The multidimensional spatiotemporal feature matrix of the historical track is represented; the edge relationship matrix E is obtained based on the fact that any two historical track points are connected by an edge, any historical track point is connected to any current frame measurement point by an edge, and any two current frame measurement points are connected by an edge; a pyramid-shaped dynamic graph G = (V, E) is constructed based on the node feature matrix V and the edge relationship matrix E. S3. Construct an interactive Transformer-GAT autoencoder network; S4. Input the pyramid-shaped dynamic graph G=(V,E) obtained in step S2 into the interactive Transformer-GAT autoencoder network constructed in step S3 for training. S5. Perform target tracking based on the trained interactive Transformer-GAT autoencoder network.

2. The method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 1, characterized in that, in It is by Calculate distance s and yaw angle The curvature K, velocity v, acceleration a, and heading angle θ yield a feature matrix of shape 3×6.

3. The method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 2, characterized in that, It is by Calculate distance s and yaw angle The curvature K, velocity v, acceleration a, and heading angle θ give the shape as The characteristic matrix of .

4. The method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 3, characterized in that, The formula for calculating the encoding part of the interactive Transformer-GAT autoencoder network is as follows: LTE (l-1) =LL(TE (l-1) ) Among them, GE (l-1) TE represents the output of the (l-1)th GAT coding layer. (l-1) η represents the output of the (l-1)th Transformer coding layer. G This represents the weighting coefficients in the output of the GAT encoder. Let E represent the node feature input of the l-th GAT coding layer, E represent the edge relation matrix of the graph construction result, GEl(·) represent the l-th GAT coding layer, and LL(·) represent the linear layer. (l-1) η represents the output of the Transformer encoding result of layer l-1 after passing through a linear layer. T This represents the weighting coefficients in the output of the Transformer encoder. TEl(·) represents the node feature input of the l-th Transformer coding layer.

5. A method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 4, characterized in that, The calculation formula for the decoding part of the interactive Transformer-GAT autoencoder network is as follows: TD (4) =TD4(LTD (3) ) Where GDl(·) represents the l-th GAT decoding layer, Let E represent the node feature input of the l-th GAT decoding layer, and let E represent the edge relation matrix of the graph construction result. (4) This represents the output of the 4th GAT coding layer, GD (l-1) This represents the output of the (l-1)th GAT decoding layer. TDl(·) represents the node feature input of the l-th Transformer decoding layer, LL(·) represents the linear layer, and TE represents the linear layer. (4) This represents the output of the 4th Transformer coding layer, LTD. (l-1) This represents the output of the (l-1)th GAT decoding layer.

6. The method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 5, characterized in that, TE (4) and GE (4) The weighted fusion is calculated as follows: TGE=η G ×GE (4) +n T ×TE (4) Where, η G η represents the weighting coefficients of the GAT encoder output. T These represent the weighting coefficients in the output of the Transformer encoder. The fused TGE result is passed through a global max pooling layer and three linear layers to obtain the prediction result for the current frame.

7. A method for tracking maneuvering targets in cluttered environments based on an interactive Transformer-GAT autoencoder network according to claim 6, characterized in that, Step S4 employs a hybrid supervised loss function during training, expressed as: in, It is the first The true coordinates of the frame, Loss CP V represents the current frame waypoint prediction loss function. (i,j) This represents the j-th feature of the i-th node. This represents the j-th feature of the i-th node in the node feature matrix reconstructed by the GAT decoder layer. This represents the j-th feature of the i-th node in the node feature matrix reconstructed by the Transformer decoder layer.

Images