A radar multi-target intelligent tracking method based on filter association interfeed

Abstract

The application discloses a radar multi-target intelligent tracking method based on filter correlation mutual feeding, and belongs to the technical field of radar multi-target tracking. The method is based on a track timing feature extraction module of a circuitous scanning Mamba network, which can effectively extract target motion characteristics and obtain preliminary state estimation. Then, a space-time correlation and state correction module based on a dynamic sparse cross-attention mechanism is used, and based on information entropy theory, clutter is adaptively filtered out, so that the reconstruction of the correlation weight and the correction of the target state are dynamically completed. Subsequently, the correlation and filtering tasks rely on a multi-layer progressive architecture for recursive interaction, realizing mutual feeding optimization of filtering and correlation. Finally, a state estimation and correlation scoring module is used to output target state estimation and measurement correlation scores in parallel. Experiments show that the tracking accuracy and stability of the proposed method are better than those of comparative methods in the scene where frequent maneuvering and dense clutter coexist, and the method still maintains good performance in the generalization test where the number of targets is doubled. The core contribution of the method is to break through the traditional serial decoupling paradigm through the mutual feeding design of correlation and filtering, and the dynamic sparse cross-attention mechanism uses information entropy to adaptively filter out redundant measurements, showing good tracking robustness in the scene where dense clutter and frequent maneuvering coexist.

Description

Technical Field

[0001] This invention belongs to the field of radar multi-target tracking technology, specifically relating to a radar multi-target intelligent tracking method based on filter correlation feedback. Background Technology

[0002] Radar multi-target tracking is a core technology in fields such as air defense early warning, air traffic control, and unmanned aerial vehicles. Compared to single-target tracking, multi-target tracking faces complex challenges such as an unknown number of targets and high uncertainty in the source of measurement data and targets. Traditional model-driven methods rely on precise prior assumptions, decomposing multi-target tracking into multiple single-target filtering problems and determining the attribution relationship between measurements and targets based on data association algorithms. However, in cluttered scenarios, data association matrix decomposition can easily lead to "combinatorial explosion," drastically increasing computational complexity and making it difficult to guarantee real-time tracking. In addition, traditional filtering methods rely on the assumption of a fixed motion model, which can lead to insufficient adaptability of the preset motion model when the target is highly maneuverable, resulting in decreased tracking accuracy or even target loss. Existing multi-target tracking methods generally treat filtering and association as two independent modules, with only one-way information transmission between them—filtering provides a prediction threshold for association, while the association result is fed back to update the filtering state. There is a lack of deep coupling and collaborative optimization mechanisms between the two, and potential information has not been effectively mined, thus limiting target tracking performance.

[0003] In recent years, artificial intelligence technologies such as deep learning and reinforcement learning have been gradually applied to the field of radar multi-target tracking. Researchers have attempted to utilize temporal networks and convolutional neural networks to extract target motion features, reducing the tracking system's dependence on fixed motion models and avoiding motion model mismatch problems. However, existing deep learning-based tracking methods mostly focus on replacing or enhancing single aspects, failing to fundamentally overcome the "information barrier" between filtering and correlation, and failing to construct a two-way feedback closed-loop framework to truly achieve co-evolution and mutual enhancement of filtering accuracy and correlation reliability. Therefore, there is an urgent need for... Summary of the Invention

[0004] To address the shortcomings mentioned above, a novel intelligent multi-target tracking method based on filter correlation and mutual feedback is explored. By constructing a two-way information interaction channel between filter state estimation and data correlation decision-making, a collaborative optimization mechanism is formed in which "correlation of correlation state guides correlation, and correlation result corrects correlation state". This enables high-precision, high-reliability, and strong real-time intelligent multi-target tracking in complex environments with strong clutter, high maneuverability, and densely intersecting multiple targets.

[0005] The technical solution of this invention is a radar multi-target intelligent tracking method based on filter correlation feedback, the method comprising:

[0006] Step 1: Given the historical associated track sequence of each target under the radar observation environment and the candidate measurement set of the current frame, normalize the input data and perform high-dimensional feature mapping on it;

[0007] Step 2: Establish a Track Temporal Feature Extraction (TTFE) module based on a detour-scanning Mamba network, which aims to extract target maneuver features containing real motion abrupt changes from historical tracks and output an initial target state estimate;

[0008] Step 3: Establish a Spatio-Temporal Association and State Correction (STASC) module based on dynamic sparse cross-attention. The initial target state estimate is used as the guiding query, and the current frame candidate measurement is used as the key and value to perform cross-attention calculation to obtain the initial attention weight distribution.

[0009] Step 4: Introduce information entropy theory to quantify the correlation uncertainty of the initial attention weight distribution, and perform dynamic sparse truncation to filter out redundant clutter interference and generate high-confidence sparse attention weights.

[0010] Step 5: Extract equivalent observation features using sparse attention weights to correct the target state estimate, and use the corrected target state as a feedback signal to input to the next level to guide target association, thus establishing a multi-feedback closed-loop collaborative optimization mechanism for association and filtering.

[0011] Step 6: Based on the mutual feedback and collaborative optimization results in Step 5, after lightweight network inference and inverse normalization processing in the last time step, the final multi-target tracking state estimation and measurement association matching decision are output in parallel.

[0012] Furthermore, the specific method of step 1 is as follows:

[0013] Step 1.1: For the j-th target, calculate the mean of its historical track sequence position state. and standard deviation :

[0014] ;

[0015] In the formula, L is the length of the historical track, and k is the time step number. This represents a scalar value indicating the location dimension. Let ε be the actual position component of target j at time k, and let ε be a minimal constant scalar to avoid division by zero.

[0016] Then calculate the statistics for the velocity dimension. and Combined into a complete mean statistic Sum of standard deviation statistics ;

[0017] Step 1.2: Track the target's historical flight path Set and current frame candidate measurements Normalization transformation of the set:

[0018] ;

[0019] In the formula, and This is the normalized set of historical tracks and current frame measurements;

[0020] Step 1.3: Normalize the... and Each layer maps to a high-dimensional feature space, resulting in a feature space of dimension . High-dimensional historical trajectory characteristics and high-dimensional candidate measurement embedding .

[0021] Furthermore, the specific method of step 2 is as follows:

[0022] Step 2.1: Process the high-dimensional historical track features output from Step 1. A local window is constructed using one-dimensional convolution to obtain extended track features. It is then sent to a roundabout scan of the Mamba network for mapping processing:

[0023] ;

[0024] In the formula, Represents the mapping function for Mamba networks. The kernel size used for the intermediate feature matrix output by Mamba. The step size is 1, and the fill size is 1.

[0025] Step 2.2: The roundabout scanning strategy performs three alternating updates at each time step. The first update utilizes information from the previous time step. Perform a pre-update, at which point the previous time step Not yet introduced Information; the second time step using previously integrated historical information. A second update is performed to transfer historical information and achieve local smoothing; the third update introduces the next time step. The information is corrected to form a two-way information flow verification;

[0026] Step 2.3: At each time step, take the last processed update result as the output, and map the indices as follows:

[0027] ;

[0028] The final output is a high-dimensional time series feature. This feature reflects both the target's maneuvering evolution and serves as a preliminary state estimate.

[0029] Furthermore, the specific method of step 3 is as follows:

[0030] Step 3.1: Using the high-dimensional temporal features output in Step 2 and the current frame measurement embedding output from step 1 Multiply by the corresponding learnable projection matrix respectively , , Construct a query matrix Key matrix Sum matrix The calculation formula is:

[0031] ;

[0032] Step 3.2: Calculate the query matrix With all key matrices The scaled dot product of the transpose of the vector is used to evaluate the spatiotemporal similarity scores between each candidate metric and the target trajectory, resulting in the attention score matrix. :

[0033] ;

[0034] In the formula, This is the scaling factor;

[0035] Step 3.3: For the attention score matrix... Perform a softmax normalization operation to obtain the initial attention weights for all candidate measurements in the current frame with respect to the target. .

[0036] Furthermore, the specific method of step 4 is as follows:

[0037] Step 4.1: Based on the initial attention weight distribution obtained in Step 3, extract the weight value corresponding to the i-th candidate metric. The uncertainty of the correlation in the point-to-point navigation matching process is quantified using the principle of information entropy:

[0038] ;

[0039] In the formula, M is the information entropy scalar of the attention weights, and M is the scalar of the number of candidate measurements in the current frame;

[0040] Step 4.2: Based on information entropy and the maximum possible entropy value under uniform distribution. The ratio of these values ​​is used to calculate the sparsity metric s of the attention weights, in order to comprehensively assess the confidence level of the match.

[0041] ;

[0042] Step 4.3: Dynamically determine the number of high-confidence measurements to retain based on the sparsity s. :

[0043] ;

[0044] In the formula, and Here, σ represents the preset upper and lower limits of the retention quantity, and σ is a learnable parameter scalar for control sensitivity; based on Generate a sparse mask matrix, force the attention scores of unselected clutter to a minimum and renormalize them to obtain the sharpened sparse attention weights. This enables adaptive clutter filtering.

[0045] Furthermore, the specific method of step 5 is as follows:

[0046] Step 5.1: Under the multi-head mechanism of the cross-attention layer, each head independently and in parallel performs the point-to-point association process, utilizing the sparse attention weights generated in Step 4. Log-value matrix Perform weighted summation to extract equivalent observation features that approximate the real target. :

[0047] ;

[0048] Step 5.2: The equivalent observation features generated by each attention head are spliced ​​and linearly fused, and then used as compensation information, they are superimposed on the input query in the form of additive correction through the residual structure to complete the correction of the target state estimate;

[0049] Step 5.3: Use the corrected state output of this layer as the query input for the next layer's attention mechanism, guiding the deep association process of the next layer. This allows the association weights to be recalibrated, enabling filtering and association to be synergistically optimized in a hierarchical progression. The average of the final generated association weights is used as the high-dimensional association result output by the STASC module. The state estimate of the last layer is taken as the high-dimensional state estimate output by the module. .

[0050] Furthermore, the specific method for step 6 is as follows:

[0051] Step 6.1: For the state estimation output, convert the high-dimensional state estimate output from Step 5... By transforming back to the target state dimension using a linear layer, we obtain the state estimation sequence in the normalized space. :

[0052] ;

[0053] In the formula, and These are used as the weight matrix and bias of the output layer, respectively;

[0054] Step 6.2: Reuse the mean value saved in Step 1.1 and standard deviation By performing inverse normalization and restoring the physical dimensions, the final state estimation matrix is ​​obtained. :

[0055] ;

[0056] Step 6.3: For the data association output, convert the high-dimensional association result finally output in Step 5. The input is fed into a lightweight multilayer perceptron to obtain the association score:

[0057] ;

[0058] In the formula, The associated score vector; MLP is a multilayer perceptron, consisting of two fully connected layers and nonlinear activation;

[0059] Step 6.4: Integrate the association score vectors of all targets to form an association score matrix, and use the Hungarian algorithm to further perform global hard matching of cross-target wayfinding, finally completing the complete wayfinding association decision.

[0060] In summary, the beneficial effects that can be obtained by the present invention are:

[0061] This invention proposes a roundabout scanning Mamba strategy, which effectively reduces maneuver misjudgments caused by noise interference and accurately captures the true maneuver evolution patterns. It introduces dynamic sparse cross-attention based on information entropy, effectively solving the problem of dense clutter distributing the association weight of the real target and improving the confidence of the association. It overcomes the decoupling constraints of key components, constructing a bidirectional feedback channel between association and filtering, enabling collaborative optimization of the two tasks and significantly reducing tracking errors in frequent maneuvering multi-target scenarios. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of a multi-target tracking scenario and the actual trajectory and measurement distribution in Embodiment 1 of the present invention.

[0063] Figure 2This is a line graph comparing the root mean square error of position and velocity of each tracking method under different clutter densities in Embodiment 1 of the present invention.

[0064] Figure 3 This is a comparison chart of multi-target trajectory tracking results under conditions of frequent target maneuvers in Embodiment 2 of the present invention.

[0065] Figure 4 This is the result of multi-target tracking under the condition of increased target number in Embodiment 3 of the present invention.

[0066] Figure 5 This shows the actual trajectory and measurement distribution of the target in a Monte Carlo experiment in Embodiment 3 of the present invention.

[0067] Figure 6 This is a graph showing the change of single-frame optimal subpattern assignment (OSPA) distance over time for each method in Embodiment 3 of the present invention. Detailed Implementation

[0068] The invention will now be explained in conjunction with simulation experiments:

[0069] I. Experimental Conditions

[0070] This invention employs multiple Monte Carlo simulation experiments to verify the robustness and tracking accuracy of the proposed method. The simulations generate 16,000 random multi-target tracking scenarios, each lasting 30 seconds, with a radar sampling interval of 1 second and a fixed number of targets (5). Target motion modes include CV, CA, and CT, randomly switching every 10 seconds. Radar measurements are generated by superimposing noise onto the actual target positions; the position noise is independent in the x and y directions, and the noise standard deviation is... Clutter points are uniformly generated within the region [−500m, 500m] near the actual measurement location of each target, and the number of clutter points generated near the actual measurement location of each target follows a Poisson distribution with parameter λ=8.

[0071] Regarding network parameter settings, input dimensions With output dimension All are set to 4, embedding dimension Set to 256, multi-head attention count The learning rate is set to 4, and the dropout rate is set to 0.1. The network architecture consists of 3 layers of meandering Mamba and 2 layers of dynamic sparse cross-attention, using the AdamW optimizer, with an initial learning rate of 5×10⁻⁶. -4The weight decay coefficient is 0.01, and the learning rate is dynamically adjusted during training using a cosine annealing learning rate scheduler. The network training adopts an end-to-end mode, and the total loss function is a weighted sum of the root mean square error (RMSE) reflecting state error and the binary cross entropy (BCE) reflecting association error.

[0072] II. Experimental Content and Results

[0073] Experiment 1: Ablation analysis of network sub-modules.

[0074] To verify the independent effectiveness and synergistic effect of the TTFE and STASC modules in improving target tracking accuracy, this experiment set up a tracking environment with 5 maneuvering targets, with a total duration of 80s and a radar sampling interval of 1s. The target's motion mode switched every 20s. Measurement noise with a standard deviation of 60m and process noise of 0.001 m / s² were added to the simulation scene. Clutter points were uniformly generated in the region [-500m, 500m] near the actual measurement position of each target, and their number followed a Poisson distribution with parameter λ=12 to test the correlation robustness of the method in a dense clutter environment. The scene parameter settings are shown in Table 1.

[0075] Table 1

[0076]

[0077] In the above test scenario, 300 Monte Carlo simulations were executed independently. The specific ablation configuration and results are shown in Table 2.

[0078] Table 2

[0079]

[0080] Experimental results show that when only the basic feature extraction is replaced with the detour scanning Mamba module of this invention, the mean velocity RMSE decreases from 84.13 m / s to 76.29 m / s, proving that this strategy can more accurately capture the true maneuvering evolution of the target. When only the dynamic sparse cross-attention module based on information entropy is introduced, the mean position RMSE drops significantly from 212.12 m to 162.04 m, effectively demonstrating that the method described in this invention adaptively filters candidate measurements and effectively removes redundant measurements. Furthermore, when the complete architecture of this invention, TTFE, is used in conjunction with the STASC module, the mean position error and mean velocity error reach their optimal values, confirming the dual advantages of the synergistic feedback between correlation and filtering.

[0081] Experiment 2: Comparative Experiment of Multi-Target Tracking During Frequent Maneuvers

[0082] To examine the tracking method described in this invention's ability to capture frequent maneuvers in cluttered environments and its robustness, a tracking scenario with five targets was constructed. Each target's duration was 100 seconds, and the target motion mode switching frequency was increased to once every 10 seconds. The radar sampling interval was 1 second. The standard deviation of the measurement noise was set to 60 m in both the x and y directions, and the clutter density was set to follow a Poisson distribution with parameter λ=8. The initial states and segmented motion modes of the five targets in the scenario are detailed in Table 3.

[0083] Table 3

[0084]

[0085] To facilitate objective comparison, the experiments selected the traditional model-driven algorithms IMM-CKF-AJPDA and IMM-GM-PHD, as well as the baseline network DeepAF, which is also based on data-driven computing. Each method was independently run for 500 Monte Carlo simulations, and the experimental results are shown in Tables 4 and 5.

[0086] Table 4

[0087]

[0088] Table 5

[0089]

[0090] Experimental results show that the MF-STASF method of this invention, with its strong noise resistance of dynamic sparse attention and the maneuverability of the detour scanning strategy, reduces the overall position error by approximately 16.5% to 24.4% and the velocity error by more than 26.0% compared to the comparative methods. Especially in complex situations where target tracks intersect and dense clutter coexist, the method of this invention still maintains the lowest tracking error and the fastest recovery speed, effectively suppressing error accumulation.

[0091] Experiment 3: Generalization ability test under the condition of doubling the number of targets

[0092] To verify the generalization ability of the MF-STASF method of this invention when directly applied to unknown and complex scenarios with doubled target numbers and more extreme maneuvering amplitudes without retraining, the pre-trained model was directly applied to a large-scale tracking scenario containing 10 maneuvering targets. The total observation time of the scenario was 60 seconds, the radar sampling interval was 1 second, the target motion mode was switched every 20 seconds, and extreme maneuvering states with accelerations as high as ±30 m / s² and turning rates as high as ±25° / s were introduced. Specific scenario parameter settings are shown in Table 6.

[0093] Table 6

[0094]

[0095] This experiment also used position, velocity RMSE, and single-frame OSPA distance as evaluation metrics to conduct 500 Monte Carlo simulations. The results are shown in Table 7.

[0096] Table 7

[0097]

[0098] The results show that despite the dual challenges of doubling the number of targets and dramatically increasing maneuverability, the proposed method still exhibits excellent generalization stability. Its mean position RMSE and mean velocity RMSE are improved by approximately 18.1% and 25.1% respectively compared to the baseline data-driven method DeepAF. Furthermore, the proposed method maintains the most stable and extremely low error distribution throughout the entire time period. The bidirectional feedback architecture of correlation and filtering proposed in this invention is further demonstrated to possess powerful adaptive adjustment and accurate tracking capabilities even when dealing with complex, unknown, and extreme scenarios that exceed the training distribution.

Claims

1. A radar multi-target intelligent tracking method based on filter correlation feedback, the method comprising: Step 1: Given the historical associated track sequence of each target under the radar observation environment and the candidate measurement set of the current frame, normalize the input data and perform high-dimensional feature mapping on it; Step 2: Establish a trajectory temporal feature extraction module based on a detour-scanning Mamba network, which aims to extract target maneuver features containing real motion abrupt changes from historical trajectories and output an initial target state estimate; Step 3: Establish a spatiotemporal association and state correction module based on dynamic sparse cross-attention. Use the initial target state estimation as the guiding query and the current frame candidate measurement as the key and value to perform cross-attention calculation to obtain the initial attention weight distribution. Step 4: Introduce information entropy theory to quantify the correlation uncertainty of the initial attention weight distribution, and perform dynamic sparse truncation to filter out redundant clutter interference and generate high-confidence sparse attention weights. Step 5: Extract equivalent observation features using sparse attention weights to correct the target state estimate, and use the corrected target state as a feedback signal to input to the next level to guide target association, thus establishing a multi-feedback closed-loop collaborative optimization mechanism for association and filtering. Step 6: Based on the mutual feedback and collaborative optimization results in Step 5, lightweight network inference and inverse normalization are performed in the last time step to output the final multi-target tracking state estimation and measurement association matching decision in parallel.

2. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 1, characterized in that, The specific method for step 1 is as follows: Step 1.1: For the j-th target, calculate the mean of its historical track sequence position state. and standard deviation : ； In the formula, L is the length of the historical track, and k is the time step number. This represents a scalar value indicating the location dimension. Let ε be the actual position component of target j at time k, and let ε be a minimal constant scalar to avoid division by zero. Then calculate the statistics for the velocity dimension. and Combined into a complete mean statistic Sum of standard deviation statistics ; Step 1.2: Track the target's historical flight path Set and current frame candidate measurements Normalization transformation of the set: ； In the formula, and This is the normalized set of historical tracks and current frame measurements; Step 1.3: Normalize the... and Each layer is mapped to a high-dimensional feature space through a linear layer, resulting in a feature space with dimension . High-dimensional historical trajectory characteristics and high-dimensional candidate measurement embedding .

3. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 2, characterized in that, The specific method for step 2 is as follows: Step 2.1: Process the high-dimensional historical track features output from Step 1. A local window is constructed using one-dimensional convolution to obtain extended track features. It is then sent to a roundabout scan of the Mamba network for mapping processing: ； In the formula, Represents the mapping function for Mamba networks. The kernel size used for the intermediate feature matrix output by Mamba. The step size is 1, and the fill size is 1. Step 2.2: The roundabout scanning strategy performs three alternating updates at each time step; the first update utilizes information from the previous time step. Perform a pre-update, at which point the previous time step Not yet introduced Information; the second time step using previously integrated historical information. A second update is performed to transfer historical information and achieve local smoothing; the third update introduces the next time step. The information is corrected to form a two-way information flow verification; Step 2.3: At each time step, take the last processed update result as the output, and map the indices as follows: ； The final output is a high-dimensional time series feature. This feature reflects both the target's maneuvering evolution and serves as a preliminary state estimate.

4. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 3, characterized in that, The specific method for step 3 is as follows: Step 3.1: Using the high-dimensional temporal features output in Step 2 and the current frame measurement embedding output from step 1 Multiply by the corresponding learnable projection matrix respectively , , Construct a query matrix Key matrix Sum matrix The calculation formula is: ； Step 3.2: Calculate the query matrix With all key matrices The scaled dot product of the transpose of the vector is used to evaluate the spatiotemporal similarity scores between each candidate metric and the target trajectory, resulting in the attention score matrix. : ； In the formula, This is the scaling factor; Step 3.3: For the attention score matrix... Perform a softmax normalization operation to obtain the initial attention weights for all candidate measurements in the current frame with respect to the target. .

5. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 4, characterized in that, The specific method for step 4 is as follows: Step 4.1: Based on the initial attention weight distribution obtained in Step 3, extract the weight value corresponding to the i-th candidate metric. The uncertainty of the correlation in the point-to-point navigation matching process is quantified using the principle of information entropy: ； In the formula, M is the information entropy scalar of the attention weights, and M is the scalar of the number of candidate measurements in the current frame; Step 4.2: Based on information entropy and the maximum possible entropy value under uniform distribution. The ratio of these values ​​is used to calculate the sparsity metric s of the attention weights, in order to comprehensively assess the confidence level of the match. ； Step 4.3: Dynamically determine the number of high-confidence measurements to retain based on the sparsity s. : ； In the formula, and Here, σ represents the preset upper and lower limits of the retention quantity, and σ is a learnable parameter scalar for control sensitivity; based on Generate a sparse mask matrix, force the attention scores of unselected clutter to a minimum and renormalize them to obtain the sharpened sparse attention weights. This enables adaptive clutter filtering.

6. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 5, characterized in that, The specific method for step 5 is as follows: Step 5.1: Under the multi-head mechanism of the cross-attention layer, each head independently and in parallel performs the point-to-point association process, utilizing the sparse attention weights generated in Step 4. Log-value matrix Perform weighted summation to extract equivalent observation features that approximate the real target. : ； Step 5.2: The equivalent observation features generated by each attention head are spliced ​​and linearly fused, and then used as compensation information, they are superimposed on the input query in the form of additive correction through the residual structure to complete the correction of the target state estimate; Step 5.3: Use the corrected state output of this layer as the query input for the next layer's attention mechanism, guiding the deep association process of the next layer. This allows the association weights to be recalibrated, enabling filtering and association to be synergistically optimized in a hierarchical progression. The average of the final generated association weights is used as the high-dimensional association result output by the STASC module. The state estimate of the last layer is taken as the high-dimensional state estimate output by the module. .

7. The intelligent multi-target tracking method for radar based on filter correlation feedback as described in claim 6, characterized in that, The specific method for step 6 is as follows: Step 6.1: For the state estimation output, convert the high-dimensional state estimate output from Step 5... By transforming back to the target state dimension using a linear layer, we obtain the state estimation sequence in the normalized space. : ； In the formula, and These are used as the weight matrix and bias of the output layer, respectively; Step 6.2: Reuse the mean value saved in Step 1.1 and standard deviation By performing inverse normalization and restoring the physical dimensions, the final state estimation matrix is ​​obtained. : ； Step 6.3: For the data association output, convert the high-dimensional association result finally output in Step 5. The input is fed into a lightweight multilayer perceptron to obtain the association score: ； In the formula, The associated score vector; MLP is a multilayer perceptron, consisting of two fully connected layers and nonlinear activation; Step 6.4: Integrate the association score vectors of all targets to form an association score matrix, and use the Hungarian algorithm to further perform global hard matching of cross-target point navigation, and finally complete the complete point navigation association decision.

Images