Transform-based radar track association method

Abstract

The present application relates to the technical field of radar track association, and more particularly to a radar track association method based on transform, which comprises the following steps: obtaining current frame radar point cloud data, performing feature embedding processing on the current frame radar point cloud data to obtain embedded features, adding time coding to the embedded features to form current frame point features; splicing the current frame point features with historical point features stored in a historical feature memory to construct an input sequence of a Transformer encoder, performing space-time context coding and feature enhancement on the input sequence based on the Transformer encoder to obtain enhanced point features; based on the enhanced point features, performing processing through a classification network and screening according to a preset classification threshold to obtain a suspected target detection result, wherein the detection result comprises original observation data and deep feature representation; according to the suspected target detection result and an existing track list, performing track prediction and multi-modal similarity matching, and updating the track list by using an optimal matching algorithm; and extracting and outputting the final state information of the track from the updated track list.

Description

Technical Field

[0001] This invention relates to the field of radar trajectory association technology, and more specifically to a radar trajectory association method based on transform. Background Technology

[0002] The widespread application of drones and low-altitude aircraft has placed higher demands on radar target tracking technology. Doppler radar, capable of providing point cloud data containing position and velocity information, is a crucial means of target detection and tracking. However, in practical applications, radar echoes not only contain real targets but also a large number of false targets due to ground clutter, weather interference, and electromagnetic noise, posing a significant challenge to traditional tracking methods. Existing target association methods are mainly based on traditional algorithms such as nearest neighbor and probabilistic data association. These methods perform well when target density is low, but their association error rate increases significantly in dense clutter environments, and they struggle to handle frequent target obstruction and cross-maneuvering.

[0003] Furthermore, traditional methods rely heavily on manually designed features and association rules, which are poorly adaptable to complex and ever-changing environments and cannot effectively utilize the rich spatiotemporal context information in point clouds. Although some studies in recent years have attempted to introduce machine learning methods into data association, most methods can only process local information and lack the ability to collaboratively perceive global spatiotemporal relationships, resulting in limited improvement in association performance in complex scenarios.

[0004] Therefore, there is an urgent need for a novel association method that can fully utilize the spatiotemporal characteristics of point clouds and has strong context awareness capabilities to improve the accuracy and robustness of target tracking in complex environments. Summary of the Invention

[0005] In view of this, the present invention provides a radar trajectory association method based on transform, which aims to solve the technical problems of insufficient association accuracy in dense clutter environments, poor adaptability to complex target motion patterns, and difficulty in effectively utilizing global spatiotemporal information in existing radar target tracking technologies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A transform-based radar trajectory association method includes the following steps:

[0008] S1. Obtain the radar point cloud data of the current frame, perform feature embedding processing on it, and add time encoding to the embedded features to form the point features of the current frame;

[0009] S2. Concatenate the current frame point features with the historical point features stored in the historical feature memory to construct the input sequence of the Transformer encoder. Perform spatiotemporal context encoding and feature enhancement on the input sequence based on the Transformer encoder to obtain enhanced point features.

[0010] S3. Based on the enhanced point features, the suspected target detection results are obtained by processing through a classification network and filtering according to a preset classification threshold. The detection results include the original observation data and depth feature representation.

[0011] S4. Based on the suspected target detection results and the existing trajectory list, the trajectory list is updated by trajectory prediction and multimodal similarity matching, and the optimal matching algorithm is used.

[0012] S5. Extract and output the final status information of the trajectory from the updated trajectory list.

[0013] Furthermore, before S1, system initialization is also included. The system initialization includes defining system hyperparameters, including feature dimension, time sliding window size, target classification threshold and maximum number of temporarily lost frames for trajectory, and initializing an empty list of active trajectories.

[0014] Furthermore, in S2, the historical feature memory is managed through a first-in-first-out (FIFO) queue mechanism, and the oldest historical feature in the historical feature memory is covered by the current frame point feature.

[0015] The Transformer encoder comprises multiple coding layers, each of which performs a multi-head self-attention mechanism and a feedforward neural network operation, and each submodule is equipped with residual connections and layer normalization.

[0016] Furthermore, the multi-head self-attention mechanism calculates a high-dimensional correlation score between points in the input sequence to determine the amount of information that should be absorbed from other points in the input sequence when updating the current point feature. This information is used to enhance the spatiotemporal context representation of the point feature.

[0017] Furthermore, in S3, the classification network is a feedforward neural network used to distinguish between the target and clutter. The feedforward neural network takes the enhancement point features as input, outputs a confidence score between 0 and 1 through the Sigmoid activation function, and compares it with the preset classification threshold.

[0018] Furthermore, in S4, the trajectory association includes using its state estimator to predict the expected position of each existing trajectory in the existing trajectory list at the current moment, so as to provide prior location information for association.

[0019] Furthermore, in S4, the multimodal similarity matching includes constructing a cost matrix. The calculation of the cost matrix is ​​based on the cosine similarity between the latest depth features of the existing trajectory and the depth feature representation of the suspected target detection result, as well as the spatial distance between the predicted position of the existing trajectory and the position of the original observation data of the suspected target detection result.

[0020] Furthermore, the optimal matching algorithm is the Hungarian algorithm, which processes the cost matrix to obtain three sets of results globally: successfully matched trajectory-detection pairs, unmatched existing trajectories, and unmatched new detection points.

[0021] Furthermore, in step S4, updating the trajectory list includes:

[0022] For a successfully matched trajectory, update its state estimator and depth feature representation with the corresponding suspected target detection result, and reset the temporary loss counter; for an unmatched existing trajectory, increment the temporary loss counter, and if the number of consecutive unmatches exceeds the maximum number of temporary loss frames, remove it from the trajectory list.

[0023] Furthermore, the updated trajectory list also includes: for the new unmatched detection point, creating a new trajectory and adding it to the trajectory list;

[0024] In step S5, the final state information includes a unique identifier for the trajectory, the filtered estimated three-dimensional position, the estimated three-dimensional velocity, and its confidence level.

[0025] Compared to existing technologies, this invention can jointly model multi-frame point cloud data through a Transformer encoder-decoder architecture, and realize global interaction between points using a self-attention mechanism. It can capture long-distance dependencies (such as the correlation between historical trajectories and the current point), thereby effectively distinguishing real targets from clutter and significantly reducing the false association rate. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a radar trajectory association method based on transform according to the present invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See appendix Figure 1 This invention is achieved through the following technical solution: First, a spatiotemporal feature extraction network is constructed using a Transformer-based encoder-decoder architecture to jointly process multi-frame radar point cloud data; a self-attention mechanism is used to model the global relationship between points, capturing long-distance dependencies and spatiotemporal context information; then, a learnable trajectory query vector and enhanced point features are used for cross-attention calculation to achieve end-to-end association between target trajectories and point traces; finally, a multi-task learning framework is used to simultaneously complete point trace classification and trajectory association tasks. This method leverages the advantages of Transformer in long sequence modeling and global relationship awareness, while avoiding the complexity of manually designing association rules through an end-to-end learning approach, significantly improving association performance and robustness in dense clutter and complex maneuvering environments.

[0030] Step 1: Radar point cloud data preprocessing and feature embedding, system initialization:

[0031] Initialize an empty list of active tracks. Define system hyperparameters: The feature dimension is 128.

[0032] The time sliding window size is 5 frames. This is the time range that the Transformer can "see" and is key to the associated trajectory. The target classification threshold is 0.7. Values ​​higher than this are considered suspected drones. The maximum number of frames temporarily lost for a trajectory is 3. A trajectory is deleted if it fails to match for 3 consecutive frames. Load the parameters of a pre-trained Transformer model (which has learned to distinguish between drones and clutter).

[0033] Data preprocessing and feature embedding (processing the current frame):

[0034] The original point cloud of the current frame is converted into high-dimensional features and timestamped to prepare for subsequent spatiotemporal association.

[0035] Input: Point cloud of the current frame , All points detected by the radar at all times Each point (3D coordinates + Doppler velocity). Historical feature memory M: a queue storing the most recent... The final point features after processing by the Transformer at each time step.

[0036] Processing point clouds at the current time:

[0037] 1. Feature embedding: Projecting low-dimensional point data into a high-dimensional feature space, enabling the model to learn more complex patterns;

[0038] formula:

[0039] in:

[0040] For the first The original input feature vector of each radar point For the transformed number High-dimensional feature embedding of points, To embed the weight matrix, This is the bias vector.

[0041] 2. Add time encoding: Add the same temporal identifier to all points in the current frame so that the model knows that these points belong to the same time.

[0042] 3. Constructing the model input sequence: The new features of the current frame are concatenated with the historical features in memory M to form the input of the Transformer. This sequence contains features from all points within the time window;

[0043] Output:

[0044] The concatenated sequence is fed into the Transformer encoder. The memory is also updated: the oldest time-sequence features in memory are overwritten with the latest sequence (FIFO queue).

[0045] Step 2: Spatiotemporal context coding

[0046] By leveraging the self-attention mechanism of the Transformer, all points can interact globally. Points in the current frame "query" points in historical frames through attention weights to find the most relevant points, thereby determining whether they are a persistent target or newly appearing clutter.

[0047] Input: The output from step one, which is the concatenated feature sequence, containing point features for all times within the time window.

[0048] Processing procedure:

[0049] The sequence is processed through a multi-layer Transformer encoder. Each layer performs two core operations:

[0050] Multi-head self-attention mechanism:

[0051] This is the core of this step. Each point plays three roles: Each point in the sequence calculates its relevance score with all other points in the sequence. This score determines how much information should be absorbed from other points when updating the current point's features. If a point in the current frame finds that its features are highly similar to those of a point in a historical frame (resulting in a high calculated attention weight), then it will absorb more information from that historical point, thereby enhancing its own feature representation. This usually means that it is likely the current observation of some persistent target. If a point in a historical frame receives high attention from multiple points in the current frame, it indicates that the trajectory it represents is being continued and confirmed.

[0052] Feedforward neural networks:

[0053] Following the attention mechanism, the features of each point undergo further nonlinear transformation and feature enhancement through a feedforward neural network.

[0054] Each submodule is equipped with residual connections and layer normalization to ensure the stability of the training process and the effective propagation of gradients.

[0055] Output:

[0056] Enhanced feature sequence: After processing through all layers, the features of each point contain rich spatiotemporal context information. The points in the current frame not only know their own observations, but also know the degree of their correlation with historical trajectories.

[0057] Update memory: Replace the initial features in memory with the final processed features from the current time step. This means that memory always stores the most essential and representative feature representations from each time step, providing high-quality historical information for processing in the next time step.

[0058] Step 3: Suspected Drone Detection (Classification)

[0059] Objective: Based on the enhanced point features after Transformer enhancement, intelligently determine whether each radar point in the current frame is suspected to be a UAV target, thereby filtering out clutter and noise.

[0060] Input: The output from step two, i.e., the complete feature sequence after spatiotemporal context encoding. From this sequence, we extract only the feature subset belonging to the current time t. These features have been fused with the contextual information of historical frames and have stronger discriminative power.

[0061] Processing procedure:

[0062] For each point feature at the current time, perform the following operations:

[0063] Classification decision: The deep features of each point are fed into a lightweight classification network. This network is a two-layer feedforward neural network (MLP) responsible for learning how to distinguish between drone targets and clutter from the features.

[0064] Confidence Score: The output of the classification network is activated by a sigmoid function, producing a confidence score between 0 and 1. This score represents the probability that the point is a drone; the higher the score, the greater the likelihood that it is a drone.

[0065] Threshold filtering: Set the classification threshold to 0.7. Only points with a confidence score exceeding this threshold will be retained as suspected drone detection results.

[0066] Output:

[0067] The current frame detection result set contains information on all points identified as suspected drones. Each detection result is a tuple containing: raw observation data, i.e., the original measurement value of the point in the radar coordinate system, including 3D coordinates and Doppler velocity information. This data will be used for subsequent target state estimation and trajectory filtering. The deep feature representation is a high-dimensional feature vector of the point after deep encoding by a Transformer. This feature is rich in spatiotemporal context information and will be used for the next step of data association, calculating the similarity with existing trajectories.

[0068] Step 4: Trajectory Association and Update

[0069] Objective: To correctly match the suspected drone detection points identified in the current frame with the existing flight trajectories in the system to achieve continuous target tracking.

[0070] enter:

[0071] Current detection results: The output from step three contains the set of all points in the current frame that are classified as suspected drones, with each point containing its original observation data and depth features.

[0072] Existing Track List: The collection of all drone tracks currently being tracked in the system. Each track contains a unique identifier, the latest feature representation, a state predictor (for estimating position and velocity), and a timeout counter.

[0073] Processing procedure:

[0074] Trajectory prediction: For each existing trajectory, its expected position at the current moment is predicted using its state estimator. This provides prior location information for subsequent association.

[0075] Similarity calculation: Construct a cost matrix to calculate the matching cost between each trajectory and each detection point. The main basis is:

[0076] Feature similarity: Calculate the cosine similarity between the latest features of the trajectory and the depth features of the detection point. The higher the similarity, the lower the matching cost.

[0077] Location proximity: Combining the spatial distance between the predicted location and the detected location, the closer the distance, the greater the probability of a match.

[0078] Optimal matching: The Hungarian algorithm is used to process the cost matrix to find the globally optimal matching scheme, resulting in three sets of results:

[0079] Successfully matched trajectory-detection pairs;

[0080] Existing trajectories that do not match;

[0081] Unmatched new detection point.

[0082] Track update:

[0083] For successfully matched trajectories: update the state estimator with the detected location, update the trajectory feature representation with the latest features of the detected points, and reset the temporary loss counter.

[0084] Unmatched tracks: Increase the temporary loss count. If the number of consecutive unmatches exceeds the threshold, the target is considered to have disappeared and removed from the tracking list.

[0085] Unmatched detection: As a newly emerging suspected drone, create a new trajectory and add it to the tracking list.

[0086] Output:

[0087] Updated list of active tracks: Contains all drone tracks that are still being tracked, with the status and characteristics of each track updated accordingly.

[0088] Step 5: System Output

[0089] Input: The output from step four, i.e., the updated list of trajectories.

[0090] Processing procedure:

[0091] Iterate through all active trajectories and extract the filtered state information for the current time step from the state estimator of each trajectory, including smoothed 3D position and 3D velocity estimates. Organize this information according to a predetermined format.

[0092] Output:

[0093] The final tracking result at the current moment: a structured list containing complete information on all tracked suspected drones, with each record containing:

[0094] Track ID: A unique identifier for the target;

[0095] Estimated position: Filtered position in the X, Y, and Z coordinate directions;

[0096] Estimated velocity: velocity components in the X, Y, and Z directions;

[0097] Confidence level: The degree to which the trajectory is believed to exist;

[0098] This output provides accurate target movement information for subsequent situational awareness, threat assessment, and decision response.

[0099] This invention uses a Transformer encoder-decoder architecture to jointly model multi-frame point cloud data and utilizes a self-attention mechanism to achieve global interaction between points. It can capture long-distance dependencies (such as the correlation between historical trajectories and the current point), thereby effectively distinguishing real targets from clutter and significantly reducing the false association rate.

[0100] An end-to-end joint learning framework enhances adaptability to complex motion patterns. Traditional methods rely on manually designed association rules and features, making it difficult to handle complex scenarios such as frequent target occlusion and cross-maneuvering. This invention uses learnable trajectory query vectors and point features for cross-attention calculation, achieving end-to-end optimization through multi-task learning (point classification + trajectory association). This enables the model to adaptively learn the target's motion patterns and appearance features without the need for manually designed complex rules, thus improving tracking robustness in maneuvering targets and occluded scenarios.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0102] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A radar trajectory association method based on transform, characterized in that, Includes the following steps: S1. Obtain the radar point cloud data of the current frame, perform feature embedding processing on it to obtain embedded features, add time encoding to the embedded features to form the point features of the current frame; S2. Concatenate the current frame point features with the historical point features stored in the historical feature memory to construct the input sequence of the Transformer encoder. Perform spatiotemporal context encoding and feature enhancement on the input sequence based on the Transformer encoder to obtain enhanced point features. S3. Based on the enhanced point features, the suspected target detection results are obtained by processing through a classification network and filtering according to a preset classification threshold. The detection results include the original observation data and depth feature representation. S4. Based on the suspected target detection results and the existing trajectory list, the trajectory list is updated by trajectory prediction and multimodal similarity matching, and the Hungarian algorithm is used. The multimodal similarity matching includes constructing a cost matrix, which is calculated based on the cosine similarity between the latest depth features of the existing trajectory and the depth feature representation of the suspected target detection result, as well as the spatial distance between the predicted position of the existing trajectory and the position of the original observation data of the suspected target detection result. S5. Extract and output the final status information of the trajectory from the updated trajectory list.

2. The radar trajectory association method based on transform according to claim 1, characterized in that, Before S1, system initialization is also included. The system initialization includes defining system hyperparameters, including feature dimension, time sliding window size, target classification threshold and maximum number of temporarily lost frames for trajectory, and initializing an empty list of active trajectories.

3. The radar trajectory association method based on transform according to claim 1, characterized in that, In step S2, the historical feature memory is managed through a first-in-first-out (FIFO) queue mechanism, and the oldest historical feature in the historical feature memory is covered by the current frame point feature. The Transformer encoder comprises multiple coding layers, each of which performs a multi-head self-attention mechanism and a feedforward neural network operation, and each submodule is equipped with residual connections and layer normalization.

4. The radar trajectory association method based on transform according to claim 3, characterized in that, The multi-head self-attention mechanism calculates a high-dimensional correlation score between points in the input sequence to determine the amount of information that should be absorbed from other points in the input sequence when updating the current point feature. This information is used to enhance the spatiotemporal context representation of the point feature.

5. The radar trajectory association method based on transform according to claim 1, characterized in that, In step S3, the classification network is a feedforward neural network used to distinguish between targets and clutter. The feedforward neural network takes the enhancement point features as input, outputs a confidence score between 0 and 1 through the Sigmoid activation function, and compares it with the preset classification threshold.

6. The radar trajectory association method based on transform according to claim 1, characterized in that, In step S4, trajectory association includes using the state estimator of each existing trajectory in the existing trajectory list to predict its expected position at the current moment, so as to provide prior location information for association.

7. The radar trajectory association method based on transform according to claim 1, characterized in that, The Hungarian algorithm processes the cost matrix to obtain three sets of results globally: successfully matched trajectory-detection pairs, unmatched existing trajectories, and unmatched new detection points.

8. The radar trajectory association method based on transform according to claim 1, characterized in that, In step S4, updating the trajectory list includes: For a successfully matched trajectory, update its state estimator and depth feature representation with the corresponding suspected target detection result, and reset the temporary loss counter; for an unmatched existing trajectory, increment the temporary loss counter, and if the number of consecutive unmatches exceeds the maximum number of temporary loss frames, remove it from the trajectory list.

9. The radar trajectory association method based on transform according to claim 8, characterized in that, Updating the trajectory list also includes: for new detection points that do not match, creating new trajectories and adding them to the trajectory list; In step S5, the final state information includes a unique identifier for the trajectory, the filtered estimated three-dimensional position, the estimated three-dimensional velocity, and the confidence level.

Images