An integrated processing method for target detection and tracking in radar multi-frame images

Abstract

The present invention relates to a radar multi-frame image target detection and tracking integrated processing method, which belongs to the field of radar target detection. It includes: 1) constructing an inter-frame echo image data set; 2) constructing an improved Yolov5 detection and tracking network based on the constructed inter-frame echo image data set when realizing target detection and tracking; 3) using the inter-frame echo image data set obtained after preprocessing in step 1 as input, and pre-training the improved Yolov5 detection and tracking network for iterative optimization; 4) motion parameter estimation, extracting the motion characteristics of the target; 5) model parameter migration training; 6) inputting the processed signal into the improved Yolov5 detection and tracking network after model parameter migration training for target detection and tracking. The present invention can realize target detection in complex environment background and target diverse motion tracking, and improve the algorithm generalization ability and detection and tracking ability.

Description

Technical Field

[0001] The invention relates to a radar multi-frame image target detection and tracking integrated processing method, belonging to the field of radar target detection. Background Art

[0002] As an effective marine monitoring tool, radar is widely used in fields such as maritime transportation, maritime rescue and fishery detection. Target detection and tracking are the main tasks of maritime surveillance radar. However, due to the complex environmental background, interference from clutter or noise, and extreme target motion, reliable and robust target detection and tracking methods are still a key technology that needs to be studied urgently. Therefore, a method with good adaptability to moving targets in complex and changing environments is needed. Traditional detection and tracking methods are usually based on statistical theory, but due to the diversity of complex environments, these methods are difficult to achieve high-performance detection and tracking, and are prone to missed detection and loss of targets. In contrast, deep learning has demonstrated stronger generalization capabilities, and its application in the field of radar is also expanding. By utilizing multi-frame data, deep learning can achieve detection and tracking integration, and improve the accuracy and stability of target detection and tracking.

[0003] There are many difficulties and problems in radar image detection, including: complex background and clutter interference. The complex background of the ocean environment and the clutter in the radar echo can easily lead to false detection; the diversity of target motion. The motion forms of marine targets are complex and changeable, which increases the difficulty of tracking; scale change. The scale change of the target will affect the accuracy of detection and tracking; occlusion and rotation. When the target is occluded or rotated, the tracker is prone to failure. At present, deep learning has shown significant advantages in solving these problems. Target detection algorithms can be divided into one-stage algorithms and two-stage algorithms: one-stage algorithms directly extract features from the image and use classifiers or regressors to predict the location and category of objects, such as YOLO, SSD, etc. These algorithms have fast detection speed and low computational complexity, but the detection accuracy is relatively low. The two-stage algorithm divides target detection into two subtasks: target proposal and target classification. First, extract the area that may contain the target (target proposal), and then classify these areas to determine whether they contain the target and its category. Common two-stage algorithms include Faster RCNN, etc. These algorithms have higher detection accuracy, but slower speed and higher computational complexity.

[0004] In terms of target tracking, although deep neural networks have advantages in accuracy, they are less robust. This is because existing algorithms have almost no separate modules in the design of deep neural networks to complete the state estimation of the target. At present, trackers still rely on classification to perform multi-scale searches to estimate targets, but this method has difficulty dealing with inaccurate or even failed tracking results caused by changes in target scale. The shortcomings of existing methods include: lack of robust state estimation modules, and low robustness of existing tracking algorithms when dealing with target scale changes, fast motion and complex backgrounds. High computational complexity, especially the two-stage detection algorithm, although it has high accuracy, but the computational complexity and speed are not ideal. Poor adaptability to environmental changes, traditional methods are difficult to cope with complex and changing environments. Although deep learning has improved, its adaptability still needs to be further improved.

[0005] Therefore, it is an urgent problem to realize the integration of detection and tracking by using multi-frame data and further improve the adaptability of the algorithm in complex environments and diverse targets. Summary of the invention

[0006] The purpose of the present invention is to solve the deficiencies of the above-mentioned prior art and provide a radar multi-frame image target detection and tracking integrated processing method to achieve complex environment background target detection and target diverse motion tracking, and improve the algorithm generalization ability and detection and tracking ability.

[0007] The present invention provides a radar multi-frame image target detection and tracking integrated processing method, which is characterized by comprising the following steps:

[0008] Step 1: Preprocess the radar echo signal and construct an inter-frame echo image data set to provide input data for subsequent target detection and tracking;

[0009] Step 2: When realizing target detection and tracking, based on the constructed inter-frame echo image dataset, a detection and tracking network based on improved Yolov5 is constructed; the detection and tracking network based on improved Yolov5 is divided into two parts: 1) Improved Yolov5 detection module: Combine the Transformer module and the attention module CBAM with Yolov5 to process the inter-frame sequence images to realize target detection; 2) DeepSORT tracking module: Realize the tracking of radar targets and use the detection results output by the improved Yolov5 detection module to track targets;

[0010] Step 3: Taking the inter-frame echo image dataset obtained after preprocessing in step 1 as input, the pre-trained improved Yolov5 detection and tracking network is iteratively optimized. During the training process, the network weights are continuously adjusted to improve the accuracy of target detection, and the pre-trained weight parameters are obtained;

[0011] Step 4: Motion parameter estimation, extracting the motion characteristics of the target to provide basic data for subsequent analysis;

[0012] Step 5: Model parameter transfer training: Use the pre-trained weight parameters obtained in step 3 to initialize the weights of the improved Yolov5 detection and tracking network, adapt to the new task through global fine-tuning, and change the source domain and target domain at the same time to enhance the generalization ability of the network;

[0013] Step 6: Input the processed signal into the improved Yolov5 detection and tracking network after model parameter migration training for target detection and tracking to complete the final target detection and tracking task.

[0014] Preferably, the specific steps of step 1 are as follows:

[0015] The radar echo data is preprocessed into a P-display image, and an inter-frame echo image data set is constructed. The extracted data is used to draw a P-display image. The P-display image is a two-dimensional image that shows the distance and azimuth of the radar relative to its location:

[0016] (1);

[0017] (2);

[0018] Where Dist_bin is the distance from the grid point AOEP is the angle data, X, Y These are the x and y coordinates when drawing the P display image.

[0019] Preferably, the specific steps of step 2 are as follows:

[0020] (1) Improved Yolov5 detection module:

[0021] ①Introduction of Transformer

[0022] The Transformer structure is introduced into the prediction head of YOLOv5, and the self-attention mechanism is used to enhance the model's feature extraction and context information capture capabilities.

[0023] (3);

[0024] in, are query, key, and value matrices respectively, is the dimension of the bond matrix;

[0025] ②Multi-scale feature fusion

[0026] Feature extraction: Multi-scale features are extracted through different convolutional layers. YOLOv5 uses the CSPNet (CrossStage Partial Networks) structure, which can effectively extract multi-level features.

[0027] Feature pyramid construction: Fusing shallow features (rich in details) with deep features (contextual information) using upsampling and downsampling techniques;

[0028] Upsampling and fusion: Use upsampling to resize shallow features to align them with deep features. The aligned features are concatenated or added to enhance feature expression capabilities.

[0029] ③ References to the Convolutional Attention Mechanism

[0030] The algorithm neck network adopts a mixed-domain convolutional attention module CBAM module, which is divided into two parts: one is channel attention, and the other is spatial attention. The CBAM module uses the intermediate feature map and the module to infer the attention of the image along the channel and spatial dimensions in sequence, and then combines the attention layer with the feature map for adaptive feature modification. The formula is as follows:

[0031] (4);

[0032] (5);

[0033] In the formula, For sigmoid operation, 7*7 represents the size of the convolution kernel. Represents the input feature map, AvgPool, MaxPool and MLP represent average pooling, maximum pooling and multi-layer perceptron respectively. CBAM is a lightweight module that can be integrated into the CNN architecture and can be trained in an end-to-end manner.

[0034] (2) DeepSORT tracking module

[0035] DeepSORT detects the target in each frame based on the output of the improved Yolov5 detection module, represents and describes the target using multi-feature fusion technology, and tracks the target using the SORT algorithm. The workflow is as follows:

[0036] Initialization: Detect the target in the first frame, create a corresponding trajectory for each target, initialize the motion variable of the Kalman filter, and predict the corresponding target box through the Kalman filter. The trajectory state at this time is unproven;

[0037] IOU matching: Perform IOU matching on the current frame’s target detection box with the target box predicted by the trajectory in the previous frame, and calculate its cost matrix:

[0038] (6);

[0039] Hungarian algorithm: The cost matrix is ​​used as the input of the Hungarian algorithm to obtain the linear matching result;

[0040] Loop update: Repeat IOU matching and Hungarian algorithm steps until a confirmed trajectory appears or the video frame ends;

[0041] Cascade matching: Kalman filter prediction is performed on the confirmed trajectory, and its target box is cascade matched with the detection (matching is performed using appearance features and motion information).

[0042] Rematching: After cascade matching, the unmatched tracks and detections are matched again by IOU, and the Hungarian algorithm is used for linear matching. The matching results are the same as before in three cases;

[0043] Loop processing: Repeat the steps of cascade matching and re-matching until all video frames are processed.

[0044] Preferably, the specific steps of step 3 are as follows:

[0045] Using the Adam gradient descent method, given the parameters θ ,gradient g t , momentum parameter β 1. Learning rate parameter β 2 and constant , first, an exponentially weighted moving average is performed to estimate the first-order moment of the gradient m t and the second moment v t :

[0046] (7);

[0047] Then, the parameters are updated according to the corrected first-order and second-order moments:

[0048] (8);

[0049] in, t is the number of current iterations, is the learning rate.

[0050] Preferably, the specific steps of step 4 are:

[0051] The DeepSORT algorithm uses a Kalman filter to estimate the state of each target, assuming that the state vector of the target is:

[0052] (9);

[0053] In the formula, The goal is The position of the frame (center point coordinates). is the size (area) of the target. is the aspect ratio of the target, is the velocity component of the target;

[0054] Kalman filter prediction:

[0055] (10);

[0056] (11);

[0057] in, is the forecast error covariance matrix, is the process noise covariance matrix;

[0058] Kalman filter update:

[0059] (12);

[0060] (13);

[0061] (14);

[0062] in, is the observation matrix, R is the observation noise covariance matrix, is the Kalman gain.

[0063] Through the update step of the Kalman filter, the velocity component of the target can be obtained Estimated value:

[0064] (15);

[0065] course is the direction angle of the target movement, which is calculated by the velocity component:

[0066] (16);

[0067] in, Is the two-variable inverse tangent function, used to calculate the four-quadrant inverse tangent angle.

[0068] Preferably, the specific steps of step 5 are:

[0069] The pre-trained model parameter file initializes the weights of the model network, replacing the original random initialization operation, and performs global fine-tuning. The rest of the training process proceeds as usual. Transfer learning focuses on the active domain. and a target domain The case where the source domain uses express, , Represents data samples and corresponding motion target category labels respectively, and the target domain uses Represents. Given a source domain and learning tasks , target domain and learning tasks ,The purpose of transfer learning is to obtain the source domain and learning tasks The knowledge in the target domain can help improve the prediction function in the target domain of learning, including or , the overall process of model migration is: ,in, represents the model's pre-trained weights in the source domain, Represents the model target domain training weights.

[0070] Preferably, the specific steps of step 6 are:

[0071] First, assume that the input image dataset is preprocessed to ensure that the image is suitable for the model input; then, use the pre-trained improved YOLOv5 detection and tracking model to detect the target on the preprocessed image and obtain the bounding box, confidence and category label of each target; then, initialize the DeepSORT tracking module and initialize the detected targets into the DeepSORT tracking module; by detecting the target in each subsequent frame and updating the DeepSORT tracking module, it is possible to track multiple targets and assign a unique ID to each target; finally, draw the detected target bounding box and tracking ID on the original image, generate a visualization result, and save the detection and tracking results to a file or database for subsequent analysis.

[0072] Compared with the prior art, the radar multi-frame image target detection and tracking integrated processing method of the present invention has the following beneficial effects:

[0073] (1) The present invention has effective detection and tracking performance for changing environments and moving targets. The invention targets the inter-frame information of moving targets, where the inter-frame information contains the target's motion trajectory, which can help the model better understand the dynamic changes of the target. The introduction of the Transformer module can capture global features and long-distance dependencies, which is particularly effective when processing a wide range of fields of view. The temporal continuity of the previous and next frames is used to reduce the probability of false detection and missed detection, and improve the stability and accuracy of detection. The CBAM module is a lightweight attention module, which is divided into two parts: channel attention and spatial attention, and is designed to improve the selectivity of feature representation. It can effectively cope with background changes, and by combining inter-frame information, it can more accurately separate the target and background.

[0074] (2) The present invention has strong generalization ability. The introduction of transfer learning enables the network to have better generalization ability.

[0075] (3) Integrated tracking and detection processing: Combining the target detection results of the improved YOLOv5 with the DeepSORT target tracking module can achieve real-time and high-precision target detection and tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 :It is the overall flow chart of the present invention (improved Yolov5 target detection and tracking network);

[0077] Figure 2 : An improved Yolov5 target detection module;

[0078] Figure 3 : Transormer module;

[0079] Figure 4 : Convolutional attention mechanism module

[0080] Figure 5 :Flowchart of DeepSORT tracking module. DETAILED DESCRIPTION

[0081] The following will be combined with the attached Figure 1-5 The specific steps of the present invention are described in detail.

[0082] A radar multi-frame image target detection and tracking integrated processing method of this embodiment includes the following steps:

[0083] 1. Radar echo signal preprocessing

[0084] The radar echo data is preprocessed into a P display image and the extracted data is drawn into a P display image (PlanPosition Indicator, PPI). The PPI image is a two-dimensional image that shows the distance and direction of the radar relative to its location:

[0085] (1);

[0086] (2);

[0087] Where Dist_bin is the distance grid point and AOEP is the angle data. X, Y are the x and y coordinates when drawing the PPI graph.

[0088] 2. Target detection and tracking integrated processing model

[0089] Compare with Figure 1 ,The constructed detection and tracking network model structure is divided into two parts: the improved yolov5 target detection module and the deepsort tracking module.

[0090] (1) Improved Yolov5 target detection module

[0091] Aiming at the complex environment background, clutter or noise interference, extreme target motion and inter-frame information problems in radar signal detection, the present invention has made a number of improvements based on YOLOv5, such as Figure 2 As shown in the figure, Transformer and CBAM are combined to improve the performance of the model. The specific improvements are as follows:

[0092] 1) Introduction of Transformer

[0093] Enhanced feature extraction and context information capture. In radar signal detection, complex background and clutter interference make feature extraction difficult, affecting target recognition. In order to extract radar inter-frame information, the Transformer structure is introduced into the prediction head of YOLOv5, and the model's feature extraction and context information capture capabilities are enhanced through the self-attention mechanism:

[0094] (3);

[0095] in, are query, key, and value matrices respectively, is the dimension of the key matrix.

[0096] This paper designs a convolutional Transformer module by integrating CNN and Transformer. Its structure is as follows: Figure 3As shown in the figure. In the convolutional Transformer module, in order to enable the Transformer to learn global representation through spatial inductive bias, the convolutional construction method is used to construct the convolutional Transformer module, so that the network has the local capture ability and global modeling ability of CNN. This module is divided into two parts: local modeling and global modeling. In local modeling, in order to better deal with the change of target scale during tracking and thus reduce the robustness of the algorithm, a deformable convolution layer is used to encode local spatial information, and then the number of feature channels is adjusted through the convolution layer. In global modeling, the features from local modeling are expanded into non-overlapping image blocks. By using the Transformer to encode the relationship between image blocks, the features at different positions are weighted and fused to obtain a richer and more accurate feature representation, thereby improving the accuracy and robustness of target tracking. Then the image block is folded back to the feature map to realize global feature modeling, and then the number of feature channels is adjusted to the original size through the convolution layer. Secondly, the local modeling features are spliced ​​with the global modeling features through the shortcut branch, and the two types of feature information are fused using the convolution layer. Finally, the input features are added to the output of the convolutional layer to obtain the final output. This residual connection method can optimize the network structure and avoid information loss and feature degradation.

[0097] 2) Multi-scale feature fusion

[0098] The detection accuracy of objects of different sizes in complex backgrounds is low. Introducing multi-scale feature inputs in the Transformer enables the model to better handle objects of different sizes. Multi-scale feature fusion improves the detection accuracy of the model at various resolutions, thereby improving the detection ability of objects of different sizes. This improvement is particularly suitable for small object detection in high-density and complex scenes.

[0099] 3) References to the Convolutional Attention Mechanism

[0100] Moving target detection and complex background processing: In radar signals, extreme target motion and inaccurate inter-frame information processing affect the detection effect. By introducing the CBAM module and combining the attention mechanism, the focus on moving targets is enhanced. The CBAM module can better process inter-frame information and moving targets in complex backgrounds through channel and spatial attention mechanisms, improving detection accuracy. These improvements comprehensively improve the performance of the model in radar composite interference signal detection, making it more suitable for practical application scenarios of complex and high-density target detection.

[0101] The algorithm neck network adopts the mixed domain convolutional attention mechanism (CBAM) module. It is a simple but effective attention module, which is divided into two parts: one is channel attention, such as Figure 4The first is spatial attention, as shown in the gray box; the second is spatial attention, as shown in the white box in Figure 4. CBAM uses the intermediate feature map and modules to infer the attention of the image along the channel and spatial dimensions, and then combines the attention layer with the feature map for adaptive feature modification. The relevant formula is as follows:

[0102] (4);

[0103] (5);

[0104] In the formula, For sigmoid operation, 7*7 represents the size of the convolution kernel. Represents the input feature map, AvgPool, MaxPool and MLP represent average pooling, maximum pooling and multi-layer perceptron respectively. CBAM is a lightweight module that can be integrated into the CNN architecture and can be trained in an end-to-end manner. Using CBAM, attention regions can be extracted to help Yolov5 suppress interference information and focus on useful target objects.

[0105] (2) DeepSORT target tracking module

[0106] DeepSORT detects the target in each frame based on the output of the improved YOLOv5, and uses multi-feature fusion technology to represent and describe the target, and uses the SORT algorithm to track the target. Figure 5 , the workflow is as follows:

[0107] Initialization: Detect the target in the first frame, create a corresponding trajectory for each target, initialize the motion variable of the Kalman filter, and predict the corresponding target box through the Kalman filter. The trajectory state at this time is unproven;

[0108] IOU matching: Perform IOU matching on the current frame’s target detection box with the target box predicted by the trajectory in the previous frame, and calculate its cost matrix:

[0109] (6);

[0110] Hungarian algorithm: The cost matrix is ​​used as the input of the Hungarian algorithm to obtain the linear matching result;

[0111] Loop update: Repeat IOU matching and Hungarian algorithm steps until a confirmed trajectory appears or the video frame ends;

[0112] Cascade matching: Kalman filter prediction is performed on the confirmed trajectory, and its target box is cascade matched with the detection (matching is performed using appearance features and motion information).

[0113] Rematching: After cascade matching, the unmatched tracks and detections are matched again by IOU, and the Hungarian algorithm is used for linear matching. The matching results are the same as before in three cases;

[0114] Loop processing: Repeat the steps of cascade matching and re-matching until all video frames are processed.

[0115] 3. Input the training data set to iteratively optimize the improved Yolov5 network;

[0116] The specific steps of step 3 are:

[0117] When performing iterative optimization training, the Adam gradient descent method is used, and the given parameters θ ,gradient g t , momentum parameter β 1. Learning rate parameter β 2 and constant First, an exponentially weighted moving average is performed to estimate the first moment of the gradient m t and the second moment v t :

[0118] (7);

[0119] Then, the parameters are updated according to the corrected first-order and second-order moments:

[0120] (8);

[0121] in, t is the number of current iterations, is the learning rate.

[0122] 4. Motion parameter estimation

[0123] In the DeepSORT algorithm, by tracking a moving target in multiple frames, the target's motion parameters, including speed and heading, can be estimated. The following are the steps and related formulas for estimating target motion parameters using the DeepSORT algorithm.

[0124] The DeepSORT algorithm uses a Kalman filter to estimate the state of each target. Assume that the state vector of the target is:

[0125] (9);

[0126] In the formula, is the position of the target in the frame (center point coordinates). is the size (area) of the target. is the aspect ratio of the target. is the velocity component of the target.

[0127] Kalman filter prediction:

[0128] (10);

[0129] (11);

[0130] in, is the forecast error covariance matrix, is the process noise covariance matrix

[0131] Kalman filter update:

[0132] (12);

[0133] (13);

[0134] (14);

[0135] in, is the observation matrix, R is the observation noise covariance matrix, is the Kalman gain.

[0136] Through the update step of the Kalman filter, the velocity component of the target can be obtained Estimated value:

[0137] (15);

[0138] course is the direction angle of the target movement, which can be calculated by the velocity component:

[0139] (16);

[0140] in, Is the two-variable inverse tangent function, used to calculate the four-quadrant inverse tangent angle.

[0141] 5. Model parameter transfer training. First, the parameters of the improved YOLOv5 are initialized using the pre-trained model using a multi-frame dataset from the source domain, instead of random initialization, thereby improving the starting point of model training. Then, the model is adapted to new task requirements by global fine-tuning on the target domain. The process of transfer learning involves transferring knowledge between the source domain and the target domain to enhance the generalization ability of the model. This method not only improves the accuracy and robustness of detection, but also accelerates the training process, solves the problem of data scarcity, effectively copes with the complexity of the marine environment, and ensures the effect of real-time target tracking.

[0142] The specific steps of step 5 are:

[0143] When performing model migration, the pre-trained model parameter file initializes the weights of the model network, replacing the original random initialization operation, and performs global fine-tuning, and the rest of the training process proceeds as usual. and a target domain The case where the source domain uses express, , Represents data samples and corresponding motion target category labels respectively, and the target domain uses Represents. Given a source domain and learning tasks , target domain and learning tasks ,The purpose of transfer learning is to obtain the source domain and learning tasks The knowledge in the target domain can help improve the prediction function in the target domain of learning, including or The overall process of model migration is: ,in, represents the model's pre-trained weights in the source domain, Represents the model target domain training weights.

[0144] 6. Preprocess the real-time radar echo signal and input it into the trained network for testing to complete target detection and tracking. The specific steps of step 6 are:

[0145] First, assume that the input P-display image is preprocessed to ensure that the image is suitable for model input. Next, the pre-trained improved YOLOv5 model is used to detect the object on the preprocessed image to obtain the bounding box, confidence, and category label of each object. Then, the DeepSORT tracking algorithm is initialized and the detected objects are initialized into the DeepSORT tracker. By detecting the object in each subsequent frame and updating the DeepSORT tracker, we are able to track multiple objects and assign a unique ID to each object. Finally, the detected object bounding box and tracking ID are drawn on the original image, a visualization result is generated, and the detection and tracking results are saved to a file or database for subsequent analysis.

[0146] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is determined by the attached claims.

Claims

1. A radar multi-frame image target detection and tracking integrated processing method, characterized in that The following steps are involved: Step 1: Preprocess the radar echo signal and construct an inter-frame echo image data set to provide input data for subsequent target detection and tracking; Step 2: When realizing target detection and tracking, based on the constructed inter-frame echo image dataset, a detection and tracking network based on improved Yolov5 is constructed; the detection and tracking network based on improved Yolov5 is divided into two parts: 1) Improved Yolov5 detection module: Combine the Transformer module and the attention module CBAM with Yolov5 to process the inter-frame sequence images to realize target detection; 2) DeepSORT tracking module: Realize the tracking of radar targets and use the detection results output by the improved Yolov5 detection module to track targets; Step 3: Taking the inter-frame echo image dataset obtained after preprocessing in step 1 as input, the pre-trained improved Yolov5 detection and tracking network is iteratively optimized. During the training process, the network weights are continuously adjusted to improve the accuracy of target detection, and the pre-trained weight parameters are obtained; Step 4: Motion parameter estimation, extracting the motion characteristics of the target to provide basic data for subsequent analysis; Step 5: Model parameter transfer training: Use the pre-trained weight parameters obtained in step 3 to initialize the weights of the improved Yolov5 detection and tracking network, adapt to the new task through global fine-tuning, and change the source domain and target domain at the same time to enhance the generalization ability of the network; Step 6: Input the processed signal into the improved Yolov5 detection and tracking network after model parameter migration training for target detection and tracking to complete the final target detection and tracking task.

2. A radar multi-frame image target detection and tracking integrated processing method according to claim 1, characterized in that The specific steps of step 1 are as follows: The radar echo data is preprocessed into a P-display image, and an inter-frame echo image data set is constructed. The extracted data is used to draw a P-display image. The P-display image is a two-dimensional image that shows the distance and azimuth of the radar relative to its location: （1）； （2）； Where Dist_bin is the distance from the grid point AOEP is the angle data, X, Y These are the x and y coordinates when drawing the P display image.

3. A radar multi-frame image target detection and tracking integrated processing method according to claim 1, characterized in that The specific steps of step 2 are as follows: (1) Improved Yolov5 detection module: ①Introduction of Transformer The Transformer structure is introduced into the prediction head of YOLOv5, and the self-attention mechanism is used to enhance the model's feature extraction and context information capture capabilities: （3）； in, are query, key, and value matrices respectively, is the dimension of the bond matrix; ②Multi-scale feature fusion Feature extraction: Multi-scale features are extracted through different convolutional layers. YOLOv5 adopts CSPNet structure, which can effectively extract multi-level features. Feature pyramid construction: Fuse shallow features with deep features, using upsampling and downsampling techniques; Upsampling and fusion: Use upsampling to resize shallow features to align them with deep features, and then concatenate or add the aligned features to enhance feature expression capabilities; ③ References to the Convolutional Attention Mechanism The algorithm neck network adopts a mixed-domain convolutional attention module CBAM module, which is divided into two parts: one is channel attention, and the other is spatial attention. The CBAM module uses the intermediate feature map and the module to infer the attention of the image along the channel and spatial dimensions in sequence, and then combines the attention layer with the feature map for adaptive feature modification. The formula is as follows: （4）； （5）； In the formula, For sigmoid operation, 7*7 represents the size of the convolution kernel. Represents the input feature map. AvgPool, MaxPool, and MLP represent average pooling, maximum pooling, and multi-layer perceptron, respectively. CBAM is a lightweight module that can be integrated into the CNN architecture and can be trained in an end-to-end manner. (2) DeepSORT tracking module DeepSORT detects the target in each frame based on the output of the improved Yolov5 detection module, and uses multi-feature fusion technology to represent and describe the target. The workflow is as follows: Initialization: Detect the target in the first frame, create a corresponding trajectory for each target, initialize the motion variable of the Kalman filter, and predict the corresponding target box through the Kalman filter. The trajectory state at this time is unproven; IOU matching: Perform IOU matching on the current frame’s target detection box with the target box predicted by the trajectory in the previous frame, and calculate its cost matrix: （6）； Hungarian algorithm: The cost matrix is ​​used as the input of the Hungarian algorithm to obtain the linear matching result; Loop update: Repeat IOU matching and Hungarian algorithm steps until a confirmed trajectory appears or the video frame ends; Cascade matching: Kalman filter prediction is performed on the confirmed trajectory, and its target box is cascade matched with the detection; Rematching: After cascade matching, the unmatched tracks and detections are matched again by IOU, and the Hungarian algorithm is used for linear matching. The matching results are the same as before in three cases; Loop processing: Repeat the steps of cascade matching and re-matching until all video frames are processed.

4. A radar multi-frame image target detection and tracking integrated processing method according to claim 1, characterized in that The specific steps of step 3 are as follows: Using the Adam gradient descent method, given the parameters θ ,gradient g t , momentum parameter β 1. Learning rate parameter β 2 and constant , first, an exponentially weighted moving average is performed to estimate the first-order moment of the gradient m t and the second moment v t : （7）； Then, the parameters are updated according to the corrected first-order and second-order moments: （8）； in, t is the number of current iterations, is the learning rate.

5. A radar multi-frame image target detection and tracking integrated processing method according to claim 4, characterized in that The specific steps of step 4 are: The DeepSORT algorithm uses a Kalman filter to estimate the state of each target, assuming that the state vector of the target is: （9）； In the formula, The goal is The coordinates of the center point of the frame. is the target size, is the aspect ratio of the target, is the velocity component of the target; Kalman filter prediction: （10）； （11）； in, is the forecast error covariance matrix, is the process noise covariance matrix; Kalman filter update: （12）； （13）； （14）； in, is the observation matrix, R is the observation noise covariance matrix, is the Kalman gain; Through the update step of the Kalman filter, the velocity component of the target is obtained Estimated value: （15）； course is the direction angle of the target movement, which is calculated by the velocity component: （16）； in, Is the two-variable inverse tangent function, used to calculate the four-quadrant inverse tangent angle.

6. A radar multi-frame image target detection and tracking integrated processing method according to claim 1, characterized in that The specific steps of step 5 are: The pre-trained model parameter file initializes the weights of the model network, replacing the original random initialization operation, and performs global fine-tuning. The rest of the training process proceeds as usual. Transfer learning focuses on the active domain. and a target domain The case where the source domain Use the expression, , Represents data samples and corresponding motion target category labels respectively, and the target domain uses Indicates that, given the source domain and learning tasks , target domain and learning tasks ,The purpose of transfer learning is to obtain the source domain and learning tasks The knowledge in the target domain can help improve the prediction function in the target domain of learning, including or , the overall process of model migration is: ,in, represents the model's pre-trained weights in the source domain, Represents the model target domain training weights.

7. A radar multi-frame image target detection and tracking integrated processing method according to claim 1, characterized in that The specific steps of step 6 are: First, assume that the input image dataset is preprocessed to ensure that the image is suitable for model input; then, use the pre-trained improved YOLOv5 detection and tracking model to perform target detection on the preprocessed image to obtain the bounding box, confidence and category label of each target; Then, the DeepSORT tracking module is initialized and the detected targets are initialized into the DeepSORT tracking module; by detecting targets in each subsequent frame and updating the DeepSORT tracking module, multiple targets can be tracked and a unique ID is assigned to each target; finally, the detected target bounding box and tracking ID are drawn on the original image, a visualization result is generated, and the detection and tracking results are saved to a file or database for subsequent analysis.

Images