Small target tracking method based on anchor-free detection network and feature re-fusion module

Abstract

The present invention relates to a small target tracking method based on an anchor-free detection network and a feature re-fusion module, belonging to the technical field of multi-target tracking. The method comprises the following steps: S1: using a drone to collect images of a target object, and naming the collected images according to the format of a Pascal VOC data set; S2: preprocessing the images; S3: designing the size of a convolutional neural network input image, the number of recognition types, the number of training times, the learning rate, and the number of network layers according to computer memory, GPU capabilities, and the recognition effect and training speed required by users; S4: improving the network structures of a feature extraction network part, a detection part, and a re-recognition part to obtain an improved monitoring model; S5: training the network model, and saving the trained network model weights after the training is completed; and S6: using the trained network model weights for recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-target tracking and relates to a small target tracking method based on an anchor-free detection network and a feature re-fusion module. Background Art

[0002] The main task of multiple object target tracking (MOT) is to continuously locate multiple targets in a video while maintaining their respective identification numbers and recording their motion trajectories. Multi-object tracking is widely used in computer vision. For example, in autonomous driving, it can sense surrounding vehicles to assist the driver in decision-making. In drones, both civilian and military applications, MOT can continuously track targets within a specified area. Currently, most tracking algorithms are based on the Tracking-by-Detection (TBD) paradigm. This involves first obtaining the target's position through a detection network. Based on the detection results, an association model is then used to associate the detection boxes belonging to the same target in previous and subsequent frames to obtain the target's motion trajectory, ultimately completing the tracking of the target object. Therefore, the algorithm's tracking performance depends almost entirely on the quality of the detection network. However, most current detection networks use an anchor-free approach to extract features, resulting in excessive parameter redundancy, slowing the model's execution speed and limiting real-time performance. Furthermore, in anchor-besed detection models, most anchor boxes are negative, with only a small number of positive examples. This imbalance in positive and negative examples results in poor generalization. To address this issue, Law et al. proposed CornerNet, the first anchor-free detection algorithm. This network locates objects by detecting the upper left and lower right corners of all objects in an image, significantly improving detection but failing to clearly describe the detailed internal features of the objects. Zhou et al. proposed CenterNet, which uses corner points and a Gaussian distribution to find the center point of an object, thereby obtaining internal feature information. To further improve the real-time performance of tracking algorithms, Wang et al. proposed the JDE algorithm, which combines one-stage detection with person re-identification to accelerate inference. In 2020, Zhang combined the CornerNet network with the one-step framework of JDE to propose the FairMOT algorithm. This method addresses the misalignment of identity embedding features in the JDE algorithm. However, FairMOT can cause overlapping center points of multiple objects in densely packed and mutually occluded situations, resulting in excessive ID switches during tracking. Furthermore, the anchor-free model used by the FairMOT algorithm requires a large number of computational parameters, making its detection of small objects insufficient, leaving significant room for improvement. In addition, when extracting the ReID features of the target, tracking networks such as FairMOT directly extract features at the center of the target detection frame in the heat map. However, since the feature center of the target may not be at the center of the detection frame, the ReID features at this location cannot accurately represent the appearance information of the target. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a small target tracking method based on an anchor-free detection network and a feature re-fusion module. By adopting a higher-resolution anchor-free detection network to extract more position information and contour information of the target, in addition, by effectively adjusting the feature center position of REID, target appearance information that is more suitable for key point detection can be extracted to reduce the frequency of ID switching. Therefore, the present invention uses the high-resolution anchor-free detection network YOLOX as a shared detection and re-identification feature extraction network, and proposes a new feature fusion module on this basis. By further integrating multi-level features while retaining their semantic hierarchy, and through non-local refinement operations, it can effectively focus on long-distance dependencies and can immediately re-detect objects after they disappear. Finally, by adding an additional REID center position offset branch to adjust the REID feature center of the target, the features of each frame of the saved target are more representative, so as to reduce the number of missed targets and the number of ID switches, thereby improving the overall performance of the algorithm.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] A small target tracking method based on an anchor-free detection network and a feature re-fusion module includes the following steps:

[0006] S1: Use a drone to collect images of the target object and name the collected images according to the format of the Pascal VOC dataset;

[0007] S2: preprocess the image;

[0008] S3: Design the convolutional neural network input image size, number of recognition types, number of training times, learning rate, and number of network layers based on computer memory, GPU capabilities, and the user's required recognition effect and training speed;

[0009] S4: Improve the network structures of the feature extraction network, detection network, and re-identification network to obtain an improved monitoring model;

[0010] S5: Train the network model and save the trained network model weight pt after the training is completed;

[0011] S6: Use the trained network model weights for recognition.

[0012] Furthermore, the image preprocessing in step S2 includes:

[0013] S21: Image marking: marking the target in the collected image, marking the location and category name of the target;

[0014] S22: Image augmentation: If the images taken by the user in S1 do not meet the number of images required to identify a single category of objects, the images will be augmented using scaling, flipping, translation, and mosaic data augmentation methods.

[0015] S23: Divide the data set: divide the amplified images and labeled files into a training set, a test set, and a validation set.

[0016] Furthermore, step S3 specifically includes: first, pre-training 60 epochs on the CrowdHuman dataset using the self-supervised learning method, and then training 60 epochs on the training set of the CHUK-SYSU, PRW, and MOT16 datasets. The validation set is the training set in the 2DMOT15 dataset. The initial learning rate is 0.0001, and the learning rate for the next 20 epochs is 0.00001. The batch_size is set to 4, the Adam optimizer is used to obtain the optimal network parameters, and siou is used as the activation function.

[0017] Furthermore, the network structure of the feature extraction network described in step S4 is improved, specifically an improved detection network YOLOX based on YOLOv5, and a feature fusion module is added;

[0018] The YOLOX network structure consists of three parts: backbone network, neck and detection head;

[0019] The detection head uses a decoupling method to process the classification loss and bounding box loss with different detection branches respectively, and uses an anchor-free method to predict the prediction box;

[0020] The backbone network uses CSPDarknet. The entire backbone consists of residual convolutions, including multiple CSPLayer structures. CSPLayer is divided into two parts. The main branch stacks N residual blocks according to the depth set by the network. The residual branch is directly spliced with the main branch after only a small amount of processing. After passing through the convolution layer and the BN layer, it is activated by the SiLU activation function and input into the next layer of CSPLayer structure. The residual block is also divided into two parts. The main body is divided into 1×1 convolution and 3×3 convolution. The residual edge uses jump connection to directly combine the input and output of the backbone. The feature map with a resolution of 160×160 obtained by downsampling four times in the original backbone network is added to the fusion network.

[0021] The neck adopts the feature pyramid structure of FPN-PAN, where FPN is responsible for transmitting semantic information from bottom to top, and PAN is responsible for transmitting position information and edge information from top to bottom;

[0022] The feature maps of four layers with different resolutions after the fusion network are input into the feature re-fusion module. The feature maps P2_out, P3_out, P4_out and P5_out output by the FPN-PAN network are subjected to linear interpolation or maximum pooling operations to obtain a feature map of the same size as P4_out. Then, all feature maps are added and averaged to obtain a unified feature map x that balances high semantics and positioning information. Non-local operations are used on the feature map x to calculate the response of a certain position as the weighted sum of the features of all positions. Then, a deep neural network is used to capture long-range dependencies. Finally, the obtained feature map y is restored to the feature map size before fusion and re-fused with the corresponding feature maps to obtain the fused feature maps E2, E3, E4 and E5. The formula is as follows:

[0023]

[0024]

[0025] Where k represents the kth feature map, i and j represent different positions;

[0026] Adopt embedded Gaussian non-local attention to refine feature maps;

[0027] Calculate the pairwise relationship between i and all other possible associated positions j, function f(x i ,x j )as follows:

[0028]

[0029] where θ(x i )=W θ x i and There are two embedded parts;

[0030] Construct a function g(x) for calculating the eigenvalue of the input feature at position j j ):

[0031] g(x j )=W g x j

[0032] Where W g It is a learnable weight matrix implemented by a 1×1 convolutional layer in the spatial domain;

[0033] Calculate the normalization parameter S(x):

[0034]

[0035] After y is scaled accordingly, the feature maps {y_1,y_2,y_3,y_4,y_5} are obtained, which are added to the corresponding original feature maps P2_out,…,P5_out to obtain the fused feature maps E2,…,E5. These feature maps will be used for the final decision of the network:

[0036]

[0037] Furthermore, the network structure of the detection part and the re-identification part is improved, including: adding an additional ReID feature extraction position offset branch to adjust the ReID feature extraction position of the target, and the detection branch calculates the target center position of the current frame. i =(o i x ,o i y ), the ReID feature extraction offset branch outputs the position offset component e i =(e i x ,e i y ), the center position of the physical frame plus the offset component is the ReID feature extraction center of the target Then the center c is extracted by the feature of target i i The four adjacent coordinates of the point are interpolated using bilinear interpolation to obtain its eigenvector F i .

[0038] Furthermore, step S5 specifically includes setting parameters for the configuration files train.py and Classes.txt of the improved detection and tracking model network, placing the set Classes.txt file and the improved detection and tracking model into a computer with a configured environment, and using the labeled images in the training set and the validation set for training. During the training process, the images divided in the test set are placed in the computer for testing to obtain the effect of training at each stage, and the trained network model weights pt are saved after the training is completed.

[0039] Furthermore, step S6 specifically includes preparing the image to be tracked on the computer, changing the configuration file model_path and the image path to be tracked in detect.py, executing detect.py, loading the network model weights after running, and finally saving the image of the target identification and tracking result on the computer.

[0040] The beneficial effects of the present invention are:

[0041] (1) When performing target tracking, an improved detection tracking network is used to train the data set. The trained tracking model not only has high tracking accuracy but also has a small number of parameters.

[0042] (2) When performing target tracking, an improved detection and tracking model is used to train the image dataset. The trained model can track dense small targets more accurately and can also accurately track target objects that are clustered and overlapped in a large area.

[0043] (3) It can be applied in actual real-time target tracking tasks and has the characteristics of fast speed and strong real-time performance.

[0044] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0046] Figure 1 This is the overall flow chart of the small target tracking algorithm based on the anchor-free detection network and feature re-fusion module of the present invention;

[0047] Figure 2 Schematic diagram of the feature fusion module structure;

[0048] Figure 3 Schematic diagram of ReID feature extraction. DETAILED DESCRIPTION

[0049] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0050] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0051] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0052] See also Figures 1 to 3 The present invention provides a small target tracking algorithm based on an anchor-free detection network and a feature re-fusion module, the method comprising the following steps:

[0053] S1: Image acquisition: The user uses a drone to collect images of the target object and names the collected images according to the format of the Pascal VOC dataset. At the same time, three folders named Annotations, ImageSets, and JPEGImages are created;

[0054] S2: Image preprocessing:

[0055] S2-1: Image labeling: In the image collected in S1, use the image labeling tool LabelImg to label the target in the image, marking the target's location and category name;

[0056] S2-2: Image augmentation: If the images taken by the S1 user do not meet the requirement of 3,000 images for identifying a single category of objects, the images will be augmented using scaling, flipping, translation, and mosaic data augmentation methods;

[0057] S2-3: Divide the dataset: Divide the amplified images and labeled files into training, test, and validation sets; the training set, test set, and validation set account for 60%, 20%, and 20% respectively;

[0058] S3: Set model hyperparameters: Design the size of the convolutional neural network input image, the number of recognition types, the number of training times, the learning rate, and the number of network layers based on computer memory, GPU capabilities, and the recognition effect and training speed required by the user;

[0059] This paper first uses self-supervised learning to pre-train the model on the CrowdHuman dataset for 60 epochs. It then trains the model for another 60 epochs on the training set of the CHUK-SYSU, PRW, and MOT16 datasets, with the validation set being the training set of the 2DMOT15 dataset. The initial learning rate is 0.0001, and the learning rate is 0.00001 for the next 20 epochs. The batch_size is set to 4, and the Adam optimizer is used to obtain optimal network parameters. The activation function, siou, is used because it has both unbounded and lower bounded properties, is smooth, and is non-monotonic, making it more effective in deep networks.

[0060] S4: Improve the network structure of the feature extraction network part, detection part and re-identification part of the original detection model. The improvement process is as follows:

[0061] S4-1: YOLOX is an improved version of YOLOv5, with higher detection accuracy than YOLOv5. Its network structure is divided into three parts: backbone network, neck and detection head. The backbone network and neck still have the same network structure as YOLOv5, but its detection head adopts an uncoupling method, processing the classification loss and bounding box loss with different detection branches respectively. In addition, it also uses an anchor-free method to predict the prediction box, thereby improving its detection accuracy and speed. The backbone network adopts CSPDarknet. The entire trunk is composed of residual convolution, which is mainly composed of multiple CSPLayer structures. CSPLayer is divided into two parts. The main branch stacks N residual blocks according to the depth set by the network. The residual branch is directly spliced with the main branch after only a small amount of processing. After passing through the convolution layer and the BN layer, it is activated by the SiLU activation function and input into the next layer of CSPLayer structure. The residual block is also divided into two parts. The main body is divided into 1×1 convolution and 3×3 convolution. The residual edge uses a jump connection to directly combine the input and output of the trunk. The advantage of the residual network is that it can increase the network depth while avoiding network degradation and improving detection accuracy. The present invention adds the feature map with a resolution of 160×160 obtained by downsampling four times in the original backbone network to the fusion network to strengthen the fusion of shallow features and improve the detection ability of small targets.

[0062] S4-2: YOLOX's feature fusion network adopts the feature pyramid structure of FPN-PAN. FPN is responsible for transmitting semantic information from bottom to top, and PAN is responsible for transmitting position information and edge information from top to bottom, so that the fused feature map has richer information. In addition, in order to ensure that the feature map has high resolution and rich semantic information at the same time, the present invention inputs the four layers of feature maps of different resolutions after the fusion network into the feature re-fusion module, so that the final feature map has high resolution and high semantic information, so as to achieve effective small target detection in complex environments. The feature maps P2_out, P3_out, P4_out and P5_out output from the FPN-PAN network are subjected to linear interpolation or maximum pooling operations to obtain feature maps of the same size as P4_out, because the present invention believes that the P4 feature map in the middle layer not only contains complete semantic information, but also does not introduce a large number of calculation parameters. After that, all feature maps are added and averaged to obtain a unified feature map x that balances high semantic and positioning information. In addition, in order to suppress background noise as much as possible and further refine the balanced feature map, the present invention uses non-local operations on the feature map x to calculate the response of a certain position as the weighted sum of the features of all positions. Use deep neural networks to capture long-distance dependencies instead of focusing on only one part to improve the network's ability to detect small targets. Finally, the obtained feature map y is restored to the size of the feature map before fusion, and re-fused with the corresponding feature map to obtain fused feature maps E2, E3, E4 and E5. At this time, the feature map will have stronger positioning capabilities and high-level semantic features. The formula is as follows: x is the feature map after re-fusion, and y is the feature map obtained after non-local refinement operation; in this step, the present invention uses embedded Gaussian non-local attention to refine the feature map. Function f(x i ,x j ) is used to calculate the pairwise relationship between i and all other possible associated positions j. Function g(x j ) is used to calculate the eigenvalue of the input feature at position j. S(x) is the normalization parameter. i )=W θ x i and There are two embedded parts, W g is a learnable weight matrix, which is actually implemented by a 1×1 convolutional layer in the spatial domain. After y is scaled accordingly, the feature maps {y_1, y_2, y_3, y_4, y_5} are obtained. These feature maps are added to the corresponding original feature maps P2_out, …, P5_out to obtain the fused feature maps E2, …, E5. These feature maps will be used for the final decision of the network. The specific formula is as follows:

[0063]

[0064]

[0065]

[0066] g(x j )=W g x j

[0067]

[0068]

[0069] S4-3: For the problem of inconsistency between the target physical detection frame and the ReID feature center, the present invention adjusts the target ReID feature extraction position by adding an additional ReID feature extraction position offset branch to extract more accurate ReID feature information. Figure 1 As shown in the part, the detection branch calculates the target center position o of the current frame i =(o i x ,o i y ), the ReID feature extraction offset branch outputs the position offset component e i =(e i x ,e i y ), the center position of the physical frame plus the offset component is the ReID feature extraction center of the target Then the center c is extracted by the feature of target i i The four adjacent coordinates of the point are interpolated using bilinear interpolation to obtain its eigenvector F i , compared to using o i =(o i x ,o i y ) Extracted features, F i It can retain more key appearance features of the target, thereby improving the tracking performance of the model and reducing the number of ID switching.

[0070] S5: Training the network model: Set the parameters of the configuration files train.py and Classes.txt of the improved detection and tracking model network, put the set Classes.txt file and the improved detection and tracking model into the computer with the configured environment, and use the labeled images in the training set and validation set for training. During the training process, put the images divided in the test set into the computer for testing to obtain the effect of each stage of training. After the training is completed, save the trained network model weight pt;

[0071] S6: Use the trained network model weights for recognition: Prepare the image to be tracked on the computer, change the model_path configuration file in detect.py and the image path to be tracked, execute detect.py, load the network model weights after running, and finally save the image of the recognition and tracking target result on the computer;

[0072] S7: The improved detection and tracking algorithm is evaluated using common evaluation indicators in the field of multi-target tracking to verify its superiority. These indicators mainly include tracking accuracy (Multiple Object Tracking Accuracy, MOTA), positioning accuracy (Multiple Object Tracking Precision, MOTP), false negative rate (False Negatives, FN), false positive rate (False Positives, FP), tracking speed (Frame Per Second, FPS), and identity switch count (Identity Switches, IDs). MOTA combines the data of FN, FP, and IDs, and can directly reflect the performance of the algorithm. It is the main evaluation indicator. Its calculation formula is as follows, where GT represents the number of ground truth boxes. When the MOTA value is negative, it means that the tracking has failed. If it is positive and the closer it is to 1, the better the tracking effect.

[0073]

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A small target tracking method based on an anchor-free detection network and a feature re-fusion module, characterized by: The following steps are involved: S1: Use a drone to collect images of the target object and name the collected images according to the format of the Pascal VOC dataset; S2: preprocess the image; S3: Design the convolutional neural network input image size, number of recognition types, number of training times, learning rate, and number of network layers based on computer memory, GPU capabilities, and the user's required recognition effect and training speed; S4: improving the network structures of the feature extraction network, the detection network, and the re-identification network to obtain an improved monitoring model; the improvement of the network structure of the feature extraction network described in step S4 is specifically an improved detection network YOLOX based on YOLOv5, and a feature fusion module is added; The YOLOX network structure consists of three parts: backbone network, neck and detection head; The detection head uses a decoupling method to process the classification loss and bounding box loss with different detection branches respectively, and uses an anchor-free method to predict the prediction box; The backbone network uses CSPDarknet. The entire backbone consists of residual convolutions, including multiple CSPLayer structures. CSPLayer is divided into two parts. The main branch stacks N residual blocks according to the depth set by the network. The residual branch is directly spliced with the main branch after only a small amount of processing. After passing through the convolution layer and the BN layer, it is activated by the SiLU activation function and input into the next layer of CSPLayer structure. The residual block is also divided into two parts. The main body is divided into 1×1 convolution and 3×3 convolution. The residual edge uses jump connection to directly combine the input and output of the backbone. The feature map with a resolution of 160×160 obtained by downsampling four times in the original backbone network is added to the fusion network. The neck adopts the feature pyramid structure of FPN-PAN, where FPN is responsible for transmitting semantic information from bottom to top, and PAN is responsible for transmitting position information and edge information from top to bottom; The feature maps of four layers with different resolutions after the fusion network are input into the feature re-fusion module. The feature maps P2_out, P3_out, P4_out and P5_out output by the FPN-PAN network are subjected to linear interpolation or maximum pooling operations to obtain a feature map of the same size as P4_out. Then, all feature maps are added and averaged to obtain a unified feature map x that balances high semantics and positioning information. Non-local operations are used on the feature map x to calculate the response of a certain position as the weighted sum of the features of all positions. Then, a deep neural network is used to capture long-range dependencies. Finally, the obtained feature map y is restored to the feature map size before fusion and re-fused with the corresponding feature maps to obtain the fused feature maps E2, E3, E4 and E5. The formula is as follows: Where k represents the kth feature map, i and j represent different positions; Adopt embedded Gaussian non-local attention to refine feature maps; Calculate the pairwise relationship between i and all other possible associated positions j, function f(x i ,x j )as follows: where θ(x i )=W θ x i and There are two embedded parts; Construct a function g(x) for calculating the eigenvalue of the input feature at position j j ): g(x j )=W g x j Where W g It is a learnable weight matrix implemented by a 1×1 convolutional layer in the spatial domain; Calculate the normalization parameter S(x): After y is scaled accordingly, the feature maps {y_1,y_2,y_3,y_4,y_5} are obtained, which are added to the corresponding original feature maps P2_out,…,P5_out to obtain the fused feature maps E2,…,E5. These feature maps will be used for the final decision of the network: The improvement of the network structure of the detection part and the re-identification part includes: adding an additional ReID feature extraction position offset branch to adjust the ReID feature extraction position of the target, and the detection branch calculates the target center position of the current frame. i =(o i x ,o i y ), the ReID feature extraction offset branch outputs the position offset component e i =(e i x ,e i y ), the center position of the physical frame plus the offset component is the ReID feature extraction center of the target Then the center c is extracted by the feature of target i i The four adjacent coordinates of the point are interpolated using bilinear interpolation to obtain its eigenvector F i ; S5: Train the network model and save the trained network model weights after the training is completed; S6: Use the trained network model weights for recognition.

2. The small target tracking method based on the anchor-free detection network and feature re-fusion module according to claim 1, characterized in that: The image preprocessing in step S2 includes: S21: Image marking: marking the target in the collected image, marking the location and category name of the target; S22: Image augmentation: If the images taken by the user in S1 do not meet the number of images required to identify a single category of objects, the images will be augmented using scaling, flipping, translation, and mosaic data augmentation methods. S23: Divide the data set: divide the amplified images and labeled files into a training set, a test set, and a validation set.

3. The small target tracking method based on the anchor-free detection network and feature re-fusion module according to claim 1, characterized in that: Step S3 specifically includes: first, pre-training 60 epochs on the CrowdHuman dataset using the self-supervised learning method, and then training 60 epochs on the training set of the CHUK-SYSU, PRW, and MOT16 datasets. The validation set is the training set of the 2DMOT15 dataset. The initial learning rate is 0.0001, and the learning rate for the next 20 epochs is 0.00001. The batch_size is set to 4, the Adam optimizer is used to obtain the optimal network parameters, and siou is used as the activation function.

4. The small target tracking method based on the anchor-free detection network and feature re-fusion module according to claim 1, characterized in that: Step S5 specifically This includes setting parameters for the configuration files train.py and Classes.txt of the improved detection and tracking model network, placing the set Classes.txt file and the improved detection and tracking model into a computer with a configured environment, and using the labeled images in the training set and validation set for training. During the training process, the images divided in the test set are placed into the computer for testing to obtain the effect of training at each stage. After the training is completed, the trained network model weights pt are saved.

5. The small target tracking method based on the anchor-free detection network and feature re-fusion module according to claim 1, characterized in that: Step S6 specifically includes preparing the image to be tracked on the computer, changing the configuration file model_path in detect.py and the image path to be tracked, executing detect.py, loading the network model weights after running, and finally saving the image of the target identification and tracking result on the computer.

Images