A tracking method based on trajectory adaptive small target detection

Abstract

The application discloses a tracking method based on trajectory adaptive small target detection and concretely relates to the technical field of target detection; a two-way aggregation multi-scale pyramid structure network suitable for unmanned aerial vehicle (UAV) aerial small targets is designed, so that the network improves the probability and frequency of multi-scale feature fusion, and higher detection precision is obtained; an SF-C2f feature extraction module is designed, the limitations of traditional methods in processing complex environments are overcome, part of the calculation amount is reduced, a loss function suitable for UAV aerial small targets is designed, the sensitivity of the model to small target detection is effectively improved, the adaptability to complex backgrounds is improved, and the network is more suitable for the detection of UAV aerial small targets; an AFK trajectory adaptive Kalman filtering algorithm for target tracking is designed, which is suitable for the nonlinear situation of the motion trajectory of UAV aerial small targets. Through the fusion method, the UAV aerial small target detection and tracking task in various backgrounds can perform well.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a tracking method based on trajectory adaptive small target detection. Background Technology

[0002] With the development of information and intelligent technologies, drones, with their advantages of low cost, small size, and flexible operation, have largely compensated for the shortcomings of traditional image data acquisition. Meanwhile, compared to stationary cameras, the maneuverability of drones gives aerial imagery greater flexibility, enabling the capture of scenes inaccessible from the ground. Research on the detection and tracking of small targets in drone aerial photography is also deepening, and target detection and tracking algorithms are now a key focus in the field of computer vision. Current detection and tracking technologies struggle to adapt to changing environments, making drone aerial photography a valuable application. However, because drones fly at high altitudes, the objects they capture are more complex, resulting in smaller and denser targets in aerial images, posing a greater challenge to existing algorithms in target detection and tracking. Therefore, this invention proposes a tracking method based on trajectory-adaptive small target detection, constructing a detection network suitable for small targets in drone aerial photography and designing a trajectory-adaptive target tracking algorithm to achieve the detection and tracking of small targets in drone aerial photography.

[0003] Chinese patent publication number "CN115953431A" is titled "Multi-target Tracking Method and System for UAV Aerial Video". This invention utilizes a multi-scale pixel-by-pixel target detection network, MSPNet, to extract target categories and bounding boxes from video frames. It then further extracts the apparent feature vectors of the targets using a multi-granularity fusion feature extraction network, MaskMGN. The apparent feature vectors of the trajectory are calculated using a time-series and detection confidence-based weighted moving average (TCMWA) method. Finally, the apparent feature vectors are correlated with the Kalman motion model to obtain the multi-target tracking results. While this method uses a target detection network based on ResNet50, it does not fundamentally solve the difficulty of detecting small targets in complex backgrounds. Furthermore, the target tracking algorithm does not consider the trajectory tracking problem of non-linearly moving objects. Therefore, designing a target detection method with better performance and higher accuracy in complex UAV aerial photography backgrounds, and a tracking method more suitable for non-linearly moving targets, is the direction this invention urgently needs to research. Summary of the Invention

[0004] The technical solution of this invention to solve the above-mentioned technical problems is to provide a tracking method based on trajectory adaptive small target detection, comprising the following steps:

[0005] Step 1. Prepare the dataset: Prepare a dataset of images of small targets taken by drone aerial photography;

[0006] Step 2. Constructing the target detection network model: Construct a small target detection network for UAV aerial photography, including a backbone network, a neck network, and four detection heads; the backbone network includes 5 convolutional downsampling modules, 4 SF-C2f feature extraction modules, and an SPPF spatial pyramid pooling module; the SF-C2f feature extraction modules improve feature extraction performance through PConv layers and SimAM attention modules; the neck network includes an upsampling module, an SF-C2f feature extraction module, and a convolutional module, and adopts a bidirectional aggregated multi-scale pyramid structure network to improve the frequency of multi-scale fusion;

[0007] Step 3. Train the network model: Input the prepared dataset into the object detection network model for training, including preprocessing small target images captured by the drone, feature extraction, classification and regression operations, and verifying the model performance through test images;

[0008] Step 4. Design the loss function: By constructing the output image of the object detection network model and labeling it, a composite loss function is adopted, including class classification loss and bounding box regression loss. The class classification loss adopts the contrast loss function and the cross-entropy loss function, and the bounding box regression loss adopts the DFL loss function plus the WIoUv3 loss function.

[0009] Step 5. Construct a target tracking network model: Design a target tracking algorithm, including cascaded matching, calculating the cost matrix, using the Hungarian algorithm for matching, and using the AFK adaptive Kalman filter algorithm to predict and update the target trajectory, so as to realize real-time tracking of small targets captured by UAV aerial photography;

[0010] Step 6. Save the model and select evaluation metrics: Evaluate the accuracy and efficiency of the target tracking algorithm.

[0011] Further, in step 2, the SF-C2f feature extraction module includes: passing through the first convolutional layer, then being divided into two parts, one part being processed by multiple bottleneck modules, concatenated with the other part and passed through the second convolutional layer, and then outputting the result through the SimAm attention module;

[0012] The bottleneck module includes: the input is divided into two parts. The first part passes through the PConv layer and then through the convolutional layer, where batch normalization and ReLU activation functions are added. This is then added to the other part of the input data to output the result.

[0013] The SimAM module accurately assesses the linear separability of a single feature within the same channel from other features by defining an energy function, thereby accurately determining the importance of each feature; the energy function is defined as follows:

[0014]

[0015] In the formula, t and x i w represents the target feature information and other feature information in the channel, respectively. t and b t λ represents the linear transformation weights and biases of t, respectively; index i represents the spatial dimension order, λ is a hyperparameter, and M is the number of all feature information in a single channel.

[0016] Furthermore, in step 2, the neck network includes: extending the PAN-FPN architecture upwards, utilizing the spatial information in the shallow feature maps to retain the high-resolution spatial details needed for detecting small targets, and adding pointwise convolutional layers during the feature map fusion process.

[0017] Furthermore, in step 5, the adaptive Kalman filter algorithm uses an adaptive forgetting factor to control the weight of past data, thereby improving the algorithm's adaptability to parameter changes.

[0018] Furthermore, in step 4, the WIoUv3 loss function in the bounding box regression loss function uses a dynamic non-monotonic mechanism to evaluate the quality of the anchor box, thereby improving the model's ability to locate objects, which is especially suitable for small target detection tasks in UAV aerial photography scenarios.

[0019] Compared with the prior art, this application has the following beneficial effects:

[0020] 1. This invention designs a novel target detection network and trajectory adaptive target tracking algorithm based on bidirectional aggregation and multi-scale fusion. The target detection network includes a backbone network, a neck network, and four detection heads, which is more suitable for detecting small targets in UAV aerial photography and better solves the technical challenges of real-time small target detection. The trajectory adaptive target tracking algorithm better solves the problem of missed detection in UAV target trajectory tracking.

[0021] 2. This invention incorporates an SF-C2f feature extraction module, a bidirectional aggregated multi-scale pyramid structure network, and a novel composite loss function into the target detection network structure. This improves the probability and number of multi-scale feature fusions, making it more suitable for detecting small targets in UAV aerial photography. The overall design of the target detection network enhances its feature extraction capabilities and target detection accuracy. It also exhibits better adaptability to complex backgrounds and can better distinguish between different target categories, making it highly suitable for small target detection in UAV aerial photography.

[0022] 3. This invention designs an AFK adaptive Kalman filter algorithm. The core idea is to use an adaptive forgetting factor to control the weight of past data, thereby improving the algorithm's adaptability to parameter changes. This allows it to better cope with the nonlinear changes in the motion trajectory during target tracking and prevent missed detections.

[0023] 4. The trajectory-adaptive UAV aerial photography small target detection and tracking method designed in this invention has a certain scene generalization ability. This algorithm performs well in different scenarios and has strong scene adaptability. It has good adaptability and generalization ability for small target detection and tracking in different scenarios. Attached Figure Description

[0024] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the steps of a tracking method based on trajectory adaptive small target detection as described in this invention.

[0026] Figure 2 This is a schematic diagram of the target detection network model of the present invention;

[0027] Figure 3 This is a schematic diagram of the backbone network structure of the present invention;

[0028] Figure 4 This is a schematic diagram of the SF-C2f feature extraction module of the present invention;

[0029] Figure 5 This is a schematic diagram of the neck network structure of the present invention;

[0030] Figure 6 This is a schematic diagram of the bidirectional aggregated multi-scale pyramid structure network of the present invention;

[0031] Figure 7 This is a flowchart of the target tracking algorithm of the present invention. Detailed Implementation

[0032] This invention proposes a tracking method based on trajectory adaptive small target detection, aiming to design a tracking method to solve the problem of missed detection in human-machine target trajectory tracking.

[0033] The tracking method based on trajectory adaptive small target detection proposed in this invention will be described below in specific embodiments:

[0034] Example 1:

[0035] A tracking method based on trajectory adaptive small target detection, such as Figure 1 As shown, it includes the following steps:

[0036] Step 1. Prepare the dataset: Prepare a dataset of images of small targets taken by drone aerial photography;

[0037] Step 2. Constructing the target detection network model: Construct a small target detection network for UAV aerial photography, including a backbone network, a neck network, and four detection heads; the backbone network includes 5 convolutional downsampling modules, 4 SF-C2f feature extraction modules, and an SPPF spatial pyramid pooling module; the SF-C2f feature extraction modules improve feature extraction performance through PConv layers and SimAM attention modules; the neck network includes an upsampling module, an SF-C2f feature extraction module, and a convolutional module, and adopts a bidirectional aggregated multi-scale pyramid structure network to improve the frequency of multi-scale fusion;

[0038] Step 3. Train the network model: Input the prepared dataset into the object detection network model for training, including preprocessing small target images captured by the drone, feature extraction, classification and regression operations, and verifying the model performance through test images;

[0039] Step 4. Design the loss function: By constructing the output image of the object detection network model and labeling it, a composite loss function is adopted, including class classification loss and bounding box regression loss. The class classification loss adopts the contrast loss function and the cross-entropy loss function, and the bounding box regression loss adopts the DFL loss function plus the WIoUv3 loss function.

[0040] Step 5. Construct a target tracking network model: Design a target tracking algorithm, including cascaded matching, calculating the cost matrix, using the Hungarian algorithm for matching, and using the AFK adaptive Kalman filter algorithm to predict and update the target trajectory, so as to realize real-time tracking of small targets captured by UAV aerial photography;

[0041] Step 6. Save the model and select evaluation metrics: Evaluate the accuracy and efficiency of the target tracking algorithm.

[0042] Further, in step 2, the SF-C2f feature extraction module includes: passing through the first convolutional layer, then being divided into two parts, one part being processed by multiple bottleneck modules, concatenated with the other part and passed through the second convolutional layer, and then outputting the result through the SimAm attention module;

[0043] The bottleneck module includes: the input is divided into two parts. The first part passes through the PConv layer and then through the convolutional layer, where batch normalization and ReLU activation functions are added. This is then added to the other part of the input data to output the result.

[0044] The SimAM module accurately assesses the linear separability of a single feature within the same channel from other features by defining an energy function, thereby accurately determining the importance of each feature; the energy function is defined as follows:

[0045]

[0046] In the formula, t and x i w represents the target feature information and other feature information in the channel, respectively. t and b t λ represents the linear transformation weights and biases of t, respectively; index i represents the spatial dimension order, λ is a hyperparameter, and M is the number of all feature information in a single channel.

[0047] Furthermore, in step 2, the neck network includes: extending the PAN-FPN architecture upwards, utilizing the spatial information in the shallow feature maps to retain the high-resolution spatial details needed for detecting small targets, and adding pointwise convolutional layers during the feature map fusion process.

[0048] Furthermore, in step 5, the adaptive Kalman filter algorithm uses an adaptive forgetting factor to control the weight of past data, thereby improving the algorithm's adaptability to parameter changes.

[0049] Furthermore, in step 4, the WIoUv3 loss function in the bounding box regression loss function uses a dynamic non-monotonic mechanism to evaluate the quality of the anchor box, thereby improving the model's ability to locate objects, which is especially suitable for small target detection tasks in UAV aerial photography scenarios.

[0050] Example 2:

[0051] A tracking method based on trajectory adaptive small target detection specifically includes the following steps:

[0052] S1, Prepare the dataset: Prepare a suitable dataset of images of small targets taken by drone aerial photography;

[0053] The image dataset, VisDrone2019, was collected by the AISKYEYE team at the Machine Learning and Data Mining Laboratory of Tianjin University. The benchmark dataset comprises 288 video clips, consisting of 261,908 frames and 10,209 still images, captured by various drone cameras. It covers a wide range of locations (from 14 different cities across thousands of kilometers in China), environments (urban and rural), objects (pedestrians, vehicles, bicycles, etc.), and densities (sparse and crowded scenes).

[0054] S2, Construct a network model for object detection: a small target detection network for drone aerial photography, such as... Figure 2 As shown, it includes a backbone network, a neck network, and four detection heads: the specific structure of the backbone network is as follows. Figure 3Specifically, it includes 5 convolutional downsampling modules, 4 SF-C2f feature extraction modules, and an SPPF spatial pyramid pooling module. Each convolutional downsampling module consists of a convolutional layer, an activation function, and a batch normalization layer. The convolutional downsampling module is used to reduce the size of the feature map. The structure of the SF-C2f feature extraction module is as follows: Figure 4 As shown, the input layer passes through the first convolutional layer, then is divided into two parts. One part is directly passed to the output, while the other part is processed by multiple bottleneck modules. The two parts are concatenated, passed through the second convolutional layer, and then output through the SimAm attention module. The bottleneck modules are as follows: Figure 4 As shown, the SimAM module consists of a series of convolutional layers, including one PConv layer and two convolutional layers. The input is divided into two parts. The first part passes through the PConv layer, then through the convolutional layers, and is followed by batch normalization and ReLU activation. This is then added to the other part of the input to output the result. PConv performs convolution operations on a subset of certain channels while leaving other channels unchanged, thus utilizing the high similarity of feature maps between different channels. This reduces computational complexity. To maintain memory continuity, the start or end sequence of consecutive channels is used to represent the entire feature map. Subsequently, the convolutional layers are merged into PConv to effectively utilize information from all channels, generating an effective receptive field similar to a T-shaped convolution on the input feature map, reducing memory access and computational load. The SimAM module accurately evaluates the linear separability of a single feature within the same channel from other features by defining an energy function, thereby accurately determining the importance of each feature. The energy function is defined as follows:

[0055]

[0056] In the formula, t and x i w represents the target feature information and other feature information in the channel, respectively. t and b t Let be the linear transformation weights and biases of t, respectively. Let index i represent the spatial dimension order, λ be a hyperparameter, and M be the number of all feature information in a single channel.

[0057] The SimAM structure is designed based on neuroscience theory to define an energy function for identifying key neurons and calculating attention weights accordingly. In neuroscience, information-rich neurons often exhibit firing patterns different from surrounding neurons and can inhibit neighboring neurons, a phenomenon known as spatial inhibition. These important neurons are identified using an energy function defined by an equation.

[0058]

[0059] As shown in the above equation, as energy decreases, the uniqueness between neuron t and its surrounding neurons gradually increases, thus making neurons t with lower energy levels more specific and important. The importance of neurons can be determined through… The calculation is performed. When the importance of neurons is evaluated using an energy function, key features are further refined through scaling operations. In summary, the SF-C2f feature extraction module reduces computational load, improves the efficient management of channel information during feature fusion, and effectively prevents the random distribution of features across different channels, thereby improving feature extraction performance.

[0060] Neck network structure such as Figure 5 As shown, it consists of an upsampling module, an SF-C2f feature extraction module, and a convolution module; the connection of the bidirectional aggregated multi-scale pyramid structure network in object detection is as follows: Figure 6 As shown, the PAN-FPN architecture is first extended upwards to utilize spatial information in shallow feature maps to preserve the high-resolution spatial details needed for detecting small targets. Secondly, the first and second layers of the backbone are located in the shallow part of the network, with limited receptive fields, and their spatial location and semantic information may be insufficient, leading to relatively cluttered information. Therefore, in the feature maps... Figure 1 and characteristics Figure 2 To address this issue, a pointwise convolutional layer is added during the fusion process. The pointwise convolutional layer first extracts relevant information from the feature map and filters out irrelevant information for object detection without adding unnecessary parameters, thus refining and enhancing shallow feature information. Finally, an additional downsampling module is added between the bottom-up and top-down paths to increase the frequency of multi-scale fusion, aggregating the rich spatial details in the shallow feature map and the robust semantic information in the mid-level feature map.

[0061] S3, Training the Network Model: Input the dataset prepared in S1 into the object detection network model in S2 for training. First, preprocess the small target images captured by the drone; second, extract features from the images using the improved object detection backbone network to obtain feature maps of the input data, and perform classification and regression operations at each point in the feature maps; next, apply the pre-trained object detection network model to the test images, and obtain the prediction results from the detection head detection from the output feature maps; finally, obtain the final object detection output results through computation.

[0062] S4, Design the loss function: By constructing and labeling the output image of the object detection network, the model parameters are considered pre-trained when the number of training iterations reaches a predetermined threshold or the loss function value falls within a given range. This pre-trained image is then stored, thus obtaining the object detection result. During training, the loss function consists of two parts: classification loss and bounding box regression loss. The regression loss function uses the focus distribution and object bounding box loss function, while the classification loss function uses the contrastive loss function and the cross-entropy loss function. The choice of loss function significantly impacts the model's performance, accurately reflecting the difference between predicted and true values ​​and providing correct feedback on the model's quality.

[0063] Contrast loss function and cross-entropy loss function are selected as classification loss functions to determine the category of target object, and DFL loss function plus WIoUv3 loss function are selected as bounding box regression loss function to determine the position and size of each anchor box.

[0064] The loss function in S4 uses a designed composite loss function, including class classification loss and bounding box regression loss, as shown in the following formula:

[0065] f loss =a1f BCELoss +a2f CLoss +a3f DFLoss +a4f BBRLoss ;

[0066] Category classification loss: The classification loss function uses both contrastive loss and cross-entropy loss. The formula for the binary cross-entropy loss function is as follows:

[0067]

[0068] By calculating the cross-entropy between the predicted probability distribution and the true probability distribution, the difference between the two can be characterized more accurately. When the difference between the model output and the true label is large, the loss function value rises faster, which helps to speed up the convergence of the model.

[0069] The contrastive loss function is used to make positive sample pairs (similar sample pairs) close in the representation space, while negative sample pairs (dissimilar sample pairs) are far apart. Its formula is:

[0070]

[0071] Among them, y i These are labels, where 1 represents a positive sample pair and 0 represents a negative sample pair, D i It is the Euclidean distance between sample pairs, and margin is a hyperparameter representing the minimum distance between negative sample pairs.

[0072] Boundary Regression Loss: The regression loss function uses the distribution focus and target bounding box loss function (f BBRLoss In this invention, the DFL loss function is used for the distributed focus loss and the WIoUv3 loss function is used for the target bounding box loss function to determine the position and size of each anchor box. The formulas for DFL and WIoUv3 are as follows:

[0073] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ));

[0074]

[0075] W i H is the intersection length of the predicted bounding box and the actual bounding box. i S is the intersection width of the predicted bounding box and the actual bounding box. i It is the joint set of predicted boxes and actual boxes, where α and β are hyperparameters, β is the degree of difference between predicted boxes, and x is the hyperparameter. gt and y gt Here, x and y represent the coordinates of the predicted bounding box, and W represents the coordinates of the actual bounding box. g and H g This represents the actual length and width of the frame.

[0076] This invention selects WIoUv3 for the target bounding box loss. On one hand, WIoUv3 combines some advantages of EIoU and SIoU, aligning with the design principles of a superior loss function. On the other hand, WIoUv3 employs a dynamic non-monotonic mechanism to evaluate anchor box quality, making the model focus more on anchor boxes of average quality, thus improving the model's object localization ability. For target detection tasks in UAV aerial photography scenarios, the high proportion of small targets increases the detection difficulty, and WIoUv3 can be dynamically optimized.

[0077] S5, Constructing the target tracking network model: Based on the DeepSort target tracking algorithm, this paper designs the target tracking algorithm. The algorithm flow is as follows: Figure 7As shown, firstly, the actual and predicted positions of small targets captured by the UAV in the k-th frame image are cascaded for matching, resulting in unmatched trajectories, target regions, and initially successful matches. A cost matrix based on cosine distance is calculated, and detection boxes with excessively large cosine matrices are removed. Then, the Hungarian algorithm is used, employing the Mahalanobis distance between the AFK Kalman predicted trajectory and the actual frame as the cost matrix. The trajectory is then matched with the detection boxes, and the result is output to the detection boxes, thus obtaining the optimal matching result. Due to the delay inherent in the Hungarian algorithm, in every 15-frame loop detection, if no match is found within 15 frames, it is removed.

[0078] The S5 algorithm is designed based on the idea of ​​adaptive forgetting factor recursive least squares method. The core idea is to use adaptive forgetting factor to control the weight of past data, thereby improving the algorithm's adaptability to parameter changes. This allows it to better cope with the nonlinear changes in motion trajectory during target tracking.

[0079] The input to the adaptive Kalman filter in this paper includes the initialization measurement matrix z. k Confidence level c of the measured value, and the predicted average value. variance Observation matrix H k Noise covariance R k Kalman gain K k An adaptive forgetting factor μ is used; the output is the final predicted average. and variance The Kalman filter measurement update formula is as follows:

[0080]

[0081] The adaptive Kalman filter process involves first predicting the target state, then matching the tracking and detection results, and finally... The detection results of the new match are updated, the covariance of the measurement noise is used to represent the scale of the measurement noise, and finally the Kalman gain K is calculated. k .

[0082] Interchange-of-Use (IOU) matching is used to match within the initial mismatch and the initial mismatched target region, determining whether the two matches were successful, the second match was successful, and the third match was successful. First, with only one frame, the IOU distance between the target and the frame to be matched is calculated to remove detection boxes with excessive deviations. Based on this, the Hungarian algorithm is used to match moving targets at a cost of the distance between the AFK Kalman prediction path and the actual detected frame, and the corresponding matching results are provided.

[0083] S6. Save the model and select evaluation metrics: Select the best evaluation metrics to measure the accuracy of the target tracking algorithm. Recall, precision, and mean precision are all options to measure the accuracy and efficiency of the target tracking algorithm, thus evaluating its performance. First, the processed image is input into the target detection network designed in this invention to generate a real-time target candidate set. Then, this set is input into the target tracking model designed in this invention to generate target detection boxes, enabling real-time tracking of small targets captured by UAV aerial photography.

[0084] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A tracking method based on trajectory adaptive small target detection, characterized in that, Includes the following steps: Step 1. Prepare the dataset: Prepare a dataset of images of small targets taken by drone aerial photography; Step 2. Constructing the target detection network model: Construct a small target detection network for UAV aerial photography, including a backbone network, a neck network, and four detection heads; the backbone network includes 5 convolutional downsampling modules, 4 SF-C2f feature extraction modules, and an SPPF spatial pyramid pooling module; the SF-C2f feature extraction modules improve feature extraction performance through PConv layers and SimAM attention modules; the neck network includes an upsampling module, an SF-C2f feature extraction module, and a convolutional module, and adopts a bidirectional aggregated multi-scale pyramid structure network to improve the frequency of multi-scale fusion; Step 3. Train the network model: Input the prepared dataset into the object detection network model for training, including preprocessing small target images captured by the drone, feature extraction, classification and regression operations, and verifying the model performance through test images; Step 4. Design the loss function: By constructing the output image of the object detection network model and labeling it, a composite loss function is adopted, including class classification loss and bounding box regression loss. The class classification loss adopts the contrast loss function and the cross-entropy loss function, and the bounding box regression loss adopts the DFL loss function plus the WIoUv3 loss function. Step 5. Construct the target tracking network model: Design the target tracking algorithm, including cascaded matching, calculating the cost matrix, using the Hungarian algorithm for matching, and using the AFK adaptive Kalman filter algorithm to predict and update the target trajectory; Step 6. Save the model and select evaluation metrics: evaluate the accuracy and efficiency of the target tracking algorithm; In step 2, the SF-C2f feature extraction module includes: passing through the first convolutional layer, then being divided into two parts, one part being processed by multiple bottleneck modules, concatenated with the other part and passed through the second convolutional layer, and then outputting the result through the SimAm attention module; The bottleneck module includes: the input is divided into two parts. The first part passes through the PConv layer and then through the convolutional layer, where batch normalization and ReLU activation functions are added. This is then added to the other part of the input data to output the result. The SimAm attention module precisely evaluates the linear separability of a single feature within the same channel from other features by defining an energy function; the energy function is defined as follows: In the formula, and These represent the target feature information and other feature information in the channel, respectively. and They are respectively The linear transformation weights and biases; index i represents the spatial dimension order. is a hyperparameter, and M is the number of all feature information in a single channel.

2. The tracking method based on trajectory adaptive small target detection according to claim 1, characterized in that, In step 2, the neck network includes: extending the PAN-FPN architecture upwards, using the spatial information in the shallow feature maps to retain the high-resolution spatial details needed to detect small targets, and adding pointwise convolutional layers during the feature map fusion process.

3. The tracking method based on trajectory adaptive small target detection according to claim 1, characterized in that, In step 5, the adaptive Kalman filter algorithm uses an adaptive forgetting factor to control the weight of past data.

4. The tracking method based on trajectory adaptive small target detection according to claim 1, characterized in that, In step 4, the WIoUv3 loss function in the bounding box regression loss function uses a dynamic non-monotonic mechanism to evaluate the quality of the anchor box.

Images