Target tracking model based on enhanced target detection algorithm

Abstract

The invention discloses a target tracking model based on an enhanced target detection algorithm, and relates to the technical field of real-time target tracking. A CD-YOLO detector is constructed; a robust multi-target tracker is constructed; constructing a collaborative detection-tracking optimization mechanism; evaluating and optimizing the performance; according to the method, a deformable convolution and coordinate attention mechanism is introduced into YOLOv10, so that the perceptual ability of the model to geometric deformation and spatial information is enhanced; the CD-YOLO and the enhanced StrongSORT tracker are combined, so that the detection precision and the tracking stability of the system in complex scenes such as dense crowds, serious shielding and variable visual angles are remarkably improved; a collaborative optimization mechanism between detection and tracking is established, so that a detection result and tracking feedback can be mutually enhanced, a closed-loop robust tracking process is formed, and the overall precision and stability are remarkably improved on the premise that the real-time performance of the system is not affected.

Description

Technical Field

[0001] This invention relates to the field of real-time target tracking technology, and in particular to a target tracking model based on an enhanced target detection algorithm. Background Technology

[0002] Multi-object tracking, a core task in computer vision, aims to achieve continuous localization, identity association, and trajectory generation of multiple objects in a video stream. With the development of deep learning technology, object detection has evolved from traditional manual feature-based methods to a series of models such as YOLO based on CNNs, achieving significant improvements in detection accuracy and real-time performance. Object tracking has also evolved from the early SORT to DeepSORT and StrongSORT, which combine appearance features, significantly enhancing identity preservation and occlusion adaptation capabilities. However, in complex scenarios such as dense crowds, frequent occlusion, and similar object appearances, challenges remain, including bounding box drift, frequent ID switching, and trajectory interruptions.

[0003] To address the challenges of multi-target detection and tracking in complex scenarios, the mainstream solution currently is to construct a collaborative optimization framework, leveraging advanced detectors and trackers to jointly improve overall performance. Since target detection and target tracking are not inherently independent tasks, the quality of detection results directly impacts the accuracy and stability of tracking. Collaborative modeling of both can significantly reduce false detections, missed detections, and ID switching issues. However, most existing methods rely solely on single features or static matching mechanisms, making them ill-suited for environments with high-density occlusion and frequent dynamic interactions, particularly exhibiting significant bottlenecks in identity preservation and target re-identification.

[0004] Current multi-target detection and tracking technologies face three key challenges: severe occlusion and viewpoint changes lead to target bounding box drift or loss; highly similar targets are prone to ID switching, resulting in discontinuous tracking trajectories; and the lack of dynamic adaptation mechanisms makes it difficult to adjust detection and matching strategies in real time according to scene complexity. Summary of the Invention

[0005] The purpose of this invention is to provide a target tracking model based on an enhanced target detection algorithm, which solves the problems of severe target occlusion, frequent identity switching, and insufficient detection stability in densely populated scenes in the existing technology.

[0006] To achieve the above objectives, this invention provides a target tracking model based on an enhanced target detection algorithm, the construction method of which includes the following steps: Step 100: Collect pedestrian video data and perform preprocessing; Step 200: Construct the CD-YOLO detector; The CD-YOLO detector is based on the YOLOv10 detection framework and introduces deformable convolution and coordinate attention mechanisms. Step 300: Construct a robust multi-target tracker; The robust multi-target tracker adopts the enhanced StrongSORT framework, which integrates a high-performance appearance feature extractor, an exponential moving average strategy, and an adaptive Kalman filter mechanism. Step 400: Construct a collaborative detection-tracking optimization mechanism; The collaborative detection-tracking optimization mechanism is as follows: the high-confidence bounding boxes generated by the CD-YOLO detector are used as the initialization and state update input of the robust multi-object tracker. The robust multi-object tracker outputs feedback information, which is used to assist the CD-YOLO detector in feature enhancement in complex regions. Bounding boxes with a confidence level higher than a set threshold are designated as high-confidence bounding boxes; otherwise, they are designated as low-confidence bounding boxes. Feedback information includes the probability of the target's existence and a predicted trajectory. Step 500: Set initial training parameters, use a high-density pedestrian dataset for model training and testing, and conduct performance analysis and comparison using evaluation metrics.

[0007] Furthermore, deformable convolution specifically involves introducing a spatial sampling offset. Introducing spatial sampling offset The mathematical expression for deformable convolution is as follows: ; in, This represents any point on the output feature map. This represents the offset of each point in the convolution kernel relative to the center point. This represents the weights at the corresponding positions of the convolution kernel. Indicates the spatial sampling offset. Indicates on the input feature map The element value at position , Indicates the output feature map The element value at a given position is obtained by convolving the kernel with the input feature map. This indicates the sampling region of the convolution kernel.

[0008] Furthermore, the mathematical expression for the coordinate attention mechanism is: ; in, These are input features. For coordinate attention, For a 1x1 convolution, It is the Sigmoid activation function. For horizontal average pooling, This is the average pooling in the vertical direction.

[0009] Furthermore, the high-performance appearance feature extractor is specifically designed to improve the appearance feature extractor by replacing the traditional CNN network with a BOT, using ResNeSt50 as the backbone model, and pre-training on the dataset.

[0010] Furthermore, the exponential moving average strategy specifically replaces the function library mechanism with EMA, and smooths feature updates by introducing an inertia term, as shown in the following expression: ; in, This represents the appearance state of the i-th trajectory in frame t. This indicates updating the appearance state of the i-th trajectory in frame t-1. This represents the appearance embedding feature of the i-th trajectory in frame t. This is the momentum term.

[0011] Furthermore, the adaptive Kalman filtering mechanism specifically involves using an improved NSA Kalman filter to adaptively adjust the measurement noise covariance matrix based on the detected confidence level, as expressed below: ; in The preset constant measurement noise covariance matrix, Let k be the confidence level of the detection in state k. To measure the noise covariance matrix.

[0012] Furthermore, the robust multi-target tracker also includes fusing appearance features and motion features to construct a cost function for target association. The expression for the cost function is as follows: ; Where C is the cost matrix. For appearance cost, For the cost of exercise, These are the weighting factors for the cost matrix.

[0013] Furthermore, the feedback information is used to assist the CD-YOLO detector in feature enhancement in complex regions, specifically as follows: For low-confidence target regions reported by the robust multi-target tracker, feature enhancement is used to strengthen the feature representation of these regions. For the target's trajectory prediction area, the region of interest of the detector is adjusted to optimize the allocation of detection computational resources.

[0014] Furthermore, the target tracking method includes the following steps: Step 101: Collect pedestrian video data and perform preprocessing to obtain a video stream; Step 102: Input the video stream into the CD-YOLO detector, model the pedestrian's pose changes and occlusion features through deformable convolution, and enhance the expression of spatial location information using the coordinate attention mechanism, and output the detection results including bounding boxes, confidence scores and target categories; Step 103: Extract appearance features for each bounding box using the ResNeSt50 BOT-ReID network, and smooth the update of appearance features using an exponential moving average strategy; predict the motion state of each trajectory using an NSA Kalman filter; construct a cost function by fusing appearance features and motion features between the detection results and the predicted trajectory; and use the Hungarian algorithm of Vanilla Matching to solve for the global optimal matching result to complete the data association, outputting the trajectory ID, current position, historical trajectory and status information of each target.

[0015] Therefore, the target tracking model based on the above-mentioned enhanced target detection algorithm adopted in this invention has the following beneficial effects: 1. Deformable convolution and coordinate attention mechanisms are introduced in YOLOv10 to enhance the model’s ability to perceive geometric deformation and spatial information, enabling the detector to have stronger localization accuracy and occlusion adaptability in complex environments. 2. By combining the CD-YOLO detector with the robust multi-target tracker of the enhanced StrongSORT, a detection-tracking collaborative optimization framework is formed, which can reduce the number of ID switching times per 100 frames by about 23 times and reduce trajectory fragmentation by about 18%. This significantly improves the detection accuracy and tracking stability of the system in complex scenarios such as dense crowds, severe occlusion, and changing viewpoints, and breaks through the limitations of traditional detection-tracking models in terms of robustness and continuity. 3. A collaborative optimization mechanism between detection and tracking was established, enabling detection results and tracking feedback to reinforce each other and form a closed-loop robust tracking process, which significantly improves overall accuracy and stability without affecting the system's real-time performance.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a target tracking model based on an enhanced target detection algorithm according to the present invention; Figure 2 This is a schematic diagram of the structure of a target tracking model based on an enhanced target detection algorithm according to the present invention; Figure 3 This is a schematic diagram of the CD-YOLO network architecture of a target tracking model based on an enhanced target detection algorithm according to the present invention; Figure 4This is a StrongSORT structure diagram of a target tracking model based on an enhanced target detection algorithm according to the present invention; Figure 5 This is a diagram illustrating the target tracking process of a target tracking model based on an enhanced target detection algorithm according to the present invention. Figure 6 This is a comparison chart of the accuracy of CD-YOLO of the present invention with YOLOv8 and YOLOv10; Figure 7 This is a comparison chart of the recall rates of CD-YOLO, YOLOv8, and YOLOv10 of the present invention; Figure 8 This is a comparison chart of mAP@50 between CD-YOLO of the present invention and YOLOv8 and YOLOv10; Figure 9 This is a comparison chart of mAP@50-95 between CD-YOLO of the present invention and YOLOv8 and YOLOv10. Detailed Implementation

[0018] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0019] Please see Figure 1-5 A target tracking model based on an enhanced target detection algorithm is constructed using the following steps: Step 100: Data Acquisition and Preprocessing; In embodiments of the present invention, data acquisition and preprocessing specifically involve: acquiring pedestrian videos from multi-scene monitoring data, including high-density scenes such as streets, shopping malls, and subways; dividing the video into frames of fixed-resolution images and performing illumination equalization, histogram equalization, and Gaussian filtering for noise reduction; and using the MOT17 dataset to label pedestrian bounding boxes and identification IDs to construct training, validation, and test sets.

[0020] Step 200: Construct the CD-YOLO detector; The CD-YOLO detector is formed by introducing Deformable Convolution (DCN) and Coordinate Attention (CA) mechanisms on the basis of the YOLOv10 detection framework. The improved CD-YOLO detector can achieve more stable and accurate target detection results in dense crowds and complex environments.

[0021] The CD-YOLO detector consists of a Backbone, a Neck (feature fusion network), and a Head. The Backbone extracts basic image features and enhances feature representation by combining deformable convolution and coordinate attention mechanisms. The Neck integrates multi-layer features to construct a multi-scale feature pyramid with top-down and bottom-up approaches. The MSF module (Multi-Scale Feature Fusion Module) further fuses and weights the features output from the Neck at different scales to enhance the detection capability for pedestrians at different scales and those with occlusions. The Head module outputs the final bounding box, confidence score, and target category based on the fused features, thus achieving the target detection function.

[0022] The deformable convolution and coordinate attention mechanisms of the CD-YOLO detector are as follows: Deformable convolution: Deformable convolution, by dynamically learning sampling offsets, can more flexibly capture the spatial features of a target, focusing on the region or target of interest. Deformable convolution introduces learnable spatial sampling offsets. This enables the convolution kernel to adaptively adjust the receptive field, improving its adaptability to pose changes and occluded targets.

[0023] Introducing learnable spatial sampling offset The mathematical expression for deformable convolution is as follows: ; in, This represents any point on the output feature map. This represents the offset of each point in the convolution kernel relative to the center point. This represents the weights at the corresponding positions of the convolution kernel. Indicates the spatial sampling offset. Indicates on the input feature map The element value at position , Indicates the output feature map The element value at a given position is obtained by convolving the kernel with the input feature map. This indicates the sampling region of the convolution kernel.

[0024] Coordinate attention mechanism: Coordinate attention is a highly efficient attention mechanism that improves the performance of lightweight networks by embedding positional information into channel attention. It generates position-sensitive attention weights by separately encoding global information in the horizontal and vertical directions, thereby enhancing feature representation capabilities.

[0025] The mathematical expression for the coordinate attention mechanism is: ; in, These are input features. For coordinate attention, For a 1x1 convolution, It is the Sigmoid activation function. For horizontal average pooling, This is the average pooling in the vertical direction.

[0026] Step 300: Construct a robust multi-target tracker; The robust multi-target tracker employs the enhanced StrongSORT framework, fusing appearance and motion features for target association. This enables it to maintain trajectory continuity and identity consistency even under frequent occlusion and complex motion patterns. By optimizing the data association strategy through a cost function combining appearance and motion information, and introducing camera motion compensation (ECC) to eliminate nonlinear motion errors, the matching cost becomes more accurate, significantly reducing ID switching and trajectory breakage. During the matching process, a gating mechanism is used to filter out unreasonable matching pairs, improving the reliability of the association.

[0027] The robust multi-target tracker integrates a high-performance appearance feature extractor (BOT+ReID), an exponential moving average (EMA) strategy, and an adaptive Kalman filter (NSA) mechanism, as detailed below: High-performance appearance feature extractor: The traditional CNN network is replaced with BOT (Bag of Tricks) to improve the appearance feature extractor. ResNeSt50 is used as the backbone model and it is pre-trained on the DukeMTMCreID dataset. These features are crucial for distinguishing and recognizing different targets.

[0028] Exponential Moving Average (EMA) strategy: This replaces the library mechanism with EMA (Exponential Moving Average), introducing an inertial term to smooth feature updates. EMA helps reduce the impact of noise and outliers, making tracking more stable. The appearance state of the i-th trajectory in frame t is updated using EMA. The expression is as follows: ; in, This indicates updating the appearance state of the i-th trajectory in frame t-1. This represents the appearance embedding features currently detected. =0.9 represents the momentum term. The EMA update strategy fully utilizes inter-frame feature change information and can effectively suppress detection noise.

[0029] Adaptive Kalman Filtering Mechanism: For the target's motion state, this invention employs an improved NSA (Neural Network-based Appearance) Kalman filter, suitable for nonlinear complex motion. It can adaptively change the noise based on the detection confidence level to adapt to more complex motion patterns. This filter adjusts the noise based on the detection confidence level. Adaptive adjustment of measurement noise covariance matrix The expression is as follows: ; in The preset constant measurement noise covariance matrix, Let k be the confidence level of the detection in state k. To measure the noise covariance matrix.

[0030] When the noise in the detection results is low, When the noise level approaches zero, the system will detect higher state update weights; conversely, when the noise level is high, Close to the original The value is reduced accordingly, thus minimizing the impact of detection.

[0031] Integrating appearance and motion features for target association: Traditional DeepSORT algorithms are limited by using only motion information as a threshold during the data association stage. This invention, however, employs a cost function that integrates appearance and motion features. The expression for the cost function is as follows: ; Where C is the cost matrix. For appearance cost, For the cost of exercise, The weighting factor for the cost matrix is ​​set to 0.98 in this embodiment, emphasizing the dominant role of appearance information in the matching process while taking into account the auxiliary role of motion information. This weighting strategy performs better in scenarios with occlusion and complex crowd interactions.

[0032] The current frame's detection is matched with existing trajectories, and the data association and matching strategy are optimized: Traditional DeepSORT uses a cascade matching priority based on trajectory age to limit the matching order, which may reduce the matching effect in complex scenarios. To address this, this invention adopts a unified matching priority strategy (Vanilla Matching) to replace the traditional Cascade cascade matching method. It uses the Hungarian algorithm to solve for the globally optimal matching result, ensuring that all trajectories participate in the association synchronously without bias towards new or old trajectories, thereby effectively improving the ID recovery rate and matching accuracy in high-density scenarios.

[0033] Step 400: Construct a collaborative detection-tracking optimization mechanism; To improve the overall performance of multi-target detection and tracking systems, this invention proposes a collaborative optimization mechanism for stable data interaction between a CD-YOLO detector and a robust multi-target tracker. This mechanism leverages the complementary strengths of both to mitigate issues such as missed detections, false detections, and ID drift. The specific scheme is as follows: Collaborative data interaction mechanism design: High-confidence bounding boxes generated by the CD-YOLO detector are used as inputs for the initialization and state update of the robust multi-object tracker, ensuring the accuracy of the robust multi-object tracker in the target generation and state maintenance stages. To this end, this invention designs a confidence threshold filtering module, which only inputs bounding boxes with confidence levels higher than a set threshold into the robust multi-object tracker, reducing the risk of false matching. The robust multi-object tracker generates feedback information based on trajectory state and motion consistency. This feedback information includes the probability of target presence and trajectory prediction, used to assist the CD-YOLO detector in performing feature enhancement and candidate region re-evaluation in complex regions (such as occlusion overlap areas and areas of varying illumination), improving detection accuracy and robustness.

[0034] Feature enhancement and dynamic region adjustment: For low-confidence target regions reported by the robust multi-target tracker, feature enhancement is used to strengthen the feature representation of the region, thereby improving the CD-YOLO detector's ability to perceive targets with weak features. For the target motion trajectory prediction area, the region of interest (ROI) of the CD-YOLO detector is adjusted to optimize the allocation of detection computing resources and improve detection efficiency and accuracy.

[0035] Step 500: Performance evaluation and optimization.

[0036] To comprehensively verify the performance of the multi-target detection and tracking system of this invention, the publicly available standard MOT17 dataset was used as the test benchmark, and multiple classic indicators were combined for quantitative evaluation to ensure the objectivity and comparability of the results. The specific implementation scheme is as follows: Performance index definition and calculation: Multi-target tracking accuracy (MOTA): By comprehensively considering three error factors—false negatives (FN), false positives (FP), and identity switches (IDSW)—the MOTA metric is defined to measure the overall tracking accuracy of the system. The formula for calculating multi-target tracking accuracy is: ; in, , , These represent the number of missed detections, false detections, and ID switching counts in frame t, respectively. Let be the number of real targets in frame t. For multi-target tracking accuracy.

[0037] Identity Recognition F1 Score (IDF1): Measures the system's performance in maintaining consistency of the target identity. It is calculated using correctly identified identity tracking positives (IDTP), incorrectly identified false positives (IDFP), and false negatives (IDFN), and is defined as follows: ; Mean precision (mAP) and recall: Using an IoU threshold range of 0.5-0.95, the detection performance of the detector is evaluated under different stringent standards, comprehensively reflecting the accuracy of localization and classification.

[0038] High-order tracking performance metrics (HOTA): These metrics combine detection accuracy and correlation accuracy to reflect the overall tracking quality and help verify the robustness of the system.

[0039] Experimental Design and Comparison Methods: Training and testing video sequences provided by the official MOT17 platform were used, covering various complex scenes such as dense crowds, severe occlusion, and varying lighting conditions. The current mainstream object detection models YOLOv8 and YOLOv10, along with a multi-object tracking combination scheme based on StrongSORT, were used as the baseline for comparison. Under unified hardware environment and parameter settings, the algorithm and comparison method of this invention were run, and the detection results and tracking trajectories for each frame were recorded. The aforementioned metrics were then calculated.

[0040] Results Analysis and Optimization Strategies: For different scenario characteristics, the system dynamically adjusts weight parameters and confidence thresholds to achieve adaptive performance optimization, further improving the stability and real-time response capabilities of detection and tracking. Combined with an online analysis mechanism for mismatched and missed samples, the model parameters and association strategies are iteratively optimized to continuously narrow the error boundary and improve the overall robustness of the system.

[0041] Please see Figure 6-9 The experiment proves the following: The training process of this invention was conducted under the Ubuntu 20.04 operating system, using CUDA 11.6 and PyTorch 1.12 for model development and training. The experimental environment configuration is as follows: GPU: NVIDIA GeForce RTX 4060 GPU, 8GB VRAM; CPU: Intel Core i7-12700H; RAM: 32GB; Deep learning framework: PyTorch 1.12; All neural network models in this invention are trained using the AdamW optimizer with an initial learning rate of 0.001 and a cosine annealing strategy. The training period is 300 epochs. The batch size is set to 16, and an early stopping mechanism is used to prevent overfitting. Data augmentation is performed using Mosaic (probability 0.8), MixUp (probability 0.2), and HSV-adjusted losses. The CIoU Loss function is used. The tracking part uses the standard StronSORT. Evaluation metrics include average accuracy mAP@0.5:0.95, standard accuracy mAP@0.5, inference speed (FPS), multi-object tracking accuracy (MOTA), identity F1 score (IDF1), and high-order tracking accuracy (HOTA).

[0042] In target tracking, the CD-YOLO-StrongSORT method proposed in this invention is compared with the following two mainstream recommendation algorithms: DeepSORT: A classic multi-target tracking method based on detection and Kalman filtering.

[0043] StrongSORT: An improved version based on DeepSORT, enhancing feature extraction and data association capabilities.

[0044] Experimental results: Table 1 shows the comparative experimental results of different target tracking methods on the MOT17 dataset. The evaluation metrics include: IDF1 (identity F1 score), MOTA (multi-target tracking accuracy), and HOTA (higher-order tracking accuracy).

[0045] Table 1. Comparison Results of Multi-Target Tracking

[0046] As shown in Table 1, StrongSORT combined with the CD-YOLO detector achieves leading performance in multiple metrics, especially the IDF1 metric, which is 4.5 percentage points higher than the baseline DeepSORT, significantly enhancing the identity preservation capability. At the same time, it also surpasses other combinations in MOTA and HOTA metrics, reflecting the overall improvement in tracking performance.

[0047] Model ablation experiment: To further verify the contribution of each model component, an ablation experiment was conducted, in which some key modules were removed, and the changes in recommendation performance were observed, as shown in Table 2.

[0048] Table 2 Ablation experiments of key components of CD-YOLO-StrongSORT

[0049] The results show that: The coordinate attention module significantly improves identity preservation (IDF1 metric). The deformable convolution module contributes substantially to the overall tracking accuracy (MOTA). The fusion of the two modules in CD-YOLO achieves synergistic effects, improving the HOTA metric by 0.2-0.3 compared to the single-module version, thus enhancing the system's robustness and stability.

[0050] Based on the two experiments, the combination of StrongSORT and CD-YOLO achieves the best balance between accuracy and robustness in multi-target tracking tasks through the synergistic optimization of detection and tracking, making it particularly suitable for complex and high-density real-world scenarios.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A target tracking model based on an enhanced target detection algorithm, characterized in that, The construction method includes the following steps: Step 100: Collect pedestrian video data and perform preprocessing; Step 200: Construct the CD-YOLO detector; The CD-YOLO detector is based on the YOLOv10 detection framework and introduces deformable convolution and coordinate attention mechanisms. Step 300: Construct a robust multi-target tracker; The robust multi-target tracker adopts the enhanced StrongSORT framework, which integrates a high-performance appearance feature extractor, an exponential moving average strategy, and an adaptive Kalman filter mechanism. Step 400: Construct a collaborative detection-tracking optimization mechanism; The collaborative detection-tracking optimization mechanism is as follows: the high-confidence bounding boxes generated by the CD-YOLO detector are used as the initialization and state update inputs of the robust multi-object tracker. The robust multi-object tracker outputs feedback information, which is used to assist the CD-YOLO detector in feature enhancement in complex regions. Bounding boxes with a confidence level higher than a set threshold are designated as high-confidence bounding boxes; otherwise, they are designated as low-confidence bounding boxes. Feedback information includes the probability of the target's existence and a predicted trajectory. Step 500: Set initial training parameters, use a high-density pedestrian dataset for model training and testing, and conduct performance analysis and comparison using evaluation metrics.

2. The target tracking model based on an enhanced target detection algorithm according to claim 1, characterized in that, Deformable convolution specifically involves introducing a spatial sampling offset. Introducing spatial sampling offset The mathematical expression for deformable convolution is as follows: ； in, This represents any point on the output feature map. This represents the offset of each point in the convolution kernel relative to the center point. This represents the weights at the corresponding positions of the convolution kernel. Indicates the spatial sampling offset. Indicates on the input feature map The element value at position , Indicates the output feature map The element value at a given position is obtained by convolving the kernel with the input feature map. This represents the sampling region of the convolution kernel.

3. A target tracking model based on an enhanced target detection algorithm according to claim 2, characterized in that, The mathematical expression for the coordinate attention mechanism is: ； in, These are input features. For coordinate attention, For a 1x1 convolution, It is the Sigmoid activation function. For horizontal average pooling, This is the average pooling in the vertical direction.

4. A target tracking model based on an enhanced target detection algorithm according to claim 3, characterized in that, The high-performance appearance feature extractor is specifically designed to improve the appearance feature extractor by replacing the traditional CNN network with a BOT, using ResNeSt50 as the backbone model, and pre-training on the dataset.

5. A target tracking model based on an enhanced target detection algorithm according to claim 4, characterized in that, The exponential moving average strategy specifically replaces the function library mechanism with EMA, and smooths feature updates by introducing an inertia term, as shown in the following expression: ； in, This represents the appearance state of the i-th trajectory in frame t. This indicates updating the appearance state of the i-th trajectory in frame t-1. This represents the appearance embedding feature of the i-th trajectory in frame t. This is the momentum term.

6. A target tracking model based on an enhanced target detection algorithm according to claim 5, characterized in that, The adaptive Kalman filtering mechanism specifically employs an improved NSA Kalman filter, adaptively adjusting the measurement noise covariance matrix based on the detected confidence level, as expressed below: ； in The preset constant measurement noise covariance matrix, The confidence level of the detection at state k. To measure the noise covariance matrix.

7. A target tracking model based on an enhanced target detection algorithm according to claim 6, characterized in that, The robust multi-target tracker also includes fusing appearance features and motion features to construct a cost function for target association. The expression for the cost function is as follows: ； Where C is the cost matrix. For appearance cost, For the cost of exercise, These are the weighting factors for the cost matrix.

8. A target tracking model based on an enhanced target detection algorithm according to claim 7, characterized in that, Feedback information is used to assist the CD-YOLO detector in feature enhancement in complex regions, specifically as follows: For low-confidence target regions reported by the robust multi-target tracker, feature enhancement is used to strengthen the feature representation of these regions. For the target's trajectory prediction area, the region of interest of the detector is adjusted to optimize the allocation of detection computational resources.

9. A target tracking model based on an enhanced target detection algorithm according to claim 8, characterized in that, The target tracking method includes the following steps: Step 101: Collect pedestrian video data and perform preprocessing to obtain a video stream; Step 102: Input the video stream into the CD-YOLO detector, model the pedestrian's pose changes and occlusion features through deformable convolution, and enhance the expression of spatial location information using the coordinate attention mechanism, outputting detection results including bounding boxes, confidence scores and target categories; Step 103: Extract appearance features for each bounding box using the ResNeSt50 BOT-ReID network, and smooth the update of appearance features using an exponential moving average strategy; predict the motion state of each trajectory using an NSA Kalman filter; construct a cost function by fusing appearance features and motion features between the detection results and the predicted trajectory; and use the Hungarian algorithm of Vanilla Matching to solve for the global optimal matching result to complete the data association, outputting the trajectory ID, current position, historical trajectory and status information of each target.

Images