Research On Key Algorithm Of Multi-Target Tracking Based On Deep Learning

With the continuous development and advancement of information technology,the demand for smart cities and social public security issues is increasing.Video surveillance and analysis technology plays an increasingly important role in it.Especially with the increasing number of urban cameras and other monitoring terminals,a large number of video data will be generated every day,which need to be processed.Traditional methods of manual monitoring and analysis have many cases of missing alarms,false positives and other errors.The implementation efficiency is relatively low.Therefore,the research of intelligent video monitoring analysis and understanding technology(including detection,recognition,tracking etc.)has a very important research significance and practical application value.Multi-target tracking is one of the most critical research directions,which has been widely concerned by researchers at home and abroad.Based on the existing mainstream detection-based multi-target tracking system framework,this thesis studies and designs a multi-object tracking system based on deep learning,which is applied to the tracking and recognition of multiple pedestrians in the video surveillance scenarios.Focusing on improving the accuracy and real-time performance of the system,it has a positive scientific research significance.In view of the visual detection part,this thesis proposes a lightweight depth network model applied to the target detection task.Researching on the network structure and loss function of the algorithm,the thesis makes detailed adjustments to the feature extraction backbone network structure to enrich the depth feature information and spatial location information contained in the extracted feature maps,making the network model more suitable for the object detection task.The design of loss function is mainly focused on the two sub-tasks of classification and regression of the detection algorithm.It optimizes and adjusts based on the traditional cross-entropy classification loss and smooth L1 regression loss to improve the robustness of visual detection model for the existing problems such as imbalance of positive and negative sample instances and detection in dense occlusion scenarios.What's more,the network structure of the detection model is constrained and simplified to ensure the real-time performance of the detection system.By comprehensively using the motion feature information and depth appearance feature of targets in the scene,this thesis proposes a multi-target tracking algorithm based on metric learning.Firstly,the Kalman filter algorithm is applied to build the motion prediction model for the targets,and the state prediction results of the existing targets in the new image and the target detection results are matched by IOU metrics matrix.Moreover,appearance features of moving targets are obtained through a lightweight appearance feature extraction model,and then,the cosine metric matching is performed with the depth appearance feature database maintained by the tracked targets.Lastly,the similarity measurement results based on the combined-features are associated to determine the ID information of the moving targets in the video sequence.The whole data association stage of multi-target tracking is completed.Based on the above work,the visual detection part and the data association module of multi-object tracking are integrated in this thesis,so a multi-target tracking system based on deep learning is designed and applied to the video monitoring scene,which mainly tracks and identifies the moving pedestrians in the scene.Meanwhile,in order to further improve the tracking accuracy and real-time performance of the proposed system,more improved optimization strategies are introduced to adjust and design the module algorithm of the system.The algorithm and the multi-target tracking system proposed are verified and evaluated on the VOC dataset,MOT dataset and actual scene data respectively.The experimental results show that the visual detection model can achieve 81.8%mAP recognition rate on the VOC dataset,and the AP value of pedestrian reaches 85.6%.The data association algorithm of multi-target tracking based on metric learning can achieve higher tracking recognition accuracy and real-time performance.The running speed of the tracking stage can reach 40fps.Eventually,the system is tested and evaluated in the video sequence of actual surveillance scene to verify the effectiveness of the proposed system.