Research And Implementation Of Multi-target Tracking Algorithm Under Air-ground Observation

Multi-object tracking(MOT)algorithms under aerial and ground observation conditions have wide applications in both civilian and military domains.Improving the accuracy and robustness of multi-object tracking algorithms is of great significance.This thesis is based on YOLOX object detection algorithm and Bytetrack object tracking algorithm to improve the accuracy and real-time performance of multi-object tracking algorithms in practical applications,and deploy the improved algorithms on end devices for aerial and ground tracking.The research work and contents of this thesis include:(1)Swin Transform is introduced as the feature extraction network for the YOLOX detector to enhance detection accuracy.An improved CBAM attention mechanism is incorporated into the feature fusion network.The CIOU and AFL loss functions are used as network training loss functions.Experimental results on the Vis Drone2019-DET dataset show that compared to the YOLOX model,the YOLOX-SW model achieves an improvement of approximately 1.08% in m AP(IOU=0.50)and approximately 1.6% in average recall rate(max Dets=100).The proposed YOLOX-SW model improves the detection accuracy.(2)Based on the convolutional kernel dynamic selection mechanism(SKC)and the OSNet network,this thesis proposes the SKC-OSNet network as the target appearance feature extraction model for the Bytetrack tracking algorithm.Experimental results on the Vis Drone2019-MOT test dataset show that the Bytetrack tracking method using the SKC-OSNet target appearance feature extraction model improves the MOTA by approximately 1.3%,MOTP by approximately 0.9%,and IDF1 by approximately 2.4%compared to previous methods.(3)This thesis proposes a new tracking target state vector representation to improve the issue of inaccurate bounding box prediction during the tracking process.Furthermore,to mitigate the influence of camera rigid motion on the target position and motion trajectory during multi-object tracking,this thesis introduces camera motion compensation to correct the predicted bounding box position.Finally,the NSA Kalman filtering algorithm is introduced to address the impact of unstable noise during the tracking process.The proposed improvements in this thesis reduce IDS by approximately 596 compared to the Bytetrack algorithm on the Vis Drone2019-MOT test dataset,enhancing the robustness of multi-object tracking.(4)Tensor RT is utilized to accelerate the improved algorithm models,which are ultimately deployed on the NVIDIA Jetson AGX Xavier for verifying the algorithm’s accuracy in practical applications.Additionally,a multi-object tracking visualization software is developed based on the Py QT5 library.