Research On Multi-object Tracking Algorithm Based On Joint Detection

Multi-object tracking is one of the important branches in the field of computer vision.Its main task is to locate the objects of interest in a video frame,and then associate the same objects across different frames with a unique identifier,thus forming a tracking trajectory for each object throughout the entire sequence.Currently,it has been widely used in fields such as intelligent monitoring,autonomous driving,and human-computer interaction.However,the mainstream detection-based multi-object tracking algorithms currently separate the detection and tracking processes,leading to redundant computation and difficulty in improving tracking speed.Additionally,in practical tracking scenarios,there are significant differences in objects scales which make it challenging for algorithms to achieve performance balance when tracking objects of different scales.This thesis focuses on addressing these issues through the following main approaches:1.Designed a multi-object tracking algorithm based on joint detection,which integrates the target detection task and embedded feature extraction task into an end-toend network model.This solves the problem of redundant computation in detection-based tracking algorithms and increases the tracking speed of the algorithm by at least 9FPS.2.Using feature pyramid network structure to fuse information from feature maps at different depths in the backbone network improves the algorithm’s ability to track objects of different scales by enabling low-level feature maps to contain more high-dimensional semantic information from deep-level feature maps.Meanwhile,the model also uses appearance and motion features of objects for data association,enhancing the robustness of the algorithm during tracking.3.Optimizing the post-processing of feature maps,using a multi-branch feature map processing network to handle different anchor boxes for the same anchor point,so that stable tracking of objects can be achieved even when occlusion occurs between objects.In addition,this thesis also optimizes the competition issue between tasks in multi-task learning,reducing the mutual influence between object detection and embedded feature extraction tasks in multi-object tracking algorithms.4.An analysis and comparison of different feature map fusion methods were conducted,comparing the basic feature pyramid structure with various fusion methods such as bidirectional feature fusion and complex bidirectional feature fusion.These feature map fusion methods help improve the performance of object detection tasks,thereby enhancing the tracking effect of multi-object tracking algorithms.The above research content has been verified through theoretical deduction and experimental analysis.The experimental results show that the multi-object tracking algorithm based on joint detection,after feature map post-processing optimization,improves at least 1.2% in MOTA index.Moreover,after changing the feature map fusion method,the number of times that object identity changes during the tracking process decreased from 995 to 823.