Research On Multi-objective Tracking Algorithm Based On Vision

Vision-based object detection and tracking algorithm is a research hotspot in computer vision tasks,and have a wide range of application prospects in intelligent surveillance,national defense and military,automatic driving,sports and many other fields.According to the number of tracking targets,this type of algorithm can be divided into two types: single-target tracking and multi-target tracking.This thesis mainly studies multi-target tracking algorithms.After decades of research by scholars,the multi-object tracking algorithm has made great progress,but there are still some problems and challenges,such as the tracking algorithm with high tracking accuracy tends to track slowly,which is the biggest obstacle to the multi-object tracking algorithm from the laboratory to the practical application.Therefore,this thesis focuses on the research of multi-objective tracking algorithms that can not only ensure tracking accuracy,but also effectively improve the real-time performance of the tracking.The specific work of this thesis is as follows:(1)This thesis sorts out the theoretical knowledge of deep learning commonly used in multi-target tracking algorithms,summarizes the motion models and appearance models commonly used in multi-target tracking algorithms to extract target features,and introduces the common matching algorithms,datasets and evaluation indicators in multitarget tracking algorithms.(2)In this thesis,a multi-target tracking algorithm based on multi-scale feature extraction network is studied,and the core module of the algorithm is multi-scale feature extraction network.After comprehensively considering the differences in feature extraction between the target detection task and the target appearance embedding feature extraction task,the network uses different modules to better balance the relationship between these two subtasks,so the network can output the detection results and target appearance embedding features that meet the requirements of the multi-target tracking algorithm at the same time,thereby effectively improving the real-time performance of the multi-target tracking algorithm.Experiments show that the algorithm balances tracking speed and tracking accuracy well.(3)In this thesis,a multi-objective tracking algorithm based on threshold separation association strategy is studied.Although the multi-target tracking algorithm has high tracking accuracy after adopting the target appearance embedded feature,it also limits the tracking speed of the multi-target tracking algorithm.Therefore,the algorithm eliminates the task of appearance embedding feature extraction,and then adopts a high-performance detection algorithm and a data association strategy based on threshold separation to ensure the speed and accuracy of tracking.The data association strategy based on threshold separation only uses the location information and motion information of the target,but its performance is not inferior to the data association strategy that uses the embedded feature of the target appearance.