Research On Multiple Object Tracking Based On Spatio-temporal Feature

In recent years,the research and application of deep learning technology have made the research of related tasks in the field of computer vision flourish.The proposal and development of attention mechanism has further deepened the study of visual tasks and achieved many remarkable results.As one of the basic visual tasks,Multiple Object Tracking not only needs to detect all objects of interest appearing in the video,but also needs to record and maintain their motion trajectory information,which has great application value in actual scenarios such as intelligent security and smart city.However,the video used in the study will frequently appear target occlusion phenomenon in actual scenarios,which brings great challenges to the research of Multiple Object Tracking.Generally speaking,combining specific spatio-temporal feature in video can solve the object tracking in occlusion scenes to a certain extent.Therefore,aiming at the problem of video Multiple Object Tracking,this thesis uses attention mechanism and other technologies to carry out research on Multiple Object Tracking based on spatio-temporal features,and the main work content is as follows:(1)Aiming at the spatial interaction relationship between objects and the temporal dependence of the object trajectory sequence within a period of time,this thesis proposes a historical spatiotemporal feature modeling method.And relying on the mainstream Multiple Object Tracking framework based on the attention mechanism,the thesis proposes a Multiple Object Tracking method based on historical spatio-temporal feature modeling,which replaces the traditional detection positioning by querying the positioning,and realizes end-to-end process of Multiple Object Tracking.In the MOT17 dataset,the Multiple Object Tracking precision(MOTP)is about 2.3% higher than baseline method,and the number of object identity switching(IDsw)is reduced by 403.The experimental results show that the method effectively avoids the wrong object trajectory association in the tracking process.(2)This thesis further proposes a Multiple Object Tracking algorithm combined with trajectory prediction on the existing basis.When the object is occluded and fail to query the positioning,this algorithm will utilize the position prediction by trajectory prediction module to realize the subsequent tracking process,which effectively avoids the interruption of object trajectories and switching of target identities.In the MOT17 data set,the number of target identity switches is further reduced,and finally reduced by about656 compared with the baseline method.At the same time,the tracking accuracy(MOTA)is about 0.8% higher than that of existing method.In addition,the proportion of the Most Tracked objects(MT)increased by about 0.6% compared with the baseline method,and increased by about 3.3% compared with the existing method.This shows that the algorithm can effectively maintain the trajectory information of the object in complex scenes.(3)This thesis designs and implements a Multiple Object Tracking result visualization system,which is used to visualize the output results of the Multiple Object Tracking algorithm proposed in this thesis.Users can upload videos on the web page to directly obtain the tracking results for performance analysis of the algorithm.