Research On Anti-occlusion Multiple Object Tracking Based On Trajectory Prediction

Multiple Object Tracking(MOT)is one of the most challenging tasks in computer vision.The main task is to detect and locate each object and associate the identity information of the same object in different frames.multiple object tracking has important academic research value and application prospects.From the perspective of academic research,multiple object tracking involves many disciplines such as image processing,machine vision and multimedia technology.From the perspective of practical application,multiple object tracking has been applied in many fields such as monitoring,virtual reality,artificial intelligence and automatic driving.Although researchers have proposed a large number of multiple object tracking algorithms,occlusion is still a difficult problem to be solved in multiple object tracking algorithms.In this master’s thesis,we mainly focus on the problem of detection loss of occluded objects,as well as the problem of multiple object tracking under multiple occlusions such as partial occlusion,severe occlusion and complete occlusion.The main work is as follows:(1)Aiming at the problem of occluded object detection loss,a tracking algorithm is proposed to improve occluded object detection,which is called IDet Track.First,the high and low detection frames are reserved,and then the motion similarity between the track prediction frame and the low detection frame,as well as the similarity between the space-time feature of the historical track and the low detection feature,is used to distinguish the low detection frame as the tracking object or the background,finally achieving the goal of recovering the tracking object of the low detection and filtering out the background false detection.Official challenge comparison experiments were conducted on multiple data sets of MOT Challenge,and the algorithm results were analyzed.Compared with the Deep Sort algorithm,the tracking accuracy(MOTA)index has improved by 10.3%,and at the same time,it has achieved better performance in many other evaluation indexes.(2)To solve the problem of local occlusion,an online multi-target tracking algorithm based on space-time features is proposed,called STCTracker.Combining space-time feature matching and pedestrian movement trajectory prediction,the relationship between the features around the target and the target is modeled,and the local correlation module is designed to use the topological structure information between the target and the surrounding environment.The spatial local correlation module is extended to the time dimension,and multi-scale correlation learning is carried out between different frames of temporal information to enhance the pedestrian data association effect.Official challenge comparison experiments were conducted on multiple data sets of MOT Challenge,and the algorithm results were analyzed.On the MOT16 data set,the tracking accuracy(MOTA)index achieved75.2%,and the identification F1(IDF1)value reached 70.0%.Compared with most advanced algorithms,in MOTA,IDF1,MT(Mostly Tracked),ML(Mostly Lost),FP(False Positive),FN(False Negative),ID Sw.(ID Switch)and other indicators were significantly improved.(3)To solve the problem of serious occlusion,a secondary multiple object tracking algorithm,called Re-Tracker,is proposed by matching the spatial position graph.When the object feature is seriously occluded,consider the spatial position relationship between the object and the adjacent object,and use the spatial position relationship of the historical track and the spatial position relationship of the current detected object to match the graph to associate the ID identity of the seriously occluded object.To solve the problem of complete occlusion,a hypothesis frame continuous tracking mechanism is proposed.First,use the determined trajectory prediction box as the target hypothesis box to continue tracking.When the target appears in the field of vision again during this period,the ID identity will remain unchanged.If the target does not appear again,the assumed track and target feature information will be discarded to ensure the tracking system operation speed.The ablation experiment was completed in the MOT17 data set,which showed the effectiveness of each module.An official challenge comparison experiment was conducted on the MOT16,MOT17 and MOT20 data sets.The final algorithm in this master’s thesis achieves 78.0% tracking accuracy and 78.3% IDF1 value on MOT16 data set,78.5% tracking accuracy and 78.0% IDF1 value on MOT17 data set,and72.8% tracking accuracy and 73.4% IDF1 value on MOT20 data set.Compared with most advanced algorithms,the algorithm in this master’s thesis achieves the best performance in the key evaluation indicators of multi-target tracking,MOTA and IDF1,and achieves performance advantages in multiple evaluation indicators such as MT,ML,FP,FN,ID Sw.