Research And Application Of Pedestrian Multiple Object Tracking Based On Deep Learning

With the development of computer vision,it has become an essential part in human life.MOT(Multiple Object Tracking)for pedestrian can be used not only in urban transportation and campus life but also in fields like automatic driving,remote supervision and dangerous working place.With the improvement of the accuracy of deep learning detectors,TBD(Tracking By Detection)becomes the main algorithm in MOT.MOT for pedestrian based on deep learning involves pedestrian detection,feature extraction,similarity calculation and data association.Pedestrian detection and data association are the main factors that affect the MOT algorithm.In this paper we discuss two aspects: pedestrian detection algorithm and occlusion problem solving strategies,which aims to apply our algorithm in complex scenes and improve its long-time tracking ability.First,this paper introduces the development,categories and advantages and disadvantages of object detection algorithms using deep learning and gives the reasons for our selected object detection algorithm.Then we introduce the current researches and frequently used database in MOT.Second,in order to address the low accuracy problem of MOT algorithm in complex scenes,Our proposed MOT algorithm is based on Joint Detection and Embedding algorithm.A parallel-structured attention mechanism is proposed to improve the extraction ability,feature pyramid structure is added into backbone to advance feature fusion and main factors are considered when designing backbone network.Experiments on MOT16 show that the proposed algorithm can achieve 57% MOTA and 56.3% IDF1.Third,to avoid ID switch problem after occlusion occurs,this paper proposes a MOT algorithm for occlusion,which contains a self-adapted Kalman filting and neighboring IOU association.The self-adapted Kalman filting is designed to dynamically adjust the pedestrian velocity factor by judging their motion status.Then a prior velocity formula is proposed to calculate the distance between detection box and tracking box.Neighboring IOU association algorithm regards IOU as cost function for data association,which considers the neighboring targets for data association to find related target in next frame.Experiments on MOT17 Faster RCNN show that the proposed algorithm can achieve 49.7% MOTA and 21.6% IDF1.Finally,to show the performance of MOT algorithm,the innovative points above are combined together to compare with the other mainstream algorithms such as DeepSort and JDE and top-one algorithms in recent years.Experiments on MOT16 show that our algorithm can achieve 65.1% MOTA and 62.3% IDF1,indicating that our algorithm has excellent tracking performance and practical value in real scenarios.