Research On Algorithm And Application Of Multi-pedestrian Tracking In Videos

The monitoring systems in Skynet project are interconnected,and the surveillance videos are uploaded to the server for unified processing.However,it is difficult for the low-level video understanding technology to process the constantly generated massive video data,which still heavily relies on human eyes for judgment.Multi-pedestrian tracking is a video understanding technology closest to people’s livelihood.It can continuously track the trajectory of the target pedestrian in the videos without switching or losing the trajectory identity.It is widely used in the fields of epidemic tracing,intelligent security,and automatic driving.In this thesis,two method are proposed to solve the problem of multi-pedestrian tracking in single shot scene,including detector,re-identification,and tracker.The main innovations and works of this thesis are as follows:In this thesis,based on the improvement of Deep SORT algorithm commonly used in industry,a two-step method is proposed to track occluded pedestrians.By mixing spatio-temporal attention mechanism with global-local occlusion coefficient self-adaptation.Experiments on MOT16 data set show that MOTA is 65.7,which is 7% higher than the original algorithm,which proves that the proposed two-step method can better track occluded pedestrians.In this thesis,an end-to-end one-step multi-pedestrian tracking method is proposed.To improve the tracking speed,it processes simultaneously in the same network the pedestrian detection and re-identification,whose data features between the two are shared..In this paper,firstly,the pedestrian movement model and the center point detection are established to get the optimal state estimation of pedestrians.Then,the pedestrian recognition model based on deep feature fusion uses Mahalanobis distance and cosine distance to enhance the pedestrian identification ability.Finally,the Hungarian algorithm is applied for online data association.The one-step method is tested on MOT15,16,and 17 data sets;and MOTA is 63.5,72.4 and70.9,respectively.Among them,the IDF1 index is the best,and the real-time video tracking rate is maintained.The results indicate that the proposed one-step method can effectively improve the multi-pedestrian tracking rate,which is closer to the actual demand.