The Research Of Visual Multiple Object Tracking Technology In Video Surveillance

As one of the typical applications of computer vision technology,video surveillance system has been widely used in civil and military fields including road safety,public security,et al.Pedestrian detection and online multi-target tracking in complex scenes are valuable and significant research directions.In order to achieve accurate and fast pedestrian detection and multi-target tracking algorithm in monitoring scenario,the algorithms of target detection,motion estimation,data association and apparent modeling are deeply researched and analyzed.In this paper,several improvements on the above algorithms are proposed to form a complete pedestrian multi-target tracking system design.For the system,the following work is completed:First,the multi-scale pedestrian detection algorithm under complex scenes are deeply researched.The shortcomings of current common pedestrian detection algorithms are compared and analyzed.We use YOLOv3(You Only Look Once)algorithm for target detection which builds a feature pyramid network based on fully convolutional residual network and uses multi-layer features to detect multi-scale targets.We carry out clustering optimization on anchor boxes to further improve the detection accuracy,reduce missed detection and false alarm.Then,based on the pedestrian detection results,Kalman filter is used to estimate and predict the motion and scale states of the targets.Considering the uncertainty caused by false detection,missed detection and motion state estimation,Hungarian algorithm is adapted to associate detection results with trajectory targets.When using raditional Euclidean distance between the target centers as distance metric,cost matrix is sensitive to scale of the targets and the distance threshold setting is not flexible.We propose using IOU(Intersection Over Union)metric between boundary boxes based on tracking gate,which is calculated by the Mahalanobis distance between state prediction and the observation.The proposed optimization of cost matrix effectively improves the accuracy and stability of multi-target tracking.Finally,we adapt the basic idea of apparent modeling in pedestrian re-identification,optimize traditional softmax classifier for cosine similarity and realize the effective extraction and matching of the appearance features.In order to reduce mismatch and identity switches caused by object interaction and occlusion,we propose the pedestrian multi-target tracking algorithm combining appearance feature and spatial information.Considering the lack of effective feature updating mechanism in multi-target tracking field,a sparse updating mechanism with inter-target occlusion judgment is proposed to effectively prevent tracking drift.