Detection Of Multi-Camera Pedestrian Spatial-Temporal Trajectory Outliers In Geographic Scene

With the continuous development of intelligent surveillance equipment,a large amount of surveillance video data has been collected and stored in the database.Mining effective information from massive video target trajectories has become a current research hotspot.Detection of video target trajectory outliers is an important topic of video target trajectory data mining.The goal is to find a small number of trajectories in the trajectory data set that are significantly different from other trajectory data.However,in recent years,the video surveillance system has developed from a single camera to a multi-camera cooperation system,resulting in the emergency of a large number of multi-camera video target trajectories.Its characteristics have a spacing partition on the one hand,showing the characteristics of segmented continuous;On the other hand,the distribution of trajectories in the field of view of different cameras is also uncertain.The traditional trajectory outliers detection algorithm is limited to continuous trajectories in 2D map space,and does not comprehensively measure and analyze the local and global differences between multi-camera video object trajectories,which will lead to a large number of false detection or missed detection of trajectory outliers.Therefore,there is an urgent need to develop a video object trajectory outliers detection method with multi-camera video target motion analysis on the basis of considering the spatio-temporal correlation of multi-camera moving video objects.In view of the above problems,this paper proposes three methods for detection of trajectory outliers of multi-camera video targets defined from different perspectives,including the method based on trajectory vectorization,the method based on clustering and the method based on video scene segmentation.The above methods are based on video object-geographic scene fusion,anomaly detection is performed on cross-camera video object trajectory data.The algorithm based on vectorization converts the trajectory data into trajectory vectorizations,and provides a specific measurement method to judge trajectory outliers;the clustering-based method provides the similarity measurement between the trajectories,and calculates the distance between the trajectory and the cluster center to judge trajectory outliers;the method based on video scene segmentation makes full use of the background of the video target to judge trajectory outliers.The specific work of this paper is mainly discussed from 4 aspects:1.Use computer video object cross-camera re-recognition technology to realize cross-camera ID association of video objects in geographical scenes;use geographic information system(GIS)related technologies to realize video objects-integrated information organization of geographical scenes,and provide a basis for follow-up work.2.Propose a trajectory outliers detection method based on trajectory vectorization.Through trajectory vectorization and vector neighborhood generation,the method further builds an isolated forest to detect position trajectory outliers and detect velocity trajectory outliers through vector neighborhood comparison.3.Propose a trajectory outliers detection algorithm based on spatio-temporal grouping trajectory hierarchical clustering.The method performs Semi-supervised clustering on the video object trajectories within the camera group and cross-camera groups,and then the trajectory outliers are judged by comparing the distance between the trajectories and their cluster centers.4.Propose a trajectory outliers detection method based on video scene segmentation.This method uses FCN((Fully Convolutional Networks)to realize the pixel-level classification of geographical scenes,and forms specific spatio-temporal trajectory data by combining dynamic and static scenes around the trajectory,and further builds a decision-making model to detect trajectory outliers.