Moving Vehicle Detection And Continuous Tracking Based On Large-scale Scene With Multiple Videos

In the construction of smart city, firstly we must realize the perception of city, among the rest, video sensor has been applied widely because of its advantage such as all-weather, intuition, high covering, real time and so on. Currently, the video monitoring in urban give priority to road traffic monitoring and security monitoring, owing to narrow-field, single function and unable to realize continuous and real-time tracking, it doesn’t play a proper role in the perception of city. According to the video sensor all over the city, this paper studies the key problem among continuous tracking of the target based on large-scale scene with multiple cameras. The main work and achievement is as follows:(1)The idea about continuous tracking of vehicle based on large-scale scene with multiple cameras is suggested. Firstly this paper realize the vehicle detection with single camera and determine the target tracked; secondly extract the feature of the target, realize the tracking with single camera; finally associate targets from multiple cameras through the algorithm of target handoff and realize continuous tracking of vehicle with multiple cameras thereby.(2)Moving vehicle detection based on optimized GMM. Classical GMM mainly has defects in three respects such as without overall consideration about the weight and matching degree of distribution in the selection of matching gaussian distribution, constant ρ in the updating of model parameters and making the distribution with highest priority only as the background. Based on GMM, this paper optimizes the ways about selection of matching gaussian distribution, model parameters’ updating and background display. The tests try out that the improved GMM algorithm can extract clear vehicle’s profile, it has strong adaptability to changes of the scene and even it can work well with the disturbance such as hard light, rustling leaves and so on. So it has practical value.(3)Moving vehicle tracking by self-adaptive Mean Shift based on Kalman and combined feature histogram of color and texture with single camera. In the process of vehicle tracking, this paper firstly build a novel combined feature histogram about the color and LBP histograms as the description of the tracked object; secondly Kalman filter is used to estimate target motion and forecast the location of target in the current moment; finally mean shift based on combined feature histogram is used to detect the matching target in the candidate region and at the same time the kernel-bandwidth varies from the moment information from the candidate object area. The improved adaptive mean shift tracking algorithm can overcome the interference of similar goals and target deformation and Minimize iterative search time, even if the target is partial occlusion it can obtain good tracking results.(4)Moving vehicle tracking based on FOVL with multiple cameras. Since the essentials of tracking of moving vehicle with multiple camera is target handoff, this paper mainly study the target handoff under the background of overlapped view from the target handover algorithm based on vision line. Firstly this paper uses the projective invariants to get the field of view lines between the adjacent cameras, secondly the target is judged that it is in the FOV or not through the distance between the target and FOVL, finally the matching target is confirmed by the Euclidean distance and SIFT feature matching.