Refined Segmentation And Tracking Of Vehicles Under Complicated Environment

Vehicle segmentation and tracking is a research hotspot in the field of image processing and computer vision,which realizes the extraction of regions of interest and their trajectories.Also,the monitoring and analysis of traffic information such as the speed,steering information,illegal behavior and the congestion extent of moving objects is completed by vehicle detection and tracking.Because of its important research value and prospect,technologies of detection and tracking have been applied widely to intelligent traffic systems,security monitoring,urban transport planning and so on.However,due to the co-movement of shadows of the moving vehicle,which results in the adverse effects of foreground adhesion and contour distortion,the follow-up vehicle analysis will be severely affected.Meanwhile,with the popularity of the camera network,the vision of the camera cluster is gradually expanding,and the demand for tracking in the wider area is growing day by day.But the massive data generated by the camera network hinders the rapid development of this application.Aiming at the above problems,this paper proposed an algorithm for precise vehicle segmentation as well as a calculation method of candidate targets applied to the large-scale area vehicle tracking based on recent methods,and demonstrates the validity of the two by experiments.The main contents and contributions are as follows:First,we carefully studied the existing target detection and tracking algorithm frameworks,principles and implementation methods by reading a large number of relevant literatures both domestic and overseas.We analyzed the advantages and disadvantages existing in traditional methods,and select the appropriate method considering the applications.Second,a precise segmentation method based on spatial-temporal multi-feature fusion was proposed to solve the problem of foreground adhesion and contour distortion caused by moving shadow.The probability map of foreground was obtained by color,physical model and texture,and combining with the temporal correlation of videos,we achieved good foreground segmentation while eliminating the shadow.Compared with the experimental data of other shadow elimination methods,the advantages of the foreground segmentation method proposed in this paper are proved.Finally,in order to solve the problem of high complexity calculation of multi-camera network,a method of vehicle target candidates extraction for spatial-temporal model was proposed for a large-scale target tracking.Based on the statistical information,the topology of the network was approximately recovered while using Gaussian mixture model to estimate the transfer time distribution between camera pairs,and then the candidate target was obtained.Last,this paper reflected the good performance of this method in terms of narrowing the search range and the hit rate by experiments and the comparison with various methods.And it laid a solid foundation for the follow-up work such as vehicle matching and tracking.In summary,this paper proposed a refined vehicle segmentation method of foreground regions,and it effectively improved both the recall and precision of shadow,completely obtaining segmentations of targets;at the same time,we put forward a method to extract candidate targets in the application of the wide area tracking,greatly reduce the search range and the computational complexity.