Detection And Tracking Of Moving Targets Based On Stereo Scene Flow

Detection and tracking of the moving targets in the video is a challenging and hot topic in the field of computer vision.It is the basis of the advanced research such as recognition, classion cation and behavior analysis of the moving targets. Detection and tracking of the moving targets is also widely used in scenarios of real life,such as intelligent security, the perception of environment in Modem driver assistance system,intelligent transportation,but it still has many problems. It is of great theoretical and actual significance to do research about the detection and tracking of the moving objects.Based on this point, this thesis detect and track moving targets based on stereo vision system,and our main work is divided into detection and tracking of the moving targets. Firstly, we obtain the objects motion area which is set as area of interest by using the frame differential method. Secondly, the Delaunay triangulation is built by considering the pixels which have high significant degree as supporting points. Then we calculate the depth information though two stereo match, and acquire the motion information of the moving targets by optical flow algorithm based on the Delaunay triangulation at the same time. The scene flow information can be obtained by fiising the depth informatioa and motion information. In order to extract the targets,the scene flow of the moving targets is clustered by using Mahalanobis distance. We track the targets based on the kalman filter algorithm and GNN algorithm. The position of objects in the next frame is predicted by the kalman filter algorithm.In order to obtain the trajectory of the targets, we reduce the searching area and match the observation and the prediction of the targets by using the GNN algorithm. The trajectoiy is used to improve the detection so that it can make a total detection and tracking of the moving targets in time domain.Lastly,we use different dataset to test the effectiveness of the algorithm based on MATLAB. The experiments indicate that it ’works very well in much complex scene such as junction, target crowed, target occlusion.