Dynamic Scene, Moving Target Tracking Algorithm

Moving object tracking in video sequences is one of the most challenging research topics in the area of computer vision. It has great research and application values in the areas of modern industry, military, navigation and spaceflight. Moving objects tracking based on video can be divided into two types according to the static or dynamic background. We mainly study the detecting and tracking algorithms of moving objects with the dynamic background in this dissertation. The main works in this dissertation are as follow:1,The traditional moving object detection methods based on feature matching have the disadvantages of large computational complexity, threshold should be decided by the experience and can't catch up with the real-time computation. In order to overcome these limitations of this method, object detection method based on SIFT is proposed in this dissertation. The keypoints are deteced by the SIFT algorithem. After matching several times, the stable key-points that belong to the moving object are left and most key-points that belong to the background are filtered.2,We find that the key-points of the same moving object have nearly the same speed, but key-points of different moving objects have different speeds. So, the detected key-points are clustered by the cluster method based on the speed vector connectivity. Thus, we get the keypoints of different moving objects.3,The traditional object tracking methold based on feature matching shows low robustness during tracking. In order to overcome this limitation, moving object tracking and keypoint eigenvectors matching method based on GMM(Gaussian Mixture Models) is proposed. GMM is modeling by eigenvectors of the keypoints and the Gaussian Mode center is calculated. If the minimal Mahalanobis Distance between the candidate keypoints and the Gaussian Mode center is smaller than the given threshold, then the corresponding keypoint is the matching keypoint that we want to find.4,In this dissertation, Increase Incremental Principal Component Analysis (IPCA) algorithem is used to update the GMM Model in order to reduce the computational complexity of the algrithem.The corresponding Gaussian Model and its center data are updated by the eigenvectors of matched keypoint.5,Based on above methods, the moving object detecing and tracking system based on dynamic background is designed and implemented.After code optimization, the algorithm proposed in this dissertation can greatly reduce the time used in object detection and tracking. It can most achieve real-time efficiency. The experience results show good results of the object detecing and tracking.