Research On Technologies Of Moving Object Tracking In Video

Moving object tracking in video is a fundamental problem of computer vision, has broad application prospects, and can be applied to video surveillance, video analysis, robotics, intelligent human computer interaction, etc. The main purpose of moving object tracking is to acquire the object's trajectory and motion parameters in video, e.g., location, scale, etc. The algorithm of moving object tracking mainly comprises two key components: appearance modeling and motion modeling. This dissertation studies the two problems seperately. Concerning appearance modeling, the main methodology adopted is joint modeling of foreground (i.e., the tracked object) and background. Methods of object tracking based on online feature selection and incremental two dimensional linear discriminant analysis are proposed respectively. Concerning motion modeling, object tracking based on graph modeling is proposed by taking ball tracking in broadcast soccer video as an example. Finally, based on the proposed moving object tracking methods, a system which can convert broadcast soccer video into 3D cartoon animation is developed. Specifically, the main contents of this dissertation are as follows:Firstly, object tracking based on online feature selection is proposed. It treats object tracking as foreground and background pixel classification, adopts Bayes error rate to evaluate each feature's discrimination ability, and selects several features having smallest Bayes error rates to construct ensemble of Bayesian classifiers. Then, ensemble of Bayesian classifiers is used to assign pixels in the current frame probabilities belonging to foreground pixels. Finally, particle filter is employed to track the object in probability map. A new observation model of particle filter simultaneously considering region and boundary information is proposed. Furthermore, with the help of integral image data structure, observation model of particle filter can be computed rapidly.Secondly, object tracking based on incremental two dimensional linear discriminant analysis is proposed. It treats object tracking as foregound and background region classification problem, directly perform linear discriminant analysis on two dimensional image matrix, and extract subspace which seperates foreground and background. Due to directly operating on image matrix, this method has high computational efficiency. Whereas traditional linear discriminant analysis needs to change two dimensional image matrix into one dimensional vector, and then performs linear discriminant analysis, which results in costly matrix computation. Moreover, we develop a method to incrementally update the subspace, which further lower the memory cost and computational cost. Finally, particle filter is used to infer the object's motion parameters.Thirdly, object tracking based on graph modeling is proposed. Object tracking is usually initialized by object detection methods. The fundamental hypothesis is that the object's pattern can be seperated from its surrounding background sufficiently. However, for some objects, e.g., the ball in broadcast soccer video, it is hard to extract effective features to detect the ball in a single video frame. The strategy adopted here is to identify the object's candidate regions in several consecutive frames, and then use graph to construct the relationship between candidate regions. Specifically, each candidate region corresponds to a node of the graph, and each edge links two candidate regions in adjacent frames. Each node is assigned a weight to represent the probability to be the object's region; similarly, each edge is assigned a weight to represent the probability that the two nodes on the edge correspond to the same region. Finally, Viterbi algorithm is used to extract the optimal path of the graph as the object's trajectory. This process is called short-term tracking. Then, it is used to initialize a Kalman filter to perform long-term tracking. In the process of tracking, the tracked region is verified to determine whether tracking is failure, and short-term tracking is restarted if failure happens.Fourthly, a system which can convert broadcast soccer video into 3D cartoon animation is developed. When the sports event is broadcasted, multiple cameras are usually deployed around the field. However, at the same time only one camera's video is available to viewers. Viewers hope to watch the game from other viewpoints. Moreover, after major sports games, some web portals provide cartoon animation of goal events. However, it is time consuming and labor tedious, and only a single viewpoint is provided. Based on the proposed object tracking methods, this dissertation employs computer vision and computer graphics techniques to develop a system which can generate 3D cartoon animation of soccer games. It allows users to watch the game from different viewpoints.