Different Model-based Moving Target Tracking Algorithm

The tracking of moving object based on video is one of the important issues in the field of computer vision, and has a wide application in industry military and civil field, including product exploitation, military guidance, and video detection control system in traffic or bank. Hence, the study on object tracking algorithms is of higher practical significance.It is worth noting that although a lot of research results have been recently obtained on the problem of object tracking, there exist many problems remaining open, eg., the case that the object is occluded in the movement, the problem of interference in the process of object tracking. Aiming at these problems, this thesis has made some research from several aspects including multi-feature extraction, state model construction, and template update. This main works are as follows:(1) In general, only adopting a single color feature cannot adapt to the change of environment and can not meet the requirement of robustness and real-time tracking. Hence, this paper proposes a multi-feature based tracking algorithm to solve the problem of object detection in a complex background with occlusion and deformation. In this algorithm, the features of both color and gradient are utilized, in which the gradient feature containing a certain amount of structural information may achieve effective moving object tracking. Experiment results show that the present algorithm has better robustness and accuracy compared with a single feature.(2 In the dynamic environment of sports competition, aiming at the occlusion and deformation during the process of moving target tracking, this paper firstly establishes the dynamic Bayesian network state model, and then combines with the improved particle filter algorithm to realize the object tracking. By compared with some conventional algorithms such as KF and EKF, it is shown from experiment results that the proposed algorithm in this paper can well handle nonlinear data structure of video image, and has better real-time tracking, robustness and accuracy in the moving object tracking.(3) To solve the problem of model update in object tracking, this paper propose a LLE based template update algorithm. Firstly, LLE algorithm is utilized to process video images. And then, the time series model and BP neural network are combined to set up a combination forecasting model. Finally, the mapping capability of RBF neural network is utilized to reflect the higher dimensional space. Because only using single-step prediction can not guarantee the accuracy in the complex environment with occlusion, this paper fuse the time series model and BP neural network to achieve multi-step prediction and make up the defect of time series model. The experiment results show that using this method can ensure accuracy and robustness in the moving object tracking.