Research On The Visual Tracking For The Guidance Of Fast Vehicles

This thesis is focused on the research on the visual tracking algorithm for the guidance of fast vehicles. Based on the Compressive Sensing Theory, a new compressive measurement matrix is designed for the reduction of the image feature dimensions, and a new scale adaptive real-time compressive tracking algorithm is proposed.Firstly, the visual tracking methods based on compressive sensing theory(ie. compressive tracking) is studied and summarized, and the drawback of the constructing method of traditional compressive measurement matrixs is analyzed. A new constructing method is proposed in the the thesis. The first and the most important issue in the compressive tracking problem is the construction of the measurement matrix. The traditional measurement matrix constructing methods is summarized and the latest rectangular-blocked measurement matrix constructing method is analyzed and discussed. The drawback of this measurement matrix, which the distribution functions of each element value in the matrix is not identically distributed, is analyzed theoretically and verified experimentally. This shortcoming is equivalent to the weakening of the influence of the upper left portion of the image, and the strengthening of the influence of the bottom right portion of the image. To solve this problem a looped rectangular-blocked measurement matrix constructing method is proposed. The range of the rectangular number within each convolution kernel of the measurement matrix is analyzed, and the common range given. Then the shortcoming of the looped rectangular-blocked measurement matrix in the satisfaction of the RIP conditions is discussed,which is the reduction of its probability to satisfy the RIP conditions, because of the rectangular-block division of the measurement matrix to reduce the computational complexity. Experiments show that the problem of which the distribution functions of each element value in the matrix is not identically distributed has been resolved, and our algorithm has better robustness of tracking.Then, the issue of the scale adaption of the window in compressive tracking is studied. A scale adaptive window modle is proposed, and the scale adaptive window searching algorithm is achieved based on particle filter algorithm. Since the scale of the target in the image is changing greatly with the approaching to the target of the vehicles, the scale adaption of tracking algorithm is always a difficult problem. To solve the shortcoming of the traditional compressive tracking method which is not adaptive in scale, the issue of the scale adaption of the window is studied and the scale adaptive window model is established. The scale adaptive window searching algorithm is achieved based on particle filter algorithm, and the algorithm has better robustness and a good real-time performance. Considering the maneuvering characteristic of fast vehicles, the issue of the target missing in the camera view is iscussed, such as the target being mostly covered or leaving the camera view. The influence of target missing on the window parameters and the target descriptor is analyzed, and a special treatment of target disappearance is considered in our algorithm, to improve the robustness. Experiments show that our algorithm has good adaption to scale changes of the target in the image, good robustness to target missing, and a good real-time performance.