| Computer vision is developing very rapidly in recent years. It is now widely used in military and civil applications. Target tracking, as an important branch of computer vision, fuses many front-line science including computer science, image processing, pattern recognition, target detection, artificial intelligence and automatic control. And it is widely used in the areas of military investigation, robot navigation, safety inspection, traffic management, medical diagnostics and meteorological detection.The Mean Shift algorithm which will be studied in this paper is one of the most popular algorithm in the field of target tracking. Because of many advantages such as good real-time ability, robust for occlusion and target distortion, it attracts a high degree of attention and have a wide range of applications. However, the Mean Shift algorithm has a few defects, such as the failure of tracking when the target in the consecutive frames only gets small part overlapped,the lack of template updating,the fixed size of the kernel window,the lack of diversity when starting at one particle and poor anti-interference ability in complex environment.Some improved algorithms against the defects of Mean Shift algorithm are introduced in this paper. Based on the theoretical analysis of Mean shift: (1) an improved algorithm combined with the Kalman filter is brought forward;(2) And variable scale and deformation prediction algorithms are used to achieve the self-adaptive size of kernel window; (3) In order to overcome the lack of template updating, the multi-template algorithm is fused to improve the mean shift algorithm; (4) What's more, the particle filter algorithm is used to improve particles' diversity of the Mean Shift algorithm. Finally; (5) based on the Theory of Discriminative Features, the separation of target and background is increased, so that the Mean Shift algorithm may keep excellent real-time ability, robustness and accuracy in a variety of complexity environment. |
PDF Full Text Request