Research And Improvement Of Video Target Tracking Algorithm Based On Mean Shift

Target tracking is a core subject in the research field of computer vision which has received the extensive attention of scholars in recent years. Target tracking technology provides the underlying object and analysis foundation for higher video understanding and scene interpretation, so it has important research value. However, factors such as the diversity of the tracked target characteristics and the complexity of the environment restrict target tracking algorithms’ performance. Therefore, it is still a challenging work to design a kind of stable and reliable moving target tracking method currently.In this thesis, we have researched the target tracking development and current research situation, then several target tracking algorithms based on Mean Shift algorithm are studied and analyzed, and a new improved algorithm is proposed. The new improved algorithm can adaptively adjust the tracking box’s size according to the target’s geometry information, thus improving the tracking performance. The main work of this thesis is summarized as follows:(1)The common technologies of moving target tracking were studied. The research mainly includes target representation, target feature selection, target tracking method classification and mathematical morphological processing.(2)This paper introduces the principle and the implementation-framework of Mean Shift target tracking algorithm and corrected background-weighted histogram Mean Shift algorithm. Then we take comparison experiment on several video sequences. Experimental results show that CBWH tracking algorithm outperforms traditional Mean Shift tracking algorithm and better adapts to the variations in the environment.(3)The prominent characteristic data of moving target can be taken as the basis of judgment target scale. Therefore, we propose an improved tracking algorithm: adopting CBWH tracking algorithm to track the location of the target, then generating a color probability distribution with target background-weighted model in RGB color space and calculating the invariant moment to resize the tracking window in the next frame. Experimental results show that the proposed method performs a certain progress comparing to CBWH. It improves the accuracy of the target scale positioning and spatial positioning, and has better stability.