Study On Mean-Shift Algorithm In Visual Object Tracking

Target tracking technology is one of the important research topics in the field of computer vision. It has a wide range of applications in human life, military reconnaissance, industrial production, medical diagnostics, traffic management, and many other aspects. This study has practical significance in human life and engineering application.In vision-based target tracking algorithm, the classic Mean-Shift algorithm with its advantages, such as scientific and effective theory, simple operation, easy to implement, better tracking performance and so on, has been the focus of many scholars. However, there are also a lot of defects. For example, the target model is contaminated with the interference of background information, which results in targeting deviation. And classic methods always regard the center of the target which is manually determined as the initial value of the iteration, but it’s not the true center of the target. So the tracking for fast-moving target can not be achieved effectively. The algorithm has the ability to deal with edge occlusion, but when the center of the target is seriously occluded, it is difficult for the algorithm to track accurately. It regards color as the global statistics, so the effectiveness of the tracking method is reduced when the target and background has small color discrepancy or similar probability density estimation. To solve the above problems, this paper does an in-depth research on the Mean-Shift algorithm and proposes improvements.For the situation that colors is effective feature in the target tracking, we propose an improved Mean-Shift algorithm based on the characteristic color. First in order to suppress the interference of the background color blended into target model and locate the target more accurately, we use the characteristic color with which the target differ from the background. Then using uniform kernel function to describe the target model, which enhances target-occlusion ability. After that calculating the centroid of the improved object model and conducting mean shift using this centroid as initial iteration value, we can quickly locate the target, which solves the problem that algorithm can not effectively track on small targets. Compared with the traditional Mean-Shift algorithm, the accuracy, real-time performance and robustness of the improved algorithm have been improved.For the situation that the target’s color is close to the background’s color or they have similar probability density estimation, the algorithm enhances target characterization ability by using local invariant feature. The specific work includes:First research the improved Mean-shift algorithm based on extracting texture and color features by the wavelet transform. Then research the Mean-shift algorithm by combining corner feature and color feature. After that propose an improved Mean-shift algorithm by using edge pixels of the target. In this paper, we analyze and compare the performance of these three methods by experiments. The texture-based algorithm has the highest accuracy, the edge-based algorithm has certain positioning deviation and the best real-time performance, and the corner-based algorithm has little positioning deviation.