Abstract:

The tracking-learning-detection (TLD) algorithm is a novel long-term target tracking algorithm, however, the detectors do not fully consider contour changes in the target tracking process, and window-based scanning affects the efficiency. We introduce the evolution mechanism to improve the TLD algorithm based on level set. The multi-scale level set method combines the edge information with the region information, which improves the adaptation and precision of target tracking while effectively overcomes gray uneven images. According to the test results of the outline, we introduce the motion detection operator of the target direction to estimate the movement direction of the target and its position in the current frame, which can reduce the scan window while improving the ability of target identification. Experimental results show that the improved method can enhance the tracking speed, and has stronger adaptability and higher tracking accuracy.

Key words: target tracking, tracking-learning-detection (TLD), multi-scale level set, adaptability, motion detection