Abstract:

Particle filters can suffer from particle scarce after a number of iterations. Illumination, face mask and rotation problems can also lead to decrease in tracking accuracy of traditional colors, even tracking failure and other issues. To solve these problems abovementioned, we propose an adaptive likelihood distribution (ALD) method, which can adaptively adjust the likelihood distribution state and increase the overlap region of the prior distributions and likelihood distribution according to the size of the noise factor, thus effectively improving the stability of the filter and reducing the number of resampling. As for the low tracking accuracy or failure moment, we use the local ternary patterns (LTP) to identify the area to be tracked, and utilize the effective particle dynamic threshold to reduce particle resampling frequency. The tracking continues by using new determined templates. Experimental results show that the proposed algorithm has a smaller number of samples and can effectively track the target with occlusion, rotation and other conditions in comparison with the traditional algorithm.

Key words: particle filter, adaptive likelihood distribution, local binary patterns, local ternary pattern, occlusion, resampling