Particle filter-based target tracking algorithm with adaptive multi-feature fusion

Particle filter is a nonlinear filtering algorithm based on Bayesian estimation, which can well deal with non-linear, non Gaussian system parameter estimation and state filtering problem. In addition, as it allows integration of a variety of features, it is widely used in target tracking. This paper selects the best two features from color and shape-texture features according to their abilities of distinguishing the target from its background to describe the target. Taking the histograms of the two features as two target models, we can get two estimated results by using particle filter algorithm. If the two estimated results are similar, it means the selected features are effective and the result is reliable. On the contrary, it means that one or two of the selected features fail and target tracking fails. When tracking fails, the reliability of the estimated output of the previous frame is chosen to decide whether to return the previous frame to re select features and estimate in current frame again. Only when the two estimated results are similar to update the target model, thus can ensure the target model does not have a big offset. Experimental results show that the proposed algorithm is better than other particle filter algorithms using a single feature or two fixed features.