A particle filter tracking algorithm based on multi-feature clustering

A particle filter tracking algorithm based on multi-feature clustering is proposed. To address the issues such as the diversity of target features, the difference between methods of feature distribution description, and the arbitrariness of feature spatial structure, the multi-features representation of target model is unified into a clustering computing framework. The mean shift based feature space analysis approach is employed to adaptively calculate the clusters in any arbitrarily structured feature space. Based on the clusters, a target probability density estimation method, which is efficient and accurate, is proposed to represent the target model. The distance between the reference target and the candidate is calculated by the similarity measure of kernel density estimation, and is taken as important information for observation in particle filter system. To efficiently enhance the utilization rate of particles, an improved particle propagation model is presented. The object tracking experiments are performed on many real image sequences by using the LUV color features and the LBP (local binary pattern) texture features. Experiment results show that the proposed algorithm can obtain high tracking accuracy and strong robustness, meet real-time demand, and provide better tracking performance comparing with other typical algorithms.