Visual tracking based on multiple instance learning particle filter

The tracking by detection algorithms treat visual tracking as the on-line object and its local surround background classification problem. The main shortcoming of the algorithms is the template drift due to the online self-learning mechanism of the visual tracker. To overcome the problem, a novel online Multiple Instance Learning (MIL) particle filter visual tracking algorithm is proposed. Main contributions of our work are: Firstly, we introduce online MIL Boosting algorithm in particle filter visual tracking framework to deal with the problem of target appearance model online learning by noisy labeled samples and to evaluate the importance weight for each particle; Secondly, the particle set, which represents the probability distribution density of the tracked target state, is utilized to construct the online training positive bag for the MIL Boosting classifier; At last, some experimental results show the proposed algorithm is a robust and accuracy tracking algorithm.