Visual Target Tracking Algorithm Based On Online Learning

Visual tracking as one of the hot research topic in the artificial intelligence field has been widely applied in lots of field including civilian and military application domain. The object of visual tracking is to give the vision sense function to machine. In some complexity environment which people cannot manage. The machine with vision sense may exchange information to environment replacing human beings. So far, the existing tracking approaches can be classified to two groups:the tracking by feature fusion algorithm and the tracking by detection algorithm. This paper makes a deep study of the both algorithm above and proposes two improved algorithm in particle filter (PF) frame work.Firstly, the proposed tracking algorithm rests on a basis of proposed distribution to fusion multiple features in the frame of two levels. For each time-step in the tracking process, the first level of PF:transmitting particle set according to motion model and proposed distribution is q(xki|xk-1i,zi:k)=p(xki|xk-1i).then choosing color feature or edge feature as observed value to construct observation likelihood function and acquiring the posterior probability distribution of the tracked target after re-sampling. The second level of PF:adding the observed information form first PF into present proposed distribution to importance sample, the proposed distribution is q(xki|xk-1i,zi:k)=p(xki|xk-1i,zi:k), combined with the feature which is not be used in first level PF, we can get the posterior probability distribution of the second level PF and determinate the accurate state of the target.The template update method used in this paper considers the two parts of template updates:The original target template and another continuously updated sports template. This method can compensate for the lack of a single template as a target template effectively: sports template which has been updated continuously can meet variation of the goal and environment. The original template is to ensure that the object tracked is the target.Secondly, we proposed a kind of online multi-instance learning (MIL) particle filter tracking algorithm. This algorithm adds online boosting algorithm and multi-instance learning into the tracking algorithm based on particle filter. The target observation model is constructed by the classifier which has been online updated, thus avoiding the problem of complex observation model constructed in the feature fusion based tracking algorithm. Because of multi-instance learning is to mark the package (the package contains instances). So it can effectively avoid the samples of the noisy labeled and to improve the robustness of the algorithm.Finally, experimental validation of the two tracking algorithm is evaluated by the data from the CAVIAR database and video sequences of continuous vehicle. Specific circumstances are divided into not block and partial occlusion pedestrian tracking and high-speed vehicles of a single color and rich colors tracking in complex context. The experimental results show that the proposed algorithm has better tracking performance.