When more video cameras are employed in a wide range of applications, how to understand this huge amount of video data in an automatic way becomes a urgent task. Visual tracking is such a strong tool used to abstract the high-level information, such as human activity recognition, traffic information accumulation, security event detection, etc. However, efficiency is still one major issue limiting tracking algorithms in many real-time applications. Consequently a great deal of research effort has been focused on the design of an efficient searching strategy and a discriminative measuring method for a good tracker. Under the probabilistic framework, we notice that intermediate measurements and the correlations among the multiple targets are valuable information for the generation of samples, which have not been fully utilized before. Therefore, the objective of this research is to exploit how to improve the searching efficiency by integrating these two factors into the sampling in several applications.
We first start with a single target tracking, and we propose to update the proposal distribution by dynamically incorporating the most recent measurements and generating particles sequentially, where the contextual confidence of the user on the measurement model is also involved. In addition, the matching template is divided into non-overlapping fragments, and by learning the background information, only a subset of the most discriminative target regions are dynamically selected to measure each particle, where the model update is easily embedded to handle fast appearance changes. The two parts are dynamically fused together such that the system is able to capture abrupt motions and produce a better localization of the moving target in an efficient way.
Then we extend our attention to the case of multiple targets, and we consider a special case where targets are highly correlated, demonstrating a common motion pattern with individual variations. We propose to update the sampling distribution by incorporating both the most recent observations and the correlation information. The correlation is either known a priori or learned from the previous tracking results. In this way, the observation of a single target is multiplexed statistically through mutual correlation for other targets, and the correlation serves as both a prior information to improve the efficiency and a constraint to prevent trackers from confusing or drifting. Experiments on both synthetic and real-world data verify the effectiveness of the new algorithm and demonstrate its superiority over existing methods.