Transductive local exploration particle filter for object tracking

Robust tracking of non-rigid objects in a dynamic environment is a challenging task. This paper presents a particle filter solution for non-stationary color tracking using a transductive local exploration algorithm. The target model is represented by a non-parametric density estimation and the similarity measure is based on a metric derived from mutual information. We employ a transductive inference to update the target model dynamically. Combining confidently labeled data and weighted unlabeled data, the proposed transductive inference offers an effective way to transduce object color model through the given observations in non-stationary color distributions. Better proposal distributions containing new observations are obtained through a method of local exploration. Targets can be tracked well despite severe occlusions or clutter. The way the transductive adaptable object model and local exploration particle filter are combined plays a decisive role in the robustness and efficiency of the tracker. In the presented tracking examples, the new approach successfully coped with target appearance variations, severe occlusions and clutters.