An Implementation of a Robust Visual Object Tracking System

Robust visual tracking system with high level of accuracy against complex tracking scenarios containing many visual attributes has been a challenging research topic in computer vision field. Among different types of visual attributes, scale variation is considered as one of the most difficult problems. Existing tracking methods either fail to handle great change of target size, or employ exhaustive scale estimation with the cost of high computational load. This paper presents a method of accurately adapting the change in target scale which employs dense spatio-temporal context for localizing target position, with the intention to increase tracking performance while maintaining low computational cost. In particular, the tracking task for determining target position is computed by utilizing the spatio-temporal context relationship between the target and its surrounding regions. Then the spatial correlation is analyzed to update target position in subsequent frames. In the meantime, the model also estimates the change of target by applying a scale filter, which learns the change in appearance of a set of various value of scales. Finally, the proposed method is evaluated by using the image sequences in TB100 dataset with different performance evaluation tests.