CS-LoFT: Color and scale adaptive tracking using max-pooling with bhattacharyya distance

Robust and accurate visual tracking is needed for many computer vision applications from video summarization to visual surveillance. Visual tracking remains to be a challenging task because of factors such as changing object appearance, illumination variations and shadows, partial and full occlusions, camera motion, distractors, and scale changes. Recently our group proposed a Likelihood of Features Tracking (LoFT) system that fuses multiple sources of information about target and its environment to perform robust single object tracking. LoFT has been shown to successfully track objects under different scenarios from full motion video to wide-area motion imagery. In this paper, we extend the LoFT framework with color information and multi-scale feature sets selection to boost its performance. Our extension of color information is incorporated through Color Name (CN) mapping scheme. Multiscale features computation, and scale selection through frame level max-pooling is also performed to adapt the tracker to scale changes. Experiments on VOT 2015 tracking benchmark, which includes video sequences with significant color and scale variations, show significant performance improvement over baseline LoFT both in accuracy and robustness 11% and 23% respectively.