Visual Tracking Using Kernelized Correlation Filter with Conditional Switching to Median Flow Tracker

The correlation filters (CF) have been extensively used in object tracking due to its robustness and attractive computational speed. However, the CF are more sensitive to object deformation because they are trained using the spatial features. Besides, updating the filter template with slightly drifted or occluded samples increase the probability of tracking failure. In contrast, the median flow tracker is complementary to the correlation techniques and is fast, robust to occlusion and deformation, but sensitive to illumination variation. In this paper, we exploit the advantage of correlation and optical flow based trackers to achieve drift free tracking. Hence, we apply the CF-based tracker to track an object and switch to the modified median flow tracker during the drift conditions. The combined model is optimized to cope up with the fast appearance change and recover from drifting. We also propose an adaptive feature selection process to select the most discriminative feature/features among colour name and histogram of oriented gradient features based on object separation from the background in intensity and colour channels. The proposed tracker updates the filter template dynamically, depending on the appearance of an object using an adaptive learning rate to track the object irrespective of occlusion, motion blur, and deformation. The scale of object is estimated using Lucas-Kanade homography method. The experiments are carried out using challenging video sequences from a standard object tracking benchmark dataset and show the best performance among the state-of-the-art techniques.