Visual Tracking Algorithm Based on Improved Kernelized Correlation Filters

During the tracking, kernelized correlation filters cannot handle this situation well when the target occluded seriously, changes scale and goes out of view. To solve the above problem, we propose an improved visual tracking approach based on kernel correlation filters. First and foremost, to locate the target and estimate the target scale, two correlation filters are employed to tackle the corresponding problems. Meanwhile, to memory the target appearance, we use an independent target appearance correlation filter conservatively updated. Second, we combine the SVM detector to re-detection when tracking failure occurs caused by the target undergoes heavy occlusion or missing in video stream. Third, to solve model drifts due to heavily surface variations of the target, we apply a novel adaptive model updating strategy to update the correlation filters and classifier. Extensive experimental results demonstrates that the proposed tracker achieves an excellent performance on the OTB2013 benchmark dataset, comparable to the state-of-the-art methods.