Deep Deblurring Correlation Filter for Object Tracking

Motion blur is a quite tricky issue in object tracking community. In recent years, many trackers have been explored to address this issue without sensational performance. This paper proposes a novel correlation tracking framework with Recurrent Neural Network (RNN) deblurnet and proposal detection to handle motion blur and heavy occlusion in object tracking tasks. We take advantage of high efficient Kernelized Correlation Filter (KCF) tracker, a typical method that exploits the circulant structure and the kernel trick to enhance the performance, and furthermore incorporate two regression methods in it. We employ RNN as our baseline of deblurnet, and introduce residual block and ConvLSTM in our deblur network to improve the result of deblurring. In addition, we suggest an edge information based rectification system to overcome the challenge of target occlusion. Finally, we update the model adaptively in term of the feedback from high-confidence tracking results to avoid the model degradation. Extensive experimental results demonstrate our tracker outperforms several state-of-the-art trackers on the OTB-2015 and VOT-2016 datasets.