Double-Channel Object Tracking With Position Deviation Suppression

The object tracking methods based on multi-domain convolutional neural network (MDNet) commonly fail to track in the case of background clutter. A novel double-channel object tracking (DCOT) is proposed to solve this problem. The discriminative correlation filter (DCF), which has strong discriminative power of low-level features, is employed for the position deviation suppress of the samples generated from MDNet. Firstly the pre-trained deep network is used to learn and classify the target and background in the video frames. If the tracked position of the DCF is judged to be correct, we delete the target candidate samples with high position deviation from MDNet. The position deviation is measured by the distance between the tracked positions of the DCF and MDNet. Finally, MDNet and DCF are updated with a robust update strategy. The experiments are performed on OTB-100 and VOT-2016. The overlap precision and distance precision of DCOT on OTB-100 are 92.2% and 69.5%, respectively, which are higher than those of MDNet by 1.3% and 1.7%. The results of DCOT in background clutter are higher than those of SANet by 0.2% and 2.8%, respectively. DCOT is also superior to other state-of-the-art trackers on VOT-2016.