Stably Adaptive Anti-Occlusion Siamese Region Proposal Network for Real-Time Object Tracking

Abstract

Siamese region proposal network has made remarkable achievements in visual object tracking because of its balanced accuracy and speed. However, it regards tracking as a local one-shot detection task, which lose the power of updating the appearance model online thereby cannot handle the object-occlusion, fast motion and out-of-view situations. To tackle this problem, we propose a method that combines adaptive Kalman filter with Siamese region proposal network (Anti-occlusion-SiamRPN) to make full use of the object spatial-temporal information. Specifically we first extract target features through deep network and then uses adaptive Kalman filter to predict target trajectory in these difficult scenarios. Further this trajectory is used to select the candidate area of the next frame for Siamese region proposal network, which improve the searching mechanism. In this way, the introduction of adaptive Kalman filter makes the tracking process online learning which makes up for the disadvantage that Siamese region proposal network can only track offline. In addition, a hard example discrimination method (HEDM) is proposed to estimate whether the occlusion occurs and how seriously it is, which also improve Kalman filtering mechanism to make it update adaptively. Our method being evaluated with the speed of 80 FPS on five widely-applied challenging benchmarks including OTB2013, OTB2015, OTB50, VOT2016 and VOT2018. The extensive experimental results demonstrate our method achieves state-of-the-art effects and great improvement in comparison to other trackers.