An Adaptive Kalman-Correlation Based Siamese Network Tracker for Visual Object Tracking

Object tracking is important in a variety of applications from surveillance to robotic vision and traffic monitoring. Because of its importance, there has recently been a lot of research and developments in this field. Meanwhile, since the deep convolutional neural networks has shown its impressive potential, Siamese networks have also drawn increasing attention. However, the trackers may fail when there are rapid motions, occlusions, and similar objects in the video. To address the limitation and improve the robustness, this paper takes advantages of both the Kalman filter and the correlation filter, and further develop an adaptive Kalman-Correlation based Siamese network (AKC-SiamTracker). AKC-SiamTracker can automatically make different adjustment strategies to adjust the detected position of the original tracker based on the adaptive influence coefficient decider. The fully connected Siamese network (SiamFC) and Siamese region proposal network (SiamRPN) are selected as the baseline models. Evaluation of our method is carried out on OTB dataset. The promising results have shown better performance and robustness compared to the baselines and other state-of-the-art models. To the best of our knowledge, our work is the first time to propose the adaptive Kalman-Correlation based Siamese tracker.