Object Tracking Algorithm with Two-way Parallel Fully-convolutional Siamese Networks

Fully-convolutional Siamese Networks (SiamFC) tracker usually achieves good performance in real-time visual tracking. Its accuracy may suffer from the complex background in some tracking scenarios. However, in other situations including target occlusion, deformation, and so on, the background information can help infer the location of the target. To improve the overall performance of SiamFC under multiple scenarios, a new algorithm with two-way parallel architecture is proposed. One branch uses the traditional SiamFC, and the other employs a modified SiamFC which can suppress the background information causing interference with the target in the template. The final matching results are selected by using the average location error (ALE) as an evaluation metric to describe the confidence level of the matching results, so the robustness of the algorithm can be enhanced under various object tracking scenarios. Testing experiments are performed on the OTB dataset as well as the VOT2017 dataset. The experimental results show that the proposed algorithm can achieve better performance when compared to the original SiamFC tracker.