A robust tracking architecture using tracking failure detection in Siamese trackers

In recent years, due to the impressive performance on the speed and accuracy, the Siamese network has gained a lot of popularity in the visual tracking community. However, both the spatiotemporal correlation of adjacent frames and confidence assessment of the results of the classification branch are missing in the offline-trained Siamese tracker. In this paper, a robust tracking architecture is proposed to implement the tracking failure detection and make better tracking decisions for the Siamese tracker. It consists of two stages including tracking failure detection and proposal re-selection. Firstly, a Siamese tracker is adopted as the baseline, and a tracking failure detection mechanism is proposed based on motion estimation of object via optical flow. It can timely supervise the reliability of the tracking system. Secondly, when the tracking failure occurs, the proposal selection strategy is optimized with spatiotemporal information to re-select more reasonable results. The overall mechanism can guide the tracker to handle target drift problem by tracking failure detection and proposal re-selection. Several representative Siamese trackers are utilized to validate the effectiveness of our approach. Furthermore, the performance of our approach is demonstrated based on extensive experiments on popular benchmarks, which can improve the robustness of the model.