Siamese Tracking with Adaptive Template-Updating Strategy

Abstract

Recently, we combined a contour-detection network and a fully convolutional Siamese tracking network to initialize a new start-up of vehicle tracking by clicking on the target, generating a contour proposal template instead of using a fixed bounding box. Tests on the OTB100 and Defense Advanced Research Projects Agency (DARPA) datasets proved that our method outperformed the state of the art and effectively solved the partial-occlusion problem. However, the current Siamese tracking method uses the target in the first frame as a template during the whole tracking period, and leads to the failed tracking of target deformation. In this paper, we propose a new template-update method and reconstruct the whole tracking process with a template-updating module. To be specific, the innovative adaptive template-updating module is comprised of a neural contour-detection network and a target-detection network. Experiment results on the DARPA dataset prove that our new tracking algorithm with the template-updating strategy prominently improved tracking accuracy regarding the deformation condition.

Keywords

• object tracking
• Siamese network
• template-updating
• deep learning