Practical Tracking Method based on Best Buddies Similarity

Abstract

Visual tracking is a crucial skill for bionic robots to perceive the environment and control their movement. However, visual tracking is challenging when the target undergoes nonrigid deformation because of the perspective change from the camera mounted on the robot. In this paper, a real-time and scale-adaptive visual tracking method based on best buddies similarity (BBS) is presented, which is a state-of-the-art template matching method that can handle nonrigid deformation. The proposed method improves the original BBS in 4 aspects: (a) The caching scheme is optimized to reduce the computational overhead, (b) the effect of cluttered backgrounds on BBS is theoretically analyzed and a patch-based texture is introduced to enhance the robustness and accuracy, (c) the batch gradient descent algorithm is used to further speed up the method, and (d) a resample strategy is applied to enable the BBS to track the target in scale space. The proposed method on challenging real-world datasets is evaluated and its promising performance is demonstrated.