Robust and adaptive keypoint-based object tracking

摘要

AbstractObject tracking is a fundamental ability for a robot; manipulation as well as activity recognition relies on the robot being able to follow objects in the scene. This paper presents a tracker that adapts to changes in object appearance and is able to re-discover an object that was lost. At its core is a keypoint-based method that exploits the rigidity assumption: pairs of keypoints maintain the same relations over similarity transforms. Using a structured approach to learning, it is able to incorporate new appearances in its model for increased robustness. We show through quantitative and qualitative experiments the benefits of the proposed approach compared to the state of the art, even for objects that do not strictly follow the rigidity assumption.Keywords: Object trackingpose estimationlearningkeypointsreal-time tracker NotesNo potential conflict of interest was reported by the authors.The GPUs used for this research were donated by the NVIDIA Corporation. This paper is selected as the "Cutting Edge of Robotics in Japan" by the Editorial Committee of Advanced Robotics.1 ‘Bounding’ is somewhat misleading since it is not necessary for the box to be a bound on the object in the image. However, we adopt it for this paper since it is commonly used in the field.