Part-based tracking for object pose estimation

Object pose estimation is crucial in human-computer interaction systems. The traditional point-based detection approaches rely on the robustness of feature points, the tracking methods utilize the similarity between frames to improve the speed, while the recent studies based on neural networks concentrate on solving specific invariance problems. Different from these methods, PTPE (Part-based Tracking for Pose Estimation) proposed in this paper focuses on how to balance the speed and accuracy under different conditions. In this method, the point matching is transformed into the part matching inside an object to enhance the reliability of the features. Additionally, a fast interframe tracking method is combined with learning models and structural information to enhance robustness. During tracking, multiple strategies are adopted for the different parts according to the matching effects evaluated by the learning models, so as to develop the locality and avoid the time consumption caused by undifferentiated full frame detection or learning. In addition, the constraints between parts are applied for parts detection optimization. Experiments show that PTPE is efficient both in accuracy and speed, especially in complex environments, when compared with classical algorithms that focus only on detection, interframe tracking, self-supervised models, and graph matching.