Camera Pose Self-calibration-based View-invariant Trajectory Analysis with Monocular Vision

Mo-Yi Zhang ； Qiu-Yu Zhang ； Hong-Xiang Duan ； Hui-Yi Wei

《電腦學刊》 31卷2期 (2020/04) Pp. 212-226

https://doi.org/10.3966/199115992020043102018

摘要

Detecting view-invariant trajectory features is a fundamental task in many human-computer interactions. Existing approaches often rely on motion models or behavioral models, which have strict constraints, and the accuracy of detection cannot satisfy the requirements. Thus, we propose a new practical method for extracting the view-invariant features of the trajectory. The most innovative feature of our proposed approach is a camera pose self-calibration model. By extracting sparse feature points from some video frames and matching them across frames, our model can compute the camera pose related to the motion plane. In experiments, we analyzed our method in terms of the correctness, effectiveness, computational efficiency, robustness, and error obtained. In addition, the proposed method and new calibration model obtained greater accuracy at trajectory analysis compared with a previously proposed motion estimation method and a calibration method. The new method obtained satisfactory performance in a gesture drawing experiment. The proposed model can be applied widely in view-invariant trajectory analysis.

關鍵字

camera pose self-calibration ； gesture interaction ； monocular vision ； motion trajectory analysis ； structure from motion ； view-invariant