具單目視覺距離量測之演示學習仿人機器人系統

Abstract

本論文主要貢獻在於提出了一種基於單目視覺距離測量的仿人機器人演示學習系統。該系統結合基於數據驅動的人體動作識別和使用鏈結向量和虛擬關節的機器人運動控制，使仿人機器人可以模仿人類的動作，此外，提出了一種具單目測距方法的類視覺里程計，該方法中提出了兩種數學模型，可使用相機平面視圖圖像和不同的相機姿態下的圖像進行距離測量，所提出的方法不需要高精度雙目攝像頭或額外的傳感器來測量距離。這種方法可以應用於各種應用領域，如物料搬運、監視和自動車輛系統，具有低成本和易於實施的額外優勢。最後，ㄧ些實際的實驗證明我們的系統對於不同的相機姿態和環境的條件下具有一定的準確度和穩定性。

This thesis proposes a humanoid robot learning from demonstration system with monocular-vision-based distance measurement. The system combines a data-driven-based human action recognition and a robot motion control based on link vectors and virtual joints. This enables the humanoid robot to imitate human actions. Moreover, two mathematical models are proposed for distance measurement using flat view images and different camera poses. The proposed method does not require high-precision binocular cameras or additional sensors to measure the distance. Furthermore, this method can be utilized in various applications, such as material handling, surveillance, and autonomous vehicles, with the added advantages of low cost and ease of implementation. Finally, some practical experiments show that our system is accurate and robust to different camera poses and environmental conditions.