基於影像掃描之即時單攝影機3D人體姿態估測

Abstract

近年來，電腦輔助之人類行為認知的應用越來越廣，這些應用包括事件偵測、行為分析、姿態辨識和人機互動等方面。當應用在人機互動時，我們可以利用不同的姿勢或姿態來表示不同的命令或指令，如作遊樂器的控制。為了能在三度空間中做姿態的辨識，通常會加入控制感應器，增添人機互動應用的臨場感跟多元性。但是多數3D遊戲控制需使用昂貴的設備，或是要求使用者使用特殊的控制感應器，容易造成局部運動傷害，且方便性不足。為了改善上述情形，我們希望以電腦視覺的方法，在單一攝影機下，不使用控制感應器，即能自動的辨識人體3D姿態，讓使用者只需擺動肢體，即可在三度空間中進行人機互動。如此可降低局部運動傷害，也使人機互動應用的空間更廣、更真實。本研究成果期望在架設有一般網路攝影機的簡單平台下，提供使用者人機互動的介面。

Recently, the computer-aided applications in human behavior recognition are becoming more and more popular, which include event detection, behavior analysis, posture recognition, and human-machine interaction. For human-machine interaction, e.g., for game control, various gestures or postures can be used to indicate different orders or commands. In order to distinguish behaviors, e.g., for game control, remote controllers are often used to make users feel as if they were in a real-world environment, like the sports field. Compared with the traditional game controllers, the above approach adds more immediate reactions and varieties. However, most 3D game controllers require expensive equipments or ask user put on sensors, which are often inconvenient, and may result in exercise injuries. To improve the above situations, we hope to develop computer vision method by using monocular camera to recognize 3D human postures. The user can then move the limbs in the 3D space to perform the human-machine interaction. In this way, the exercise injuries can be reduced while enabling more realistic application. This research hopes to employ computer vision to make the human-machine interaction system more affordable on ordinary webcams.