基於相容性身體部位組態的隨意姿勢人體偵測研究

Abstract

在本篇論文中，我們會著重於靜態影像中的隨意姿勢以及視角的人體偵測。針對此議題，近年來的主流代表作大多只能透過引進更多的人體姿勢模板來做偵測，這樣無疑會大幅度提升運算量!對此，其實只要限制所偵測的目標必須有很高的機率重複出現在各種不同的姿勢以及視角，即可迴避此問題發生!在此，我們限制所偵測的目標為四肢、臉、頭以及軀幹。對比基於相同想法的幾篇相關論文，我們提出幾個不同的觀點。第一點，我們認為可以藉由假設四肢是由數個大小位置略有差異的片段所組合而成的，來提高對於四肢形變的容忍度。第二點，頭與軀幹的形變可以透過使用可形變身體部位模型來增加容忍度。第三點，只討論有偵測到的身體部位所扮演的角色，可以更好的應對遮蔽現象所帶來的負面影響!第四點，影像中的區域型資訊以及人體四肢的特質，可以輔助我們減少所需要偵測的範圍，達到加速的目的。

In this thesis, we focus on the detection of human with arbitrary poses in different view-points in static images. To handle this issue, recently representative works need to produce lots of detectors to cover the cases of human with arbitrary poses in different view-points. In this way, the computation cost will be increased exponentially. To prevent this dilemma, we restrict body parts for detection to be limb, head, face or torso, which have high probability to be observed in arbitrary poses and view-points. Compared to related works in the literature, several different opinions are proposed. Firstly, a patch based approach is proposed to model the limb instead of parallel lines or well-segmented half limb used in related works. Secondly, a strong classifier with the “Deformable Part Model” proposed by Felzenszwalb et al. [1] is adopted to cover more variation on head-torso shape, instead of using the rectangular shape assumption for torso. Thirdly, we consider configuration inference as a label assignment problem, instead of a model fitting problem, in order to handle the limitation caused by occlusion or missing parts. Finally, instead of exhaustive search, segmentation information and native property of limb are adopted to reduce the searching space.