摘要
這篇論文主要討論以視覺為基礎的人體動作分析，其中包含以模型為主的手部與走路的人物肢體動作分析，以及以外觀為主的人物步伐姿態識別。首先在多重視野下利用三維手部模型追蹤手腕的轉動與手指彎折的動作，並提出改良式的粒子濾波器－可分離狀態粒子濾波器來解決高維度的難題。透過整合多重視野下擷取出的多個特徵，提出的動作追蹤系統可以有效擷取手部動作參數並解決遮蔽的問題。 第二個系統為可追蹤走路的人物肢體動作的系統。首先建構符合結構與運動學的三維人體模型，使用輸入影像中人體的形狀與彩色直方圖作為觀測。我們的系統利用事先訓練的關節角度空間相關性與時間約束改良了傳統的退火式粒子濾波器的效果，特別是在人體走路當中肢體發生遮蔽的情況時，改良效果更明顯。 另外我們提出一個以外觀為主的步伐姿態辨識系統，擷取動態與靜態資訊來進行人物步行路徑和身分的識別。將這兩種特徵轉換至低維度空間以便進行人物識別。為了解決訓練影片與測試影片中，人物走路速度可能不同的問題，我們提出了一個混合人物識別演算法來選定最有效的特徵。每當測試特徵向量進來時，利用最鄰近分類法則進行路徑確認與人物識別。本系統利用CASIA步伐姿態資料庫進行評估，實驗結果證明的確可獲得極高的辨識率。
並列摘要
This thesis will discuss the vision-based human motion analysis. The research field contains the model-based hand and walking human body motion analysis, and the appearance-based human gait recognition. A 3-D hand model is developed to track the rotating hand with finger movements in multi-view. To solve the difficulty of high dimensionality, a new modified version of particle filter - separable state based particle filtering (SSBPF) is proposed. Then, by integrating different features in different view angles, the proposed motion tracking system can capture the hand motion parameter effectively and solve the self-occlusion problem of the finger motion. We also propose a system to track the walking body parameters in the videos from different perspectives. A 3D human model with structural and kinematic constraints is constructed. The shape and color histogram of the image is obtained as the observation. Our system improves the traditional annealed particle filter (APF) with the pre-trained joint angle spatial correlation and the temporal constraint, especially when self-occlusion occurs. A gait analysis method is also developed to extract the dynamic and static information from the input video for walking path determination and human identification. The extracted dynamic and static feature is then transformed into lower dimensional embedding space. A hybrid human ID recognition based on the walking velocity of human object is proposed, which can select the more effective feature. Given a test feature vector, the nearest neighbor classifier is applied for walking paths determination and human identification. The proposed algorithm is evaluated on the CASIA gait database, and the experimental results demonstrate a highly acceptable recognition rate.
參考文獻
[1] M. Bray, E. Koller-Meier, and L. Van Gool, “Smart Particle Filtering for 3D Hand Tracking,” in Proc. IEEE FGR 2004.
[2] M. Bray, E. Koller-Meier, and L. Van Gool, “Smart Particle Filtering for High-dimensional Tracking,” Computer Vision and Image Understanding, vol. 106, Apr. 2007.
[3] D. Huan and E. Charbon, “3D Hand Model Fitting for Virtual Keyboard System,” IEEE Workshop on Applications of Computer Vision 2007.
[4] Y. Wu, Y. L. Lin, and T. S. Huang, “Capturing Natural Hand Articulation,” ICCV 2001.
[5] A. Erol, G. Bebis, M. Nicolescu, R. D. Boyle, and X. Twombly, “A Review on Vision-Based Full DOF Hand Motion Estimation,” CVPR 2005.