在實驗的部份，我們藉由Kinect設備收集了幾種常見的單人行為與多人行為的深度影像與骨架資訊，丟入動作森林演算法進行訓練，並組合比較各種參數之效果，以及相對於原始演算法的改良程度。從結論來說，此一訓練模型下成功的適應了三維空間與人體骨架相關的特徵，並且能在即時運行 (30fps) 的條件下有效的分類出不同的人類行為模式。 Human action recognition is one of the most important issues in computer vision. In this thesis, we plan to design a general approach to recognize human behavior. The approach is implemented based on a pre-collected action database, which is extracted by the depth images to form the sequences of skeletons, trained by the proposed Action Forests (AF) model. The proposed AF model extends the random forest algorithm by using different decision functions to fit the skeletal features in the 3D space. The system achieves real-time classification result without the limitation constrained by background and camera position.
Experiments were conducted on various examples to verify the validity of the proposed method. Several human behaviors with single-character actions and two-person interactions were collected to train the AF model. The skeletal features were retrieved from the depth sensor Kinect. In the experiments, we investigate the effects of several training parameters in AF. Experimental results demonstrate that the proposed AF model can learn the skeletal features efficiently and run at 30 frames per second on action classification with high accuracy.
