Automatic detection of human fall events is a challenging but important function of the real-time surveillance system. The goal of the proposed system is to develop a frame-by-frame fall detection system based on real-time RGB-D camera devices. The proposed system is composed of a complex off-line learning stage which combines several novel machine learning techniques and a series of on-line detection processes. A background subtraction method based on iterative normalized-cut segmentation algorithm is proposed to identify the pixel-wise human regions rapidly. The silhouettes are extracted to measure the pose similarity between different samples. Manifold learning algorithm reduces the feature dimensions and several discriminant analysis techniques are applied to model the final human fall detector. The experimental database contains 65 color video and corresponding depth maps. The experimental results based on a leave-one-out cross-validation testing show that our proposed system can detect the fall events effectively and efficiently.