Falls in elderly remain a very important public health care issue. The wearable devices based on tri-axial accelerator proves to be an effective tool for fall detection in the recent years. In this paper, we propose an approach to distinguish falls and normal activities of daily living (ADL). A novel method 3D Kernel Principal Component Analysis (3D KPCA) to improve (Kernel Principal Component Analysis) KPCA based feature extraction is developed which can extract the statistical features without the loss of 3D data structure information. Additionally, the threshold techniques and AdaBoost are combined for prediction. The fall detection based on KPCA and 3D KPCA algorithm for feature extraction is firstly proposed in the paper and the experiment conducted on the public database (UCI) shows the efficiency of the approach.