The recognition based on vision to extract features from abnormal human behaviors usually utilize straightforward sharp movement information or traditional PCA methods. The former lacks of information and the latter has ignored nonlinear information in data. Therefore, this paper will use KPCA in recognizing abnormal human behaviors to solve the aforementioned problems. Since KPCA has some defects in extracting feature abnormal behaviors, W2KPCA-KNN algorithm is proposed, which is to do weighting in both feature extraction and classification respectively. While retaining behavioral information in the image, it improves recognition accuracy and satisfies the technical requirements for abnormal behavior recognition system. The experimental comparisons show that this algorithm outperforms traditional KPCA and K-Nearest Neighbor classifier on both feature extraction and classification.