Review of Abnormal Human Behavior Detection

This paper systematically reviews video-based human anomaly behavior detection technologies. First, it defines and categorizes anomalous behaviors (individual anomalies such as falls and violent acts; group anomalies such as fights and stampedes). Next, it analyzes five types of feature extraction methods (motion information, motion trajectories, spatiotemporal interest points, skeletal information, and deep learning features) and discusses their strengths and weaknesses. In terms of detection approaches, it compares the performance and applicability of supervised, semi-supervised, and unsupervised learning, highlighting key challenges such as the dependency of supervised learning on labeled data and the poor generalization of unsupervised methods. The evaluation framework based on ROC curves, AUC, accuracy (ACC), and true positive rate (TPR) is introduced. Finally, future research directions are outlined, including multimodal fusion, self-supervised learning, lightweight model design, and privacy protection, along with their potential applications in intelligent security and smart healthcare. This paper provides a systematic reference for both theoretical innovation and practical implementation.