A Novel Framework for Anomaly Detection in Video Surveillance Using Multi-feature Extraction

Abstract

In this paper we present a novel framework based on multi-feature extraction for anomaly detection in video surveillance which global anomaly and local anomaly are detected separately. To detect global anomaly, we define kinetic energy Ek and compute the first derivative of Ek and then derive a global anomaly score of each test frame. As for local anomaly detection, three kinds of local anomaly are defined namely appearance anomaly, location anomaly and velocity anomaly where different kinds of features are extracted respectively and finally fused into a unified framework. At last, an improved Normality Sensitive Hashing method is proposed to classify abnormal instances from normal instances. The experiment results demonstrate that our method can detect global and local anomaly with a comparative performance.