Automated Video Surveillance Anomaly Detection with a Deep Reinforcement Learning Framework

Anomaly detection in automated video surveillance is regarded as one of the most essential challenges to address, with the objective of identifying various real-world irregularities. This paper presents an innovative method for anomaly detection utilizing deep reinforcement learning. In recent years, deep reinforcement learning has demonstrated considerable success across diverse applications involving complex data, such as robotics and gaming, by emulating human learning through experience. Typically, state-of-the-art techniques categorize a video as either normal or abnormal without identifying the precise location of the anomaly within the input video, primarily due to the use of unlabeled clip-level data during training. Our focus is on adapting prioritized Dueling deep Q-networks to tackle the anomaly detection issue. This model is designed to assess anomalies in video clips by leveraging video-level labels to enhance detection accuracy.