Deep Reinforcement Learning for Real-world Anomaly Detection in Surveillance Videos

Surveillance videos are considered as a very important mechanism in a smart city project. In recent years, a method called deep learning was combined with reinforcement learning techniques to learn useful representations for the problems with high dimensional raw data input. In this paper, we develop a Deep Q Learning Network (DQN) to localize anomalies in videos by enabling the agent to learn how to detect and recognize the abnormalities in videos. Our idea is inspired by multiple instance learning (MIL) techniques based on common share features with reinforcement learning. We consider normal and abnormal videos as bags and the selection of videos clips as actions. In our DQN architecture we will design a fully connected layer, which compute probability for each video segment in both positive (anomalous) and negative(normal) bags indicating how likely a clip is containing an anomaly. Our method is applied to a new large-scale dataset of 128 hours of videos called UCF-Anomaly-Detection-Dataset, it is about of 1900 long and untrimmed real-world surveillance videos, with 13 cases of realistic anomalies.