Motion-Constrained Generative Adversarial Network for Anomaly Detection

Abstract

AbstractIncreased volume of video data made human monitoring impossible and necessitated automated abnormal event detection system. Abnormal event detection involves learning normal patterns and identifying any deviation as abnormal. In this paper, we propose a generative adversarial network (GAN) based abnormal event detection system. GAN is trained to generate normal frames, and fails to generate the frames with abnormal events. Usually adversarial loss and appearance loss are used to train GANs. In this paper, we propose to impose motion constraint using optical flow and acceleration during generator training. Our experiments on standard datasets prove that motion is an important cue for abnormal event detection.KeywordsAbnormal event detectionGenerative adversarial network (GAN)Motion constraint