Chong Yong Shean Tay Yong Haur

Abstract

We present an efficient method for detecting anomalies in videos. Recentapplications of convolutional neural networks have shown promises ofconvolutional layers for object detection and recognition, especially inimages. However, convolutional neural networks are supervised and requirelabels as learning signals. We propose a spatiotemporal architecture foranomaly detection in videos including crowded scenes. Our architecture includestwo main components, one for spatial feature representation, and one forlearning the temporal evolution of the spatial features. Experimental resultson Avenue, Subway and UCSD benchmarks confirm that the detection accuracy ofour method is comparable to state-of-the-art methods at a considerable speed ofup to 140 fps.