ANOMALY DETECTION IN SURVEILLANCE VIDEOS USING DEEP RESIDUAL NETWORKS

Lucas P. Cinelli,E. Silva,Antonio Lobato,Carlos Oliveira,Eduardo Barretto,Hugo Sadok,Thiago Cosenza

Abstract:Efficient anomaly detection in surveillance videos across diverse environments represents a major challenge in Computer Vision. This work proposes a background subtraction approach based on the recent deep learning technique of residual neural networks capable of detecting multiple objects of different sizes by pixel-wise foreground segmentation. The proposed algorithm takes as input a reference (anomalyfree) and a target frame, which should be temporally aligned, and outputs a segmentation map of same spatial resolution. Experiments show competitive performance in the studied dataset, as well as real-time capability.