Anomalous Event Detection in Surveillance Videos using Spatio-temporal Autoencoders

Surveillance cameras are proliferating, and millions of devices are being used to capture endless footage of surveillance videos. With the advancements in computer vision and deep learning, we can now contemplate these videos to detect anomalies. In this paper, we propose to identify anomalies by using Spatiotemporal autoencoders. The autoencoders based on 3-D convolutional networks will identify anomalies in surveillance footage based on spatiotemporal features. Our architecture consists of two components, an encoder for spatial feature extraction, and a decoder for the reconstruction of frames. Then, abnormal events are identified based on reconstruction loss. We used the Avenue dataset and UCSD dataset for training and evaluation.