An Examination on Autoencoder Designs for Anomaly Detection in Video Surveillance

Current anomaly detection methods for video surveillance find anomalies effectively enough; however, it comes at a high computational cost and specific hardware resources demanding. In counterpart, other video analysis tasks such as video action recognition now employ techniques that reduce the need for higher computational cost. Some of those techniques can be helpful for video anomaly detection. Therefore, this paper explores the effectiveness of the potential concepts of distillation and joint spatiotemporal training, adapted to two novel convolutional autoencoder architectures for anomaly detection in video surveillance. Our experimental results show the feasibility of reducing the computational resources requirements with smaller architectures (only $~6K$ trainable parameters), competing and outperforming current methods in challenging benchmarks.