Publication

Abstract : Artificial intelligence in the field of security and surveillance is evolving so rapidly and video anomaly detection (VAD) is an essential activity that seeks to discover anomalous events in large-scale video streams. In this study, a memory module-enhanced convolution neural network (CNN)-based unsupervised anomaly detection technique is presented. Improving anomaly detection’s generalization is the main goal, with a special emphasis on human-centric events like fights and robberies. Our method adds a new updating methodology to the memory adaptors to extract representative patterns of normal data, and then enhances discriminating power with specially designed losses for these memory adaptors like compactness and separateness losses. Using an end-to-end attention-based (memory adaptors mimics attention) and a encoder-decoder architecture, Further we expand our approach to include skeletal sequences, enabling the collaborative learning of hidden representations of occluded regions. The efficacy of our proposed method is confirmed by trail evaluations carried out on benchmark datasets, such as Avenue and ShanghaiTech, which produce Area Under the Curve (AUC) scores of 82% and 73% for appearance based anomaly detection module and 89% and 78% for Skeleton Trajectories. Finally, handle human-centric anomaly occurrences by smoothly integrating our VAD system with a video captioning model, improving interpretability and providing explanation captions for anomalies that are discovered. Code and features are released at GitHub ramk06122000/VAD.