A Novel Hybrid Approach of Autoencoding and Explainable AI for Video Anomaly Detection

The core video surveillance systems serve as the cornerstone for ensuring safety and security across public and office environments. Meanwhile, watching live feeds from lots of cameras involves inefficiency at times, though the traditional systems lack interpretability. This thesis presents a hybrid approach of developing an advanced version of video anomaly detection by proposing autoencoders with Explainable AI (XAI). It is trained and tested on the UCSD Pedestrian dataset, which contains preprocessing steps including frame extraction, resizing, and normalization. Reconstruction errors are calculated from the autoencoder to compare the original frames and reconstructed ones and identify as anomalous any one with the high error counts. For increased usability, SHapley Additive exPlanations (SHAP) is included, providing interpretability, and letting the user understand the nature and severity of the detected anomalies. It shows robust performance in complex environments with high accuracy. Real-time visual feedback helps users quickly identify anomalous frames, reducing the need for constant human supervision while improving response times.