Explainable Action-Recognition Based Approach For Unsupervised Video Anomaly Detection
Résumé
In recent years, there has been a growing focus on videoanomaly detection from unlabeled data, raising the task of UnsupervisedVideo Abnormal Detection (UVAD). However, most of the existingapproaches predominantly focus on global features derived from entireframes, while overlooking local features associated with individual objects.This oversight can result in sub-optimal performance and loss ofsemantics and explainability of the video scene provided by local features.In this paper, for the task UVAD, we introduce a Global-Local ExplainableNetwork (GLE), which focuses on local features and is based on aMulti-Instance Learning (MIL) method. The proposed approach not onlyoutperforms UVAD state-of-the-art approaches, but also provides explanationsabout the anomaly by leveraging the rich information within localfeatures. Our experiments demonstrate that GLE achieves the state-ofthe-art performance in both detection and explanation. GLE achieves upto 6.22 improvement in the AUC for abnormal event detection on twowidely used datasets, UCF-crime and ShanghaiTech. Moreover, GLE offersup to 62.94% enhancement in explanatory capabilities as validatedon the X-MAN dataset.