MC-MIL: video surveillance anomaly detection with multi-instance learning and multiple overlapped cameras

Abstract

Anomaly detection approaches have limiting aspects regarding the representativeness of the information since the video data is captured from a single perspective and may not distinguish all relevant aspects of the scene. The lack of sufficient labeled data is also a challenging aspect of building video anomaly detection approaches. Although multiple instance learning (MIL) has been explored extensively in the weakly supervised video anomaly detection (WS-VAD) literature since it is less hungry for labeled data, there are no studies that exploit multiple overlapping camera views to provide wider representativeness of vision data under MIL assumption. In this work, we show the performance of the video anomaly detection task can be improved by using multiple cameras to capture spatiotemporal information from different perspectives. We propose the approach MC-MIL (Video Anomaly Detection with Multiple Overlapped Cameras and Multiple Instance Learning) framework, which consists of a training scheme with multiple cameras under multiple instance learning for video anomaly detection. We specialize our proposed framework for the two-camera case as a proof of concept for performance evaluation. Due to the lack of datasets for this task, we relabeled the multiple-camera PETS-2009 benchmark dataset for the anomaly detection task from multiple overlapped camera views to evaluate the MC-MIL algorithm. The result shows a significant performance improvement in the AUC ROC score compared to the single-camera configuration and with the literature.