A Scene-Adaptive Framework for Pose-Oriented Abnormal Event Detection

For intelligent surveillance systems, abnormal event detection (AED) automatically analyses monitoring video sequences and detects abnormal objects or strange human actions at the frame level. Due to the shortage of labelled data, most approaches for AED are based on reconstruction or prediction models in a semi-surprised manner. However, these methods may not generalize well to an unseen scene context. To address this, we present a pose-oriented scene-adaptive frame-work for AED. In this framework, we propose synergistic pose estimation and object detection, which integrates human poses and object detection information well to improve pose information accuracy. Subsequently, the enhanced pose sequences are taken into a spatial-temporal graph convolutional network to extract the geometric features. Finally, the features are embedded in a clustering layer to classify the type of actions and calculate the normality scores. For evaluation, the proposed framework is tested on video sequences with unseen scene context across from UCSD PED1 & PED2 and ShanghaiTech Campus datasets. The performance analysis and the results compared with other state-of-the-art works confirm the robustness and effectiveness of our proposed framework for cross-scene AED.