Confidence-Aware Anomaly Detection in Human Actions

Anomaly detection in human actions from video has been a challenging problem in computer vision and video analysis. The human poses estimated from videos have often been used to represent the features of human actions. However, extracting keypoints from the video frames are doomed to errors for crowded scenes and the falsely detected keypoints could mislead the anomaly detection task. In this paper, we propose a novel GCN autoencoder model to reconstruct, predict and group the poses trajectories, and a new anomaly score determined by the predicted pose error weighted by the corresponding confidence score associated with each keypoint. Experimental results demonstrate that the proposed method can achieve state-of-the-art performance for anomaly detection from human action videos.