Point Cloud Video Anomaly Detection Based on Point Spatiotemporal Autoencoder

Video anomaly detection is the task of localizing anomalies in space and/or time in a video, which has great potential to enhance safety in the production and monitoring of special areas. Previous works have made significant progress in RGB modality, but its redundant semantic information may breach the privacy of residents or patients. The 3D data obtained by depth camera and LiDAR can accurately locate anomalous events in 3D space and protect personal privacy through its sparsity. In this study, we propose Point Spatio-Temporal Auto-Encoder (PSTAE), a framework that can be used to detect anomalies that occur in point cloud videos. We introduce PSTOp and PSTTransOp to model human dynamics in point cloud videos. To measure the reconstruction loss of the proposed framework, a shallow feature extractor is introduced. To alleviate the issue of low reconstruction loss of some anomalous inputs caused by the excessively strong generalization ability of auto-encoder, we propose an anchor frame discard operation and form an asymmetric auto-encoder structure. Experimental results on the TIMo dataset show that our method outperforms the representative depth modality-based methods in terms of AUROC and sets a new state-of-the-art on the TIMo dataset. These results suggest the potential of point cloud modality in video anomaly detection.