Local-Scene-Guided Recursive Reconstruction for Skeleton-Based Video Anomaly Detection via Diffusion Model

Video anomaly detection(VAD) is one of the important branches in the field of computer vision. Existing VAD methods are primarily categorized into skeleton-based frameworks and RGB-based frameworks. For skeleton-based video anomaly detection, we propose a novel framework named Local Scene-Guided Recursive Reconstruction (LSGRR). The framework initially employs the forward diffusion process of a diffusion model to generate stochastic noise for the action sequence. Subsequently, LSGRR need to restore this sequence. It utilize the prediction results of the previous action and the stochastic noise of the current action to restore the action. Then, this process is carried out recursively. The intermediate feature representations from the Spatial-Temporal Graph Convolutional Network (ST-GCN) are utilized as one of the generative conditions for the diffusion model and are continuously updated throughout the process of reconstructing the sequence. Additionally, we also extract local scene features while masking out the foreground. The features will also serve as a part of the diffusion conditions. We validate our framework on four datasets: HR-Avenue, HR-STC, HR-UBnormal, and UBnormal, and have achieved commendable results across all of them.