VQ-VAD: Discrete Motion Learning for Human-centric Video Anomaly Detection

A novel approach for video anomaly detection (VAD), named Vector-Quantized Video Anomaly Detection (VQ-VAD), has been introduced in a paper on arXiv (arXiv:2608.05069). This technique tackles issues in VAD, including the limited occurrence of anomalies and the visual diversity found in surveillance videos, by emphasizing pose-based analysis to reduce visual clutter and address privacy issues. Unlike traditional pose-based methods that represent human actions in continuous latent spaces, VQ-VAD utilizes Vector-Quantized GAN (VQ-GAN), initially designed for image creation, to learn discrete motion representations from keypoint sequences. This enables the development of a motion codebook based solely on normal behavior patterns. The framework prioritizes human-centric analysis, enhancing the reliability of behavior evaluation by capturing concise motion patterns. The paper may also have been presented or published in other venues, contributing to advancements in AI-driven surveillance and anomaly detection with potential uses in security and monitoring systems.