MSPL-VAD: Multi-Stage Pseudo-Labeling framework for Video Anomaly Detection

Video anomaly detection is challenging due to the rare occurrence, high variability, and difficulty in defining anomalies. Traditional methods rely on manual annotations, struggle with generalization, and are computationally expensive. To address the aforementioned issues, we propose a Multi-Stage Pseudo-Labeling framework for Video Anomaly Detection (MSPL-VAD). Our framework integrates Global Pseudo-Labeling (GPL) with Deep Embedded Clustering for coarse separation of normal and anomalous instances, followed by Local Pseudo-Labeling (LPL) using Kernel Density Estimation to refine segment-level labels with greater precision. Furthermore, we also introduce a self-attention-based anomaly detector that can be effectively applied to segments of an unseen test video, producing anomaly predictions at both the segment and frame levels. To validate the effectiveness of the proposed MSPL-VAD, we have conducted experiments on UCSD Ped2, CUHK Avenue, ShanghaiTech, and UCF-Crime. Experimental results demonstrate that our approach achieves superior performance than state-of-the-art results.