RLST-VAD: Robust Lightweight Semantic Transmission for Edge Video Anomaly Detection

Abstract

With the increasing demand for real-time video anomaly detection (VAD) in edge intelligent applications, efficiently transmitting task-relevant semantic information under limited bandwidth and dynamically varying wireless channels has become a critical challenge. Existing video semantic transmission methods lack collaborative optimization with VAD tasks, making it difficult to preserve anomaly-related semantic features. In this paper, a Robust Lightweight Semantic Transmission system for Edge VAD (RLST-VAD) is proposed. The RLST-VAD adopts a parallel architecture for video reconstruction and VAD, enabling real-time temporal anomaly detection while maintaining semantic fidelity for video reconstruction. Furthermore, a robust lightweight convolution-based SNR-adaptive module is proposed to adaptively protect locally critical semantic features under varying channel conditions, while substantially reducing computational complexity and parameter overhead for edge deployment. Experimental results show that RLST-VAD improves anomaly detection with respect to the Area Under the Curve (AUC) by 16% compared to the sequential design employing fully connected SNR-adaptive module, while reducing the parameters of the proposed SNR-adaptive module by 75.5%.