Semantic Communication for Video Surveillance in Railway Intrusion Detection

In recent years, video surveillance has played an increasingly important role in railway intrusion detection. Due to limited computational resources, it is difficult for the surveillance devices deployed along the railway to run the entire video analysis network, necessitating the transmission of videos to edge servers over wireless channels for further processing. In this paper, we propose a video semantic communication (SC) system named Rail-SC for railway intrusion detection. The goal is to achieve high detection accuracy while reducing transmission bandwidth by transmitting task-relevant semantics, thereby enabling efficient communication. To avoid the additional computational and storage overhead of training and storing multiple models for specific signal-to-noise ratio (SNR) values, we design a plug-and-paly SNR-adaptive module to achieve dynamic encoding and decoding based on SNR. By incorporating this module, Rail-SC can perform well across a wide range of SNRs, indicating the robustness to various wireless channel conditions. Experimental results demonstrate that, compared to traditional approaches that combine video transmission with downstream intrusion detection, Rail-SC achieves a significant gain in detection accuracy under various SNRs, which is more pronounced in the low SNR range. Particularly, when SNR is 0 dB, the accuracy gain is up to 19.26%, demonstrating the effectiveness of Rail-SC.