Semantic Communication Empowered Collaborative Perception in Constrained Networks

Abstract

Traditional collaborative perception methods typically focus on optimizing perception performance in ideal wireless transmission conditions. However, in real-world constrained networks, it is crucial for collaborative perception schemes to alleviate network load while ensuring performance robustness. To address this challenge, this letter introduces S2CP, a Semantic Communication empowered Collaborative Perception framework. Within the S2CP, we propose the Multi-scale Dilated Cross-Attention (MDCA) module to effectively extract task-oriented valuable semantic features for transmission, thereby minimizing data transmission overhead and improving perception performance. Furthermore, to mitigate feature distortion during wireless transmission, we develop a pre-training strategy utilizing Masked AutoEncoders (MAE) to enhance the robustness of S2CP. Experimental results demonstrate that S2CP significantly enhances perception performance while substantially reducing network transmission volume.