Vision-Radar Semantic Communication for Target-Oriented Visual Sensing: Image Reconstruction and Dynamic State...

Target-oriented sensing image processing and analysis are important for intelligent surveillance and dynamic sensing applications, where wireless systems need to recover task-relevant visual and state information from heterogeneous sensing data. This study develops a vision-radar semantic communication system for joint target image reconstruction and dynamic state estimation. The transmitter jointly encodes scene-level visual observations and target-associated radar echoes through vision-radar semantic extraction and fusion, lightweight contextual semantic encoding, and communication-aware semantic transmission. The receiver uses a multi-task semantic decoder to reconstruct the target image and estimate normalized distance, angle, and velocity. Experiments are conducted under adaptive signal-to-noise ratio (SNR)-modulation operating points and evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and normalized state-estimation root mean square errors (RMSEs). The developed system achieves an average PSNR of 24.57 dB, SSIM of 0.8896, Normalized Distance RMSE of 0.0120, Normalized Angle RMSE of 0.0059, and Normalized Velocity RMSE of 0.0015, outperforming a continuous-channel training variant on all reported average metrics. Ablation studies show that the composite reconstruction objective and adaptive task-balancing scheme provide favorable visual-sensing tradeoffs. These results indicate that vision-radar semantic communication can support target-oriented visual sensing and dynamic state recovery within a unified sensing-oriented wireless framework.