An FPGA-Assisted WiFi-CSI Framework for Human Pose Estimation in Search-and-Rescue Environments

Abstract

Recent advances in wireless sensing show that commodity WiFi can support human perception indoors. This study proposes WiFi-based human pose estimation for rescue scenarios where cameras fail due to smoke, darkness, occlusion, or privacy constraints. Channel State Information (CSI) from standard WiFi links is processed through a two-stage GAN pipeline. First, windowed CSI tensors are mapped into pose-oriented representations, such as keypoint heatmaps or latent pose maps. Second, a refinement GAN estimates physically consistent joint coordinates using adversarial, reconstruction, bone-length, and kinematic losses. CSI denoising, normalization, and multi-link aggregation improve robustness under multipath conditions. For real-time embedded deployment, the trained PyTorch model is compiled using hls4ml and implemented on the AMD Kria KR260 FPGA platform. This enables low-latency, contactless, and lighting-independent pose estimation in GPS- and vision-degraded rescue environments.