This paper presents a fault-tolerant human detection system that integrates a 77 GHz FMCW radar module and a visible-light camera through a hybrid sensor fusion architecture. The camera subsystem employs YOLOv4-tiny to perform real-time person detection, while the radar subsystem, built around the Texas Instruments AWR1843 ES2.0 single-chip SoC, leverages on-chip signal processing to offload the host CPU and provides range, angle, and velocity measurements as a processed point cloud. Under normal operating conditions, detected bounding boxes from the camera are fused with projected radar point cloud data to estimate target distance. A three-criterion camera health monitoring mechanism continuously evaluates image quality through focus measure and edge density analysis. When the camera is deemed unreliable due to low illumination, occlusion, or hardware failure, the system autonomously transitions to a radar-only detection mode employing DBSCAN-based clustering and human signature classification, with mode switching completed within 333 ms. This hybrid approach ensures continuous human detection regardless of camera availability. Inter-module communication is established via a TCP-based real-time data stream using a lightweight JSON protocol. Experimental results demonstrate that sensor fusion raises detection confidence from 0.5 in radar-only mode at 1.7 m to 0.9 in fusion mode at 1.7 m, while providing metrically grounded distance estimates that neither modality can deliver alone. The complete system sustains approximately 30 Hz throughput on a CPU-only embedded platform, confirming that robust, fault-tolerant human detection is achievable without dedicated GPU hardware. These findings highlight the proposed system's potential for safety-critical ADAS and indoor monitoring applications.