Targeting the multi-target detection requirements of automotive millimeter-wave radars in complex traffic environments, this paper proposes an edge-side AE-CFAR detection framework based on the AWR2944 platform. The system generates an RD- \(\theta \) three-dimensional data cube through a \(3\textrm{Tx}\times 4\textrm{Rx}\) array, constructs a reconstructable background model with a lightweight autoencoder, and conducts CFAR threshold decision-making in the residual domain. Experiments conducted in an anechoic chamber, urban vehicle scenarios, artificial smoke environments and highway gantry settings demonstrate that the proposed method operates stably in scenarios with dense targets, low visibility and long-distance high-speed conditions. It supports a maximum of 50 stably detected targets, with a detection rate exceeding 92% under dense working conditions. In low visibility with a visual range of 5 meters, the radar pedestrian detection rate reaches approximately 87%, while that of cameras drops below 40%. The P95 end-to-end latency is about 68 ms, and the time consumption of AE inference is less than 5 ms per frame. The results verify that this framework boasts excellent feasibility for edge-side deployment.
