Efficient Recursive Convolutional Target Detector for FMCW Radar with Implementation on a Programmable Deep Learning Processor Unit

Advanced Driver Assistance Systems (ADAS) and autonomous driving (AD) systems are becoming increasingly vital in modern vehicles. These systems rely on precise target detection to enhance safety and performance. Frequency-Modulated Continuous-Wave (FMCW) radar sensors are central to ADAS/AD, offering high-resolution range and velocity measurements. However, traditional detection algorithms like Constant False Alarm Rate (CFAR) methods face limitations in complex, cluttered environments, especially under high interference conditions from other automotive radars. To overcome these challenges, we propose a novel Recursive Convolutional Target Detector (RCTD) algorithm that elevates detection performance while adhering to the stringent real-time and hardware constraints of ADAS/AD platforms. The RCTD algorithm utilizes a lightweight Convolutional Neural Network (CNN) that recursively processes segmented Range-Doppler (RD) maps to efficiently localize targets. This hierarchical approach reduces computational complexity and minimizes false alarm rates by concentrating computational efforts on regions of interest. We validate the RCTD algorithm through extensive simulations using realistic FMCW radar models and demonstrate its robustness across various scenarios. Furthermore, we implement the RCTD on field-programmable gate array (FPGA) hardware equipped with a Deep Learning Processing Unit (DPU), illustrating its capability to meet the latency and resource requirements of embedded ADAS/AD systems. Our results indicate that the RCTD algorithm outperforms traditional CFAR methods, achieving higher detection accuracy and lower false alarm rates, thus advancing the state of the art in FMCW radar target detection for ADAS/AD applications.