GSP-enhanced embedded hardware framework for real-time multi-class road target detection in ADAS using mmWave radar

Reliable perception of road environments remains a critical challenge for autonomous driving systems, particularly under adverse weather and low-visibility conditions where conventional sensors such as cameras and LiDAR experience significant performance degradation. While millimeter-wave (mmWave) radar offers robust sensing capabilities in such environments, effectively exploiting its sparse, noisy, and non-uniform point cloud data for accurate multi-class road object classification in real time remains an open problem. Existing methods often rely on computationally intensive deep learning architectures or handcrafted feature representations that fail to adequately preserve spatial relationships, leading to increased inference latency and power consumption. These limitations restrict their practical deployment on resource-constrained edge platforms. To address these challenges, this paper presents a real-time, multi-class road object classification framework based on mmWave radar sensing, graph-based feature extraction, and a Light Gradient Boosting Machine (LightGBM) classifier. The radar-derived three-dimensional point cloud data are projected into two-dimensional representations using top-view (TV) and front-view (FV) filtering, where the TV projection more effectively preserves spatial relationships and enhances class separability. The proposed framework is evaluated on a balanced dataset comprising 7800 radar samples uniformly distributed across 13 road object categories, including multiple vehicle types, vulnerable road users, static obstacles, and road boundaries. Experimental results demonstrate high classification accuracy of 99.00% with low inference latency of 0.98 ms and a measured power consumption of only 0.63 W. Furthermore, the complete processing pipeline is implemented using hardware-software co-design on PYNQ-ZU FPGA platform, validating its low-latency, energy-efficient performance and suitability for real-time autonomous driving applications. • Graph-based features capture structure in TV-filtered radar point clouds. • LightGBM achieves 99% accuracy with 0.98 ms latency at 0.63 W power. • PYNQ-ZU FPGA enables low-latency, energy-efficient edge AI driving.