Radar-Only 3D Object Detection for Long-Range Automotive Perception: Architectures, Datasets and Embedded Deployment

Robust environmental perception is a fundamental prerequisite for advanced autonomous driving functions, particularly demanding high-speed tasks such as overtake assist in performance vehicles and autonomous heavy-duty truck platooning. While LiDAR and camera systems are widely used, automotive radar offers unique advantages due to its extensive range, cost-effectiveness, and resilience to adverse weather and lighting conditions. However, performing accurate 3D object detection using radar-only data remains a significant challenge due to the sensor’s inherent sparsity, lack of precise geometric resolution, and susceptibility to multi-path noise (ghost points). This thesis comprehensively evaluates state-of-the-art neural network architectures for radar-only 3D object detection, specifically analyzing grid-based methods (PointPillars, PillarNet), graph-based approaches (RadarGNN), and LiDAR-radar cross-modal feature hallucination. To address the stark domain shift between existing public datasets—which often focus on high-speed highway scenarios—and dense urban environments, a custom 3,429-frame dataset was engineered utilizing a 3+1D radar configuration. The ground truth was generated through a robust, semi-automated labeling pipeline incorporating an ego-motion compensated Kalman Filter, the Hungarian algorithm for data association, and rigorous manual refinement. Extensive experiments were conducted to baseline the architectures against standard public datasets (TruckScenes, nuScenes, View-of-Delft) before executing a multi-phase fine-tuning process to adapt the models to the heavily static, urban distribution of the custom dataset. Model accuracy was quantitatively evaluated using KITTI Average Precision and nuScenes evaluation metrics. Additionally, a multi-radar Cross-Potential preprocessing algorithm was investigated for spatial noise reduction, ultimately highlighting critical limitations regarding long-range detection preservation. The experimental evaluation yields a definitive verdict: PillarNet significantly outperforms the other architectures. By leveraging a 2D sparse convolutional encoder and an anchor-free CenterHead, PillarNet achieved a peak Car Mean Average Precision (mAP) of 0.63 on the TruckScenes dataset, demonstrating superior spatial accuracy and overall detection precision. Furthermore, to bridge the gap between theoretical accuracy and real-world deployment, the PointPillars architecture was subjected to rigorous post-training acceleration using ONNX and NVIDIA TensorRT. This optimization successfully slashed steady-state inference latency to sub-millisecond levels—achieving execution times as low as 0.75 ms-demonstrating real-time feasibility on desktop-class hardware while effectively bounding dynamic peak memory consumption. Ultimately, this work delivers a holistic framework for evaluating radar-based perception, from raw signal challenges and data engineering to hardware-accelerated inference.