Inhaltszusammenfassung

Real-time object detection from Light Detection and Ranging (LiDAR) data is a critical requirement for autonomous driving and higher-tier advanced driver assistance systems. However, many state-of-the-art LiDAR detection pipelines rely on computationally demanding 3D feature extraction, voxel processing, or multi-stage regression, limiting their applicability on constrained automotive embedded platforms. This paper presents a LiDAR Bird’s Eye View (BEV) detection framework that combines an op...Real-time object detection from Light Detection and Ranging (LiDAR) data is a critical requirement for autonomous driving and higher-tier advanced driver assistance systems. However, many state-of-the-art LiDAR detection pipelines rely on computationally demanding 3D feature extraction, voxel processing, or multi-stage regression, limiting their applicability on constrained automotive embedded platforms. This paper presents a LiDAR Bird’s Eye View (BEV) detection framework that combines an optimized 3D-to-BEV encoding strategy with the YOLOv8s OBB (You Only Look Once - Oriented Bounding Box) detector as a generic high-speed 2D detection meta-architecture. Analogous to the Bird-Net(+)/BEVDetNet encoding strategy, each LiDAR sweep is represented as a three-channel BEV image comprising maximum height, mean intensity, and normalized point density, enabling direct application of YOLO-v8s-Oriented Bounding Box (OBB) to LiDAR data. Training is performed from scratch on an augmented Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) training split, with evaluation following the KITTI BEV detection protocol. For the Car class, the proposed method achieves average precision values of 88.13% / 81.61% / 77.41% for the Easy / Moderate / Hard difficulty levels at IoU 0.7, respectively. Runtime evaluation demonstrates inference latencies of 2.8 ms in FP32, 1.7 ms in FP16, and 1.5 ms in INT8 on an NVIDIA RTX 3080, with compact model sizes supporting deployment-oriented LiDAR perception. We provided a quantitative comparison of detection and computational performance, using BEVDetNet as a reference method with limited public implementation details. The results demonstrate that BEV LiDAR representations can be effectively combined with modern 2D detection meta-architectures, providing a favorable trade-off between accuracy and latency, while maintaining simplicity and modularity in architecture for real-time automotive perception.» weiterlesen» einklappen

Autoren

Klassifikation

DDC Sachgruppe:
Informatik