Camera and lidar sensor fusion for projection of 2d detections into 3d coordinate space

Engineering notes

Experimental evaluation on a real-world dataset collected from an autonomous vehicle platform demonstrates superior mean IoU compared to classical LiDAR-only geometric methods and a purely vision-based approach. The method achieves real-time performance on embedded automotive hardware, confirming its suitability for resource-constrained autonomous systems.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Accurate estimation of surrounding vehicle pose is a critical component of perception systems in highly automated vehicles (HAVs). This paper presents a computationally efficient mid-level sensor fusion algorithm for estimating oriented 3D bounding boxes of vehicles in the local coordinate frame using monocular 2D detections and ground-filtered LiDAR point clouds. The proposed method contributes to the field of simulation-driven development of autonomous systems by providing a geometry-based perception module that can be directly integrated into virtual testing environments and digital twins of highly automated vehicles. The proposed approach combines density-based clustering, projection-based camera–LiDAR association via the Hungarian algorithm, and RANSAC-based geometric line fitting to construct minimum-area oriented bounding boxes in bird’s-eye view. In contrast to deep learning–based fusion approaches, the proposed method does not require model training, large-scale annotated datasets, or high-performance GPU hardware, which simplifies deployment and improves adaptability to non-standard sensor configurations. Experimental evaluation on a real-world dataset collected from an autonomous vehicle platform demonstrates superior mean IoU compared to classical LiDAR-only geometric methods and a purely vision-based approach. The method achieves real-time performance on embedded automotive hardware, confirming its suitability for resource-constrained autonomous systems.

Links and sources

• PDF from original source

Comments