How Front Fusion Improves High-Precision Map Obstacle Detection

Introduction

Element recognition and mapping are core components of high‑precision maps, supporting localization. Traditional image‑based semantic extraction suffers from pixel‑level errors; depth information from LiDAR or stereo cameras mitigates this by providing spatial coordinates.

Calibration and Data Acquisition

Accurate calibration using IMU‑GNSS‑Camera‑LiDAR aligns LiDAR points to images within ±5 px, enabling spatial error reduction. Mapping is an offline process involving point‑cloud registration, trajectory‑based 3D reconstruction (GPU‑multithreaded ICP), and storage of tiles identified by MGRS codes.

Static and Dynamic Obstacles

Reconstructed point clouds contain static obstacles (defining the map’s core assets), fused road surface data, and dynamic obstacle trails. Static obstacles are essential for absolute coordinate systems, SLAM, and real‑time recognition.

Detection Strategies

Three detection approaches are discussed: pure LiDAR, pure image‑based (using range images or depth maps), and front‑fusion of LiDAR, radar, and camera data. Front‑fusion combines spatial information in a single perception pipeline, leveraging image textures and point‑cloud features to improve accuracy.

Front‑Fusion Algorithm Framework

The front‑fusion pipeline extracts ROI and 3D OBB from spatial data, feeds them into an image network’s RPN layer via ROIAlign, and performs regression and classification. Two implementations are described: one using traditional ROI projection, another using image‑based techniques to handle uneven point distribution.

Semi‑Supervised Spatial Annotation

Point‑cloud labeling is challenging; a semi‑supervised method uses existing annotations to automatically complete missing depth via semantic and optical flow cues, enabling large‑scale training data generation.

Conclusion

A complete front‑fusion based recognition solution for static obstacles in high‑precision maps is presented, offering higher accuracy at the cost of reduced real‑time performance.