Convex Geometry-Driven Vehicle Localization in LiDAR for Advanced Driver Assistance Systems

Engineering notes

Key topics: autonomous driving, lidar, kitti, adas. See the paper for implementation details and experimental results.

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

Precise localization of vehicle bounding boxes in LiDAR images is an important part of Advanced Driver Assistance Systems (ADAS). This paper offers a explainable solution, which is a combination of point-cloud clustering and a convex hull algorithm. The approach initially uses a clustering procedure to isolate vehicle areas by grouping LiDAR points, and the points of the same vehicle can be combined with each other, despite noise or partial occlusions. Each cluster is then subject to a convex hull which produces a minimal bounding polygon that geometrically approximates the vehicle footprint. Outlier removal, size-adaptive clustering parameters, and occlusion correction are further used to enhance bounding-box accuracy. Evaluation on the KITTI data has been performed experimentally with a mean absolute error of 0.167, root mean squared error of 0.186 and average percentage error of 8.5% variance among vehicle dimensions. The approach has a Pearson correlation coefficient of 0.9995 with ground-truth annotations, which is high. Also, scores of 1-D Intersection over Union are above 93 on average, which is also a good sign of spatial alignment. The convex-hull-based framework is proposed to be straightforward, strong and efficient in terms of calculation. Although no training data is needed, the performance obtained is high and the results are better than some of the recent learning-based detectors, which has an important practical implication of the method. Sample LiDAR frames can be run on an NVIDIA Xavier platform in less than 30 ms, which can meet the real-time latency of ADAS with low computation costs.

Links and sources

• PDF from original source

Comments