A Clustering and Convex hull-based approach for Vehicle Bounding Box Localization in LiDAR Images for Advanced Driving Assistance System

Abstract

The accurate and efficient localization of vehicle bounding boxes in LiDAR images is of great importance in Advanced Driver Assistance Systems (ADAS) for improving road safety. In this paper, a novel and interpretable method for vehicle bounding box localization in LiDAR images using a combination of point cloud clustering and an efficient convex hull algorithm is proposed. The method starts with the application of a clustering algorithm to LiDAR point cloud images to obtain potential vehicle bounding box information. This involves grouping LiDAR points of individual vehicles, even in the presence of noise and partial occlusion. Next, an efficient convex hull algorithm is applied to obtain a minimal bounding polygon for each cluster of points, representing the geometric approximation of the vehicle. For better bounding box localization accuracy, additional refinement techniques are applied to the method. These techniques include statistical outlier removal, adaptive clustering parameters for varying vehicle sizes, and corrections for occluded objects. The effectiveness of the method was tested using the KITTI dataset, with the method achieving a mean absolute error of 0.167, root mean squared error of 0.186, and an average percentage error of around 8.5% for vehicle dimensions.The approach was also able to attain a Pearson correlation coefficient of 0.9995 with the ground truth data. Additionally, the 1D IoU scores were above 93% on average, which further confirms the excellent spatial overlap. The convex hull-based approach provides several benefits with its simplicity, robustness, and efficiency. It is also model-free and geometry-driven, which is quite beneficial for safety-critical ADAS applications.