Regression with Ensemble of RANSAC in Camera-LiDAR Fusion for Road Boundary Detection and Modeling

Abstract

This paper describes a technique to perform a post-detection fusion of camera and LiDAR data for road boundary estimation tasks. To be specific, the technique takes the road boundary detection results that are generated separately from the camera and LiDAR to enhance the accuracy of the estimated road boundaries. The proposed approach can achieve a more accurate estimation in the near range than just LiDAR-based detection and in the long range than just camera-based detection. Random sample consensus (RANSAC) of linear regressions is used to create the road boundary model that is capable of reducing errors and outliers while keeping it simple, explainable, and adaptive to the road curvature. The generated linear models are then combined into a single road boundary that can be interpolated and extrapolated using a Boosting-like algorithm with a non-parametric strategy. This technique is called as RANSAC-Ensemble. The experiments show that this technique has better accuracy with comparable processing time than certain other common methods of road boundary model estimation.