ROI-based LiDAR sampling in on-road environment for autonomous driving

Abstract

The acquisition of laser range measurements (i.e. Light Detection and Ranging sensor (LiDAR)) can be a time-consuming process if a high spatial resolution is required. Hence, designing an effective sampling algorithm is essential for many laser range applications. Previous approaches, such as two-step sampling, can be useful in general situations involving indoor and less complex scenes. However, they show a deficiency in the outdoor complex environment, especially in the condition of a very low sampling rate. To address this problem, this paper proposes an ROI-based sampling algorithm in the road environment, the typical environment for ADAS (advanced driver-assistance systems). Taking the merit of existing road and object detection algorithms, i.e. YOLO, the proposed method utilizes the semantic information and effectively distribute the sample budget to maximize the reconstruction quality, especially in the objects area. Experimental results show that the proposed method significantly reduces the mean-absolute-error (MAE) in the objects area and in the overall ROI by 44.9% and 15.1% compared to the two-step sampling. In addition, it achieves robust reconstruction quality with a very low sampling rate (i.e. 1% in experiments).