Authors:
Pathum Rathnayaka
;
Young Hun Kim
;
In Gu Choi
;
Gi Chang Kim
and
Duk Jung Kim
Affiliation:
U1 Geographic Information System, 282, Hagui-ro, Dongan-gu, Anyang-si, Gyeonggi-do, Republic of Korea
Keyword(s):
Computer Vision, Geo Information Systems, Image Processing, Lane Detection, Luminance, Lane Marking Quality, Contrast Ratio, Binary Ratio.
Abstract:
The detection of lane marking candidates is crucial for autonomous vehicles and advanced driver-assistance systems ADAS as they deliver essential information for accurate lane following, localization, and route planning. Detecting these candidates becomes difficult when road conditions deteriorate or the marking paint is faded, sometimes making identification nearly impossible. While deep/machine learning (DL/ML) methods perform well in reliable detection, they often come with the need for extensive labeled datasets, substantial computational power, and intricate parameter adjustments. In this research, we present a method purely based on digital image processing to identify and evaluate lane marking candidates, thus avoiding the use of specialized reflectivity equipment (such as luminance meters) and bypassing complex DL/ML methodologies. Our pipeline initially identifies lane marking regions using a collection of image-processing techniques. It then subsequently evaluates their qua
lity using two conditional metrics: a luminance-based contrast ratio and a white-to-black pixel ratio. Each candidate is categorized as good, bad, or ambiguous by comparing these metrics to empirically determined thresholds. Evaluations conducted on large sets of road images from conventional and urban highways in South Korea confirm the effectiveness of our proposed method. The system significantly reduces the dependence on time-consuming labor-intensive annotation, high-end hardware, and DL/ML expertise. We thus claim that our lightweight, deployable method effectively addresses a significant gap in luminance-centric lane candidate quality evaluation and can serve either as an independent solution or as a supplementary option to more sophisticated DL/ML systems in GIS and ADAS applications.
(More)