Detecting Road Lanes under Extreme Conditions: A Quantitative Performance Evaluation

Abstract

Vehicle autonomy definitionally is the act of processing information gathered from the environment and acting on the decisions formed based on this information. Therefore, any autonomous paradigm can only perform as good as the quality of the information it can understand. Lane identification forms the foundation of many of the autonomous drive and driver-assist technologies. However, current methods are not always reliable, especially under the edge-cases. In this paper, we have experimentally evaluated and extended the state-of-the-art deterministic lane detection methods. Our evaluation provides experimental evidence towards their efficacy in extreme cases: real-data with sharp shadows and varying lighting that is recorded through a camera that has a limited field of view. Experimental results suggest that a method that builds similarly to human perception performs better—with an increase of 32% in its accuracy. Our hypothesis is that autonomous vehicles that can perform even under these extreme conditions will play an important role on the fully autonomous systems.