A Real-Time Approach Based on Deep Learning for Ego-Lane Detection

Abstract

Lane detection has been one of the most important tasks of autonomous driving vehicles. The main objective of lane detection is tracking the lane boundaries on a real-time basis. Recently, many researches have shown deep learning models that are capable of detecting road lanes robustly. Yet, providing testing results in the context of real-time is not commonly found. This paper aims to provide a comprehensive real-time evaluation for performing ego-lane detection based on deep learning. The lane detection is recognized here as a semantic segmentation task where a pre-trained ResUNet++ model is adopted from a prior study. The real-time evaluation approaches include testing driving video sequences, CARLA simulator environment, and finally, a hardware kit that runs the deep learning model on the NVIDIA Jetson Nano developer kit. The performance of the model was investigated in various complex environments and dynamic scenarios. Eventually, the experimental results were qualitatively and quantitatively evaluated, showing reliability and promptness of using deep learning models for the lane detection task in a real-time context.