Autonomous Driving Path Planning Under Adverse Weather Based on Generative Adversarial Networks

Abstract

Due to decreased lane perception, obstacle localization, and decreased confidence in the drivable area brought on by rain, fog, snow, and poor light, autonomous driving path planning is significantly hampered in inclement weather. This research proposes a generative adversarial network-based path planning paradigm for autonomous driving in inclement weather to improve planning resilience. First, weather-degraded driving scenarios that preserve the obstacle structure and road geometry are created via conditional adversarial generation. A weather-aware risk representation is then created based on the lane-boundary confidence, obstacle likelihood, visibility attenuation, and road-surface adhesion risk. Lastly, feasible pathways under perceptual uncertainty are produced by a stable trajectory optimization technique. The suggested approach decreased the mean path deviation from 0.47m to 0.31m, raised the obstacle avoidance success rate from 84.6% to 94.1%, and decreased the planning failure rate by 18.6% in dense fog and heavy rain, according to tests conducted in a simulated and actual adverse weather driving environment. The tests demonstrate how adversarial weather generation might enhance driving safety by extending planning to invisible settings.