Qi Weiqing ; Zhao Guoyang ; Ma Fulong ; Zheng Linwei ; Liu Ming

Abstract

Road lanes are integral components of the visual perception systems inintelligent vehicles, playing a pivotal role in safe navigation. In lanedetection tasks, balancing accuracy with real-time performance is essential,yet existing methods often sacrifice one for the other. To address thistrade-off, we introduce CLRKDNet, a streamlined model that balances detectionaccuracy with real-time performance. The state-of-the-art model CLRNet hasdemonstrated exceptional performance across various datasets, yet itscomputational overhead is substantial due to its Feature Pyramid Network (FPN)and muti-layer detection head architecture. Our method simplifies both the FPNstructure and detection heads, redesigning them to incorporate a novelteacher-student distillation process alongside a newly introduced series ofdistillation losses. This combination reduces inference time by up to 60% whilemaintaining detection accuracy comparable to CLRNet. This strategic balance ofaccuracy and speed makes CLRKDNet a viable solution for real-time lanedetection tasks in autonomous driving applications.