LiDAR-Based Lane Width Detection Performance: A Semi-Autonomous Driving Field Test Study on Mountainous Freeways

Abstract

The complex driving environment of mountainous freeways brings potential safety risks to drivers and challenges the lane detection of autonomous vehicles (AVs). Although Light Detection and Ranging (LiDAR) is widely used in AVs for its superior perception, mountainous conditions may still affect LiDAR-based lane detection (LD) performance. Therefore, evaluating and improving the readiness of mountainous freeways for LiDAR-based LD becomes critical. However, few studies have examined contributing factors to LiDAR-based LD performance in these environments. Therefore, this study aims to investigate how various roadway and environmental factors affect LiDAR-based LD performance on mountainous freeways. First, a field experiment collected 147 km of LiDAR detection data on a mountainous freeway in Hunan Province, China. Then, a hybrid CatBoost method integrating particle swarm optimization and an Autoencoder was developed to analyze nine potential contributing factors. Finally, Shapley Additive exPlanations revealed four critical scenarios significantly increasing detection failure probability: (1) the lane marking types consist of acceleration lane on the right side, deceleration lane on the right side, and emergency parking lane on the right; (2) average horizontal curvature within 200 m downstream was 0.2–1.5 km; (3) test vehicle speeds between 80–100 km/h; and (4) complex interactions of multiple variables. These findings provide insights for mountainous freeway geometric design and operational management to enhance compatibility with high-level AVs.