Deep learning based perception for autonomous car multi mode navigation in road tunnels
TL;DR
Comparative experiments demonstrate that YOLOP uniquely achieves multi-task real-time performance on the Nvidia Jetson TX2 embedded processor, with projected inference at 65 FPS using TensorRT FP16 and confirm the viability of the proposed approach for safe autonomous navigation in challenging low-light tunnel scenarios.
Abstract
This paper presents a deep learning-based multi-mode navigation framework for autonomous vehicles (AVs) operating in road tunnel environments. The proposed system integrates a panoptic perception network (YOLOP) with multi-sensor fusion (camera, LiDAR, and depth camera) to simultaneously perform lane detection, drivable area segmentation, and obstacle detection within a unified forward pass. Two operational modes are introduced: tunnel navigation mode, which enforces lane keeping and emergency braking under low-light conditions, and routine navigation mode, which extends the system with LiDAR-guided obstacle overtaking on open roads. The framework is validated on the QCar platform, a 1:10 scale AV under four controlled lighting conditions representative of tunnel environments (bright, soft glow, moderate, and dim). Comparative experiments against YOLOv11, YOLOv12, Single Shot Detector MobileNetV2, RT-DETR, and classical Hough Lines demonstrate that YOLOP uniquely achieves multi-task real-time performance on the Nvidia Jetson TX2 embedded processor, with projected inference at 65 FPS using TensorRT FP16. The results confirm the viability of the proposed approach for safe autonomous navigation in challenging low-light tunnel scenarios.