Real-Time Road Detection Implementation of UNet Architecture for Autonomous Driving

This paper presents a real-time implementation workflow of neural networks for autonomous driving tasks. The UNet structure is chosen for a road segmentation task, providing good performance for low complexity. The model is trained and validated using two datasets, KITTI (validation of the model with respect to state of art) and a local highway dataset (UHA dataset), collected by the laboratory research team. The performance of the model for road detection is evaluated using the F1 score metric. After a simulation validation on both sets, the model is integrated into a real vehicle through the RTMaps platform. The application is tested in real-time conditions, around the city, under various weather and light. Finally, the proposed model proves low complexity and good performance for real-time road detection tasks.