TFM: Autonomous driving in traffic using end-to-end deep learning

less than 1 minute read

Enrique Shinohara has successfully presented her final master’s project named “Conducción autónoma en tráfico usando aprendizaje profundo extremo a extremo”.

Information and Resources  
Work PDF
Slides PDF
Git Repository Link
Tutor JoseMaria Cañas Plaza and Sergio Paniego
In this project an end-to-end vision-based deep learning approach for an autonomous car to drive in traffic is presented. In this context the car should both keep the lane and also keep a safety distance with the front vehicle, if any. The imitation learning approach has been followed, creating a supervised dataset with the autopilot of CARLA simulator in different traffic conditions which include front vehicles. A variant of the PilotNet model has been developed, with additional dropout and the previous speed as input, and it has been (re)trained with such in-traffic dataset. The experimental results show that the model successfully drives the car in traffic conditions without losing performance in free road conditions. It also keeps safety distance with the front cars and even properly generalizes to several types of front vehicles.

A plugin for measuring the distance to the front vehicle has been created and added to Behavior Metrics, an open-source autonomous driving assessment tool.