Deep Learning Applications for Autonomous Driving

Abstract: This thesis investigates the usefulness of deep learning methods for solving two important tasks in the field of driving automation: (i) Road detection, and (ii) driving path generation. Road detection was approached using two strategies: The first one considered a bird's-eye view of the driving scene obtained from LIDAR data, whereas the second carried out camera-LIDAR fusion in the camera perspective. In both cases, road detection was performed using fully convolutional neural networks (FCNs). These two approaches were evaluated on the KITTI road benchmark and achieved state-of-the-art performance, with MaxF scores of 94.07% and 96.03%, respectively. Driving path generation was accomplished with an FCN that integrated LIDAR top-views with GPS-IMU data and driving directions. This system was designed to simultaneously carry out perception and planning using as training data real driving sequences that were annotated automatically. By testing several combinations of input data, it was shown that the FCN having access to all the available sensors and the driving directions obtained the best overall accuracy with a MaxF score of 88.13%, about 4.7 percentage points better than the FCN that could use only LIDAR data.