École Nationale Supérieure de
Techniques Avancées

Abstract

This project aimed at studying dierent supervised learning algorithms to perform road detection in images captured from a driving context. It can be divided in three main parts: in the first one, an ordinary convolutional neural network (CNN) was trained from scratch on the KITTI dataset to classify small patches of the image as road or not-road. Some variations to this base model architecture were proposed, mainly the addition of multi-scale patches and location features, and their eect on accuracy rate was evaluated. In the second part, these CNNs models were used to perform semantic segmentation through a method of sliding window. Finally, the algorithms eciency were studied, a minimalist and faster version introduced and all the models compared to state-of-the-art network in this field. This study revealed that patch-classification-based algorithms can achieve slightly worse, yet comparable accuracy rate when compared to what the state-of-the-art does. Besides, that algorithm is considerably lighter, which may make it a good choice in some applications for which memory and processing power are limited, such as some embedded systems.