| Abstrakt: | Our faculty's students annually participate at the competition called RoboTour Outdoor Delivery Contest. The robot, which the faculty has at its disposal, had been designed, built and improved in previous works. One of the rules of this competition strictly forbids the robot from leaving driveable trail. Since the previous approaches have not been accurate enough, we focus on solving the problem of recognizing driveable trails using deep learning, whose application to various areas has achieved remarkable success in recent years. In this work we firstly deal with this problem by utilizing existing convolutional neural network models for semantic segmentation and test them right at the competition in the outdoor environment. These models reach very good results and the robot is able to distinguish between driveable and non-driveable segments more accurately, which contributes to a better navigation. In order to reach higher prediction speed, we minimize the size of these models and reach more than double speedup with prediction of the road. Subsequently, we compare our models with the ones specifically designed for devices with lower computational power. Finally, we examine the possibilities of reducing the number of images needed for training, leading to less effort dedicated to labeling. Our simulations show that by making use of image clustering combined with entropy of prediction it is possible to halve the number of training data at the cost of a very little decrease in accuracy.
|
|---|