Deep learning algorithm for autonomous driving using GoogLeNet

Abstract

In this paper, we consider the Direct Perception approach for autonomous driving. Previous efforts in this field focused more on feature extraction of the road markings and other vehicles in the scene rather than on the autonomous driving algorithm and its performance under realistic assumptions. Our main contribution in this paper is introducing a new, more robust, and more realistic Direct Perception framework and corresponding algorithm for autonomous driving. First, we compare the top 3 Convolutional Neural Networks (CNN) models in the feature extraction competitions and test their performance for autonomous driving. The experimental results showed that GoogLeNet performs the best in this application. Subsequently, we propose a deep learning based algorithm for autonomous driving, and we refer to our algorithm as GoogLenet for Autonomous Driving (GLAD). Unlike previous efforts, GLAD makes no unrealistic assumptions about the autonomous vehicle or its surroundings, and it uses only five affordance parameters to control the vehicle as compared to the 14 parameters used by prior efforts. Our simulation results show that the proposed GLAD algorithm outperforms previous Direct Perception algorithms both on empty roads and while driving with other surrounding vehicles.