A Systematic Comparison of Deep Learning Architectures in an Autonomous Vehicle

Self-driving technology is advancing rapidly, largely due to recent developments in deep learning algorithms. To date, however, there has been no systematic comparison of how different deep learning architectures perform at such tasks, or an attempt to determine a correlation between classification performance and performance in an actual vehicle. Here, we introduce the first controlled comparison of seven contemporary deep-learning architectures in an end-to-end autonomous driving task. We use a simple and affordable platform consisting of of an off-the-shelf, remotely operated vehicle, a GPU equipped computer and an indoor foam-rubber racetrack. We compare a fully-connected network, a 2-layer CNN, AlexNet, VGG-16, Inception-V3, ResNet-26, and LSTM and report the number of laps they are able to successfully complete without crashing while traversing an indoor racetrack under identical testing conditions. Based on these tests, AlexNet completed the most laps without crashing out of all networks, and ResNet-26 is the most 'efficient' architecture examined, with respect to the number of laps completed relative to the number of parameters. We also observe whether spatial, color, or temporal features - or some combination - are more important for such tasks. Finally, we show that validation loss/accuracy is not sufficiently indicative of the model's performance even when employed in a real vehicle with a simple task, emphasizing the need for greater accessibility to research platforms within the self-driving community.