ABSTRACT
Research into autonomous vehicles has made great strides in the past two decades; some of the literature utilizes deep learning to train a model to steer a vehicle. However, autonomous steering algorithms have previously focused on roads with lane markings. This chapter outlines work on training a neural network to steer a vehicle on entirely unmarked roads. These models were trained on data collected by a simple webcam mounted on our research vehicle, a Polaris Gem 2 modified with a drive-by-wire system and sensor components. Training data – images of the roads – were collected by driving the vehicle along several routes with a webcam recording and saving frames labeled with the steering angle, which was used to train convolutional neural networks. Our preliminary real-world testing with the ACTor vehicle shows that a CNN using a fine-tuned prebuilt network is capable of navigating unmarked roads. After collecting more training data, we trained various fine-tuned pretrained networks as well as configurations including different recurrent structures using temporal contextualization. Pruning techniques were also experimented with to achieve a better balance of training data. Testing the models on unseen data demonstrated the ability to predict steering wheel angles to within approximately 0.2 rad on average.
