On Semantic Segmentation and Path Planning for Autonomous Vehicles within Off-Road Environments

Abstract

There are many challenges involved in creating a fully autonomous vehicle capable of safely navigating through off-road environments. In this work we focus on two of the most prominent such challenges, namely scene understanding and path planning. Scene understanding is a challenging computer vision task with recent advances in convolutional neural networks (CNN) achieving results that notably surpass prior traditional feature driven approaches. Here, we build on recent work in urban road-scene understanding, training a state of the art CNN architecture towards the task of classifying off-road scenes. We analyse the effects of transfer learning and training data set size on CNN performance, evaluating multiple configurations of the network at multiple points during the training cycle, investigating in depth how the training process is affected. We compare this CNN to a more traditional feature-driven approach with Support Vector Machine (SVM) classifier and demonstrate state-of-the-art results in this particularly challenging problem of off-road scene understanding. We then expand on this with the addition of multi-channel RGBD data, which we encode in multiple configurations for CNN input. We evaluate each of these configuration over our own off-road RGBD data set and compare performance to that of the network model trained using RGB data. Next, we investigate end-to-end navigation, whereby a machine learning algorithm optimises to predict the vehicle control inputs of a human driver. After evaluating such a technique in an off-road environment and identifying several limitations, we propose a new approach in which a CNN learns to predict vehicle path visually, combining a novel approach to automatic training data creation with state of the art CNN architecture to map a predicted route directly onto image pixels. We then evaluate this approach using our off-road data set, and demonstrate effectiveness surpassing existing end-to-end methods.