| dc.description.abstract | This research examines terrain roughness prediction for off-road autonomous vehicles as an image classification problem, labeling monocular images of upcoming drivable terrain with a measure of roughness derived from the z-axis acceleration readings taken by the vehicle's Inertial Measurement Unit. We derive eight potential roughness metrics for labeling images and select the two which were most effectively learned by our roughness classifiers: Label 6 (the standard deviation of a 1-second sampling of z-axis acceleration readings directly ahead of the image's timestamp, discretized using k-means clustering with k = 2), and Label 8 (the standard deviation of a 1-second sampling of z-axis acceleration readings directly ahead of the image's timestamp, discretized using k-means clustering with k = 4). We use deep learning, utilizing transfer learning with the ResNet50 neural network, to train models for learning these two labeling schemas. The model trained to predict Label 6 achieved 70.19% overall accuracy and 67.44% average accuracy by class, and the model trained to predict Label 8 achieved 52.92% overall accuracy and 39.93% average accuracy by class. Finally, we examine whether an attention region surrounding the upcoming drivable terrain can improve the ability of our classifiers to predict terrain roughness. In all cases, the models utilizing images with no attention region show increased performance compared to those utilizing images with an attention region, indicating that there is important contextual information in the non-path pixels of the images that assists the model in predicting upcoming terrain roughness. | |
