Impact of data augmentation on labelling confidence in deep learning terrain traversability analysis for unmanned ground vehicles

Abstract

This study investigates the impact of data augmentation techniques on the accuracy and prediction probability of deep learning-based terrain classification systems for Unmanned Ground Vehicles (UGVs) in unstructured environments. The challenge of limited datasets in such environments is addressed through the implementation and evaluation of various data augmentation methods, to enhance the accuracy and reliability of pixel-level terrain measurements. The methodology is based on the DeepLabv3+ neural network architecture for supervised learning, trained on a custom dataset collected from an outdoor environment. A systematic assessment of multiple augmentation strategies is conducted, including geometric transformations (cropping and mirroring), colour space modifications (HSV transformations), and noise injection (Gaussian noise addition). The performance of these techniques is quantified using standard metrics, such as classification accuracy and Intersection over Union (IoU), alongside an analysis of pixel-wise classification prediction probability. Results indicate that, while traditional metrics show modest improvements, the application of data augmentation significantly enhances the model's prediction probability in its measurements, particularly for critical terrain features, such as traversable paths. A detailed analysis of the prediction probability distribution is presented, showing a significant improvement in the model's confidence for correctly classified pixels. Specifically, when augmentation strategies are applied, the percentage of traversable terrain pixels classified with high confidence (> 99.7 % probability) significantly increased from 75 % to 85 %.