Scene Classification in Rural Environments for Autonomous Driving Using U-Net and ResNet

Resumo

This paper addresses the semantic segmentation of road images, a critical task for autonomous vehicle navigation, particularly in non-urban environments that present significant challenges. While much research focuses on well-maintained roads in developed countries, this study confronts the complexities of real-world conditions, such as those prevalent in developing nations, which feature vast networks of unpaved and poorly maintained roads. The core of our methodology is a neural network architecture that synergistically combines the encoder-decoder structure of U-Net with the feature extraction power of a ResNet backbone. The primary objective is the precise classification of each image pixel into one of four essential categories for navigation: background, asphalt, paved, and unpaved road. The model's training regimen involved exploring different ResNet versions (ResNet18, ResNet34, and ResNet50) as the encoder backbone to assess the impact of network depth. A key aspect of our approach was a progressive training strategy, where model versions were trained on images of varying resolutions. The results demonstrated a significant and somewhat counter-intuitive finding: training the ResNet34-U-Net model with images at half the original resolution yielded the best overall performance, achieving the highest Dice and IoU scores. This suggests that reducing image resolution acts as an effective form of regularization, compelling the model to learn more general and robust features by ignoring minor, irrelevant details. This outcome not only enhances the model's generalization capabilities for diverse and imperfect road conditions but also carries a substantial practical advantage by reducing the computational cost of training and inference, a crucial factor for deployment on resource-constrained embedded systems in autonomous vehicles.