Unravelling the Complexity of Indian Roads: Semantic Segmentation with LinkNet-UNet for Autonomous Vehicle Scene Understanding

Abstract

As autonomous vehicles navigate through complex environments, understanding the scene poses a significant challenge. Traditional computer vision-based methods struggle to segment complex driving scenarios, making deep learning techniques increasingly popular. Semantic segmentation is essential for scene understanding in India’s cluttered and diverse roads, where structure is often lacking. In this paper, we evaluate the performance of deep learning-based architectures including Linknet and Unet for semantic segmentation using the driving dataset of India’s variety and cluttered roads, IDD-Lite. Through a series of experiments, we examine the combined effect of Linknet and Unet on IDD-lite. By harnessing the localization prowess and contextual understanding inherent in both models, our unified approach raises the bar for scene recognition tasks. Our approach is designed to work on Indian roads and prioritizes precision, efficiency and adaptability to various environmental conditions. Experimental results show our ensemble model has a MIoU of 0.69 and F1 score of 0.9. This is better than conventional ensemble methods and a big jump forward for semantic segmentation in autonomous driving systems for Indian roads.