DeepLabv3Att: Integrating Attention Mechanisms in DeepLabv3 for Enhanced Road Segmentation
Md Sabbir Hossain, Mostafijur Rahman, Mumtahina Ahmed, Md. Mohsin Kabir, M. F. Mridha, Jungpil Shin
Abstract
Road segmentation is a critical task in autonomous driving technology and advanced driver-assistance systems. This article introduces a model, DeepLabv3+Att, which enhances the standard DeepLabv3 by integrating attention mechanisms to improve segmentation accuracy. Our proposed model focuses on salient regions of the input images, thereby effectively distinguishing between road and non-road areas. We evaluate DeepLabv3+Att on the challenging KITTI Road dataset, where it achieves a mean Intersection over Union score of 97.72% and a Maximum F-score of 99.45%. The attention mechanisms incorporated in our model enable more precise segmentation by emphasizing relevant features, leading to superior performance. These results demonstrate the potential of DeepLabv3+Att in enhancing road segmentation tasks and its applicability in real-world driving scenarios.