Residual balanced attention network for real-time traffic scene semantic segmentation

Abstract

Intelligent transportation systems (ITS) are among the most focused research in this century. Actually, autonomous driving provides very advanced tasks in terms of road safety monitoring which include identifying dangers on the road and protecting pedestrians. In the last few years, deep learning (DL) approaches and especially convolutional neural networks (CNNs) have been extensively used to solve ITS problems such as traffic scene semantic segmentation and traffic signs classification. Semantic segmentation is an important task that has been addressed in computer vision (CV). Indeed, traffic scene semantic segmentation using CNNs requires high precision with few computational resources to perceive and segment the scene in real-time. However, we often find related work focusing only on one aspect, the precision, or the number of computational parameters. In this regard, we propose RBANet, a robust and lightweight CNN which uses a new proposed balanced attention module, and a new proposed residual module. Afterward, we have simulated our proposed RBANet using three loss functions to get the best combination using only 0.74M parameters. The RBANet has been evaluated on CamVid, the most used dataset in semantic segmentation, and it has performed well in terms of parameters’ requirements and precision compared to related work.