Efficient lightweight residual network for real-time road semantic segmentation

Abstract

Intelligent transportation system (ITS) is currently one of the most discussed topics in scientific research. Actually, ITS offers advanced monitoring systems that include vehicle counting, pedestrian detection. Lately, convolutional neural networks (CNNs) are extensively used in computer vision tasks, including segmentation, classification, and detection. In fact, image semantic segmentation is a critical issue in computer vision applications. For example, self-driving vehicles require high accuracy with lower parameter requirements to segment the road scene objects in real-time. However, most related work focus on one side, accuracy or parameter requirements, which make CNN models difficult to use in real-time applications. In order to resolve this issue, we propose the efficient lightweight residual network (ELRNet), a novel CNN model, which is an asymmetrical encoder-decoder architecture. Indeed, in this network, we compare four varieties of the proposed factorized block, and three loss functions to get the best combination. In addition, the proposed model is trained from scratch using only 0.61M parameters. All experiments are evaluated on the popular public the cambridge-driving labeled video database (CamVid) road scene dataset and reached results show that ELRNet can achieve better performance in terms of parameters requirements and precision compared to related work.