EDRNet: Enhanced Dual-Resolution Network for Semantic Segmentation in Autonomous Driving

Semantic segmentation is a indispensable technology in autonomous driving, enabling vehicles to accurately perceive their environment and make real-time decisions. With the rapid advancements in the autonomous driving industry, real-time semantic segmentation has become a key area of research. This paper provides a comprehensive review of existing methods and introduces a novel network model for real-time semantic segmentation, termed Enhanced Dual-Resolution Network (EDRNet). The proposed model integrates several innovative components, including a Stem Block, Pyramid Pooling Module (PPM), and Feature Fusion Module (FFM). Additionally, to address the issue of class imbalance, Cross Entropy (CE) Loss with Online Hard Example Mining (OHEM) and inverse logarithmic class weights are applied during training. Extensive experiments conducted on benchmark datasets, namely Cityscapes and CamVid, validate the effectiveness and efficiency of EDRNet. The results demonstrate that EDRNet outperforms existing methods in terms of semantic segmentation accuracy while maintaining a satisfactory inference speed, positioning it as a promising solution for real-time applications in autonomous driving.