Related Experiment Video
Updated: Jan 14, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
732
A segmentation network for enhancing autonomous driving scene understanding using skip connection and adaptive
Jiayao Li1, Chak Fong Cheang2,3,4, Xiaoyuan Yu1
1School of Computer Science and Engineering, Faculty of Innovation Engineering, Macau University of Science and Technology, Taipa, Macao Special Administrative Region, China.
Scientific Reports
|October 21, 2025
Summary
This study introduces SUSC-SNet, a novel semantic segmentation network for autonomous driving. It improves road edge accuracy and object classification, enhancing scene understanding for safer self-driving vehicles.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Autonomous Systems
Background:
- Accurate semantic segmentation is crucial for autonomous driving perception.
- Current algorithms struggle with road edge contours, partial object misclassification, and small object segmentation.
Purpose of the Study:
- To propose a novel semantic segmentation network, SUSC-SNet, to address existing challenges in autonomous driving scene understanding.
- To enhance the accuracy and efficiency of identifying elements like roads, sidewalks, and vegetation.
Main Methods:
- Developed SUSC-SNet, incorporating a Skip Connection Module (SCM) for feature fusion and Multi-Branch Fusion (MFM) and Dual Branch Fusion (DBFM) modules for adaptability.
- SCM utilizes dense skip connections for aggregated semantic extension.
- MFM and DBFM employ weighted fusion for increased flexibility.
Main Results:
- SUSC-SNet demonstrated improved performance on public autonomous driving datasets.
- Achieved increases in mean intersection over union (mIoU) of 0.67% and 0.9%.
- Increased mean class accuracy (mAcc) by 0.95% and 0.67% respectively, showcasing superior segmentation capabilities.
Conclusions:
- SUSC-SNet offers significant improvements in semantic segmentation for autonomous driving.
- The network exhibits efficiency, robustness, and broad applicability.
- The proposed modules enhance feature representation and fusion for better scene understanding.
