Implementation of Joint Segmentation Semantic Understanding Method to Develop Autonomous Vehicle Driving Perception

Engineering notes

This technology holds the potential to significantly enhance safety, operational efficiency and promote inclusive mobility.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The development of autonomous vehicle technology has become a central focus in the global transformation industry. This technology holds the potential to significantly enhance safety, operational efficiency and promote inclusive mobility. A primary challenge in this domain lies in designing visual perception systems that can rapidly and accurately interpret environmental conditions in highly dynamic settings. Conventional sensor-based systems relying on LiDAR or radar are often cost-prohibitive and computationally intensive for wide-scale deployment. Therefore, leveraging monocular camera input with deep learning-based joint semantic segmentation offers a cost-effective and high-performance solution. This study aims to develop and implement a joint segmentation model based on multi-task learning (MTL) using the YOLOP architecture and its modified version of YOLOP(new) to address three critical perception tasks: object detection, drivable area segmentation, and lane line segmentation. By combining the CSPDarkNet backbone with GhostNet and integrating the SIoU loss function and SiLU activation, this research seeks to enhance computational efficiency without sacrificing accuracy. YOLOP(new) also improved lane line segmentation, though with a slight decrease in drivable area segmentation performance. In conclusion, the modified YOLOP-based MTL architecture effectively improves both the performance and efficiency of autonomous vehicle perception systems on resource-constrained devices.

Links and sources

• Official / arXiv page

Comments