Cross-Modal Supervision Based Road Segmentation and Trajectory Prediction With Automotive Radar

Abstract

Automotive radar plays a crucial role in providing reliable environmental perception for autonomous driving, particularly in challenging conditions such as high speeds and bad weather. In this domain the deep learning-based method is one of the most promising approaches, but the presence of noisy signals and the complexity of data annotation limit its development. In this letter, we propose a novel approach to address road area segmentation and driving trajectory prediction tasks by introducing Differential Global Positioning System (DGPS) data to generate labels in a cross-modal supervised manner. Then our method employs a multi-task learning-based CNN trained by radar point clouds or occupancy grid maps without any manual modification. This multi-task network not only boosts processing efficiency but also enhances the performances in both tasks, compared with single-task counterparts. Experimental results on a real-world dataset demonstrate the effect of our implementation qualitatively, achieving decimeter-level predictions within a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\text{100} \,\text{m}$</tex-math></inline-formula> forward range. Our approach attains an impressive 91.4% mean Intersection over Union (mIoU) in road area segmentation and exhibits an overall average curve deviation of less than <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\text{0.35} \,\text{{m}}$</tex-math></inline-formula> within a range of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\text{100} \,\text{m}$</tex-math></inline-formula> forward in trajectory prediction.