The paper presents a novel method for estimating the height of radar points associated with 3D objects in autonomous driving applications. The key highlights are:
Formulation of the radar height estimation problem as a sparse target regression task, where the goal is to predict a height map that aligns with the 2D bounding boxes of objects in the scene.
Introduction of a robust regression loss function, the Enhanced Huber Loss (EHL), to address the challenges of sparse target regression. The EHL incorporates a dynamic weighting factor to prioritize larger discrepancies and prevent the model from producing all-zero predictions.
Adoption of a multi-task training strategy, where the network jointly learns to estimate the height map and segment the free space in the scene. This approach helps mitigate the issue of the predicted height map reverting to all-zero values.
Extensive evaluation on the nuScenes dataset, demonstrating that the proposed learning-based height estimation method significantly outperforms the state-of-the-art Adaptive Height (AH) extension approach, reducing the average radar absolute height error from 1.69 to 0.25 meters.
Integration of the refined radar data, with the estimated height values, into existing radar-camera fusion models for object detection and depth estimation tasks. This leads to notable performance improvements, highlighting the crucial role of precise radar data in enhancing the overall perception capabilities.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Huawei Sun,H... at arxiv.org 04-10-2024
https://arxiv.org/pdf/2404.06165.pdfDeeper Inquiries