Automated Ground Truth Estimation of Vulnerable Road Users in Automotive Radar Data Using GNSS

Scheiner, Nicolas; Appenrodt, Nils; Dickmann, Jurgen; Sick, Bernhard · 2019 · Crossref

DOI: 10.1109/icmim.2019.8726801

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of annotating automotive radar data for machine learning applications in autonomous driving. Manual labeling of radar data is described as tedious, expensive, and requiring expert knowledge due to the abstract nature of radar representations compared to visual data. To mitigate this bottleneck, the authors propose an automated ground truth estimation system for Vulnerable Road Users (VRUs), such as pedestrians and cyclists, using Global Navigation Satellite System (GNSS) technology. The proposed method equips VRUs with portable GNSS modules, including receivers and Inertial Measurement Units (IMUs), while the ego-vehicle also carries a GNSS unit. The system utilizes Real-Time Kinematic (RTK) positioning to achieve centimeter-level accuracy. Although IMU data was tested, experiments revealed that the internal Kalman filter failed to handle the unsteady movements of pedestrians, causing trajectory drift; thus, the final system relies on pure GNSS positions smoothed by a moving average filter. An adaptive selection algorithm assigns radar labels by defining geometric regions around the VRU’s estimated position. For pedestrians, an ellipse is used, with dimensions adjusted based on velocity and yaw rate to account for swinging body parts. For cyclists, a rectangle is employed, adjusted for turning dynamics. Radar detections falling within these calculated regions are automatically labeled. The system was evaluated through experiments involving pedestrians and cyclists traversing an eight-shaped track, generating over 9,000 radar measurement cycles. The automated labels were compared against manual annotations by experts. The results demonstrated high accuracy, with a macro-averaged precision of 99.48% and a recall of 99.66%. Statistical analysis of the radar measurements revealed that wearing the GNSS equipment introduced minor biases. Specifically, there was a statistically significant increase in the width of the pedestrian’s radar signature and variations in the cyclist’s Doppler values. However, these changes were deemed negligible and comparable to the effects of wearing ordinary backpacks. The authors noted that manual labeling took approximately 18 minutes per scene, whereas the automated process saved significant time. The study concludes that the automated GNSS-based labeling method offers substantial time savings without compromising labeling accuracy, making it a viable alternative to manual annotation. The minor biases introduced by the equipment are outweighed by the efficiency gains. The authors recommend using this system to drastically enlarge training databases, potentially in conjunction with manual labeling to prevent model bias. Future work includes tracking multiple VRUs simultaneously and improving IMU-GNSS fusion algorithms.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• dms validation