Automated Radar Data Labeling through Computer Vision

Publikationstyp
Conference Paper
Date Issued
2023-04
Sprache
English
Author(s)
Gabsteiger, Jasmin  
Maiwald, Timo  
Wünsche, Simon
Weigel, Robert  
Lurz, Fabian  
Hochfrequenztechnik E-3  
TORE-DOI
10.15480/882.8024
TORE-URI
https://hdl.handle.net/11420/42377
Citation
IEEE Wireless and Microwave Technology Conference (WAMICON 2023)
Contribution to Conference
2023 IEEE Wireless and Microwave Technology Conference, WAMICON 2023  
Publisher DOI
10.1109/WAMICON57636.2023.10124886
Scopus ID
2-s2.0-85160946360
ISBN
9798350398649
This paper presents the combination of an existing computer vision model with a highly integrated mm-wave radar system for automated radar data labeling. A pre trained model for person presence detection is used in combination with a camera sensor to determine, whether a person is present or not. The binary prediction is used to label simultaneously measured radar data of the same scenery. The paper addresses the challenges of varying illumination conditions, camera and radar sensor aperture angles and data quality which are dominant influences at the labeling process. The system significantly decreases human effort by automating the labeling process. It is used to generate 10,000 data points, label them and train a neural network with 95.7% accuracy. Finally, a proof-of-concept, training, and evaluation of radar-based activity detection with automatically labeled data is presented. The proposed method can contribute in person presence detection under challenging light conditions.
Subjects
computer vision
radar data labeling
MLE@TUHH