People Counting Solution Using an FMCW Radar with Knowledge Distillation From Camera Data
Stephan M, Hazra S, Santra A, Weigel R, Fischer G (2021)
Publication Type: Conference contribution
Publication year: 2021
Publisher: IEEE
City/Town: NEW YORK
Conference Proceedings Title: 2021 IEEE SENSORS
DOI: 10.1109/SENSORS47087.2021.9639798
Abstract
Radar systems enable remote sensing of multiple persons within their field of view. In this paper, we propose a novel architecture to perform people counting using a 60 GHz Frequency Modulated Continuous Wave radar trained on supervised radar data and knowledge distillation performed using synchronized camera data. In the evaluation phase, only the radar encoder with Range - Doppler Images (RDI) as input is used and tested on a dataset consisting of scenarios recorded in a different setup than the training recordings with up to 6 persons present. In this paper we focus on showing the benefit of using the cross-modal camera information compared to the same unimodal model. In spite of the low-cost radar sensor, the proposed architecture achieves an accuracy of 71% compared to 58% for the test data from a different sensor with a different orientation and aspect angle, and an accuracy of 89% compared to 74% for test data from the same radar sensor when training without knowledge distillation.
Authors with CRIS profile
Michael Stephan
Lehrstuhl für Technische Elektronik (LTE)
Robert Weigel
Lehrstuhl für Technische Elektronik (LTE)
Georg Fischer
Professur für Technische Elektronik
Involved external institutions
Germany (DE)How to cite
APA:
Stephan, M., Hazra, S., Santra, A., Weigel, R., & Fischer, G. (2021). People Counting Solution Using an FMCW Radar with Knowledge Distillation From Camera Data. In 2021 IEEE SENSORS. NEW YORK: IEEE.
MLA:
Stephan, Michael, et al. "People Counting Solution Using an FMCW Radar with Knowledge Distillation From Camera Data." Proceedings of the 20th IEEE Sensors Conference NEW YORK: IEEE, 2021.
BibTeX: Download