PuSH - Publication Server of Helmholtz Zentrum München

Engel, L.* ; Bergmann, M.* ; Kammel, C.* ; Ullmann, I.* ; Eskofier, B.M. ; Vossiek, M.*

Bringing modern multi-object tracking to millimeter-wave radar target lists for indoor human tracking.

IEEE Access 14, 164508 - 164524 (2026)
Publ. Version/Full Text DOI
Open Access Gold
Creative Commons Lizenzvertrag
Accurate indoor multi-person tracking is a key enabling technology for smart-home automation and human–machine interaction. While modern deep multi-object tracking (MOT) has revolutionized computer vision through tracking-by-detection and tracking-by-attention paradigms, it remains unclear whether these concepts transfer to sparse millimeter-wave radar target lists, whose unordered and point-based measurements fundamentally differ from image data. This work closes this gap by introducing the first comprehensive study of modern deep MOT paradigms for radar target lists. Two end-to-end architectures are proposed, representing the tracking-by-detection and tracking-by-attention paradigms, and adapted specifically to object-centric radar representations. Furthermore, a large-scale benchmark comprising more than 405,000 annotated radar frames acquired with our custom RadarBox platform is presented and released together with high-fidelity ground-truth annotations generated using an Azure Kinect system and refined by a bidirectional Kalman filter. Extensive experiments compare both paradigms against optimized DBSCAN–Kalman tracking pipelines and representative deep-learning baselines while analyzing the role of temporal modeling for radar-based MOT. Both proposed architectures consistently outperform conventional and existing deep-learning approaches across localization, detection, and tracking metrics. The results demonstrate that learning object-centric spatial-temporal representations directly from radar target lists effectively overcomes the characteristic limitations of conventional density-based clustering, particularly under sparse detections, multipath propagation, and occlusions. The presented results further demonstrate that modern end-to-end multi-object tracking paradigms can be successfully transferred to millimeter-wave radar target lists, establishing learned tracking as a superior alternative to conventional clustering-based radar tracking pipelines. Code and dataset will be publicly released.
Altmetric
Additional Metrics?
[➜Log in]
Edit extra informations Login
Publication type Article: Journal article
Document type Scientific Article
Keywords Attention ; Dbscan ; Deep Learning ; Detection ; Indoor ; Kalman Filter ; Localization ; Radar ; Remote Health Care ; Remote Sensing ; Smart Home ; Tracking
ISSN (print) / ISBN 2169-3536
e-ISSN 2169-3536
Journal IEEE Access
Quellenangaben Volume: 14, Issue: , Pages: 164508 - 164524 Article Number: , Supplement: ,
Publisher IEEE
Reviewing status Peer reviewed
Institute(s) Institute of AI for Health (AIH)