• Citer

Radar Fusion and Transfer Learning for Environmentally Robust People Counting

par Cakoni, Dejvi
Président du jury Determe, Jean-François
Promoteur Horlin, François
Co-Promoteur De Doncker, Philippe
Publication Non publié, 2026-03-05

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

This doctoral dissertation addresses the critical challenge of accurate, privacy-preserving,and environmentally robust people counting for crowd safety and smart-city applications.It proposes a comprehensive radar-based framework that leverages both active Frequency-Modulated Continuous-Wave (FMCW) millimeter-wave radars and passive Wi-Fi-based radars,pushing the boundaries of single-modality, single-sensor, and single-viewpoint solutions towardscalable, multi-modal, multi-sensor systems.The thesis begins by establishing a complete simulation-to-reality pipeline using the Boulic–Thalmannkinematic model to generate synthetic micro-Doppler signatures, enabling rapid architectureexploration and transfer learning when real labeled data are scarce. It then introduces a highperformancesingle-radar counting paradigm based on feature-level fusion of micro-Dopplerspectrograms and range–time maps within a dual-branch convolutional neural network,achieving state-of-the-art performance on large-scale outdoor datasets collected over severalmonths with synchronized, privacy-compliant camera ground truth.Extensive real-world deployments with FMCW sensors demonstrate the limitations ofviewpoint-dependent training and motivate robust multi-radar solutions. Systematic experimentson cross-sensor generalization, fine-tuning, and joint multi-view training reveal thatsimultaneous exposure to diverse geometries yields highly viewpoint-invariant models.Furthermore, a significant contribution of this thesis lies in the first systematic investigationof active–passive radar complementarity and fusion. A dedicated joint measurement campaignsimultaneously illuminates the same dynamic pedestrian scene with a millimeter-wave FMCWradar and a Wi-Fi passive radar. Comparative analysis of range–Doppler representationsfollowed by feature-level and decision-level fusion shows that the hybrid system consistentlyoutperforms either modality alone, combining the high resolution radar signatures of activeillumination with the ubiquitous spatial coverage and low infrastructure cost of passive Wi-Fisensing.Through these contributions such as novel fusion architectures, large experimental datasets,reproducible sensing nodes, and end-to-end methodologies, this work establishes radar, in itsactive, passive, and fused forms, as a powerful, scalable, and ethically sound technology forreal-time crowd monitoring, paving the way for future city-wide, privacy-by-design people-flowmanagement systems.