Thesis of Raphaël Delecluse

Learning-Based Pedestrian Analysis for Transportation : From Person Re-Identification to Motion Prediction

Understanding pedestrian mobility is important for the planning and evaluation of transportation systems. Traditional surveys and counting systems generally provide limited or aggregated information, while computer vision can extract richer representations of individual movements. This thesis investigates three complementary levels of pedestrian analysis: identity continuity, trajectory prediction, and articulated motion modelling. First, a privacy-conscious person re-identification method is proposed for temporal sequences acquired by overhead RGB-D cameras. A Transformer encodes pedestrian passages, while Hungarian assignment optimizes associations between observations. The results show that depth sequences contain discriminative geometric and temporal information, reducing dependence on identifiable RGB appearance. This work is complemented by TVRID, a height-aware top-view RGB-D benchmark supporting RGB, depth-only, and cross-modal Re-ID. Second, a hierarchical Transformer is introduced for multi-pedestrian trajectory prediction. The architecture separates individual temporal encoding, multimodal fusion of trajectory and pose, and scene-level reasoning about surrounding pedestrians. Experiments show that pose and social context provide complementary information for forecasting future displacement. Finally, a predictive latent dynamics framework is proposed for full-body human motion modelling. Learned pose representations condition a non-autoregressive spatio-temporal Transformer for motion prediction and sequence completion. The results show that the latent space preserves meaningful motion information, while also identifying autoregressive latent drift as the main limitation for long-term prediction. Together, these contributions progress from privacy-conscious identity continuity to predictive representations of pedestrian trajectories and articulated body motion, providing methodological foundations for richer pedestrian analysis in transportation environments.

Jury

M. Hazem WANNOUS Professeur IMT Nord Europe Directeur de thèse, M. Jonathan WEBER Professeur des universités Université de Haute Alsace Rapporteur, Mme Sylvie CHAMBON Professeure des universités ENSEEIHT – École Nationale Supérieure d’Électrotechnique, d’Électronique, d’Informatique, d’Hydraulique et des Télécommunications Rapporteure, M. Alain TREMEAU Professeur des universités University Jean Monnet, St-Etienne Examinateur, M. Laurent GRISONI Professeur des universités Université de Lille Co-directeur de thèse, M. Hamid LAGA Professeur Murdoch University Examinateur, M. Laurent GUIMAS Explain Invité.

Thesis of the team MINT defended on 06/10/2026