1. Research Overview and Scientific Goals The primary objective of this project is to develop a robust generative framework for pedestrian trajectory reconstruction in constrained indoor environments. While outdoor GPS-based trajectory imputation has matured, indoor settings present unique challenges due to dense geometric constraints (e.g., walls, corridors). Standard probabilistic models often treat trajectories as unconstrained coordinate sequences, resulting in physically infeasible paths that "pass through" obstacles. This project proposes a map-conditioned generative framework based on conditional diffusion models to explicitly integrate environmental geometry into the trajectory synthesis process, ensuring high physical feasibility and low collision rates in applications such as security surveillance and indoor navigation. 2. Technical Methodology Our approach leverages the iterative denoising capabilities of diffusion models, enhanced by two critical components: (1) Map-Conditioning via CNNs: We integrate dense scenario maps as structured spatial inputs using a CNN-based encoder. This allows the model to "perceive" the environment and guide the denoising process toward feasible spatial paths. (2) Endpoint-Aware Mechanism: Observations (start and end points) are embedded directly into the scenario map, providing spatial anchors through image-level features to refine the reconstruction accuracy. 3. Expected Impact This research will provide a state-of-the-art solution for robust trajectory reconstruction in complex spaces. By reducing collision and invalid path rates, the framework enhances the reliability of digital twin simulations and surveillance analytics, contributing to the advancement of AI and Machine Learning research within the Swedish academic infrastructure. 4. Supervisors This research is supervised by Xiaodan Shi (Stockholm University) and Oskar Juhlin (Stockholm University).