Creation of Synthetic Datasets using Generative AI for Implementing Vehicle Access Systems based on Face and Gait Recognition

The increasing reliance on keyless access systems in modern vehicles, such as proximity sensors, wireless fobs, and smartphone-based digital keys, has raised serious security and usability concerns in the automotive industry. These systems, while convenient, are vulnerable to attacks, device loss, and user inaccessibility. In response to these limitations, the project developed by Stellantis and Centro Ricerche Fiat (CRF) with the Dipartimento di Automatica e Informatica (DAUIN) of Politecnico di Torino aims to enable secure and contactless vehicle access through biometric identification, using a combination of face and gait recognition in real-world conditions. This scenario, described as “recognition in the wild”, poses several challenges, as biometric models must cope with variations in lighting, pose, background, and occlusions. One fundamental set-back within this area of research is the lack of large-scale, diverse, and realistic datasets suitable for training and evaluating biometric systems in vehicle access contexts. This thesis work addresses that limitation by exploring the use of generative AI to synthesize realistic image and video data for training biometric machine learning models. The work details the design a pipeline for synthetic data generation using ComfyUI, a modular interface for configuring generative model workflows. The image generation phase leverages transformer-based multimodal models such as OpenAI's ChatGPT and Sora to create high-quality full-body portraits and apply identity-preserving edits. These static images are then transformed into gait videos using state-of-the-art latent video diffusion models, including WaN 2.1 and its domain-adapted VACE variants, accompanied by ControlNet modules used to guide motion with pose maps and to ensure spatio-temporal consistency. Preserving motion realism is a key aspect for the synthetic videos, which must emulate recordings from ADAS (Advanced Driver-Assistance Systems) cameras mounted on the vehicle. The generated subjects are required to maintain facial and structural coherence while displaying realistic gait patterns in environmental conditions consistent with real-world scenarios. To this end, video generation models are considered based on their ability to meet biometric constraints such as character consistency, viewpoint accuracy, and smoothness. In parallel with synthetic data creation, a relatively small dataset of real gait videos was collected directly in the laboratory under controlled conditions. These real samples serve as the starting point for creating augmented and modified versions using generative tools. Ultimately, the resulting synthetic dataset was incorporated into the multimodal recognition pipeline developed in previous stages of the project. The goal is to procure training diversity and improve recognition robustness under unconstrained access scenarios, while reducing reliance on personal data collection. The results demonstrated the feasibility of enhancing biometric recognition pipelines through synthetic data generation, offering a privacy-conscious and scalable solution for future applications in secure, keyless vehicle access systems.