Abstract
Large-scale and diverse datasets are essential for training AI models to make real-time decisions for autonomous vehicles (AVs) in intelligent transportation systems (ITS). Pedestrian intention and trajectory prediction models rely on diverse pedestrian images, but unrestricted access to these datasets poses serious security risks such as identity theft and pedestrian tracking. The challenge lies in applying privacy preservation procedures while retaining the necessary image attributes to train effective models. Existing privacy methods may protect pedestrian privacy but degrade image usability, hindering model performance.
This work focuses on implementing a five-stage pipeline that protects pedestrian privacy through face swapping while preserving essential facial attributes. It is specifically designed to meet the privacy needs of the Egy-DRiVeS dataset. Additionally, the Roop and Ghost-v2 face-swapping models are evaluated, with evidence showing that Roop outperforms Ghost-v2 in various aspects. Therefore, Roop is selected as the face-swapping model for the pipeline to balance pedestrian privacy via identity concealment and data usability through facial attribute preservation.
Blogger's Review: This study introduces a five-stage privacy protection pipeline that effectively addresses the conflict between pedestrian privacy and data usability, particularly in the realm of autonomous driving. The superiority of the Roop model offers new insights for future privacy protection technologies, and its real-world performance will be worth watching.