A New License Plate De-Identification Method for Indoor Parking Lot Datasets

Abstract

This paper addresses de-identification problem of vehicle license plates in image-based datasets crucial for autonomous driving and robotics research. Traditional methods such as blurring and masking used for privacy protection reduce data quality by obscuring key features like license plates, limiting the utility of datasets. To overcome this, this paper introduces a new method that retains original data quality while anonymizing license plates by combining detection, vertices identification, and generation of virtual license plates resembling textures of the original. The approach utilizes advanced models such as YOLOv8x and Swin Transformer, enhanced by the Weighted Box Fusion ensemble technique, to improve the detection accuracy of license plates in complex indoor parking lot environments characterized by poor lighting and varied vehicle positions. This method enhances the accuracy of detecting multiple license plates in a single image, isolating images with only one vehicle for further processing. The RetinaFace model identifies the precise positions (or 4 vertices) of license plates and the perspective transformation technique replaces the original plates with virtual ones using the identified vertices. These virtual plates are randomly generated but initially do not match the original plates' texture, making them obviously artificial. To address this, a finely adjusted CycleGAN model adapts the texture of the virtual plates to closely resemble the originals. Finally, the modified plates are merged back with the original vehicle images, ensuring that the dataset retains the visual similarity to the original images while effectively making de-identification of the license plates for privacy. The effectiveness of the proposed de-identification method is validated by comparing its performance with various deep learning models in an indoor parking environment.