This thesis describes the process of creating a dataset of labeled blurred traffic images, designed for the future training of traffic detection algorithms with built-in privacy. The blurring of images is intended to obscure the identity of individuals and other identifying data, thus ensuring privacy protection. The process includes the development of a traffic recording system using two Raspberry Pi cameras, which simultaneously recorded traffic - one providing blurred and the other sharp images. The images were then aligned using the SIFT algorithm. Clear images were labeled using the DETR algorithm, with the labels being the same for blurred images as we recorded with two cameras. The final result is a database of over a thousand blurred images with labels in COCO format. This dataset is prepared for use in further research and development of traffic detection algorithms with built-in privacy.
|
