A privacy-compliant approach to responsible dataset utilisation for vehicle re-identification

Abstract

Modern surveillance systems increasingly adopt artificial intelligence (AI) for their automated reasoning capacities. While AI can save manual labor and improve efficiency, addressing the ethical concerns of such technologies is often overlooked. One of these AI application technologies is vehicle re-identification - the process of identifying vehicles through multiple cameras. If vehicle re-identification is going to be used on and with humans, we need to ensure the ethical and trusted operations of these systems. Creating reliable re-identification models relies on large volumes of training datasets. This paper identifies, for the first time, limitations in a commonly used training dataset that impacts the research in vehicle re-identification. The limitations include noises due to writing on images and, most importantly, visible faces of drivers or passengers. There is an issue if facial recognition is indirectly performed by these black box models as a by-product. To this end, an approach using an image-to-image translation model to generate less noisy training data that can guarantee the privacy and anonymity of people for vehicle re-identification is proposed.

Graphical abstract

Keywords

Vehicle re-identification / Image-to-image translation / Responsible data generation / Artificial intelligence ethics