VRID-1: A basic vehicle re-identification dataset for similar vehicles

Abstract

Vehicle re-identification is a process of recognising a vehicle at different locations. It has attracted increasing amounts of attention due to the rapidly-increasing number of vehicles. Identification of two vehicles of the same model is even more difficult than the identification of identical twin humans. Further-more, there is no vehicle re-identification dataset that considers the interference caused by the presence of other vehicles of the same model. Therefore, to provide a fair comparison and facilitate future research into vehicle re-identification, this paper constructs a new dataset called the vehicle re-identification dataset-1 (1 VRID-1). VRID-1 contains 10,000 images captured in daytime of 1,000 individual vehicles of the ten most common vehicle models. For each vehicle model, there are 100 individual vehicles, and for each of these, there are ten images captured at different locations. The images in VRID-1 were captured by 326 surveillance cameras, and thus there are various vehicles poses and levels of illumination. Yet, it provides images of good enough quality for the evaluation of vehicle re-identification in a practical surveillance environment. In addition, according to the characteristics of vehicle morphology, this paper proposes a deep learning-based method to extract multi-dimensional robust features for vehicle re-identification using convolutional neural networks. Experimental results on the VRID-1 dataset demonstrate that it can deal with interference from vehicles of the same model, and is effective and practical for vehicle re-identification.