UVDS: A New Dataset for Traffic Forecasting with Spatial-Temporal Correlation

This paper introduces UVDS, a traffic flow dataset from the vehicle detection system (VDS) in an urban area of South Korea. Specifically, with the rapid growth of computer vision for intelligent transportation systems, using detection systems for estimating traffic flow become an emergent issue. In this study, we first discuss the main differences between UVDS and existing datasets in terms of spatial-temporal dependencies for accurate traffic prediction. Then, preliminary work for construct a graph structure of the VDS data based on the geometric information is presented. The objective is to provide a benchmark dataset for exploring the capabilities of graph neural networks for traffic forecasting. Consequently, we present baseline results by adopting state-of-the-art models in this research field and discuss some future work for exploring the UVDS dataset.