Spatiotemporal Traffic Forecasting as a Video Prediction Problem

Abstract

In this paper we propose an idea of applying video prediction methodologies for urban traffic forecasting. We state that the spatiotemporal structure of traffic data is similar to the structure of video streams, therefore developed video prediction models could be utilized for urban traffic forecasting after some modifications. Recent advances in video prediction led to development of model architectures that deal with large data in video streams and effectively extract high-level spatiotemporal dependencies. Similar problems were addressed in modern studies of urban traffic forecasting. We discuss analogies between these two application areas and discover key issues that should be solved for successful merging of methodologies. These issues are related to different spatial structures (regular grids in video streams and graph-based data in traffic flows) and different extends of spatial non-stationarity (spatial convolution rules are normally constant over the frame for video processing but have natural spatial peculiarities for traffic flows). The proposed idea is illustrated by a developed model that is based on the video prediction methodology and applied for a real-world urban traffic data. The developed model architecture includes custom graph-based rules for spatiotemporal feature learning and the support vector machine as a predictor. Obtained empirical results demonstrate that the proposed model outperforms other state-of-the-art spatiotemporal models (regularized vector autoregressive models, autoregressive integrated moving average model with exogenous spatial predictors). We consider these results as a successful proof of concept and conclude that application of complex video prediction architectures will be beneficial for spatiotemporal urban traffic forecasting.