Road Traffic Flow Prediction Using Aerial Imagery

Abstract

Technological advancements have increased the potential and feasibility of widespread drone networks. Among other tasks, monitoring road traffic flow is a task well-suited for such networks. While real-time traffic flow estimation systems have been explored at length and exist as commercial services, these systems have limited spatial reasoning and suffer in accuracy when predicting future traffic conditions. To that end, graph neural networks can account for spatial patterns, and can more effectively capture the impact of a region’s current traffic conditions on neighboring regions in the future. Our work builds on prior graph neural network architectures for traffic flow prediction. While current traffic prediction models are trained on ground-based data with limited features, we propose leveraging aerial traffic data to train spatiotemporal models with richer feature spaces. Our research makes contributions towards assembling a dataset from aerial footage and predicting traffic across a road network given aerial images from a small set of drones.