Learning Traffic as Videos: A Spatio-Temporal VAE Approach for Traffic Data Imputation

Abstract

In the real world, data missing is inevitable in traffic data collection due to detector failures or signal interference. However, missing traffic data imputation is non-trivial since traffic data usually contains both temporal and spatial characteristics with inherent complex relations. In each time interval, the traffic measurements collected in all spatial regions can be regarded as an image with more or fewer channels. Therefore, the traffic raster data over time can be learned as videos. In this paper, we propose a novel unsupervised generative neural network for traffic raster data imputation called STVAE, which works well robustly even under different missing rates. The core idea of our model is to discover more complex spatio-temporal representations inside the traffic data under the architecture of variational autoencoder (VAE) with Sylvester normalizing flows (SNFs). After transforming the traffic raster data into multi-channel videos, a Detection-and-Calibration Block (DCB), which extends 3D gated convolution and multi-attention mechanism, is proposed to sense, extract and calibrate more flexible and accurate spatio-temporal dependencies of the original data. The experiments are employed on three real-world traffic flow datasets and demonstrate that our network STVAE achieves the lowest imputation errors and outperforms state-of-the-art traffic data imputation models.