Traffic Forecasting using Modified Unified Spatio-Temporal Graph Convolutional Network for Developing City: Dhaka, Bangladesh (A Case Study)

Abstract

Deep learning models for traffic forecasting gained a lot of success in recent years. Important application in traffic domain is to predict traffic congestion after certain time window based on historical data. While most of the deep learning models are evaluated using well-known traffic dataset containing vehicle speed collected using loop detectors, those model performances were not being tested on the generated traffic dataset from google maps containing traffic density information. We demonstrate the effectiveness of Unified Spatio-temporal Graph Convolutional Network in forecasting traffic congestion based on the traffic data of a developing countries like Bangladesh which is collected from google maps. We have quantified the traffic fluctuation pattern of any road of Dhaka dataset by introducing a single metric (coefficient of variation of traffic density fluctuation) which can explain the traffic congestion fluctuation pattern within a certain time window. We have also analyzed the whole traffic network of Dhaka using centrality measures (betweenness centrality) of Graph Theory. Based on the coefficient of variation of traffic density fluctuation and betweenness centrality of each road, we built clusters of roads. Based on those clusters, we proposed modification of USTGCN for generating better prediction. Finally, the prediction results are compared with the base USTGCN framework and we have explained the factor behind model performance degradation in terms of sparsity of the datasets with which the USTGCN models are trained on.