Estimation of Urban Travel Time with Sparse Traffic Surveillance Data

The estimation of vehicle travel time in urban area is very important for route planning, commercial site selection, urban management and other issues. Most of the existing methods are based on vehicles' intensive GPS records, but due to privacy, business competition and other reasons, GPS data of all vehicles cannot be obtained, and the use of devices to upload GPS information every other period of time will cause unnecessary energy consumption. There are few methods to estimate the travel time of vehicles using the existing traffic surveillance system in urban areas. In this paper, we propose a new method to estimate vehicles' travel time based on sparse traffic surveillance records at camera-equipped intersections. In order to solve the problem that we only know the total time difference between the two cameras, while the travel time allocated on each road segments and the driving path cannot be obtained, we take Expectation-Maximization (EM) algorithm as the basic framework to update the last two variables mentioned above iteratively. In order to overcome the problem that iterative updating cannot be carried out due to the sparsity of data, we construct a generative adversarial network based on spatiotemporal characteristics (ST-GAN), which uses long short term memory networks (LSTM) to extract the time dynamic characteristics of travel time in the road network, graph convolutional networks (GCN) to extract the spatial characteristics, and then uses semi-supervised learning to complete the data of the whole road network. Finally we test on the real data of Suzhou Industrial Park (SIP), and find that the estimated travel time of our model is better than that of multiple baselines in evaluation indicators.