STDR: A Deep Learning Method for Travel Time Estimation

With the booming traffic developments, estimating the travel time for a trip on road network has become an important issue, which can be used for driving navigation, traffic monitoring, route planning, and ride sharing, etc. However, it is a challenging problem mainly due to the complicate spatial-temporal dependencies, external weather conditions, road types and so on. Most traditional approaches mainly fall into the sub-segments or sub-paths categories, in other words, divide a path into a sequence of segments or sub-paths and then sum up the sub-time, yet which don’t fit the real-world situations such as the continuously dynamical changing route or the waiting time at the intersections. To address these issues, in this paper, we propose an end to end Spatial Temporal Deep learning network with Road type named STDR to estimate the travel time based on historical trajectories and external factors. The model jointly leverages CNN and LSTM to capture the complex nonlinear spatial-temporal characteristics, more specifically, the convolutional layer extracts the spatial characteristics and the LSTM with attention mechanism extracts the time series characteristics. In addition, to better discover the influence of the road type, we introduce a road segmentation approach which is capable of dividing the trajectory based on the shape of trajectory. We conduct extensive verification experiments for different settings, and the results demonstrate the superiority of our method.