The Impact of the Spatial Sampling Resolution on the Prediction of Vehicular Mobility

This work characterizes the influence of different spatial sampling resolutions on the prediction of vehicular mobility. By assuming different spatial quantization areas over a region the vehicles move through, we characterize the distribution of trajectories for a fixed number of samples using the data available in a dataset of GPS positions sampled in an urban scenario. Both spatial resolution and the number of GPS samples per trajectory are analyzed, concluding that similar distributions of the trajectories can be obtained when the unitary dimension of the spatial area is approximately 0.05 km2 and the trajectories last approximately 5 minutes. This is of particular importance to adequate the spatio-temporal sampling variables to the dynamics of the vehicular motion, thus avoiding over-sampling or sub-sampling in the spatial and temporal domains. The paper also proposes a deep-learning approach based on recurrent neural networks to predict future positions of a vehicular trajectory, showing the influence of spatial sampling to predict single and multiple future positions of the trajectory. The accuracy and the computation time of the prediction process are evaluated, showing how the magnitude of the prediction error is influenced by the adopted spatial sampling resolutions.