Super-Resolution Vehicle Trajectory using Recurrent Time Series Imputation

Abstract: Vehicle data finds its use in a variety of applications in the fields of machine learning and data analysis. The volume of available data is limited by the frequency of data collection, and for several reasons, it can be infeasible to simply amplify this frequency. Furthermore, the increasing concern about data privacy and regulations like the General Data Protection Regulation (GDPR) has sparked a lot of interest in the domain of synthetic data generation. This thesis focuses on increasing the temporal resolution of trip data for vehicles by modifying an existing Recurrent Neural Network (RNN)-based architecture, namely Recurrent Imputation for Time Series (RITS), that treats the geospatial data as a time series. Various modifications to the existing architecture have been introduced in the pipeline to optimise the performance of the model in a spatio-temporal application. The samples of the geospatial dataset are collected from several vehicles periodically, however, due to the stochastic occurrence of various events, the dataset is essentially an unevenly spaced time series. In order to handle this nature of data collection, a novel adaptive downsampling algorithm has been introduced to prepare the data in a suitable way for the model. A baseline model that uses spatial midpoints and road information has been developed for comparison with the performance of the main modified-RITS model. The results show that the baseline model quantitatively performs better than the modified-RITS model, but the imputations of the latter are more feasible and realistic. The model partially understands the road network and can reasonably estimate the intermediate positions of the vehicle.