Vehicle trajectory map-matching method based on deep neural networks

Abstract

Map matching is very important for applications such as trajectory-based location services, route planning, and vehicle navigation. However, to preserve privacy, it has been proposed to sanitize the GPS data so that only coarse trajectory information is available. This increases the difficulty for map matching algorithms to achieve good accuracy. We use a deep neural network-based map matching method to improve the matching accuracy of complex road networks. And since the deep learning model requires a large amount of trajectory data with labels, this dissertation proposes a GPS trajectory generation algorithm based on road networks. Two kinds of trajectory sequences are generated at the same time which are GPS trajectory coordinate sequences and the corresponding road IDs where the trajectories are located. The generated trajectories are sampled and noise is added to simulate the features of real trajectories. We investigate and develop a TRANSFORMER-based trajectory map matching algorithm and evaluate to what extent an accurate trajectory can be reconstructed. Extensive experiments demonstrate that the TRANSFORMER-based trajectory map matching algorithm outperforms the classical Hidden Markov Map Matching model, not only in terms of higher matching accuracy and faster matching efficiency, but also in terms of robustness to noise and missing trajectory data. Keywords: Map matching, deep learning, trajectory generation, transformer.