Deep learning-based map matching considering road network constraints
In low-frequency or non-uniform sampling conditions, existing map matching algorithms have problems of low matching accuracy or low efficiency. In this paper, we propose a road network constrained map matching model based on deep learning ( RNCMM) . Firstly, Seq2Seq framework is used to map the low frequency track point sequence to the high frequency road segment sequence from end to end. Secondly, a fine-grained constraint mask layer is constructed according to the distance and azimuth difference between the road and the trajectory point, which is conducive to alleviating the limitations of the trajectory grid representation and improving the matching accuracy. Then, attention mechanism and multi-task learning mechanism are introduced to mine the spatiotemporal correlation between trajectory points and perform joint prediction of road segments and directions. Finally, experiments are conducted on the Porto taxi trajectory dataset and OSM road network. The results show that compared to traditional hidden Markov model ( HMM ) , the proposed algorithm can effectively improve the matching accuracy and efficiency of low-frequency floating car trajectories.
万方数据
维普
