Citation

Zhao C, Song A, Du Y, Yang B. Transp. Res. C Emerg. Technol. 2022; 142: e103787.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.trc.2022.103787

PMID

unavailable

Abstract

With the increasing deployment of roadside sensors, vehicle trajectories can be collected for driving behavior analysis and vehicle-highway automation systems. However, due to dynamic occlusions, vehicles are often lost from the view of roadside sensors, strongly affecting the data availability. To address this issue, we propose a novel deep learning framework to impute missing Trajectory data called map-embedded Graph ATtention network (TrajGAT). The framework splits the problem into two subtasks, a trajectory forecasting task based on historical data and an imputation task based on the forecasting results and real-time incomplete observational data. Temporal features are extracted and fused following an encoder-decoder architecture. To model dynamic spatial patterns, we introduce a sparse heterogeneous graph construct technique via vectorized lane-level map and a rule-based graph attention network, which can effectively capture remote dependencies and highlight key adjacency relationships. Numerical experiments based on the Argoverse imputation dataset and Lyft dataset are conducted to compare our TrajGAT and other state-of-the-art models. The results indicate that our model performs best based on various evaluation indicators and has strong robustness with different missing trajectory rates. The learned dynamic interaction can further help achieve a better understanding of the spatiotemporal dependency of vehicles in complex traffic scenarios.

Language: en

Keywords

Attention mechanism; Graph convolutional network; Lane-level map; Missing data imputation; Vehicle trajectory