Modeling Trajectory as Image: Convolutional Neural Networks for Multi-scale Taxi Trajectory Prediction
Year: 2016
Type: preprint
Abstract: Precise destination prediction of taxi trajectories can benefit efficient scheduling of taxies and accurate advertisement. Traditional prediction algorithms treat trajectories as sequences of spatial points and process them in one single predefined spatial scale. In this paper, we show that taxi trajectories can exhibit different patterns in different spatial scales, and the combination of multi-scale patterns can improve the accuracy of prediction. Based on this, we propose T-CONV to achieve higher accuracy, which models trajectories as images and adopts multi-layer convolutional neural networks to combine multi-scale trajectory patterns. Furthermore, in order to solve the sparsity problem of trajectories, we integrate multiple local convolutional fields in T-CONV to capture important specific areas of trajectories. Comprehensive experiments based on real trajectory data show that T-CONV can achieve much better results than state-of-the-art methods.
Source: arXiv (Cornell University)
Authors: Jianming Lv, Qing Li, Xintong Wang
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2603340102