DeepMG: Learning to Generate Maps from Trajectories

In this study, we aim to generate a routable road network from trajectories using a deep learning approach.

Paper

If you find our code useful for your research, please cite our paper:

Sijie Ruan, Cheng Long, Jie Bao, Chunyang Li, Zisheng Yu, Ruiyuan Li, Yuxuan Liang, Tianfu He, Yu Zheng. "Learning to Generate Maps from Trajectories". AAAI 2020.

Geometry Translation

Feature Extraction

python feature_extraction.py ../data/conf_tdrive_sample.json

Training

python train.py --name t2rnet_tdrive_sample --dataroot ../data/tdrive_sample/learning --lam 0.2 --batch_size 8 --model t2rnet --display_id -1

Inference

python test.py --name t2rnet_tdrive_sample --dataroot ../data/tdrive_sample/learning --lam 0.2 --model t2rnet

Region Concatenation

python region_concatenation.py --tile_path ./results/t2rnet_tdrive_sample/test_latest/images/ --conf_path ../data/conf_tdrive_sample.json --results_path ../data/tdrive_sample/results/

Topology Construction

Graph Extraction

python main.py --phase 1 --conf_path ../data/conf_tdrive_sample.json --results_path ../data/tdrive_sample/results/

Link Generation

python main.py --phase 2 --conf_path ../data/conf_tdrive_sample.json --results_path ../data/tdrive_sample/results/

Map Refinement

Map Matching

python main.py --phase 3 --conf_path ../data/conf_tdrive_sample.json --results_path ../data/tdrive_sample/results/

Edge Pruning

python main.py --phase 4 --conf_path ../data/conf_tdrive_sample.json --results_path ../data/tdrive_sample/results/

Requirements

DeepMG uses the following dependencies with Python 3.6

• gdal==2.3.2
• opencv==3.3.1
• rtree==0.8.3
• networkx==2.3
• scikit-image==0.16.2
• pytorch==1.1.0
• torchvision==0.3.0

Other packages can be easily installed using conda install, while the following scripts are recommended for gdal, opencv and pytorch.

conda install -c conda-forge gdal==2.3.2

conda install -c menpo opencv==3.3.1

conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch -c defaults -c numba/label/dev

Note that gdal must be installed first, and a restart might be required after all installation.

Datasets

• Trajectory Data

  • There are some open-source trajectory datasets, e.g., TDrive.
  • Data should be organized as text files. A text file can contain several trajectories (refer to sample.txt.template)

• Road Network Data

  • OpenStreetMap (OSM) is an open-source map data source
  • The Shapefile road network format can be generated by osm2rn.