Modeling Trajectory with Recurrent Neural Networks

The source code of my paper.

Hao Wu, Ziyang Chen, Weiwei Sun, Baihua Zheng, Wei Wang, Modeling Trajectories with Recurrent Neural Networks. IJCAI 2017

Directory Structure

The directory structure may be as follows

workspace (e.g., /data)
    └ dataset_name1 (e.g., porto_6k)
        ├ data
        |   └ your_traj_file.txt
        ├ map
        |   ├ nodeOSM.txt
        |   └ edgeOSM.txt
        └ ckpt
            └ CSSRNN
                ├ dest_emb
                |   └ emb_50_hid_50_deep_1
                ├ dest_coord
                |   └ emb_200_hid_50_deep_1
                └ without_dest
                    └ emb_250_hid_350_deep_3

  
codespace
    ├ config
    ├ main.py
    ├ geo.py
    ├ trajmodel.py
    └ ngram_model.py
    

Format of the Data

Map/road network data

The map is recommended to be constructed from OpenStreetMap. You can get your own road network file from OpenStreetMap by selecting the rectangle area and export the file by, say, Overpass API (the first choice in the web page). Then you will be prompted to download a data named map which is actually an XML-formatted file. To parse the raw map data, you can use the tool here. After successfully parsed the raw map data, the nodeOSM.txt and edgeOSM.txt will be automatically extracted. The format of these two files are shown as follows.

nodeOSM.txt

Format: [NodeID]\t[latitude]\t[longitude]

One node/vertex per line with increasing (continuous) ids. E.g.,

0   41.1689665  -8.6444747
1   41.1658735  -8.6444774
2   41.1670798  -8.6424338
3   41.1673856  -8.642543
4   41.1669776  -8.6417132
5   41.1676312  -8.6424866
...
1000   41.1575375  -8.6443184

, which records 1,000 vertices in the road network.

edgeOSM.txt

Format: [EdgeId]\t[StartNodeId]\t[EndNodeId]\t[k]\t[lat1]\t[lon1]\t[lat2]\t[lon2]...\t[latk]\t[lonk]

One edge (StartNode -> EndNode) per line with increasing (continuous) ids. And k refers to the number of points of a polyline representing the shape of the road (including the start and the end node). I.e., (lat1, lon1) is just the coordinate of StartNode, and (latk, lonk) is the coordinate of EndNode.

E.g.,

0   0   1326    5   41.1689665  -8.6444747  41.1688112  -8.6443785  41.1685579  -8.6440804  41.1683059  -8.6438068  41.1680768  -8.6437482
1   4   5   2   41.1669776  -8.6417132  41.1676312  -8.6424866
...
1500    1499    1494    2   41.1849529  -8.6317477  41.185196   -8.6318118

, which records 1,500 edges in the road network. Edge 0 represents an edge from node 0 to node 1326 with 5 points representing the shape of the road as a polyline. And edge 1 represents an edge from node 4 to 5 with 2 points representing the shape, which means the shape of this road is a straight line segment (i.e., the first point(41.1669776, -8.6417132) is just the coordinate of node 4 and the second point (41.1676312, -8.6424866) is just the coordinate of node 5).

Trajectory data

The format of trajectory data is very simple. The data may like as follows,

1,2,4,6,8,12,7,23,
9,2,4,18,76,42,3,78,98,54,
432,214,678,532,3,5,74,13,123,67,4,
...

Each line records several edge ids in the road network, which represents a trajectory w.r.t. the definition introduced in the paper, i.e.,

Definition 2 (Trajectory). A trajectory T in the form of r_1 → r_2 → … → r_k captures the movement of an object from r_1 to r_2 and so on to r_k along the road network G, where every two consecutive road segments are connected, i.e., \forall r_i, r_{i+1} ∈ T, r_i, r_{i+1}∈ E ∧ r_i.e = r_{i+1}.s.

The trajectory (route) is generated by the HMM map matching algorithm Hidden Markov Map Matching Through Noise and Sparseness using the raw GPS trajectory data (sequence of coordinates, geographical coordinate should be transferred into rectangular coordiante). The kernel code is implemented in C++ (see HMM_mapmatching.cpp). There remain some function to be implemented by your own. For more detail, please refer to the comments in HMM_mapmatching.cpp.

Environment and Package Dependencies

The code can be successfully run under following environments

• Python: py2/py3 compatible
• Tensorflow version: 12.0 (You may have to change some code if you want to use higher version of Tensorflow since some APIs have been changed after v1.0)
• OS: Linux (I've tried this code on Windows and there may occur some strange runtime problems.)

[TODO] I'll modify some code to make it compatible with newest API of Tensorflow.

The project will also need following package dependencies

• Numpy
• Matplotlib

Usage

• Put the codes and the config file into the code space. Leave the trajectory data and the map data in the workspace following the directory structure as above.
• Modify config file
• Run main function in main.py

Details of configuration

All model settings are included in config file or can be set through the instance of Config class. To run the model, the following settings are important and should be set according to your own dataset.

• dataset_name: give a name to your own dataset and put all stuffs w.r.t. this dataset into the directory named by this name (as the directory tree structure in the above section).
• workspace: the place you want to put all data in, e.g. \home\data. Note that you may have several datasets, e.g., with the names being, dataset1, dataset2, ... . The directory may like as follows,

/home/data/dataset1/data
/home/data/dataset1/map
/home/data/dataset1/ckpt
 
/home/data/dataset2/data
/home/data/dataset2/map
/home/data/dataset2/ckpt
...

• file_name: set this as the name of your own trajectory file. If you follow the directory tree sturcture as above, you may set the file_name as data/trajectory.txt since the file is located in the data/ directory.

If you have correctly set these fields in the config, the model will be able to run with configuration printed in the screen.

To get the details of the remaining attributes of config, please refer to main.py. Each field is described in detail in comments.