OSMDeepOD - OpenStreetMap (OSM) and Machine Learning (Deep Learning) based Object Detection from Aerial Imagery (Formerly also known as "OSM-Crosswalk-Detection").
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
docker updated Dockerfile Feb 27, 2017
src followstreet from string to bool Apr 7, 2017
tests followstreet from string to bool Apr 7, 2017
.dockerignore build and test improvements Apr 29, 2016
.gitattributes using versioneer to supply version info Jun 20, 2016
.gitignore added wms api Sep 7, 2016
.project restyled using autopep8 Apr 29, 2016
.pydevproject change methods to functions where appropriate Apr 29, 2016
.style.yapf extended converter to convert from geojson files too Jun 20, 2016
.travis.yml refactorings Nov 29, 2016
LICENSE enhanced setup.py to be more generic May 4, 2016
MANIFEST.in updated setup.py Dec 28, 2016
README.md Update README.md Aug 17, 2018
requires.dev.txt updated Dockerfile Feb 27, 2017
setup.cfg added overpass api Aug 3, 2016
setup.py Changed name in setup.py Aug 21, 2017


MIT License Codacy Badge Build Status Stories in Ready

OSMDeepOD - OSM and Deep Learning based Object Detection from Aerial Imagery

OSMDeepOD is a project about object detection from aerial imagery using open data from OpenStreetMap (OSM). The project uses the open source software library TensorFlow, with a retrained Inception V3 neuronal network.

This work started as part of a semester thesis autumn 2015 at Geometa Lab, University of Applied Sciences Rapperswil (HSR). See Twitter hashtag #OSMDeepOD for news.

Material and Publications





Getting Started

The simplest way to use the detection process is to clone the repository and build/start the docker container.

git clone https://github.com/geometalab/OSMDeepOD.git
cd OSMDeepOD/docker/
sudo docker build . -t osmdeepod
sudo docker run -it --name osmdeepod -v ./:/objects osmdeepod bash

After the previous shell commands you have started a standalone instance of OSMDeepOD and you are connected to it. If you have a nvida GPU and nvidia-docker installed, you could use the "nvidia-docker" command to run the container for automatically usage of the GPU1.

To start the detection process use the src/role/main.py2 script.

  1. Use the manger option to select the detection area and start the detection with the --standalone parameter.
python3 main.py --config ./config.ini --standalone manager 9.345101 47.090794 9.355947 47.097288

After the detection process has finished a "detected_nodes.json" file will appear with the results. If you like to use OSMDeepOD in a more parallel and distributed way have a look at the https://github.com/geometalab/OSMDeepOD-Visualize repository. There you have got the ability to use redis as a message queue and you can run many OSMDeepOD instances as workers.


The configuration works with an INI file. The file looks like the following:

Network = /path/to/the/trained/convnet
Labels = /path/to/the/label/file/of/the/convnet
Barrier = 0.99
Word = crosswalk
Key = highway
Value = crossing
ZoomLevel = 19
Compare = yes
Orthofoto = other
FollowStreets = yes
StepWidth = 0.66

Server =
Port = 40001
Password = crosswalks

BboxSize = 2000
Timeout = 5400

Some hints to the config file:

  • "Word" is the key value of the labels file
  • "Key" and "Value" builds the search Tag for OSM
  • "Compare" means compared to OSM tagged Nodes
  • "StepWidth" regulates the distance between the cut out images
  • The section REDIS should be self explanatory, this is not necessary in the standalone mode
  • "BboxSize" is the size in meters of the split large Bbox
  • "Timeout" after the expired time the job does fail

Own Orthofotos

To use your own Orthofotos you have to do the following steps:

  1. Add a new directory to src/data/orthofoto
  2. Add a new module to the directory with the name: <your_new_directory>_api.py
  3. Create a class in the module with the name: <Your_new_directory>Api (First letter needs to be uppercase)
  4. Implement the function def get_image(self, bbox): and returns a pillow image of the bbox
  5. After that you can use your api with the parameter --orthofots <your_new_directory>

If you have problems with the implementation have a look at the wms or other example.


During this work, we have collected our own dataset with swiss crosswalks and non-crosswalks. The pictures have a size of 50x50 pixels and are available by request.

Crosswalk Examples

Picture 3: Crosswalk Examples

No-Crosswalk Examples

Picture 4: No Crosswalk Examples


  • Python

    At the moment, we support python 3.5

  • Docker

    In order to use volumes, I recommend using docker >= 1.9.x

  • Bounding Box of area to analyze

    To start the extraction of crosswalks within a given area, the bounding box of this area is required as arguments for the manager. To get the bounding box the desired area, you can use https://www.openstreetmap.org/export to select the area and copy paste the corresponding coordinates. Use the values in the following order when used as positional arguments to manager: left bottom right top



  • 1: The crosswalk_detection container is based on the nvidia/cuda:7.5-cudnn4-devel-ubuntu14.04 image, may you have to change the base image for your GPU.
  • 2: For more information about the main.py use the -h option.


Big Data; Data Science; Data Engineering; Machine Learning; Artificial Intelligence; Neuronal Nets; Imagery; Volunteered Geographic Information; Crowdsourcing; Geographic Information Systems; Infrastructure; Parallel Programming.