Documentation Status

Cannot load information on name: ipa_room_segmentation, distro: electric, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: fuerte, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: groovy, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: hydro, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: indigo, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: jade, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: kinetic, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.
Cannot load information on name: ipa_room_segmentation, distro: lunar, which means that it is not yet in our index. Please see this page for information on how to submit your repository to our index.

Code

The source code for segmenting floor plans and the dataset of 20 floor maps with and without furniture can be found here: https://github.com/ipa-rmb/autopnp/tree/indigo_dev/ipa_room_segmentation

The floor maps lie in .png format under common/files/test_maps (https://github.com/ipa-rmb/autopnp/tree/indigo_dev/ipa_room_segmentation/common/files/test_maps). The relevant are the ones which are existing in several versions, shown by following name extensions:

  • _furnitures : The original map with furnitures drawn in
  • _trashbins : The original map with trashbins as blue pixels drawn in

And the combinations of these extensions, resulting in 4 versions for each map.

Additionally there are ground-truth maps, that show how a human might intersect these maps, shown by the name extension _gt. If you want to colour these ground-truth maps, to show the different segments, you can use a flood-fill algorithm for the white pixels to collect each pixel that belong to one segment, because the drawn lines seperates the white spaces from each other. Then you can colour each of this found pixels with a color that is unique for this segment. This is done in the computePrecisionRecall() function of the EvaluationSegmentation class (common/src/evaluation_segmentation.cpp) if you want it.

General Procedure

1. Change the algorithm parameters in ros/launch/room_segmentation_action_server_params.yaml in ros/launch to the wanted algorithms and settings.

  • room_segmentation_algorithm: Choose which segmentation algorithm should be used.
  • display_segmented_map: Specify if you want a window to pop up after the given map has been segmented, showing the results. To continue the program press any button in the window.
  • train_semantic, train_vrf: Booleans that show, if the feature-based or voronoi random field segmentation should be trained. For details of this training method see the further document.
  • Every algorithm has its own specific range of allowed room areas. Segments that are out of this range will not be drawn and are going to be filled by the surrounding valid rooms. Remark: The upper border in the feature-based segmentation method has a little addition. If found segments are too big they become separated by randomly putting centers in the segment and splitting it by using a wavefront algorithm. It's better to turn this off by setting the allowed maximal area very high.

  • Some algorithms have specific parameters that change some functionalities of it, which is described in detail in the .yaml file.

2. Start the action server using the launch file /ros/launch/room_segmentation_action_server.launch, which executes the /ros/src/room_segmentation_server.cpp file. If the server crashes for some reason (e.g. no valid map given by the client) it respawns with a little delay.

3. Start an action client, which sends a goal to the segmentation action server, corresponding to the MapSegmentation.action message, which lies in ipa_building_msgs/action. The goal consists of the following parts

  • input_map: The map that should be segmented, as sensor_msgs/Image. Has to be a 8-Bit single channel image, with 0 as occupied space and 255 as free space.

  • map_resolution: The resolution the map has been sampled with [meter/cell].
  • map_origin: The origin of the map in [meter] to bring the provided map coordinates in absolute coordinates, if wanted.
  • return_format_in_pixel: Boolean to indicate if the results should be returned in pixel-coordinates.
  • return_format_in_meter: Boolean to indicate if the results should be returned in absolute coordinates.
  • robot_radius: The radius of your robot. If this is set to a value greater than 0 the room centers are chosen at locally reachable areas.

4. The action server returns a result, that has the following parts

  • segmented_map: The map with the N found segments drawn in. The value of a pixel indicates its ID, which can be used to find pixels belonging to one room. The IDs start at 1 and go up to N, the return format is 32-Bit single channel, because it allows to have more rooms than the 8-Bit format.

  • RoomInformation: For each found room the min/max x/y coordinate of points belonging to this room and the center is provided. See ipa_building_msgs/msg/RoomInformation.msg for details. This can be in pixel coordinates or absolute coordinates, depending on the booleans defined above.

  • doorway_points: The fifth algorithm (voronoi random field segmentation) is capable of predicting doorways. If this algorithm is chosen, central points on a computed voronoi graph are returned that were labeled as doorways.

A client also can change the parameters of the server by using dynamic_reconfigure. This can be done using the dynamic_reconfigure_client.h file that provides a client for this purpose. With client.setConfig("param_name", value) one specific parameter can be changed from the client, see ros/src/room_segmentation_client.cpp for an example. Of course rqt_reconfigure can also be used to change specific parameters.

The algorithms are implemented in common/src, using the headers in common/include/ipa_room_segmentation. Each algorithm has its own class, so if you want to use it alone you have to include the header for the algorithm you want to use and make a new object. Then you have to execute the segmentMap() function, which segments the map.

For large maps the algorithms can take a few minutes to complete the segmentation, especially the feature-based and voronoi random field segmentation. If you want fast results, the morphological and distance segmentation are the fastest, but they might return not the best results if you have wide open spaces. To see results on different maps, see the chapter 6. Segmentation results.

Available algorithms

1. Morphological segmentation: Given the map as binary image, a binary erosion is iteratively applied. After each step, the algorithm searches for separate regions that have an area in the defined range. If it finds one, this region is saved and drawn black such that it doesn't need to be looked at any further. After all white regions are black, the algorithm stops and draws the segments in an image, then applies a wavefront algorithm to fill remaining white spaces.

2. Distance segmentation: Out of the given image a distance-map is computed that stores for each white pixel its distance to the nearest black pixel. On this distance-map a threshold operator is applied iteratively that makes every pixel above a certain threshold black and everything beneath it white. After each step the algorithm searches for segments in the given area range and stores them, then increases the threshold. After all possible thresholds have been looked at, the algorithm chooses the segmentation with the most valid segments as segmentation. After that it also draws the segments in an image and applies a wavefront algorithm.

3. Voronoi segmentation: For the given map a pruned voronoi graph is computed. The points belonging to this graph are then separated into different neighborhoods and searching in this neighborhood for the point with the smallest distance to the black pixels, called a critical point. This critical point is connected with the two closest black pixels, creating a separation of two segments.

4. Feature-based/semantic segmentation: For each white pixel a 360 degree laserscan is simulated, resulting in different distances to obstacles around this point. These laserscans are then used to compute different features, which are put into two AdaBoost algorithms. These compute the probability that the current point is a room or hallway. Depending on the probabilities a label is given to that point. The same procedure is then repeated for each white pixel in the given map, resulting in different separated regions for rooms/hallways that are treated as segments.

5. Voronoi Random Field Segmentation: First a pruned voronoi graph is computed for the given map, on that a neighborhood gets concentrated into one central point. After all points are found, neighboring relations are obtained, meaning which node is next to another depending on the voronoi graph. This gives several cliques, resulting in a conditional random field. The features for this CRF are then the weak responses of the AdaBoost classifiers for rooms, doorways and hallways. These AdaBoost classifiers use the same features as the feature-based segmentation. On this CRF then a loopy belief propagation is done, resulting in several intersected segments.

Training

In this package two algorithms are implemented that use machine-learning techniques, which of course need pre-labeled examples for training. These algorithms are the feature-based segmentation and the voronoi random field segmentation. The training procedure starts as soon, as you start the action server.

Important: The algorithms assume that the maps are given as 8-Bit single channel image, with 0 as occupied space and 255 as free space.

1. The feature-based segmentation uses two AdaBoost classifiers, one to detect hallways and one to detect rooms, based on simulated laserscannner-data. For the training of this algorithm hand-labelled data showing where rooms and where hallways are. To do so see the given files in common/files/training_maps. You can see that for different maps some regions are grey and not white or black. These grey regions show where a hallway/room is, the colour itself is not important. After you have created some training files yourself you have to specify where they lay on your file-system in the room_segmentation_action_server.launch file. When you set the train boolean to true these given maps are read in and given to the training routine.

2. The voronoi random field segmentation also uses AdaBoost algorithms to check for rooms/hallways/doorways. Additionally the weak classifier results from these algorithms are used as Features in a conditional random field, resulting in two different algorithms that need to be trained. Examples for training data for this algorithm can be seen in common/files/training_maps/voronoi_random_field_training. The data consist of the original maps, with only white and black regions and the labeled maps (training_maps) with colours showing the class of the current point. Optionally one can also provide additional data to save computation time, the precomputed maps with a voronoi graph drawn in (colour 127) and maps that show where nodes are in the given voronoi graph. If you provide one of it you also need to provide the other. After you created your personal training files you also have to specify the path to them in room_segmentation_action_server.launch. The colours provided in the training maps have the meaning

  • 77: room
  • 115: hallway
  • 179: doorway

References

If you use this work please cite our ICRA 2016 paper, article.pdf :

R. Bormann, F. Jordan, W. Li, J. Hampp, and M. H├Ągele. Room Segmentation: Survey, Implementation, and Analysis. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2016.

Segmentation results

without furniture

ground truth labeling

morphological segmentation

distance-based segmentation

Voronoi graph-based segmentation

feature-based room segmentation

voronoi random field segmentation

without_furniture.png

with furniture

with_furniture.png

Wiki: ipa_room_segmentation (last edited 2017-03-17 08:05:57 by RichardBormann)