ALLocalization - advanced
Custom localization algorithms ¶
- Advanced functionalities in ALLocalization are made available for those who want to use them in their own localization algorithms. If you are not one of them, please use standard functionalities instead.
- Those functionalities are not supported yet.
The first step to use the module is to set it up by shooting a reference panorama. The robot will perform a 360 degrees shooting by rotating its head and base to gather images. The panorama is then saved to the robot hard drive, and visual information is extracted. It is also possible to load a panorama from the robot hard drive, and use it for localization.
Once the setup phase has been done, ALLocalization gives the current robot orientation relatively to the setup origin. It also allows making the robot move to a given position in the panorama.
If the robot is equipped with a 3D camera, the reference panorama will include depth maps in addition to the images. It then becomes possible to use this module for localization purposes.
A reference panorama containing depth maps can be used to locate the robot near the origin of the panorama (within 2m). Please note that (0,0,0) corresponds to the location of the robot before the setup and that the robot will return to this location after the acquisition.
In this case, the location (x, y and theta, relatively to the setup origin) can be retrieved in two ways:
- an estimate location can be retrieved instantly, using the odometry and the last precise position known.
- a precise location can be retrieved by shooting half a panorama and comparing it to the reference. If the location is not known with enough confidence, the robot will make a half-turn and shoot the other half of the panorama. It will then return the location after the half turn.
How it works ¶
General principle ¶
ALLocalization uses information extracted from its camera images. Just like ALVisualCompass , it extracts keypoints from the panorama images, and builds a search index out of them.
When trying to estimate its position, the robot retrieves an image from its camera, then compares it to the stored panorama. Using the best match between the current image and the panorama, it can compute precisely its orientation relatively to the panorama.
The localization functionalities rely on pointcloud alignment to compute the position of the current depthmap relatively to the reference (which corresponds to the best alignment of the pointclouds).
Implementation details ¶
The panorama images must cover 360 degrees. To ensure the panorama robustness, all images overlap by half with the previous and the next. Therefore, each image which is taken later theoretically half overlaps with one of the panorama image.
In total, the panorama consists in 16 images, taken in two steps. Between the two steps, the robot turns around of 180 degrees using ALVisualCompass to ensure that the stitching is as good as possible.
In addition to those 16 images, 16 depth maps can be acquired from the 3D camera.
The visual features used here are multi-scale FAST keypoints with BRIEF descriptors, as in ALVisualCompass . They are chosen for their robustness and computational efficiency.
Pointclouds are aligned using the Iterative Closest Points (ICP) algorithm, which is a local search. We use the angle computed as explained above and the odometry as origin of the search. If getting a decent origin for the search proves to be impossible, it is also possible to find one automatically, by matching points with the help of the shape context descriptor.
Performances and limitations ¶
The performances of
depend mainly on the initial panorama
setup. Ideally, all images taken during the setup must be well textured and
easily distinguished. To detect potential problems, the robot performs a
diagnosis, by comparing all reference images to the others: the problematic
images are then raised in
The setup takes about 30 seconds to shoot the images, plus a few seconds to perform the diagnosis (which is only run once per panorama).
Regarding the 3D camera, the quantity of information retrieved will depend on the geometry of the environment. Concretely, only objects located in a range from 40cm to 5m of the robot will be usable as a reference.
Image textures ¶
The quality of the localization depends on the number of detected keypoints in the panorama, and thus on the textures. If the images are not sufficient, the robot compensates by using its odometry to avoid being lost, so the precision is guaranteed to be at least as good as the one from ALMotion .
The underlying algorithms allow the position estimation to be relatively robust to changes in the panorama.
The position estimation itself is optimized to be as quick as possible, by using hints from the robot odometry to narrow the search in the reference panorama. However, it will still work even if the odometry is no longer relevant, for example if the robot has been lifted and moved around.
The movements are performed using a closed loop based on the panorama images, similarly to ALVisualCompass . The robot first estimates its current position in the panorama, then uses the panorama frame corresponding to its target to perform a closed loop.
In practice, the robot first moves its head towards the target frame, then aligns its base in that direction, keeping the target frame in sight. If the target frame is no longer detected, the robot finishes its move using its odometry, thus guaranteeing that the final position will be at least close to the target one.
Like in the setup, only objects located in a range from 40cm to 5m will be usable for localization purposes. The robot looks up during acquisition, which allow it to mainly detect walls. It is therefore quite robust to minor changes in the environment.