realtime_urdf_filter

This package provides OpenGL-accelerated, shader-based filtering for depth images. Scene geometry is defined in URDF models, along with parameters defining the camera's location and filtering parameters. Both static environment models and articulated robot models are working beautifully, as long as TF latency is not too high.

Incoming Kinect frames are transferred to the GPU as textures, and the scene is rendered from the same point of view. As a result, we can access the measured as well as the virtual depth map in the shader, where we can define efficient comparision operations.

There are two ROS nodes that can be used out of the box:

• realtime_urdf_filter

  This is a node that subscribes to a depth map topic, and outputs the filtered depth map on /output.

• urdf_filtered_tracker

  This node is basically the openni tracker with additional functionality to pre-filter the depth image before it's sent to the skeleton tracker.

  urdf_filtered_tracker accesses the Kinect directly through OpenNI, and hooks itself in between the depth image generator node and the skeleton tracker node, so any "known" geometry that could potentially confuse the skeleton tracker (e.g. robot arms, people in background, static furniture) can be deleted.

  This has been used at Automatica 2012, where an operator was tracked to allow control of the robot arms using gesture recognition. Two robot arms needed to be filtered to allow stable skeleton tracking also for hand over tasks. Additionally, the camera was pointed towards a busy walkway intersection, so two virtual walls were added that filtered passing visitors from the depth data.

Adapting it to different scenarios

There are two example launch files provided that show basic usage and parametrization and are a good starting point.

The following rosparam parameters are supported:

• fixed_frame is used to specify the "fixed" TF link (e.g. /map, /world, etc.). This is useful to decouple tf lookups with different publising frequencies, e.g. robot and static publishers.
• camera_frame specifies the camera TF frame (e.g. /camera_rgb_optical_frame)
• camera_offset lets you specify additional offsets to the camera link. It has two components: translation (e.g. [0.0, 0.0, 0.0]) and rotation (e.g. [0.0, 0.0, 0.0, 1.0]).
• models contains a list of URDF models that are supposed to be filtered. For each, model defines the rosparam key that contains the URDF model, and tf_prefix contains, well, the tf prefix.
• depth_distance_threshold pixels with a depth difference of less than this value get filtered.
• filter_replace_value (for urdf_filtered_tracker) defines the new value that will be written to the filtered pixel depth values. Interestingly, setting 0 creates beautiful silhouettes, but the OpenNI tracker needs "background" (more distant) pixels around people. Weird. That's why we set this value to 5 meters.
• show_gui specifies whether a visualization window should pop up.

Also, the shaders in include/shaders/ can easily be adapted. The vertex shader is basically just a pass through, so the fragment shader is more interesting for adding features. The shader as of now has access to 4 color attachments, and the red channel of gl_FragData[1] is used to return the filtered image. The other attachments can be used for visualization (see show_gui).

Note: starting remotely

While this package uses offscreen rendering, it does need to connect to a X11 server to get a valid OpenGL context (even with show_gui set to false). When launching one of the nodes in this package remotely via roslaunch or similar mechanisms, it will be necessary to set a DISPLAY variable and possible turn off access control for the X server. In this case, a bash script like the following can be launched from remote:

#!/bin/bash
DISPLAY=:0
xhost +
roslaunch realtime_urdf_filter tracker.launch

License

The code is licensed under the BSD License, see the LICENSE file in the project root dir.