Combines the abilities of rail_object_detector and handle_detector to locate objects in the current scene as seen by the Kinect sensor and apply the correct translations and orientations to them.
With dependencies installed simply run roslaunch object_location object_location_engine.launch to launch the locator and its dependent nodes
The locator can be started by calling the /location_query service which will process and return a standard header and a list of detected objects with associated label, detection probability, pose, and cropped point cloud.
Note: The service call can take upwards of a minute to return. Performace is mostly bound by the YOLO object detection library used by rail_object_detector and can be sped up by enabling CUDA.
If the publish_as_tf option is set to true (default) the following will be continuously published:
- Detected objects under
/object_location - Object cropped point clouds under
/object_location_cloud/<label> - Transforms relative to the global frame (default:
map) by object label
Bottle detected, labeled, and located by this package as shown in RViz. The grayscale point cloud is the feed coming directly from the Kinect sensor while the false color area is the processed and cropped point cloud containing just the recognized object.
Combining
rail_object_detectorandhandle_detector's capabilities, this package finds the objects in the scene and uses handle data to apply the proper orientations to them. Here,rail_object_detectorfound the bottle and monkey wrench (labeled as toothbrush) whilehandle_detectorfound handle orientations for all three of the wrenches. The results were combined to form this output with the z-axis of each transform parallel to the axis of the handle (if one exists)
A closer look at the calculated transforms of the monkey wrenches in the scene.
- ROS Indigo
rail_object_detectorhandle_detector(petschekrfork)
Copyright © 2017 Ryan Petschek. Released under the MIT license. See LICENSE for more information.