ROS Message Definitions. As in the detection case, we tried to keep the message definitions for people tracking as generic as possible such that they can be re-used across a large vari- ety of different people tracking algorithms. In the following sections, we illustrate concrete implementations that we have experimented with during the project, all of which yield the following output. A TrackedPerson, according to our definition, possesses the following attributes: – track id [uint64]: An identifier of the tracked person, unique over time. – is matched [bool]: False if no matching detection was found close to the track’s predicted position. If false for too long, the track usually gets deleted. – is occluded [bool]: True if the person is physically occluded by another per- son or obstacle. False if the tracking algorithm cannot determine this. – detection id [uint64]: If is matched is true, the unique ID of the detection associated with the track in the current tracking cycle. Otherwise undefined. × – pose [geometry msgs/PoseWithCovariance]: The position and orientation of the tracked person in metric 3D space, along with its uncertainty (expressed as a 6 6 covariance matrix). The pose is relative to the coordinate frame specified in the TrackedPersons message (see below). – twist [geometry msgs/TwistWithCovariance]: The linear and possibly angu- lar velocity of the track, along with their uncertainties. – age [duration]: The time span for which this track has existed in the system. The TrackedPersons message aggregates all TrackedPerson instances that are currently being tracked: – header [std msgs/Header]: The coordinate frame is usually a locally fixed frame that does not move with the robot, such as “xxxx”. The timestamp is copied from the incoming DetectedPersons message; this is important to synchronize DetectedPersons and TrackedPersons messages to be able to look up details about a DetectedPerson identified by its detection id. – tracks [array of TrackedPerson]: All persons that are being tracked in the current tracking cycle, including occluded and unmatched tracks. 10 People Detection, Tracking and Visualization on a Mobile Robot
ROS Message Definitions. To keep a memory of which detections have been fused together, we define a CompositeDetectedPerson message (Fig. 9). Retaining this information is es- sential for components at later stages in the perception pipeline, like a human People Detection, Tracking and Visualization on a Mobile Robot 15 attribute classifier that takes as an input the regions of interest found by a vision- based person detector. If these ROIs likewise bear a corresponding detection ID, they can later on be associated with the correct tracked person such that the extracted information can be smoothed over time. CompositeDetection CompositeDetections DetectedPerson
