-
Notifications
You must be signed in to change notification settings - Fork 1
Person Detection Ideas
Here are some ways to detect people in RGB-Depth images. We need this in order to filter certain people out of images in order to protect their privacy. We may even try to associate personal possessions with their owners, and filter them similarly!
- Q: How are those people being colored in the depth image?!
People Standing on Flat Ground
Person and Body Part Detection using GPU
These ideas all start with the joints found using the openni_tracker package.
-
Cluster objects naively, then separate any clusters that are joined near a hand or foot joint
-
Do region growing from PCL using the joints (or their nearest neighbors) as seeds
-
For each joint (and perhaps some intermediate joints), capture all foreground points within a specified radius using the "neighbors within radius search" described in this PCL tutorial
-
In RGBD image space, use GrabCut to segment out the person, where the "foreground" points we give to GrabCut are taken from around the skeleton. Can use RGB, HSV, HSD (D = depth) for GrabCut, and perhaps even scale the depth nonlinearly so the usual range of person interaction (maybe 1-4 meters) has particularly high resolution.
Benjamin Choi, Cetin Mericli, Joydeep Biswas, and Manuela Veloso. Fast Human Detection for Indoor Mobile Robots Using Depth Images. International Conference on Robotics and Automation (ICRA), 2013
Hao Zhang, Christopher Reardon, and Lynne E. Parker. Real-Time Multiple Human Perception with Color-Depth Cameras on a Mobile Robot. IEEE Transactions On Cybernetics B (TMSC-B), vol. 43, no. 5, pp. 1429-1441, October 2013