Fast Detection of Partially Occluded Humans from Mobile Platforms: Preliminary Work

Abstract: The final goal of this project is to create a robust system for finding the pose, i.e. locations of arms and legs, of a human within a depth image. The system will take depth images as input and use a trained Convolutional Neural Net (CNN) to find the locations of the body segments. It will also make no assumptions about the orientation of the camera relative to the human, so the system will still function when used on cameras in any arbitrary position. In order to train the CNN, large amounts of training data must be created. This data must have depth images with each pixel labeled as one of the body parts. Creating this data manually is a time consuming task. This preliminary work involves creating a system to generate synthetic data suitable for training.


  • Organization: Nevada NASA Space Grant
  • Award #: NNX10AN23H
  • Amount: $2,500
  • Date: Jan. 20, 2015 - May 15, 2015
  • PI: David Frank
  • Co-PI: Dr. David Feil-Seifer

Supported Publications

  • Lucas, H., Poston, J., Yocum, N., Carlson, Z., & Feil-Seifer, D. Too big to be mistreated? Examining the Role of Robot Size on Perceptions of Mistreatment. In IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), page 1071-1076, New York, NY, Aug 2016. ( details ) ( .pdf )
  • Carlson, Z., Sweet, T., Rhizor, J., Lucas, H., Poston, J., & Feil-Seifer, D. Team-Building Activities For Heterogeneous Groups of Humans and Robots. In International Conference on Social Robotics (ICSR), page 113-123, Paris, France, Oct 2015. ( details ) ( .pdf )