- Abstract: This paper presents the foundation of a new system for human body segmentation. It is based on a Fully Convolutional Neural Network that uses depth images as input and produces a per-pixel labeling of the image where each pixel has been labeled as a body segment of interest or as non-person. The training data are fully synthetic which allow for large amounts of data to be generated in a relatively short period of time. By using a GPU accelerated implementation of the convolutional neural network, the system is capable of segmenting an image in 8.5 milliseconds. This work will form the basis for more robust system in the future that will be suitable for finding pose skeletons in more cluttered environments.
- Conflicts: cse.unr.edu, unr.edu, nasa.gov, flirtey.com