Section: New Results

Human-Limb Segmentation for Intelligent Mobility Assistance Robots

Paticipants: Siddhartha Chandra, Stavros Tsogkas, Iasonas Kokkinos

We developed a computer vision component  [12] to be used as part of an intelligent robotic assistant. This component exploits RGB and depth information extracted from Kinect sensors mounted on the robot, to accurately segment human limbs, using fully convolutional neural networks (FCNNs). We trained our network using an in-house Human-Limb dataset composed of video frames, and described a scheme for dynamically exploiting RGB and depth data in a single framework for training and testing. Our method demonstrated promising performance, being very efficient at the same time, with a run-time of 8 frames per second on a single GPU.