We propose an unsupervised method to extract skeleton from single-view depth videos acquired from a modality such as Microsoft Kinect/mobile stereo camera. Our algorithm takes in depth video and outputs per-frame skeleton of the articulated generic object. The presence of noise and inconsistencies in the Kinect video make the skeletonization challenging as opposed to the state-of-the-art skeletonization methods on complete point clouds. The key contributions include: (a) improved adaptation of Coherent PointDrift algorithm for RGB-D videos, (b)spectral clustering that considers the spatiotemporal property of the trajectories to encode rigid and non-rigid motions in the point cloud sequence. (c) Lastly, we overcome the limitations of previous skeletal extraction methods that utilize minimum spanning tree leading to incorrect bones and skeletal structure while extracting skeleton through the addition of boundary information of clusters.
Unsupervised Skeleton Extraction of Generic Objects from Depth Videos
We propose an unsupervised method to extract skeleton from single-view depth videos acquired from a modality such as Microsoft Kinect/mobile stereo camera. Our algorithm takes in depth video and outputs per-frame skeleton of the articulated generic object. The presence of noise and inconsistencies in the Kinect video make the skeletonization challenging as opposed to the state-of-the-art skeletonization methods on complete point clouds. The key contributions include: (a) improved adaptation of Coherent PointDrift algorithm for RGB-D videos, (b)spectral clustering that considers the spatiotemporal property of the trajectories to encode rigid and non-rigid motions in the point cloud sequence. (c) Lastly, we overcome the limitations of previous skeletal extraction methods that utilize minimum spanning tree leading to incorrect bones and skeletal structure while extracting skeleton through the addition of boundary information of clusters.
0 Comments