Fast Object Detection On Raw Depth Images
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] We propose an effective object detection system which is able to localize object in real time and outperforms the state-of-the-art results on standard Washington RGB-D benchmark datasets, designed exclusively for object detection with a depth sensor. Our system is based on a new feature Normal Vector Pattern (NVP) describe the surface of an object discriminatively by encoding the inner products and cross products between neighboring surface normals. Moreover, to speed up the system and make the real time processing possible, we further propose a new method for generating a small set of candidate object locations on raw depth, in which the relation between the depth and scale of an object is formulated and cascaded with a simplified NVP detector to reduce the search space. The proposed system can achieve 9 fps on a single laptop CPU.
Access is limited to the campuses of the University of Missouri.