Novel 3D Shape Representations

SET Associates proposes to identify and implement novel, real-time 3D shape representations for measured point-cloud data maximizing exploitation while minimizing computational complexity and bandwidth. The Interest in the use of 3D sensors for munitions autonomous agent applications is continually increasing. Common data sources include Laser Radar (LADAR), calibrated stereo cameras, and video utilizing structure from motion (SFM) algorithms. For each source, the sensor output is a range image combined with information on the sensor pointing direction and location at each frame. To emphasize the 3D structure and shape information, range images are normally transformed into a XYZ point cloud referenced to world or local coordinates. While point cloud representations are easily extracted from range imagery, they have various shortcomings for both supervised and autonomous target acquisition and recognition applications. Foremost, individual points in the data have no explicit connectivity with other points to help define surfaces and objects which are building blocks of human perception and Automatic Target Acquisition/Automatic Target Recognition (ATA/ATR) algorithms. This motivates a change in shape representation in many applications - either to facilitate human reasoning over the 3D data or enable robust autonomous information extraction. For a selected set of applications, SET will investigate and develop representations optimally suited to munitions and autonomous agents such as micro air vehicles. In this context, SET’s proposed candidate representations are compact, enabling transmission over low-bandwidth networks. Proposed representations are also computationally simple, enabling processing to be done on vehicle with limited available processing resources. We propose to identify and implement candidate 3D shape representations using appropriate synthetic or measured 3D data. SET will assess each representation’s computational complexity, bandwidth requirements, and accuracy compared to the point cloud data.