Fast3D Object Detection Using RGBD Data

With those Kinect-style equipments that can capture both color anddepth data simultaneously come to our lives, many exciting applicationsbegin to appear. However, endless needs of applications force us tocontinually search for better ways of using RGBD data. Among theseneeds,3D object detection, as the foundation of intelligent robots andaugmented reality, has received increased attention.Among all these works on3D object detection, Dominant OrientationTemplates (DOT) based methods have shown their excellent potential. Inthis paper, we extend original method, speed up the detection processwithout lose accuracy and robustness to partial occlusion. Throughanalyzing the RGB and depth data of the object template, we define andselect a few key areas on both template images. These decentralized keyareas contain both the boundary information and the surface geometryinformation thus make themselves distinct and robust to occlusion. We usethese key areas to quickly filter negative areas of the input images andspeed up the algorithm. Meanwhile, key areas on the depth map can alsobe used to speed up the3D object registration. We also proposed different strategies for preprocessing RGB anddepth data that make the input data more reliable for the subsequentdetection. In addition, we elaborate how to use GPU to speed up thecomputing.