Monocular Object Localization by Superquadrics Curvature Reprojection and Matching

This paper presents a new method for 3D object localization from a single image. It is known that single camera provide 2D image data, annihilating valuable 3D information about object and its localization in space. The main new idea is to match 2D image gradient to the reprojection of 3D curvature to retrieve objects position relative to the camera. The object parameters are a-priori known and modelled by SuperQuadrics (SQ) that enable the calculation of the analytical form of curvature. The image processing stage includes object detection and segmentation by the Histogram of Oriented Gradients (HOG) algorithm. The method proposed uses the dependencies between SQ curvature and image gradient also considering the illumination model and object contour embedded in a proper cost function. To manage local minima we propose the use of particle swarm optimization (PSO).