Scale and rotation invariant 3D object detection using spherical nonlinear correlations

摘要

Three-dimensional object recognition with scale and rotation changes is addressed. The recognition method is describedin terms of correlations between spherical surfaces. Tridimensionality is codified into range images. We used the phaseFourier transform of a range image which gives information about the orientation of the 3D object surfaces. A 3D objectorientation map containing information about all possible rotations of the object is obtained. This map distribution iscalculated using the amplitude of the phase Fourier transform for different views of the object. From that 3D objectdescription it is possible to achieve detection and rotation estimation by performing a correlation between unit sphereseven when only partial information is presented. In addition, a scale change of a rotated 3D object implies a change ofthe intensity in the unit sphere. We define correlations between the reference unit sphere and a certain target-patchplaced on the surface of the unit sphere. Various experiments are carried out to confirm the correct detection. We alsovalidate the method when other false targets were used. In addition to tolerance to scale and rotation, high discriminationagainst false targets is also achieved.