Compressed VFH descriptor for 3D object classification

Abstract

This paper presented a compressed version of viewpoint feature histogram descriptor (VFH) for object classification based on shape recognition. VFH is known for representing geometrical features (3D rotation angles) of 3D points with concatenated histogram. However this histogram is large and sparse. The proposed descriptor employ eigenvalue decomposition to extract dominant orientation features from the point cloud and using it as a descriptor instead of sparse histogram. Using dominant features maintains the main properties of the object with minimal descriptor length which applicable for recognizing the geometrical class of 3D objects (sphere, rectangle, ... etc). This descriptor was tested for two class and multi-class 3D object classification using support vector machines (SVM). The new descriptor showed promising matching results of 88% for two class classification and 83% for multiple class classification of 3D objects.