Multi-class assembly parts recognition using composite feature and random forest for robot programming by demonstration

Abstract

In robot programming by demonstration (PBD) for assembly tasks, one of the important purposes is to identify multi-class objects during demonstration. In this paper, we propose a composite feature representation method using color histogram, LBP, aspect ratio, circularity and Zernike moment, which is invariant to image translation, rotation and scale. Then Random Forest algorithm is employed to be trained as the classifier, by which the weight parameters of the composite feature are obtained simultaneously. Experimental results on 20 different kinds of objects demonstrate that our approach achieves high recognition accuracy with 99.33%. According to comparisons with other composite features and classification algorithms, the effectiveness with fewer collected samples and the efficiency using less model training time of our approach are verified. Our approach has been successfully applied in two PBD tasks - flashlight assembly and building blocks assembly.