A Recognition Algorithm for Workpieces Based on the Machine Learning

Abstract

In order to learn and grasp the predetermined workpieces for robot actively, a recognition algorithm based on machine learning is proposed. Compared with traditional algorithms, we replenish a MTSM (multi threshold space model) for getting clearer workpiece shapes. To automatically and compactly learn workpieces knowledge, both shape and gradient features are designed to express the specific object by aid of contour mask, meanwhile, the compound descriptors are fed into a SVM classifier and they are trained jointly to minimize a classification loss. Finally, we adopt density estimation to acquire the grasping point of the workpieces from MTSM. Experimental result of workpieces grasping demonstrates the effectiveness and stability in complex environment, and the proposed algorithm is robust to rotation, scaling and deformation of shapes.