A fast positioning algorithm based on 3D posture recognition

Abstract

Automatic workpiece grabbing on production line is important for improving the production efficiency in manufacturing industry. However, due to their irregular shapes, uncertain positions, and various posture changes, traditional edge detection, feature extraction and other methods are difficult to accurately identify and locate complex workpieces. In this paper, we propose an approach to recognize workpieces and determine their postures with deep learning. Based on object detection, input images containing workpieces are fed into an angle regression network which is used to determine the three-dimensional posture of workpieces. The classification, position and posture information are obtained from the final output. Experiments show that this method is more robust and can achieve higher accuracy than traditional feature extraction methods.