A Pile of Objects Detection, Classification for Grasping System Using RGB-D Faster R-CNN

• Main Authors: Wan-JenLi, 李宛臻
• Other Authors: Shu-Mei Guo
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/fr7s46

碩士 === 國立成功大學 === 資訊工程學系 === 107 === In order to reduce the production cost due to the rising wages, the manufacturing industry replaces the traditional manpower with automated production equipment to perform material processing or quality inspection. However, the feeding among machines is still a challenge. The goal of this study is to use a robot arm system to replace the part of loading materials manuually or shaking and flattening the piled semi-finished products by the vibratory feeder, resulting in an increase in the flexibility and a decrease in the cost. In this thesis, a RGB-D sensor, robot arm, and computer vision algorithms are combined to form a system for classifying, detecting and grasping objects in the pile. Three methods are applied. The first method is the detection and classification by analyzing the color and depth images using traditional computer vision methods. Next, the original Faster R-CNN is improved to multi-modal inputs which are RGB image and raw depth map. The RGB-D Faster R-CNN model has a precise detection result with fused RGB-D feature. Finally, the RGB-D Faster R-CNN is further modified to 3D RGB-D Faster R-CNN, which outputs 3D information of object directly.