딥러닝 알고리즘과 국소 커널회귀법에 기반한 비정형 소화물 탐지 알고리즘 개발

Abstract

Recent trends in industrial robot development are automation for reducing labor costs, working in dangerous environments, and increasing efficiency in repetitive tasks. Specially, automation of the picking process to acquire parcels location information in the online shopping mall logistics could not be achieved by lack of skill in the detection of the parcels. In addition, the parcels at the online shopping mall logistics center is sold by the piece and is various in sizes and shapes. Moreover, the parcels are various in packaging forms such as unpacked forms, box packaging forms and plastic packaging forms, etc. Therefore, a new detection algorithm to pick these atypical parcels is positively necessary. The purposes of this thesis are firstly to develop an atypical parcel detection algorithm based on a deep learning algorithm and a local kernel regression method using Kinect camera and secondly to design a posture controller for the end-effector of a 7-link manipulator to move the extracted the bounding box location of the parcel with a desired velocity of its end effector. To do these tasks, the followings are done. Firstly, the system used in this thesis consists of a 7-link manipulator for moving end-effector to the atypical parcels and an image processing system for detecting atypical parcels are used for this thesis. Secondly, a new detection algorithm for detecting atypical parcels is proposed. The proposed detection algorithm consists of 4 steps: detecting central pixel of bounding box of atypical parcels using a deep learning algorithm, obtaining 3D depth map of parcels using Kinect camera, detecting the edge of a parcel surrounding the central pixel of the bounding box using a local kernel regression method, and extracting central location of the bounding box. Thirdly, forward kinematics modeling and Jacobian matrix of the 7-link manipulator are described. Fourthly, moving the position of the end-effector to the extracted bounding box location of the parcel and tracking angular velocity of the end effector is controlled by a controller based on differential kinematics. Finally, experiment results are shown to verify the validity of the proposed detection algorithm method compared to Canny method and of the designed controllers results.