Deep learning based 6D pose estimation for robotics

6D pose estimation technology has been deeply used in different tasks, such as robotics grasping, autonomous driving, and augmented reality. At present, there exist different methods for tasks of 6D pose estimation, and the methods based on deep learning have the advantage of accuracy over those tra...

Full description

Saved in:
Bibliographic Details
Main Author: Huang, Li
Other Authors: Jiang Xudong
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/163364
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:6D pose estimation technology has been deeply used in different tasks, such as robotics grasping, autonomous driving, and augmented reality. At present, there exist different methods for tasks of 6D pose estimation, and the methods based on deep learning have the advantage of accuracy over those traditional methods, which has great application prospect. However, there are still many challenges in practical applications such as occlusion, truncation, illumination variation et al. Therefore, deep learning based 6D pose estimation has practical significance and research value. This dissertation mainly carried out the following work: 1. Introduce the theories of pose estimation such as coordinate transformation. Introduce the basis of deep learning and architecture of CNN. 2. In literature review, this dissertation introduces 6D pose estimation methods from different dimensions. According to different input data, they can be classified into pose from RGB images, depth/point cloud and RGBD images. According to different principles, they are based on correspondence, template, or voting. Compare and analyze different application scenes, advantages, and disadvantages of various methods. For each type of methods, the dissertation analyzes advantages of deep learning-based methods over traditional methods. 3. For the experiment part, this dissertation first introduces some typical 6D pose estimation BOPs and focuses on LineMod and YCB-Video dataset. Then, we introduce some commonly used metrics such as ADD(Average Distance of Model Points) and ADD-S(Average closest point distance). Finally, we realize Densefusion on LineMod dataset, understand its network structure, do evaluation on LineMod dataset and calculate the metrics. After three times of training and validation, the ADD is improved from 0.0113 to 0.0088, and the success rate is improved from 0.806 to 0.932, which is close to 0.953, the result of given trained model on LineMod_preprocessed dataset. Keywords: 6D pose estimation, robotics grasping, deep learning