本文为德国慕尼黑工业大学（作者：Niklas Barkmeyer）的硕士论文，共104页。

由于照明条件的可变性、可见性约束以及大量标记数据的可用性，使得机器人能够识别现实世界中的对象是一个非常重要和具有挑战性的问题。目前，这个问题已经使用诸如卷积神经网络（CNN）的分层表示来解决，这些分层表示是受人脑如何处理视觉信息的启发而设计的多级结构。作为深度学习的一个子领域，这些算法可以进行端到端的训练，并且在许多基准测试上优于其它方法。然而，CNN还没有在机器人领域得到广泛的应用。此外，现有的机器人系统没有利用上下文信息来增强其具有知识和推理系统的机器学习算法，以理解并与其环境中的已知和未知对象进行交互。

在本文中，我们提出并分析了一种结合深度学习算法和语义推理技术的方法。该系统通过在人类监督的学习过程中引入包括物体的背景信息，例如材料、形状、颜色和启示，从而丰富了机器人的视觉识别能力。我们使用深度卷积网络并训练提取高维特征上的分类器来预测这些属性。我们的实验在iCub类人机器人上进行评估，并在两个数据集上进行了测试，即iCubWorld28和新的TUM-ICS数据集。

研究结果表明，该系统提高了iCub机器人对已知和未知物体的整体识别精度和学习能力。与独立的深度学习网络相比，对已知对象的识别率从85%提高到98%，对未知对象的识别率从0%提高到65%。

Enabling robots to recognize objects in the real world is a veryimportant and challenging problem, due to variability in lighting conditions,visibility constraints, and the availability of large amounts of labeled data.This problem has currently been addressed using hierarchical representationssuch as convolutional neural networks (CNN), which are multi-stage architecturesinspired by how the human brain processes visual information. A subgroup of deeplearning, these algorithms can be trained end-to-end and outperform othermethods on many benchmarks. However, CNNs are not widely deployed in roboticsyet. Additionally, existing robotic systems do not leverage contextualinformation to enhance their machine learning algorithms with knowledge andreasoning systems to understand and interact with known and unknown objects intheir environment. In this thesis, we present and analyze a method thatcombines deep learning algorithms with semantic reasoning techniques. Thissystem enriches the visual recognition capabilities of a robot by includingcontextual information of an object such as material, shape, color, andaffordance during a human-supervised learning process. We use deepconvolutional networks and classifiers trained on extracted high-dimensionalfeatures to predict these attributes. Our experiments are evaluated on the iCubhumanoid robot and tested on two datasets, namely the iCubWorld28 and the newTUM-ICS dataset. The results show that our proposed system improves the overallrecognition accuracy and learning capabilities of our iCub robot, both forknown and unknown objects. Compared to a stand-alone deep learning network, therecognition performance increases from 85% to 98% for known objects and from 0%to 65% for unknown objects.

1 引言

2 项目背景与相关工作

3 系统设计

4 实验与结果

5 讨论

6 总结与结论

下载英文原文地址：

http://page2.dfpan.com/fs/2lc1j2f21b29f16dea7/ 

更多精彩文章请关注微信号：