基于特征学习的RGB-D目标识别算法研究

发布时间：2018-04-30 20:45

本文选题：目标识别 + 深度学习　；参考：《华中科技大学》2016年硕士论文

【摘要】：近几年随着传感器技术的发展(微软的Kinect,Intel RealSense 3D等),彩色和深度图像(RGB-D)可以同时获得。那么目标的深度数据的引入的确可以得到额外具有描述能力的信息。那么相对于传统的基于RGB的目标识别范例,如何更好的使用RGB-D的信息来进行目标识别。直观的来说,二维RGB图像并不能获取目标的三维几何特性,而深度信息的引入就带来了克服这个问题的潜力。因此,研究有效的、较强表达能力的RGB-D目标识别算法具有重要的理论意义和实用价值。为了有效的利用深度图像的三维几何信息并结合彩色二维结构信息,即从RGBD图像中获取具有较强表达能力的特征,本文旨在研究性能优异的RGB-D目标识别算法。为此,本文提出了两种基于特征学习RGB-D目标识别算法并取得了比较良好的实验效果。论文的主要内容如下:首先,针对标准的多模态输入的RGB-D目标识别任务,本文提出了一种基于深层卷积神经网络与特征编码的混合结构的RGB-D目标识别算法。利用深层卷积神经网络对于彩色模态的数据利用深层卷积神经网络提取高级的特征表达。对于深度图像利用改进的HONV特征描述子作为底层特征描述并进行Fisher Vector编码以增强特征描述。总体的特征由各个模态的特征级联得到并利用线性支撑向量机作为分类器,在通用RGB-D物体数据集上取得了最优的识别性能。其次,本文研究以无监督的方式为RGB-D图像提取表达能力强的特征,提出了基于多视角无监督学习的RGB-D目标识别算法。对于给定的一帧RGB-D图像数据,利用多视角数据表达将其转换为不用“视角”的数据形式。对于每个视角的数据送入我们提出的无监督特征学习算法(卷积匹配跟踪)进行特征提取。在公共数据集上实验结果表明本文算法中采用的每个视角数据之间的互补特性以及提出的算法的有效性。
[Abstract]:In recent years, with the development of sensor technology (Microsoft's Kinect, Intel RealSense 3D, etc.), color and depth images (RGB-D) can be obtained at the same time. Then the introduction of the target depth data can indeed obtain additional descriptive information. Then, compared to the traditional target recognition paradigm based on RGB, how to better use RGB-D Information is used for target recognition. The two-dimensional RGB image can not obtain the three-dimensional geometric characteristics of the target, and the introduction of the depth information brings the potential to overcome this problem. Therefore, the study of the effective and strong expressive RGB-D target recognition algorithm has important theoretical significance and practical value. The three-dimensional geometric information of the deep image and the color two-dimensional structure information, that is, obtain the strong expressive features from the RGBD image. This paper aims to study the RGB-D target recognition algorithm with excellent performance. In this paper, two target recognition algorithms based on feature learning based on RGB-D are proposed and good experimental results are obtained. The main contents are as follows: firstly, a RGB-D target recognition algorithm based on the mixed structure of deep convolution neural network and feature coding is proposed for the standard multimodal input RGB-D target recognition task, and the advanced convolution neural network is used to extract advanced special features of the color mode data using the deep convolution neural network. The improved HONV feature descriptor is used as the underlying feature description and Fisher Vector coding to enhance the feature description. The overall feature is cascaded by the characteristics of various modes and is used as a classifier by linear support vector machines, and the optimal recognition performance is obtained on the general RGB-D object data set. At the same time, this paper presents a RGB-D target recognition algorithm based on multi view unsupervised learning in unsupervised RGB-D image extraction. For a given frame of RGB-D image data, it is converted into a data form without "visual angle" by multi view data expression. The data for each view is sent to me. The unsupervised feature learning algorithm (convolution matching tracking) is proposed for feature extraction. The experimental results on the public dataset show the complementary characteristics of each view data used in the algorithm and the effectiveness of the proposed algorithm.

【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】