基于机器视觉的室内人物检测与场景识别
发布时间：2019-01-10 09:53
【摘要】：室内环境下的场景理解是室内移动机器人必须具备的能力之一,随着全球服务机器人行业的兴起,半结构化环境下的室内场景理解成为计算机视觉领域的研究热点,也是一个难点,其主要体现在室内环境的复杂性,识别算法的鲁棒性以及实时性上。室内场景理解包括室内环境下的目标物体检测,机器人所处环境估计,室内障碍物规避,人的检测和身份识别等。围绕上文所提出的问题,本文以室内行人和物体检测为研究内容,主要的研究和工作内容如下:1)本文详细分析了卷积神经网络的特征提取和分类方法,并将该方法进行物体识别效果与SIFT特征提取加FLANN匹配方法的物体识别效果作对比,得出在目标物体的不同观察角度与目标物体发生形变的情况下,卷积神经网络物体识别效果明显优于SIFT特征提取加FLANN匹配方法识别效果的结论。2)针对传统场景识别底层特征语义信息表达能力的不足,结合卷积神经网络,本文提出一种基于物体检测的室内场景识别方法。该方法首先采用卷积神经网络对场景中的目标进行特征提取和分类,然后基于概率模型以检测到的目标作为中间桥梁去推断当前所处的场景。与基于计算机视觉底层特征的场景识别方法相比,该方法更接近于人类对场景的认知思维。本文运用该方法对场景的五种室内场景进行场景识别分类,取得不错效果。3)为了测试机器人在室内环境下对行人检测效果和响应,本文在PR2机器人平台下基于ROS系统(Robot Operating System),采用Haar-Like特征与Ada Boost分类器实现人脸检测,并用EigenFace进行身份识别,同时用HOG(Histogram of Oriented Gradient)特征与SVM(Support Vector Machine)分类器进行人体检测,并实现机器人对行人的自主跟随。
[Abstract]:Scene understanding in indoor environment is one of the necessary capabilities of indoor mobile robot. With the rise of global service robot industry, indoor scene understanding in semi-structured environment has become a research hotspot in the field of computer vision. It is also a difficult point, which is mainly reflected in the complexity of indoor environment, the robustness of recognition algorithm and the real-time performance. Indoor scene understanding includes object detection in indoor environment, robot environment estimation, indoor obstacle avoidance, human detection and identity recognition. The main contents of this paper are as follows: 1) the feature extraction and classification methods of convolution neural network are analyzed in detail. The object recognition effect of this method is compared with that of SIFT feature extraction and FLANN matching method. The result of object recognition based on convolution neural network is obviously better than that of SIFT feature extraction and FLANN matching. 2) aiming at the deficiency of semantic information expression of traditional scene recognition underlying features, we combine convolutional neural network with convolution neural network. This paper presents a method of indoor scene recognition based on object detection. Firstly, the convolution neural network is used to extract and classify the features of the targets in the scene, and then based on the probabilistic model, the detected target is used as the intermediate bridge to infer the current scene. Compared with the scene recognition method based on the underlying features of computer vision, this method is more similar to the cognitive thinking of the scene. In this paper, we use this method to classify the scene of five indoor scenes, and get good results. 3) in order to test the robot in the indoor environment to detect the effect and response to pedestrian, In this paper, based on ROS system (Robot Operating System), Haar-Like feature and Ada Boost classifier are used to realize face detection based on PR2 robot platform, and EigenFace is used to identify human face. At the same time, the HOG (Histogram of Oriented Gradient) feature and SVM (Support Vector Machine) classifier are used to detect the human body, and the robot can follow the pedestrian autonomously.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
本文编号：2406198
[Abstract]:Scene understanding in indoor environment is one of the necessary capabilities of indoor mobile robot. With the rise of global service robot industry, indoor scene understanding in semi-structured environment has become a research hotspot in the field of computer vision. It is also a difficult point, which is mainly reflected in the complexity of indoor environment, the robustness of recognition algorithm and the real-time performance. Indoor scene understanding includes object detection in indoor environment, robot environment estimation, indoor obstacle avoidance, human detection and identity recognition. The main contents of this paper are as follows: 1) the feature extraction and classification methods of convolution neural network are analyzed in detail. The object recognition effect of this method is compared with that of SIFT feature extraction and FLANN matching method. The result of object recognition based on convolution neural network is obviously better than that of SIFT feature extraction and FLANN matching. 2) aiming at the deficiency of semantic information expression of traditional scene recognition underlying features, we combine convolutional neural network with convolution neural network. This paper presents a method of indoor scene recognition based on object detection. Firstly, the convolution neural network is used to extract and classify the features of the targets in the scene, and then based on the probabilistic model, the detected target is used as the intermediate bridge to infer the current scene. Compared with the scene recognition method based on the underlying features of computer vision, this method is more similar to the cognitive thinking of the scene. In this paper, we use this method to classify the scene of five indoor scenes, and get good results. 3) in order to test the robot in the indoor environment to detect the effect and response to pedestrian, In this paper, based on ROS system (Robot Operating System), Haar-Like feature and Ada Boost classifier are used to realize face detection based on PR2 robot platform, and EigenFace is used to identify human face. At the same time, the HOG (Histogram of Oriented Gradient) feature and SVM (Support Vector Machine) classifier are used to detect the human body, and the robot can follow the pedestrian autonomously.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
【参考文献】
相关期刊论文 前1条
1 邓中亮;余彦培;袁协;万能;杨磊;;室内定位现状与发展趋势研究(英文)[J];中国通信;2013年03期
，本文编号：2406198
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2406198.html
