1、Computer Software and TheoryAuthor:Wei WanSupervisor:A.P.Ping KuangSchool:School of Information and Software Engineering独 创 性 声 明本人声明所呈交的学位论文是本人在导师指导下进行的研究工作及取得的研究成果。据我所知,除了文中特别加以标注和致谢的地方外,论文中不包含其他人已经发表或撰写过的研究成果,也不包含为获得电子科技大学或其它教育机构的学位或证书而使用过的材料。与我一同工作的同志对本研究所做的任何贡献均已在论文中作了明确的说明并表示谢意。签名,日期,年月日关于论文使用
2、授权的说明本学位论文作者完全了解电子科技大学有关保留、使用学位论文的规定,有权保留并向国家有关部门或机构送交论文的复印件和磁盘,允许论文被查阅和借阅。本人授权电子科技大学可以将学位论文的全部或部分内容编入有关数据库进行检索,可以采用影印、缩印或扫描等复制手段保存、汇编学位论文。(保密的学位论文在解密后应遵守此规定)签名,导师签名,日期,年月日摘要I摘要目标检测是计算机视觉领域中一个富有挑战性的课题,其主要目的是从静态图片或视频中检测并定位特定的目标。它综合运用了图像处理、模式识别、人工智能和自动控制等技术,在道路交通事故预防、工厂可疑危险品警告、军事禁区监控、高级人机交互等方面有着广泛的应用。
3、由于目标所处环境复杂多变,目前还没有一种比较通用成熟的检测方法,在实际应用中目标检测算法研究的机遇与挑战并存。本论文首先分析了国内外目标检测算法的研究现状,着重介绍了当前应用较为广泛的方法,使用目标特征训练分类器进行目标分类检测。针对使用现有特征训练的分类器进行目标分类检测时存在误检率较高的缺点,本文在深度学习的基础上提出了一种基于卷积神经网络的行人目标检测算法。为了解决直接使用卷积神经网络进行滑动窗口检测效率较低的问题,本文将算法分为两大步骤:(1)疑似存在行人窗口确认;(2)行人检测。在疑似存在行人窗口确认中,本文使用融合特征作为行人的描述特征训练分类器,采用了邻近尺度特征值相似的思想构建
4、分类器金字塔,在待检图像上利用不同尺度的滑动窗口进行滑动遍历确定疑是存在行人窗口;在行人检测中,使用大量正负样本训练了一个卷积神经网络,为了更好的适应行人检测,将该卷积神经网络的拓扑结构进行改进。将疑似存在行人的窗口输入改进后的卷积神经网络进行行人检测,在保持原有的检测率的基础下降低了误检率。为了验证本文所提出算法的准确性,在 INRIA 行人数据库进行行人检测实验。分别以每个窗口和每幅图像为检测单位,统计本文算法的检测率和误检率,在平均每幅图像存在一个误检窗口的标准下,达到了 93%的检测率。检测率比使用 ACF特征训练的检测器高三个百分点,检测速度比单独使用卷积神经网络检测提升四倍以上。实
5、验结果证明了本文算法的有效性。关键字关键字:目标检测;行人检测;深度学习;卷积神经网络ABSTRACTIIABSTRACTObject detection is a challenging problem in the field of computer vision andwhich main purpose is to detect and locate specific goals from static images or video.Itis based on the technology of technology of image processing,pattern recog
6、nition,artificial intelligence and automatic control and widely used in traffic accidentprevention,suspicious warned of dangerous goods in factory,military restricted zonemonitoring and senior human-computer interaction.The current lack of a mature andgeneral method to detect object because of the e
7、nvironment is complicated.Objectdetection research exist opportunities and challenges in practical application.This thesis first analyzes the domestic and foreign research status of objectdetection algorithm,emphatically introduces the application method which are widelyused is based on the object f
8、eature trained classifier to classify object.Because of theexisting feature of the trained classifier to classify object has high false positives rate,this thesis present a pedestrian object detection algorithm based on convolution neuralnetwork on the basis of deep learning.The algorithm consists o
9、f two steps in order tosolve the low efficiency of sliding window with convolution neural network,(1)thesuspected pedestrian window confirmation;(2)the pedestrian detection.In suspectedexisting pedestrian window confirmation,this thesis use the fusion feature as thedescription of the pedestrian trai
10、ning classifier and the ideal of nearby scale featuresimilar to build classifier pyramid.On the inspected images,this thesis use differentscales of sliding window to slide traversal to confirm suspected exist pedestrian window.In the pedestrian detection,this thesis rely a large number of positive a
11、nd negativesamples to train and get a convolution neural network.In order to better adept thepedestrian detection,this thesis improve the topology of traditional convolution network.Input suspected existence of pedestrians window into the improved convolution neuralnetwork to detect the pedestrian.I
12、n order to verify the accuracy of the proposed algorithm,this thesis test pedestriandetection experiments in the INRIA pedestrian database.Separately treat each windowand each image as detection unit,this thesis statistics the detection rate and errordetection rate of the algorithm.On the standard o
13、f the existence of an error in everyimagedetectionwindow,thisthesisgets93%detectionrate.ComparedtheABSTRACTIIIexperimental results with train detector using ACF feature,under the same false positiverate,the algorithm in this thesis has 3%detection rate higher than the detector trainedfrom ACF featur
14、e and detection time less four folds than single use convolutional neuralnetworks.The experimental results certify the effectiveness of the algorithm in thisthesis.Keywords:Object Detection,Pedestrian Detaction,Deep Learning,ConvolutionalNeural Network目录IV目录第一章 绪论.11.1 研究背景及意义.11.2 国内外研究现状.21.3 本文主要内容及结构安排.4第二章 相关基础知识.62.1 行人目标检测算法原理.62.1.1 基于模板匹配的检测算法.62.1.2 基于分类的行人检测算法.72.2 典型的人体特征.92.2.1 HAAR-LIKE 特征.92.2.2 HOG 特征.102.2.3 SIFT 特征.122.3 典型分类器.152.3.1 支持向量机算法.152.3.2 ADABOOST 算法.182.4 深度学习.202.4.1 深度学习概述.202.4.2 典型的深度学习结构.212.5 本章小结.27第三章 基于融合特征的疑是行人窗口确认.283.1提取融合特征.283.2分类器训练.343.2.1训练标准尺度级
copyright@ 2008-2022 冰豆网网站版权所有
经营许可证编号:鄂ICP备2022015515号-1