摘 要

Abstract 6

1. 绪论 7

1.1 课题背景概述 7

1.2 相关技术概述 8

1.2.1 卷积神经网络技术 8

1.2.2 目标检测技术 9

1.3 国内外研究现状 13

1.4 常用术语解释 14

1.5 本文的主要贡献与创新 14

1.6 本论文的结构安排 15

2.Faster R-CNN目标检测框架 15

2.1 R-CNN系列目标检测框架的演进 15

2.1.1 R-CNN目标检测框架 15

2.1.2 SPP-Net目标检测框架 18

2.1.3 Fast R-CNN目标检测框架 19

2.2 Faster R-CNN原理概述 21

2.2.1 RPN网络 22

2.2.2 FRCNN网络 24

2.2.3 非极大值抑制NMS 25

2.3 Faster R-CNN实验结果分析 25

2.4 本章小结 27

3. 基于Faster R-CNN的JPD-CNN行人检测框架 27

3.1 关键问题 27

3.2 检测框架设计 27

3.2.1 PRPN网络 27

3.2.2 行人区域概率热区 28

3.2.3 建议区域横移扩充 29

3.2.4 联合大型分类网络与小型区域建议网络 30

3.3 本章小结 31

4. 工程实现 31

4.1 深度学习库Caffe 31

4.2 Faster R-CNN的MATLAB实现 32

4.3 JPD-CNN改进算法的实现 38

4.4 编写算法的C 版本 39

4.5 本章小结 44

5. 数据集实验 44

5.1 KITTI数据集 44

5.2 网络训练 46

5.2.1 迁移学习 46

5.2.2 训练PRPN 47

5.3 实验对比 48

5.4 本章小结 50

6. 总结与展望 50

致谢 51

参考文献 52

摘要

行人检测为目标检测技术的一项具体应用，其对于自动驾驶、智能监控等领域意义重大。近年来，深度卷积神经网络这一技术高速发展，一大批优秀的目标检测框架被相继提出，诸多经典方法已经投入商用。但对于行人检测这一具体任务，仍然存在诸多难点，如行人目标尺寸较小、遮挡严重等难以克服的问题。本文以深度卷积神经网络为基础，回顾了目标检测与行人检测的发展历史，并试图改进现有的行人检测方法。

首先，本文回顾了目标检测领域的发展，介绍了卷积神经网络的组成原理，并且基于国内外顶级期刊、会议文献分析了行人检测的国内外研究现状，为研究提供了理论基础。

其次，本文介绍R-CNN系列检测框架的演进，从R-CNN到SPP-NET，再到Fast R-CNN这一陆续推出的框架，介绍了其各个部分的原理，网络的损失函数，以及训练方法。接着引出Faster R-CNN检测框架，详细分析了RPN网络、FRCNN网络，非极大值抑制的原理，并解释了各个部分的损失函数，对这一目标检测框架提供了宏观和微观上的把控。

然后，基于行人检测中出现的问题，针对性的改进Faster R-CNN检测框架，提出了JPD-CNN行人检测框架，设计PRPN网络、行人区域概率热区获取、建议区域横移等改进方法，并将大型的FRCNN与小型的RPN网络相结合，实现高精度的行人检测，并阐明了各部分的实现原理。

接着，基于上部分引入的改进方法，采用MATLAB和C ，并利用外部库Caffe实现了改进算法，并详尽介绍了工程实现。

最后，在目前流行的KITTI、Caltech行人检测数据集对JPD-CNN进行训练，并验证了改进算法的效果。对比实验表明，JPD-CNN能够较好地处理行人区域小、遮挡、易判定为负样本的问题。同时，指出了存在的不足和未来可以改进之处。

关键词：深度卷积神经网络；目标检测；行人检测；行人区域推荐网络

Abstract

Pedestrian detection is a specific application of the object detection technology, which is of great significance in areas such as automatic driving and intelligent monitoring. In recent years, the technology of deep convolutional neural networks has developed rapidly. A large number of excellent object detection frameworks have been proposed in succession, and many classical methods have been put into commercial use. However, for the specific task of pedestrian detection, there are still many difficulties, such as the small size of the pedestrian target, and the insurmountable problems such as severe occlusion. Based on the deep convolutional neural network, this paper reviews the history of the development of object detection and pedestrian detection, and attempts to improve the existing pedestrian detection methods.

Firstly, this paper reviews the development of object detection field, introduces the principle of convolutional neural network, and analyzes the research status of pedestrian detection at home and abroad based on top journals and conference documents, and provides a theoretical basis for research.

Secondly, this paper introduces the evolution of the R-CNN series of detection frameworks. From R-CNN to SPP-NET to Fast R-CNN, this framework is introduced one after another. The principles of each part, the loss function of the network, and training are introduced. method. Then it introduced the Faster R-CNN detection framework, analyzed in detail the RPN network, FRCNN network, the principle of non-maximal suppression, and explained the loss function of each part, providing macro and micro control over this target detection framework.

Thirdly, based on the problems in pedestrian detection, we improve the Faster R-CNN detection framework and put forward the JPD-CNN pedestrian detection framework. It includes: (i) Designing PRPN network. (ii) Pedestrian area probabilistic hot zone acquisition. (iii) Proposing regional traverse and other improved methods. Furthermore, we combine a large-scale FRCNN with a small-scale RPN network to achieve high-precision pedestrian detection. We clarify the principle of realization of each part.