摘 要

近几年来，行人检测一直是计算机视觉领域的一个热门主题。本文针对行人检测这一主题，对比研究了传统的基于机器学习的方法和基于深度学习的方法。

首先，研究了基于机器学习的算法，分析了适合于行人检测的特征和分类器，这里重点比较了Haar特征和HOG特征，实验结果表明，HOG特征检测速度快，而Haar特征因其特征维数高，表示行人特征更细腻。对于分类器，本文研究了Adaboost和SVM分类器，对比研究两种基于机器学习的行人检测算法Haar Adaboost和HOG SVM，研究结果表明，对于MIT数据库，Haar Adaboost算法检测效果并不好。对于MIT训练集，HOG SVM算法检测率为84%，且检测速度为17ms每帧。对HOG SVM算法的进一步研究表明，训练集数量与检测率和训练时间基本不相关，仅与测试时间成正相关，其检测率最高可达93%，此时其平均检测速度在20ms每帧左右，故本文选这种算法与基于深度学习的算法进行比较。

其次，研究了基于深度学习的神经网络算法，先给出了用神经网络进行模式检测和用BP算法训练参数的过程。接着选择了卷积神经网络用于行人检测，研究了其卷积层、采样层及全连接层的特征。基于VOC2007行人检测数据库在GTX1080上进行实验测试，结果表明，Faster-RCNN相比RCNN和Fast-RCNN，性能有了大幅提高，检测率达到了73%，检测速度为0.2秒每帧。

最后，本文在华硕Y581L比较了Faster-RCNN和HOG SVM，发现前者在检测率上达到99%，优于后者的最高检测率，这体现了基于深度学习方法的优势。但前者检测速度为6.295秒每帧，大幅慢于后者，说明基于深度学习的行人检测算法需在检测速度上做进一步改进。

关键词：行人检测；深度学习；卷积神经网络；Haar特征；支持向量机；Faster-RCNN

Abstract

In recent years, pedestrian detection has been a hot topic in the field of computer vision. In this paper, the traditional method of machine learning and the method based on deep learning are compared with the subject of pedestrian detection.

Firstly, the algorithm based on machine learning is studied, and the characteristics and classifiers suitable for pedestrian detection are analyzed. The characteristics of Haar and HOG are compared. The experimental results show that the HOG feature is fast and the Haar feature is characterized by its characteristic dimension High, that pedestrian characteristics more delicate. For the classifier, this paper studies the Adaboost and SVM classifiers, and studies two Haar Adaboost and HOG SVM based on machine learning. The results show that Haar Adaboost algorithm is not good for MIT database. For the MIT training set, the HOG SVM algorithm has a detection rate of 84% and a detection speed of 17 ms per frame. Further research on the HOG SVM algorithm shows that the training set is basically independent of the detection rate and training time, and only with the test time is positive correlation, the detection rate of up to 93%, this time the average detection speed of about 20ms per frame , So this algorithm is chosen to compare with the algorithm based on deep learning.

Secondly, the neural network algorithm based on deep learning is studied. Firstly, the neural network is used to detect the parameters and the BP algorithm is used to train the parameters. Then, the convolution neural network is selected for pedestrian detection, and the characteristics of convolution layer, sampling layer and all connection layer are studied. Based on the VOC2007 pedestrian detection database, the experimental results show that Faster-RCNN has a significant improvement in performance compared with RCNN and Fast-RCNN. The detection rate is 73% and the detection speed is 0.2 seconds per frame.

Finally, this paper compares Faster-RCNN and HOG SVM with ASUS Y581L, and finds that the former achieves 99% of the detection rate, which is better than the highest detection rate of the latter, which reflects the advantages of deep learning method. But the former detection speed of 6.295 seconds per frame, significantly slower than the latter, indicating that deep learning based on the pedestrian detection algorithm to be further improved in the detection speed.

Key Words：Pedestrian recognition; deep learning; convolution neural network; Haar feature; support vector machine; Faster-RCNN

目 录

第1章 绪论 1

1.1 课题研究的背景、目的和意义 1

1.2 国内外研究现状 1

1.3 本文主要工作 3

1.4 论文组织结构 3

第2章 基于机器学习的行人检测 4

2.1 特征提取 4

2.1.1 Haar特征提取 4

2.1.2 HOG特征提取 6

2.1.3 Haar特征与HOG特征对比实验与分析 7

2.2 分类器检测过程 8

2.2.1 Adaboost分类 9

2.2.2 SVM分类 10

2.3 基于机器学习的行人检测对比实验与分析 10

2.3.1 实验样本选择及其标签制作 11

2.3.2 基于Haar Adaboost的行人检测实验与分析 12

2.3.3 基于HOG SVM的行人检测实验与分析 13

2.4 本章小结 14

第3章 行人检测神经网络搭建 15

3.1 人造神经元实现 15

3.1.1 感知器算法 15

3.1.2 Sigmoid神经元 16

3.2 神经网络的构建与训练 16

3.2.1 基于神经网络的模式识别 16

3.2.2 基于BP算法的神经网络训练 17

3.3 神经网络类型的选择及其基本结构研究 19

3.3.1 神经网络类型的选择 19

3.3.2 卷积神经网络的基本结构研究 19

3.4 本章小结 21

第4章 基于卷积神经网络的行人检测 22

4.1 卷积神经网络研究 22

4.1.1 RCNN检测过程 22

4.1.2 Fast-RCNN检测过程 23

4.1.3 Faster-RCNN检测过程 26

4.2基于卷积神经网络的行人检测的对比实验与分析 28

4.2.1 RCNN、Fast-RCNN与Faster-RCNN对比实验与分析 28

4.2.2 Faster-RCNN与HOG SVM对比实验与分析 29

4.3 本章小结 30

第5章 总结与展望 31

5.1 总结 31

5.2 展望 32

参考文献 33

致谢 35

第1章 绪论

1.1 课题研究的背景、目的和意义

近年来，随着计算机的普及以及自动化控制技术的高速发展。计算机视觉（computer vision）作为使用计算机对视频或图像进行检测和处理的技术，在工业生产和日常生活中起着越来越重要的作用。

如今，随着无人驾驶技术的发展，行人检测成为其中一个重要研究方向。对于无人驾驶技术而言，能够准确的检测出行人，就能为无人驾驶系统的控制提供重要的信息。