摘要行人检测是目标检测的一个分支,是计算机视觉领域的一个热点问题,在实际应用中具有很大的价值,尤其在交通部门的监控领域工作中,对提高工作效率以及减轻工作量有着巨大的作用。本文采用INRIA数据库和监控场景的图像集进行行人检测算法研究,具体工作如下。42166
首先介绍行人检测的相关背景以及发展状况;其次介绍了经典的HOG+SVM行人检测算法,接着本文实现一种基于可变形部件模型(DPM)的行人检测系统以及相关改进算法,该算法利用根滤波器、部件滤波器构建了一个包含目标全局信息和细节信息的星型模型,最后基于DPM算法存在的重复检测问题,在后处理环节进行了改进,并基于INRIA数据库和监控场景的图像集对算法进行实验比较分析,实验结果表明算法相对于Cascade_DPM算法,本文的改进在保持原有时间复杂度的基础上平均精度有所提高。
毕业论文关键词 行人检测,梯度方向直方图,支持向量机,可变形部件模型
毕业设计说明书外文摘要
Title Research on Pedestrian Detection under Surveillance Scene
Abstract
As a branch of object detection, pedestrian detection is one of research topics in the area of computer vision. It plays an important role in practical application, especially in the area of transport, which improves the work efficiency and at the same time reduces labor. The study focus on the pedestrian detection based on the image dataset of INRIA and surveillance scene. The main work are as follows.
Firstly, the paper presents a simple introduction of pedestrian detection. Then a traditional algorithm in pedestrian detection, HOG + SVM, is described. Next this paper realizes a pedestrian detection system and an improved algorithm based on deformable part model (DPM), which was constructed using a star model containing both global and local information with root filters and part filters. Finally, Towards the problem of repeating detection, this paper made a little improvement during the post-processing stage. The experimental results based on INRIA dataset and surveillance scene dataset indicates that the post-processing stage made some improvement in the average precision compared to the Cascade_DPM and at the same time kept the time increased only a little.
Keywords pedestrian detection, HOG, SVM, deformable part model
目 次
1 引言 1
1.1 课题背景及相关发展…1
1.2 本文使用图像数据库介绍…1
1.3 本文内容安排3
2 基于HOG的行人检测算法…4
2.1 HOG特征 4
2.2 支持向量机(SVM) 5
2.2.1 几何距离和函数距离5
2.2.2 最大距离准则…6
2.3 实验结果…8
2.4 本章总结9
3 基于DPM的行人检测算法… 10
3.1 DPM行人检测算法10
3.1.1 基于HOG特征的改进…10
3.1.2 模型11
3.1.3 隐性SVM…14
3.2 Cascade_DPM行人检测算法14
3.2.1 星型模型检测器的改进14
3.2.2 临界值计算…16
3.3 实验结果…17
3.4 本章总结…19
4 基于Cascade_DPM算法的改进…21
4.1 行人检测中的评价参数…21
4.2 基于Cascade_DPM算法的改进22
4.3 实验结果比较…24
4.3.1 时间复杂度比较…24
4.3.2 算法平均精度比较24
4.4 本章总结 24
结论 …26
致谢 …27
参考文献… 28
1 引言