Adaboost与SVM结合的级联分类器在行人检测中的应用

需积分: 9 89 浏览量更新于2024-09-11 收藏 641KB PDF 举报

"这篇论文提出了一种改进的级联AdaBoost分类器，即级联AdaBoost-SVM分类器，用于行人检测。通过结合Adaboost和SVM算法，实现了一个实时的单摄像头行人检测系统。该系统首先用固定大小的窗口捕获行人候选区域，然后对这些区域进行特征提取，利用基于Haar-like矩形特征的计算方法，最后通过提出的级联AdaBoost-SVM分类器完成行人检测。这种级联AdaBoost-SVM分类器能够自适应地调整级联分类器的数量，根据训练集有效地构建级联结构。实验在捕获样本数据库和PETs数据库上完成了行人检测验证。" 本文的核心知识点包括： 1. **Adaboost算法**：Adaboost（Adaptive Boosting）是一种集成学习方法，通过迭代多次弱分类器并赋予不同权重，将多个弱分类器组合成一个强分类器。在每个迭代中，Adaboost会更重视前一轮被错误分类的数据，使得下一轮的弱分类器能更好地关注这些困难样本。 2. **级联分类器**：级联分类器是一种由多个分类器组成的结构，其中每个分类器依次过滤输入，减少后续分类器的工作量。在行人检测等应用中，级联分类器通常用于快速排除非目标区域，只将少数可能包含目标的区域传递给更复杂的分类器，提高处理速度。 3. **支持向量机（SVM）**：SVM是一种二分类模型，它的基本模型是定义在特征空间上的间隔最大的线性分类器，间隔最大使它有别于感知机；SVM还包括核技巧，这使它成为实际上的非线性分类器，具有很好的泛化能力。 4. **行人检测**：行人检测是计算机视觉领域的一个重要任务，其目的是在图像或视频中自动识别出行人。该论文中，通过固定大小的窗口捕获可能包含行人的区域，然后进行特征提取和分类，实现检测。 5. **Haar-like特征**：Haar-like特征是用于对象检测的简单特征，常用于人脸、行人等的特征提取。它们由矩形结构组成，可以表示图像中的边缘、线段和区域差异。 6. **自适应调整级联分类器数量**：级联AdaBoost-SVM分类器可以根据训练数据动态调整级联的分类器数量，以达到最优的分类效果和效率平衡。 7. **实验验证**：论文通过实际的捕获样本数据库和PETs（Pedestrian Detection Evaluation Test Set）数据库进行了行人检测实验，以证明新方法的有效性和准确性。 8. **实时性**：级联AdaBoost-SVM分类器的设计考虑了实时性，能够在单一摄像头系统中实现实时的行人检测，这对于自动驾驶、监控系统等应用场景至关重要。

A Cascade Classifier Using Adaboost Algorithm and

Support Vector Machine for Pedestrian Detection

Wen-Chang Cheng

Dept. Computer Science and Information Eng,

Chaoyang University of Technology, Taichung, Taiwan,

wccheng@cyut.edu.tw

Ding-Mao Jhan

Dept. Computer Science and Information Eng,

Chaoyang University of Technology, Taichung, Taiwan,

s9827643@cyut.edu.tw

Abstract— In this paper, we improve cascade-Adaboost classifier

and propose a cascade-Adaboost-SVM classifier. It is combined

with Adaboost and SVM and real-time pedestrian detection

system with a single camera. We capture the pedestrian candidate

areas with a window of fixed size, conduct feature extraction to

candidate areas and mobile images with Haar-like rectangle

feature calculation and then, complete pedestrian by using the

proposed cascade-Adaboost-SVM classifier. As this

cascade-Adaboost-SVM classifier can adjust numbers of cascade

classifiers adaptively, it can construct cascade classifiers

effectively based on training set. Finally, we complete the

pedestrian detection experiment with the database of captured

samples and PETs database. The experimental result shows that

the cascade classifier proposed by us can get better performance

than cascade-Adaboost classifier and its accuracy can reach

99.5% and the false alarm rate is less than 1e-5.

Keywords-Background subtraction; Human detection; Object

recognition; Haar-like feature; Ensemble classifier

I. INTRODUCTION

Pedestrian detection based on computer vision acts an

important role in many applications, for example driver

assistance system, surveillance system and intelligent robot

[1-3]. The driver assistance system can avoid the collision

through pedestrian and route detection to guarantee the safety

of drivers and pedestrians effectively. For surveillance system,

the false operation caused by shadow changes and swaying of

trees can be avoided by using pedestrian detection to activate

camera to monitor and record, which can save labor cost and

reduce the space to store the videos. Secondly, route of

pedestrians can be predicted and behavior of pedestrians can

be analyzed by tracking pedestrians, so as to prevent accidents

actively. If accidents are unavoidable, the prediction and video

of pedestrians’ route can help the police know information of

suspects quickly to increase the possibility of case solving. For

the robot applications, pedestrian detection can be used for

interactivity of human-computer interface and can provide

information about location of surrounding pedestrians and

offer real time service when necessary. It can be known from

the description above that the pedestrian detection is being

actually applied to many fields.

The real-time pedestrian detection system is currently a

hot research subject. Many methods have already been

proposed [1-12]. A pedestrian detection system generally

includes three parts, namely, object segmentation, feature

extraction, and classification. According to the use of camera,

they can be divided into visible light camera and invisible light

camera. According to camera architecture, they can be divided

into single camera and multiple cameras. The cost of invisible

light equipment is generally much higher than that of visible

light camera. Multiple cameras require more equipment and

the cost is relatively high, therefore we mainly discuss the

pedestrian detection method using single visible light camera

in this paper. Based on whether or not to adopt the background

image method, object segmentation can be classified roughly

into two methods. The first is to use the background image

[1-3, 11], which helps us to get the foreground image by

subtracting the background image from the real-time input

image. Then, it segments the object according to the connected

parts on the foreground image. Finally, it performs feature

extraction and classification identification. This method needs

background reconstruction and technology update. This is

unsuitable for mobile platforms, such as driving safety

assistance systems and intelligent robots. The second is not to

adopt the background image method [1-9, 12], which segments

the input image into some image blocks with the same window

size. Then, it makes feature extraction to all image blocks.

Finally, it identifies whether the blocks belong to a pedestrian

according to the extracted feature. The advantage of this

method is that it does not need extra time for reconstruction

and update of the background image. This is suitable for

mobile platforms.

Viola and Jones [12-14] used Haar-like method for the

local feature of the pedestrian shown in the image blocks.

Moreover, they took an Adaboost classifier [15-16], which has

many weak classifiers forming a powerful one, to identify the

pedestrians. Each weak classifier makes the classification for

only one dimension within the feature vector. For a few weak

classifiers, this method can effectively reduce the complexity

and time of the calculation, but the accuracy obtained is low.

For many weaker classifiers, it provides high accuracy but

requires a long calculation time and still has high false alarm

rates. To solve the problem, Viola and Jones further proposed

cascade Adaboost classifier which is composed by the

cascading of many Adaboost classifiers. Assume there are L

layers in this classifier and the detection rate and false alarm

rate of each layer are d

and f

respectively. The detection rate

and the false alarm rate of the whole cascade-Adaboost

classifier can be defined as D=(d

)

and F=(f

)

respectively.

For example, if the detection rate d

and the false alarm ratef

are set as 99.9% and 30.0% respectively in all layers, then the

whole detection rate D and the whole false alarm rate F will be

(0.999)

> 0.99and (0.3)

< 1e-5, respectively, for training set.

However, when we conducted the cascade-Adaboost classifier,

we found that Adaboost classifiers in front layers could reach

preset targets with less weak classifiers. But those in rear

下载后可阅读完整内容，剩余5页未读，立即下载

Acala_Deng

粉丝: 0
资源: 1

Adaboost与SVM结合的级联分类器在行人检测中的应用

Real-time vehicle detection with foreground-based cascade classifier

Android代码-OpenCV – Cascade Classifier (オブジェクト認識)

Opencv 下Cascade Classifier实现人脸识别的例程

cascade-classifier-trainer:使用Haar功能为AdaBoost训练OpenCV级联分类器的脚本

Face-eye-detection-using-Haar-Cascade-classifier

Quantum dot quantum cascade photodetector using a laser structure

Rapid Object Detection using a Boosted Cascade of Simple

Rapid Object Detection using a Boosted Cascade of Simple Features

Rapid Object Detection using a Boosted Cascade of Simple Features.pdf

opencv 下 cascade classifier 实现人脸识别的例程 csdn

最新资源