FastVehicleDetectionUsingaDisparityProjectionMethod资源-CSDN文库

138 浏览量 2021-02-08 16:25:02 上传评论收藏 2.79MB PDF 举报

资源推荐

资源详情

资源评论

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 9, SEPTEMBER 2018 2801

Fast Vehicle Detection Using a

Disparity Projection Method

Jing Chen , Wenqiang Xu, Haitao Xu, Fei Lin, Yong Sun, and Xiaoying Shi

Abstract—The stereo vision could obtain the 3-D coordinate of

the detected object by computing the disparity of the correspond-

ing image points. However, on account of the time complexity

and the low robustness of the image matching algorithm, it is

seldom used in large-scale scene. This paper puts forward a new

vehicle detection method, which simpliﬁes the massive Fourier

transformation in the image matching process. The method

converts the 2-D Fourier transformation to 1-D with the dimen-

sionality reduction of reused Fourier transformation. Meanwhile,

1-D Fourier transformation of the fast image matching model

is also derived. The coarse-to-ﬁne pyramid search strategy is

used according to the gradient information of each depth map

adaptively. The adjacent area of the same depth is obtained

with a larger matching weight which improves the matching

accuracy and robustness. The model can be used in the ﬁtting

and projection transformation of road plane after background

extraction. Thus reduce the complexity of vehicle segmentation

and enhance the robustness of vehicle detection. The experimental

results show the method can effectively achieve the large-scale

and real-time detection. It is adaptive to various illumination

changes and impervious to shadow and occlusion.

Index Terms—Stereo vision, disparity feature, points matching,

phase correlation, reusing of fourier transformation.

I. INTRODUCTION

HE stable and reliable vehicle detection is the primary

task of trafﬁc image analysis. It is taken as the basis

of vehicle count, tracking, classiﬁcation, auxiliary driving,

accident detection and road behavior judgments. In the trafﬁc

detection process, the trafﬁc information obtained by the single

traditional detector is not sufﬁcient. Various sensors are usually

needed to assist to complete a test task. In addition, due

Manuscript received March 20, 2016; revised November 14, 2016,

March 28, 2017, August 13, 2017, and September 25, 2017; accepted October

3, 2017. Date of publication November 13, 2017; date of current version

September 7, 2018. This work was supported in part by the National Science

Foundation of China under Grant 61703127 and Grant 61602141, in part

by the Zhejiang Provincial Natural Science Foundation of China under

Grant LY17F020026, Grant LY18F020013, and Grant LY12F02017, in part

by the Zhejiang Province Public Welfare Technology Application Research

Project of China under Grant 2015C33067, and in part by the Public Projects

of Zhejiang Province under Grant 2013C33082 and Grant LGF18F030006.

The Associate Editor for this paper was J. M. Alvarez. (Corresponding author:

Jing Chen.)

J. Chen, H. Xu, F. Lin, and X. Shi are with the School of Computer

Science and Technology, Hangzhou Dianzi University, Hangzhou 310018,

China (e-mail: cj@hdu.edu.cn; xuhaitao@hdu.edu.cn; linfei@hdu.edu.cn;

shixiaoying@hdu.edu.cn).

W. Xu is with the College of Economics and Management, China Jiliang

University, Hangzhou 310018, China (e-mail: xwq2009@126.com).

Y. Sun is with the Department of Information Technology, Zhejiang Institute

of Communications, Hangzhou 310000, China (e-mail: sy@zjvtit.edu.cn).

Color versions of one or more of the ﬁgures in this paper are available

online at http://ieeexplore.ieee.org.

Digital Object Identiﬁer 10.1109/TITS.2017.2762718

to the characteristics of multi-sources and different natures

of test data, the integration and fusion of multiple trafﬁc

detection data have become the bottleneck. So, the visual

trafﬁc detection method becomes popular for its capability in

obtaining multiple trafﬁc parameters.

The machine vision based vehicle detection is realized

by extracting the features of road vehicles in the images.

This detection method can get different kinds of trafﬁc para-

meters and has a wide range of application. According to

the sampling conditions, the vision based vehicle detection

methods can be divided into two categories: static appearance

features based method and dynamic motion features based

method. The appearance features based method pays attention

to external physical features, including color, texture, edge,

contour and shape. Variety feature descriptors have been used

in this ﬁeld as scale-invariant feature transform (SIFT) [1],

histogram of oriented gradient (HOG) [2] and Haar-like [3].

There are also studies which extract features to detect the

vehicles by the use of deep neural networks (DNN) [4],

support vector machines (SVMs) [5], boosting [3], condi-

tional random ﬁelds (CRF) [6] and convolution neural net-

works(CNN) [7], [8]. What the motion features based method

concerned with is how to separate the foreground moving

vehicles from the static background image. Among which the

background subtraction methods [9] are most widely used. The

main background removal methods include Kalman ﬁlter [10],

single Gaussian pixel distribution [11], [12], Gaussian mixture

model (GMM) [13], and wavelets [14]. Another method of

motion features is based on the optical ﬂow [15], [16]. It is

the instantaneous speed of image pixels, which corresponds to

the moving vehicles in 3-D space. Optical ﬂow is widely used

in vehicle detection, since it is less susceptible to occlusion

issues [17].

There are several problems in above detection methods.

(1) The shadow in the detection process will affect the

accuracy directly. To solve this problem Cucchiara et al. [18]

analyzed the pixel with hue saturation. It distinguished the

different color and brightness to detect the vehicle shadow.

Ji et al. [19] effectively removed the vehicle shadow by

looking for the boundary and detecting the color feature.

Xue et al. [20] detected vehicles with the updating feature

template with mobile information. However, this method is

only suitable for situation of low vehicle ﬂow rate. Otherwise

the test results will be different from the actual situation.

(2) If the road vehicle density is large while the angle of the

sampling camera is relatively low, the vehicle is prone to be

sheltered. In view of this situation, literature [21], [22] put

See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2802 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 9, SEPTEMBER 2018

forward part-based detection models. This model decomposes

the vehicle into several parts. By studying the feature rela-

tionship between each part, the vehicle can be detected by the

features of other parts in the case of being partially sheltered.

However, this method is not suitable for the multi-scale vehicle

detection in the remote scene. (3) The inconformity of image

size resulted from the different distances between the vehicles

and the camera may lead to inaccurate feature extracting

for the reason of the resolution of long distance vehicle.

Aiming at this problem, literature [23], [24] detected vehicle

which is far from camera and sheltered by front vehicle by

introducing the local features and global features to describe

the vehicle model. Although this method can solve the multi-

scale problem in vehicle detection, only 2-D trafﬁc sampling

information can be obtained because of the limitation of the

monocular visual image detection method, which leads to the

loss of depth information of trafﬁc parameters. There may be

problems in the following information quantization process

including vehicle type, vehicle speed, trafﬁc ﬂow, and the

length of the motorcade.

The detection method based on stereo vision expands the

feature vehicle detection to 3-D space, which provides a good

way to solve these problems. A Real-Time system introduced

in [10] initializes the object list by using the 3-D model, which

could be used in the detection and tracking of vehicles in the

intersection. The work in [25] represents the dynamic scene

as a collection of rigidly moving planes with 3-D models.

Song et al. [26] solves the shadows under the conditions

of known lighting by using a Bayesian framework with

MCMC sampling. In recent years, the arisen stereo based

method [27]–[29] provides a good solution to the occlu-

sion problem encountered in the detection process, which

allows the detection process to avoid vehicle segmentation.

Vatavu et al. [30] puts forward a vehicle detection method

based on stereo vision, which could realize multiple vehicles

detection in a crowded environment where the road signs

are not visible. The vehicle detection method by utilizing

the feature of the road and disparity histogram is presented

by Lee et al. [31], which is based on the stereo method,

including obstacle segmentation, obstacle location and vehicle

veriﬁcation.

The corresponding points matching algorithm is the key

to the stereo vision method. Points matching algorithm

can be split into two categories: frequency domain match-

ing method [32], [33], [34] and time domain matching

method [35], [36]. Time domain matching method according

to regional gray or color is the most direct and simplest

method to solve the matching problem of corresponding

points. Take one point as center in an image, and then its

small window is selected. Afterwards, area, which has the

maximum similarity with above zone, is sought in another

image. The center of this area is regarded as corresponding

point. The commonly used time domain matching methods

are Sum of Absolute Differences (SAD), Sum of Squared

Differences (SSD), Normalized Cross Correlation (NCC) [37]

and so on Although the idea of area-based method is simple,

it is sensitive to noise. It may lead to larger difference of each

point pair due to inﬂuence factors such as illumination, camera

sampling angle, cameras synchronization and so on. Given

this, some cost methods combining with multiple features

are proposed, which improve the aforementioned problems

remarkably. In recent years, the matching method based on

feature learning is used to improve the illumination robustness

and matching precision. Luo et al. [38] proposes a matching

network which is able to produce very accurate results in

less than a second of GPU computation. Seki et al. [39]

leverages Convolutional Neural Networks to overcome the

limited accuracy of conventional learning based conﬁdence

measures.

The above problem can also be solved by frequency

domain matching method. When neighborhood of correspond-

ing points are in line with shift assumption, shift estima-

tion of time domain image will be converted into frequency

domain by phase matching algorithm based on Fourier Trans-

lation [32]. It may greatly enhance robustness of matching

process for environmental changes like illumination [40]. [41]

proposed a high-accuracy sub-pixel method to estimate image

disparity with 1/100-pixel accuracy. [33] compared the robust-

ness against transformation complexity, noise, missing data,

and multiple motions of POC. POC algorithm is insensitive

to image illumination differences. And it is more robust than

region-based matching algorithms. So it is used increasingly

in high-precision dense 3-D points cloud matching in recent

years [42], [43]. Such algorithm has higher matching precision,

but it is time consuming in case of larger matching quantity.

It takes 3 times of time-frequency convertible computation for

each corresponding point.

In summary, although the method of 3-D model is effective

in solving the shelter and shadow, there are still some problems

in the detection method based on stereo. Firstly, the method

of establishing the 3-D scene by using natural lighting condi-

tions needs to overcome the changeable natural environment.

Secondly, although the corresponding points matching algo-

rithm based on frequency domain is good for illumination

changing robustness, the bottleneck of the algorithm is cor-

responding to the matching points.

Because of the huge amount of dense points cloud matching,

the matching efﬁciency cannot meet the needs of real-time

sampling. Although there are a large amount of Fast Fourier

Transform (FFT) operations required by the stereo matching,

we ﬁnd that there are a lot of repetitive computations in the

FFT operation. The corresponding matching calculated amount

can be decreased greatly by reducing the dimensionality and

reusing. To overcome the drawbacks of the previous methods,

this paper proposes stereo vision matching algorithm for large

scene of road sampling. This method extracts the dynamic

background by using Gaussian mixture model, and improves

the matching efﬁciency of corresponding points by reusing

FFT model. Thus the disparity could be calculated rapidly to

obtain the 3-D coordinates of the foreground. The foreground

object is projected on the ﬁtting and sparse road plane to avoid

the complex vehicle segmentation.

This paper has the following contributions:

(1) Meeting requirements of vehicle detection in complex

environment, and solving the problems of occlusion, shadow

interference, low resolution of remote sampling resolution.

剩余12页未读，继续阅读

评论收藏

内容反馈

weixin_38626984

粉丝: 5
资源: 921

Fast Vehicle Detection Using a Disparity Projection Method

vehicle detection

Vehicle-Detection:车辆检测

vehicle_detection

Qt 5实现串口调试助手 （源工程文件、0积分下载）

【SystemVerilog】路科验证V2学习笔记（全600页）.pdf

AutoSAR标准协议4.2.2

光伏-储能并网系统仿真.rar

XCP协议的规范文档

GD32替换STM32注意事项.pdf

NPPJSONViewer.zip

蓝牙BLE协议中文版.pdf

CANoe通过CAPL脚本实现自动测试

电路分析基础第二版PDF电子书免费下载

qt样式表一键生成（花狗Fdog）

Tangent免费.rar

CMSIS-DAP使用说明及驱动.rar

VS2015安装证书，JavaScript_ProjectSystem.msi，JavaScript_LanguageService.msi

BaiduOCR.zip

Elsevier期刊word模板.zip

rpa拆包工具（小白适用）

电气类的visio模版元件库

软件需求规格说明书模板(超详细).doc

数字设计和计算机体系结构第二版奇数答案.pdf.zip

EPLAN部件库（正泰）

ISO14229汽车诊断协议文档

单相Boost功率因数校正（PFC）仿真（Simulink & Saber）

ISO26262汽车功能安全协议文档

A1点阵喷码圆点.ttf

matpower5.0b1.zip

最新资源

Qt 5实现串口调试助手（源工程文件、0积分下载）