Weather Recognition Bad on Images Captured by Vision System in Vehicle
Xunshi Yan1,Yupin Luo1,and Xiaoming Zheng2
1Tsinghua National Laboratory for Information Science and Technology(TNList), Department of Automation,Tsinghua University,Beijing100084,China
2INF Technologies,Ltd.,Beijing100086,China
yanxs06@mails.thu.edu,luo@tsinghua.edu,zheng@tsinghua.edu
Abstract.Weather recognition is widely required in many areas,and
it is also a challenging and brand-new subject.This paper propos an
approach to recognize weather bad on images captured by in-vehicle
vision system.We bring three groups of features,including histogram of
gradient amplitude,HSV color histogram,road information,and employ
an algorithm bad on Real AdaBoost,making u of the category struc-
ture to achieve the task of classification.Experiments confirm superior
performances on our datat collected from images captured by vision
system.
Keywords:Weather recognition,Vision system,Real AdaBoost,HSV
color space,Category structure.
1Introduction
Computer vision system has achieved great success in many areas,such as surveillance,navigation,driver assistance system.However,the cameras expod outside are easily influenced by bad weather.For example,pedestrian detection system in vehicle could not work at all when raindrop falls on the camera,even results rious fal-detection.Many vision systems also need to ret parameters such as lighting,rain wiper,under different weather conditions.Hence,rearch of weather recognition in vision system is in urgent demand.
Weather recognition is a brand-new subject and only a few of previous work has addresd this issu
出国留学艺术
e.Garg and Naya[1]in Columbia University focus on detecting and removing rain streaks from videos.The idea comes from moving object detection.It makes a difference between the two adjacent frames,and can give some perfect results under certain scenarios,but it is hard to satisfy the dynamic background or the situation of raindrops adherent on the cam-era.Kurihataand and Takahashi[2][3]in Nagoya University collect large mount of raindrop patches,utilize principal component analysis to make a raindrop template.They arch the global or part of images by computing the similar-ity between the raindrop template and the patches in the images.If detecting enough similar patches,the images can be identified as a rainy image.However, when there is no raindrop falling on the camera,it doesn’t work.It also facedconspiracy theory
W.Yu,H.He,and N.Zhang(Eds.):ISNN2009,Part III,LNCS5553,pp.390–398,2009.
c Springer-Verlag Berlin Heidelberg2009
Weather Recognition Bad on Images Captured by Vision System391 misclassification by some objects such as lamps.Unfortunately,both of work focus on detecting rain,and don’t put energy into different weather recognition and give a whole recognition result.
The contribution of this paper is to propo three groups of features according to images in different
weather conditions,and give an algorithm derived from Real AdaBoost,which makes full u of category structure.
The rest of paper is organized as follows.In Section2,we propo three groups of features by analyzing the images in different weather conditions.Real AdaBoost is introduced and the category structure is prented in Section3. Section4shows perfect effect of our algorithm on our datat captured by vision system in vehicle.
2Feature Selection阿加莎克里斯蒂小说
For any pattern recognition problem,it is important to lect proper features. Weather recognition from images is different from general image classification tasks.We always implement the image classification task by lecting interesting points as features or detecting the object emerging in the scene.It is impractical for our project becau under different weather conditions there can be same objects and interesting points.Hence,applying the same kind of features as general image classification tasks is not proper.We propo typical features in low vision level by analyzing the property of images under the different weather conditions.
2.1Histogram of Gradient Amplitude(HGA)
The images under different weather conditions take on different degrees of blur. In sunny days,it always makes the images sharper whereas blurred in rainy days.Especially when raindrop covers the camera,the images are always more blurred and the values of pixels in images areflatted.Gradient is a perfect tool of measuring sharpness[4][5].Generally,the larger the gradient is,the more it is possible to be sunny.We compute the amplitude of gradient according to(1) and form a histogram of gradient amplitude.
M(x,y)=
G x(x,y)2+G y(x,y)2.(1)
Fig.1shows sunny and rainy images and their corresponding histograms of gradient amplitude.Wefind that the distribution of histogram is different.It is flatter for rainy images than sunny images.There are more low value pixels in rainy images and more high value pixel in sunny images.
2.2Histogram of HSV Color Space(HSV)
According to our obrvation,brightness value is high in sunny images,and low in rainy images.There are more vivid pixels under the condition of sunny days and in
392X.Yan,Y.Luo,and X.Zheng
理学学士
Fig.1.(a)A sunny image;(b)Histogram of gradient amplitude of(a);(c)A rainy image;(d)Histogram of gradient amplitude of(c)
欧美一对一外教
Fig.2.ROI is surrounded by the red rectangle and the lected points is denoted by the red dots
contrast in rainy days[6].It is also corresponding to human knowledge.Hence,HSV color space is thought to be a good measurement tool for weather classification.We convert the image into the HSV color space,portion the hue,saturation,value into 2,3,5bins parately,and form a histogram as a group of feature.
Weather Recognition Bad on Images Captured by Vision System393 2.3Road Information(Road)
Not every patch of image contains the discriminative information for classifica-tion.Some areas of image contain more discriminative information than others and we often call it Region of Interest(ROI)[7].In images captured by camera in vehicles,the road surface is always distinctive to human eyes,and we choo the central area as ROI.10points are lected in the road area,and the mean of gray value in11×11panes which are centered at the lected points is calculated.The ten values form a vector as a group of feature.Fig.2shows the ROI which is surround by the red rectangle and the lected points are denoted by the red dots.ROI area is discriminative,but can not include all information in the images.Therefore,combining the global features and local features are our choice.
3Recognition Algorithm
In this ction,we describe our algorithm in detail.Real AdaBoost is introduced briefly in thefirst part.The category structure in weather recognition is propod in part2.
3.1Real AdaBoost
AdaBoost is a powerful algorithm in pattern recognition and is widely ud in the past ten years.Different from Support Vector Machine,Artificial Neural Network and Nearest Neighbor,AdaBoo
st combines many weak learners and forms a strong classifier.It has high accuracy and is resistant to overfitting. We choo Real AdaBoost[8][9]as our basic algorithm,which is shown to be better than discrete AdaBoost in most situations.We reproduce the algorithm procedure as follows.Suppo S={(x1,y1),(x2,y2),(x3,y3),···,(x N,y N)}as sample space,where x∈X are feature vectors and y∈{−1,+1}are labels.
D(i)=1/N,i=1,2,···,N is the initial distribution.
For t=1:T
四级英语作文万能句子
Step1:Run a CART(Classification and regression tree)and get a best weak learner,and divide the sample space X into X1,X2···X m.
Step2:Under D(i),compute
p j l=P(x i∈X j,y=l)
=
i:x i∈X j∧y i=l
D t(i)l=±1.(2)
Step3:Set weak learner∀x∈X j,h t(x)=12ln
p j+1+ε
p j−1+ε
,where j=
1,2,···,m,εis a small positive number.
Step4:Refresh the sample weights D t+1(i)=D t(i)exp[−y i h t(x i)]
Z t ,where
Z t is a normalized constant.
394X.Yan,Y.Luo,and X.Zheng Thefinal strong classifier is
H(x)=sign
T
t=1
h t(x)
.(3)
3.2Category Structure in Weather Recognition
In general image classification issues,there is no logical relation between class and all class are thought to be independent.As shown in Fig.3(a),the data of three class overlap each other and we have to u three two-class classifiers to carry out the task.However,if the data of the extracted features from different class distribute along a one-dimension ,a curve and the bound-aries of the classifiers are approximately parallel in feature space as shown in Fig.3(b),we can think there exists a special category structure and it will be helpful for developing some simple algorithms for multi-class tasks.
Fig.3.Three different color circles or ellips reprent three different class data distribution
Our weather recognition issue under the extracted features is approximately agreed with the situation in Fig.3(b).The special category structure suggests that when a sample is misclassified into a rainy sample by Sunny-Rainy classifier, it has a higher probability to be a cloudy sample while not a sunny sample. Likewi,if a sample is misclassified into a sunny sample,it is more likely to be a cloudy one.That inspires us to devi an algorithm bad on this category structure.We describe the algorithm as follows,which is shown in Fig.4.
Begin:
Step1:Train the Sunny-Rainy,Rainy-Cloudy,Cloudy-Rainy classifiers.
Step2:Input the testing samples into Sunny-Rainy classifier,and get the temporary label L temp.
Step3:If L temp is sunny,put the sample into the Sunny-Cloudy classifier and give thefinal label L fina
l which belongs to sunny or cloudy.If L temp is rainy,put the sample into the Cloudy-Rainy classifier and give thefinal label L final which belongs to cloudy or rainy.
承担的英文End.
Weather Recognition Bad on Images Captured by Vision System395
Input
Fig.4.The procedure of our algorithm bad on category structure One-vs-all makes the most basic algorithms suitable for multiple class prob-lems.Suppo there are altogether K categories.One-vs-all need train K two-class classifiers and samples should also be tested in K classifiers.If the problem has the category structure what we have analyzed,our algorithm is much simpler than one-vs-all,becau samples are only tested in log2K classifiers.Hence,
our algorithm is faster and it also has approximately the same recognition rate as one-vs-all.The larger K is,the more efficient our algorithm is.For example, when K=8,each sample is tested in8classifiers by using one-vs-all,while our algorithm only need to test3times in the ca offinding a similar category structure.Our algorithm is easy to be extended to more weather categories in future.
4Experimentsxianyan
We applied the above features and algorithms to weather recognition problem. We try to learn three concepts(Sunny,Cloudy,Rainy)and2496images are collected as databa which compris training t and testing t.All the im-ages are acquired from videos captured by vision system in vehicle.Different scenes(e from Fig.5)such as city street,highway,overpass are included in our databa.The number of images in each category are shown in Table1.The training t is randomly lected from databa.Each experiment is repeated20 times and the average result is reported.The number of nodes in CART weak learner is t3,and the number of iteration in AdaBoost is200.We run our experiment in C++code on a CPU of AMD Sempron3200+with1.5G RAM.
First we evaluate the effect of the three groups of feature.The result is shown in Fig.6.HSV is the best feature,which can be as high as even exceeding90%. HGA is also perfect and able to reach80%at its peak.Road performs not better
Table1.Training and Testing t for our experiments
Category Sunny Cloudy Rainy
N training400400400
desrt怎么读
英语资料N testing372272652
