Recognition and Tracing Scheme Study of Moving Objects by Video Monitoring System

Transcription

1 85 Recogniion and Tracing Scheme Sudy of Moving Ojecs y Video Monioring Sysem Peilong XU 1 1 The Growing Base for Sae Key Laoraory, Qingdao Universiy, No. 308, Ningxia Road, Qingdao , China. Asrac Ojecive: In his paper a recogniion and racing scheme for moving ojecs y video monioring sysem was sudied. Mehods: During moving ojecs recogniion, Muli frame sampling mehod was used o esalish he iniial ackground. Edges of he moving ojecs were drawn according o he changes of he images, and he influence facors for edges drawing were eliminaed. In he end, racing for moving ojecs could e realized y recogniion of morphological characerisics. Resuls: The experimens indicae ha, for field environmen, he numer of colleced frames eween 120 o 180 could ge eer image ackground. The shadows of he moving ojecs are he main facor which influence deecion of ojec edges, and hey could e eliminaed y Shadow edge deecion operaor. For vehicles racing, adjacen frame maching mehod could e used o reflec he ime-space ransformaion of he vehicles. Conclusion: This scheme could realize he recogniion and racing of moving ojecs y video monioring sysem effecively. Keywords: Image Recogniion, Image Tracing, Video Monioring Sysem. 1. Inroducion Video monioring sysems are widely used in fields like indusry, ranspor, securiy and miliary, and are playing more and more imporan roles. Wih ime progressing, he video monioring sysems have even reached scales of hundreds and housands ways. I is impossile o rely enirely on human o monior so many sysems. In his case, all kinds of inelligen monioring sysem emerge as he imes require. Nowadays, he monioring sysems could realize auo alarming of he invasion ojecs in cerain disances hrough all kinds of sensors and image recogniion sofware[1]. Bu his is far from enough. How o realize recogniion and racing of he moving ojecs y video monioring sysem has ecome an imporan prolem o solve. For recen years, he ojecs recogniion and racing echniques have made grea progress. In many rands of digial cameras, face recogniion funcion has een developed, and he face could e posiioned and e focused auomaically. Bu hese echniques could jus realize recogniion and comparison of some prese specific shapes, and hen rack hem. They echniques sill could no realize recogniion and racing of he muli ype and muli angle ojecs. This sudy developed a inelligen monioring sysem scheme according o sociey video monior needs, which could recognize and real imely race he arges. This scheme realized classificaion saisics and racing of he moving ojecs wih high recogniion rae. 2. Mehod for Realizing he Sysem Tradiional arge segmenaion algorihms are mainly ased on Ieraion hreshold segmenaion algorihm. The general progress is o deec edges and acquire difference images, hen wo value he daa hrough hreshold segmenaion so as o highligh he pars in he images we are ineresed in. Bu his mehod usually needs large amouns of compuaion, and demands highly ime complexiy[2]. So i is no suiale for real ime analysis for images of he moving ojecs. In order o ensure he real ime of he sysem reacion, his sudy used ackground suracion algorihm o realize quick and effecive segmenaion of moving ojecs. The key of he algorihm is ha how o esalish he ackground models. 2.1 Background suracion algorihm Usually, here are wo kinds of ackground developmen scheme for using ackground suracion algorihm. Firs, selec wo frames of adjacen images, and pu he former image as ackground of he laer one. Deec changes of he wo images using differenial operaion. Second, prese an unified ackground, and deec image changes hrough comparing all acquired real ime images wih he prese ackground image y differenial operaion. For he firs scheme, i is no need o prese ackground, so ha i is more suiale for moving monioring faciliies. Bu he arges volume i could monior is relaed o he arges movemen speeds. If he arge sopped moving afer coming ino monioring views, hen i could no e deeced y he monioring sysem. This is called "Hole" phenomenon in video monioring. For his reason, he firs

2 86 scheme is no suiale for video monioring sysems wih fixed camera. The second scheme could cusomer he prolem of "Hole" phenomenon caused y he algorihm in he firs scheme and is eer for racing moving ojecs, u his scheme need monioring sysems o esalish and real ime updae ackgrounds y hemselves. Through analysis of he aove wo schemes, our sudy ry o esalish a ackground suracion algorihm wih an uniformed ackground, and ased on his algorihm o se up a monioring sysem. Since he ackground qualiy has grea influence on he arges recogniion and racing, his sudy used a ackground esimaion mehod y muli frame difference image o se up he iniial ackground. And hrough a ackground changing esimae sraegy, he ackground updaes only on cerain condiions. To ensure he sysem could work normally in differen ligh condiions and monioring scene changes accidenally, his sudy used a filering algorihm ased on differenial wo value image processing, and his increased rousness of he algorihm effecively. 2.2 Mehod for ackground developmen and updae This scheme used improved ackground differenial mehod o separae moving ojecs wih ackground. The scheme includes wo pars which are developmen and updae of ackground Mehod for acquire ackground images The ackground images are developed y a muli-frame suracion images dynamic evaluaion mehod, and he images are colleced in accordance wih cerain ime inervals. I is supposed ha 3 coninues images named a,, c, and {B, }, {O, } are ackground and moving ojec of he frame. The processing procedure should e: 1) Separae he 3 frame coninues images ino 2 groups. The firs group include frame a and, and he second group include frame and c. Gray scale difference suracions eween frames are operaed o each pixel of he wo groups, hen he asolue values were preserved in {N, } and {N, }. See in formula (1): N I I, N I I (1) 1 a 1 c i, j i, j i, j i, j i, j i, j 2) Because {N, } and {N, } are difference values of adjacen wo frames, hey are highly similar and he hisogram shows doule peaks. The hreshold T 0 of {N, } calculaed hrough OTSU mehod could e used as he es separaion hreshold of foreground and ackground in {N, } and {N, }. Compare values in {N, } and {N, } wih T 0 separaely. If for any poin X ( i, j ), he corresponding values in {N, } and {N, } are all igger han T0, hen i could e judged ha his poin are moving in all he 3 frames, so ha he poin could e included ino moving ojecs {O, } in foreground. See in formula (2): if Di, j T0 AND Di, j T0 O i, j 0 else (2) 3) According o moving images {O, }, all pixels valued 255 are eliminaed from inpu frames, and he ress are ackground images seleced from frames{b, }. See in formula (3): Ii, j if Oi, j 0 B i, j 0 else Oi, j 255 (3) 4) Deal all colleced frames wih mehod aove, and supplemen los pars in ackground, and hen a enire iniial ackground image could e acquired Mehod for ackground image updae. Afer he monioring sysem is used for a cerain periods, he ackground would have some changes inevialy, for example influence of weaher, ligh and displacemen ecause of oher facors. If he ackground is no updaed hroughou working, i mus influence accuracy of recogniion and racing for moving ojecs[3]. This sudy could realize ackground updae hrough mehod of esalishing saisical models for each pixels, and calculaing proailiies o judge ackground changes. When a cerain pixel values were changing consanly during a cerain imes, hen he grey value of he previous ackground would e replaced y he value of his poin. Or else, he ackground would no e changed. In acual operaion, he ackground judging model is esalished y calculaing wo parameers of each images he mean value μ and variance σ. For he newly colleced sample value S for a cerain poin (x, y), if formula (4) were saisfied, he poin would e regarded o e he new ackground pixel. 2 1 ( s ) f ( s) Exp( ) T (4) T is he proailiy hreshold value in his formula, and could e se dynamically. 2.3 Recogniion and racing of moving ojecs Afer ackground was esalished, he ackground difference mehod was used o collec shapes of moving ojecs in he ackground. The mehod is o use he grey values of each pixels of he curren image o surac wih hose of he ackground. See in formula (5):

3 87 N I B (5) i, j i, j i, j Because his sysem need o recognize moving ojecs, u hrough experimens we go o know ha he moving ojecs produced shadows are imporan facors o influence sysem o recognize arges accuraely. So ha he firs sep o recognize ojecs is o eliminae influences of he shadows. Through sudies of moving ojecs, wo characers were found: one is ha he difference values of he moving ojec shadows o he ackground are smaller han he hose of ojecs hemselves o he ackground. And he oher is ha he grey values of he shadows are usually smaller han hose of he surrounding areas. Then he shadows could e eliminaed y hreshold seing. Bu only use hreshold o eliminae moving arges would cause los of deails of moving ojecs. So edge deecing ecomes a key prolem for moving ojecs recogniion Edge deecion of he moving ojecs and heir shadows In deecion of he ojecs edges, he Prewi operaor and he Soel gradien operaor are usually used. The Prewi operaor finish he calculaion y couning grey differences of each pixels wih heir adjacen poins, and hen y neighor convoluion of he model and images using he horizonal and verical vecors. This mehod has eer effec on smoohing noises, u also has shorcoming of lower posiioning precision. The Soel gradien operaor also deecs image edges using he horizonal and verical vecors, which is similar wih he Prewi operaor. Bu he Soel gradien operaor did weighed processing on posiion influences of he pixels, so ha i is more accurae in deecion. This sudy used he Soel gradien operaor o deec image edges, and he calculaion mehod is shown in formula (6). E 1 I I 1 I I j1 i1 ' i, j max y i1 i1, x x, j 1 x, j 1 4 y j1 4 xi1 (6) Through he aove calculaion, and afer image inaryzaion, we could ge he enire inary image of he whole moving ojec area. Afer his, he edges of he shadows could e ge hrough mehod elow: A window of 5 5 moves on he inary image, and if he cenral pixel grey value reached 255 in he window, hen coun convoluion of he window sugraph wih 4 sensiive one dimensional Laplasse operaor. The maximum value of he 4 convoluion asolue value is regarded as he asis of judging wheher he cenral pixel is he moving ojec edge pixel. See in formula (7): Through he aove process, shadow areas were suraced from he moion edge depiced inary image, influences of shadows on moving ojecs recogniion could e eliminaed, and he exure feaures of he moving ojecs are enirely conserved. This is he asic of realizing moving ojecs recogniion and racing Recogniion mehod for moving ojecs The sudy esing arges for his sudy is he monioring cameras se on a school gae area. In his circumsance, mos popular moving ojecs are vehicles and pedesrians[5]. Usually, he arge recogniion work hrough recognizing shapes, colors and exures feaures of he arges. Bu in his complicaed circumsance, high misjudging rae will happen if his recogniion mehod were used. So our scheme used arges recogniion mehod ased on morphology. The firs sep is o se up a hree dimensional feaure lirary for vehicles and pedesrians. In feaure depicing for vehicles and pedesrians, heir lengh, widh and shape feaures could e shown y morphological parameers. Because morphological feaures could reflec quaniaive difference of shapes of vehicles and pedesrians, i is effecive for recogniion of he arges. According o sudy needs, our monioring sysem seleced morphological feaures include he following: Size raio: Deal wih he arges as a recangular and ge heir raio of lengh, widh and heigh. Recangle filling degree: Each ojec has is own shape filling degree. We could recognize he shape of he arges according o his. Projecion rae: I is raio of A size o convex polygon area size, and his feaure could depic he irregular of he ojec edges. Eccenriciy: This is a parameer for deecing wheher he arge is concenraed or no, and hrough calculaing he eccenriciy, compac degrees of he arges could e depiced. To solve prolems of highly recogniion ime complexiy caused y vas numer of daa in feaure daaase, fas linear classifier for self adapion feaure selecion is used o realize arges fas recogniion in his sudy. The processing procedure of he fas linear classifier is ha, selec feaures wih ig differences o sar iniial classificaion for he arges, for example, large cars, small cars, pedesrians and oher, and hen self adapion selec feaures fi for furher classificaion and finish arges recogniion ( See in Fig 1 ). In occasion of oo many moving arges causes high presser for he monioring sysem, he sysem could lower is classificaion level o insure sysem recogniion speed. EC max{ BE K : p 1, 2,3,4} (7) i, j i, j p K p is he p h convoluion operaor in he formula.

4 88 Calculaing of feaure parameers for iniial classificaion value ( MR, MG,MB ), and weigh parameers for relaive changing of arge size S in funcion calculaion. And + + =1. The sysem used feaure relaive variaion o increase adapion of arge maching. Fig 1. Targe recogniion char Tracing for he moving ojecs Mehod for racing moving ojecs relies on calculaing maching degrees of moving arges eween former and laer frames. In his sudy, moving ojecs in adjacen wo frames are mached according o ojec posiion, size, average color grey value ec., o realize consisency laeling for he same arges. And he acion pah for he ojecs are recorded according o he mached posiion. The mehod is shown in formula (8): C DIS AVE AREA a, a, a, a, DIS a, X X Y Y ( ) ( ) R R c, a c, 2 c, a c, 2 M M M M AVEa, M B, a M B, AREA ( S S ) / S R, a R, G, a G, a, a a x Iniial classificaion Self adap selecing feaures ased on iniial classificaion Self adap selecing feaures ased on iniial classificaion Furher desinaion classificaion Classificaion resuls oupu y / 256 (8) In he formula, ( R x, R y ) symols image resoluion,,, show arge posiion (X c, Y c ), average color 3. Resul and analysis In he experimen, IK-HD1 ype 3CCD camera produced y Toshia company is used as image collecion faciliy. This camera owns following feaures: 1) Oupu pixel: ; 2) Oupu por: digial HD-SDI ( SMPTE 292M),DVI oupu; 3) Manual / auomaic mode whie alance seings are availale. Digial collecion faciliy is MV9300HD video capure card from WOSHI company. Is mainly parameers are : Oupu qualiy: 10 i; Compress mode: H.264;8 video collecion channels and 4 voice collecion channels. is Precision T7500 ype graphics worksaion from Dell company is used o e daa collecion plaform. This worksaion has 4 channels memory sysem and NVIDIA Quadro FX3800 display chip, which has relaively high speed for graphics processing. In sofware developmen, Microsof Visual C++(VC++) is used and Mala numerical calculaion sofware is used in programming of formulas in his sudy, hese programs are evenually compiled ino VC++ procedure for call. There are 3 compiling mehods used in his sudy: 1) Use keil compiler MCC of Mala; 2) Use Macom compiler; 3) Use COM Build ool of Mala. Among hese mehods, mehod 1 is he simples, u could no call powerful Mala image oolox. Mehod 2 is high efficien, u imperfec in supporing graphics and image funcions. Mehod 3 is fas in program working, could e used ou from he Mala environmen, and suppors almos all Mala funcions. I is also perfec in supporing graphics funcions, and is a MahWorks company recommended Mala mixed programming mehod. Therefore, The 3rd mehod was used in his sudy o develop image processing program. 3.1 Esalishmen of ackground image The calculaion resuls from he muli-frame suracion mehod indicae ha, under he same frame sampling frequency, ackground qualiy is posiively relaed o frame sampling ime. Le e sampling ime, and n e sampling frames, he ackground collecion effec is shown in Fig 2. When equals o 2 sec, he ackground image C came from image A and B showed large lanks; and when equals o 6 sec, he ackground image F came from D and E could asically fulfill ackground esalishmen requiremens.

5 IJCSI Inernaional Journal of Compuer Science Issues, Vol. 10, Issue 1, No 3, January Fig 2. Iniial ackground esalished under differen and n values (In he figure, C is he ackground esalished from A and B, = 2; F is he ackground esalished from D and E, =6) In general ransporaion and raffic condiion, when sampling rae is 5 per second, he igger value is, he more clear he ackground will e. When equals o12, he ackground is imperfec, and when equals o 36, he ackground is clear, and when equals o 72, he ackground definiion shows almos no difference comparing wih equals o 36. Afer many experimens, he resuls indicaes ha, in general ransporaion and raffic condiion, a ackground clear enough for monioring sysem could e esalished when value reach 36. Bu in condiions of oo many pedesrians or raffic jam, he value needs o ecome igger o esalish a eer ackground for he monioring sysem o recognize and rack arges effecively. 3.2 Targes recogniion and racing In arge edges deecion aspecs, in order o decrease pressers for calculaion sysem, firsly he colleced RGB images are convered ino grayscale images, and hen Soel gradien operaor is used. The arge edges are deeced hrough Mala programming. Then he images are processed wih edge funcion, and he inpu parameer of he funcion is he wo dimensional marix, indicaor sring and some numerical parameers wih resriced mehod afer imread. The edge deecion procedures used Mala are as follows: f=imread('1.jpg'); f=rg2gray(f);% conver o Gray scale map f=im2doule(f);% The funion im2doule, he value is normalized o 0 1 % Use verical Socl operaor, and selec hreshold graes auomaically [VSFAT Threshold]=edge(f, 'soel','verical'); % Edge deecion figure,imshow(f),ile(' iniial image ');% Show he iniial image figure,imshow(vsfat),ile( ' Verical image edge deecion '); % Show he edge deecion image 89 % Using he horizonal and verical Soel operaor o selec hreshold auomaically. SFST=edge(f,'soel',Threshold); figure,imshow(sfst),ile(' Horizonal and verical image edge deecion '); % Show he edge deecion image % Use specified angles of 45 degrees Soel operaor filer o specified he hreshold s45=[-2-1 0;-1 0 1;0 1 2]; SFST45=imfiler(f,s45,'replicae');% Funcion o filer arrays of arirary ypes or mulidimensional image SFST45=SFST45>=Threshold; figure,imshow(sfst45),ile(' angles of 45 degrees edge deecion') ; % Show he edge deecion image The ackground eliminaion effec for moving ojec afer edge deecion is shown in fig 3. Fig 3. The ackground eliminaion effec for moving ojec afer edge deecion During arge recogniion and racing, he sysem frame rae is 15/ second, and image resoluion is pixels. In he acual experimen, videos sored wih local hard disk were used o es he funcion of recogniion, and he experimen resuls were calculaed arificial couning and saisical classificaion resuls comparison. The resuls go from esing 1200 second images recogniion is shown in ale 1. Tale 1: Resuls of recogniion rae experimens Name of moving ojec Recogniion rae % Average recogniion ime ms Pedesrians Small cars Big uses Small uses Trucks Bicycles The experimen shows ha his sysem could finish arges recogniion under frame rae of 15 / second, and recogniion raes are higher han 87%. I also could finish arge racing according o fron and rear frames maching

6 90 degree, u he recogniion of small cars are relaively lower, ecause ha he recangular degree of he cars are usually difficul o conrol, and he sysem ofen misake hem wih small uses. In fuure sudies, his sysem could e furher developed y improving he arge morphological daaase. 4. Conclusion This sudy designed a se of moving ojecs recogniion and racing mehods according o unique requess of monioring sysem. And researchers also developed an effecive scheme in ackground esalishmen and moving ojecs shadows eliminaion regarding o he acual condiions. This sudy proved hrough experimens ha his scheme could fulfill video monioring sysem needs for moving ojecs recogniion and racing. And he recogniion rae and racing speed oh reached he design sandards. Peilong XU, orn in The auhor has achieved Maser degree from Tongji Universiy of Shanghai in He is currenly an engineer of compuer science in he Sae Key Laoraory, Qingdao Universiy. The auhor's research ineress include sofware engineering, image processing, and specral analysis. Recenly, he have pulished paper named Design and Implemenaion of Landscape Sysem for Eas and Wes Huashi Sree in Beijing Based on Virual Realiy Technology on journal Applied Mechanics and Maerials( Trans Tech Pulicaions Inc, Swizerland)ec. References [1] Weihua Liu. An image resoraion algorihm ased on image fusion, Inernaional Review on Compuers and Sofware, 2012,Vol.7, n.3,pp [2] Ao-Eleneen Z.A., Adel-Azim Gamil. An improved image segmenaion algorihm ased on MET mehod, Inernaional Journal of Compuer Science Issues, Vol.9 n.5-3, 2012, pp [3] Jun Sun, Yan Wang, Xiaohong Wu.A New Image Segmenaion Algorihm and Is Applicaion in Leuce Ojec Segmenaion, TELKOMNIKA, Vol.10, n.3,2012,pp [4] Kondapalli Varaprasad S., Chiranjeevi Manike, Mishra Kundan Kumar, Tanuja Kds. Image processing and analysis for DTMRI, Inernaional Journal of Compuer Science Issues, Vol.9 No. 1-1,2012, pp [5] Ren Mingwu, Yang Jingyu, Sun Han. Tracing oundary conours in a inary image, Image and Vision Compuing, Vol.20 No.2,2002,pp: [6] Kang Lie, Zhong Sheng, Wang Fang. A new conour racing mehod in a inary image, 2011 Inernaional Conference on Mulimedia Technology, ICMT2011,July 26, July 28, 2011,Hangzhou,China,2011,pp