V N U Journal o f Scicnce, Mathematics - Physics 25 (2009) 143-151 M otion detection and tracking algorithms in video streams R B o g u s h * , S M a lts e v , A K a s try u k , N B r o v k o , D H lu k h a u Polotsk State Umversity\ Blochin sir., 29, Novopolotsk, Belarus, Ĩ Ỉ 4 Received 30 June 2009 Abstract Moving objects detection and tracking in video stream are basic fundamental and critical tasks in many computer vision applications We have presented in this paper effectiveness increase of algorithms for moving objects detection and tracking For this, we use additive minimax similarity function Background reconstruction algorithm is developed Moving and tracking objects detection aluorilhms are modified on the basis of additive minimax similarity function Results of experiments are presented according to time expenses of the moving object detection and tracking Keywords: Movinu Objects Detection, Tracking, Background Reconstruction , Minimax Similarity l-unction Introduction M oving objects detection in video streams is a key fundam ental and critical task in m any com puter vision applications, including video surveillance, as well as people tracking, gesture recognition in hum an -m ach ine interface, traffic monitoring and so o n [l,2 ] Detection o f m oving object should be ch ara cien z ed by som e im portant features: hi^ỉh precision in case o f noise com ponents presence on the video streams; flexibility in different scenarios (indoor, outdoor) or different light conditions; ct'ficiency, in ord er for detection to be provided in real-time Basic m ethods for motion dctcction in a continuous video stream are: optical flow, frame dii’i’c rcnce and backg rou nd subtraction All o f them are based on com paring o f the cunrcnt video frame with one from the previous frames or with background The m ost w idely adopted approach for moving object detection with fixed cam cra is based on background subtraction I-or frame com parison o f a video information a row o f m easures arc used as unit for measurement ()l similarity images |4 | N orm alized correlation function IS w idely used am ong known measures of similarity H owever, the problem o f perfection the estimation m ethods o f objects similarity is rather actual, bccausc correlalion characteristics o f video sequences arc far from ideal, i.e., and charactenzed by a significant level o f secondai7 spikes and main spike inaccuracy [3] It leads to false identifications o f objcct, or am biguity o f positioning object on the image In w ork [3] attem pt to detailed analysis o f existing m ethods for m easuring various signal param eters to generate steady against various influences algorithm s o f objects similarity evaluation is undertaken Corresponding author E-mail; a.kastruk@mail.ru 143 144 R Bogush et al / VNU Journal ofScictwe, Miitiwmatics - Physics 25 (2009) N - Ỉ Ỉ Tlic analysis o f the considered m ethods testifies that 11 IS possible to speak only about quiisi optim um o f the considered similarity evaluation algorithms, d ep e n d in g on external conditions and type o f analyzed data Practically for all m ethods the basic pro b lem is an accuracy o f positioniim, w h ic h IS lim ited by the b a se w id th o f the m ain correlation pe a k a n d th e p r e s e n c e o f m tc n s n c ICV'CI o i‘ secondary spikes ibr the analysis o f the image in a mix with noise [3J E xccpt for this It IS ncccs.sary to note c o m p u tin g c o m p le x ity pro b le m s In this paper we have introduced effectiveness incrcasc o f algorithm s for m o ving objects detection and trackirm For this, we use additive minimax similarity function, w h ich posscssinu the adv an ced qualitative characteristic and in com parison with function OÍ’ n o rm alized correlalion, also prov idcs reduction o f caicLilalion com plexity, as twice Background reco nstru ction a lg o n lh m IS developed M oving and tracking objects detection algorithms arc m odified on the basis o f adtiitive mini max similarity function Also the results o f experim ents arc presented M inim ax sim ilarity function Functions o f similarity are applied for decision o f some practical p ro b lem s m a video processing; movini» objcct detection, object localisation, target tracking, reco gnition N orm alized corrclalion f u n c t i o n IS w i d e l y u s e d a m o n g k n o w n m e a s u r e s o f s i m i l a r i t y In process o f algorithm s perfection and expansion fields o f im ag es p roccssm u the correlation coefficient has undergone essential modifications, which have allow ed g en era tin g on Us basis a ro w o f m ethods measures o f similarity differing on properties and characteristics In w ork [4] presented effective family o f function similarity for im age and video processing, r h e s c functions foưns an integral similarity estimate based on sequential m in im a x analysis image elements In com parison with function o f normalized correlation, the m inim ax function provides rcducnon o calculation complexity, as mm twice We use an minimax s im ila n ly function for decision o f some problems: background reconstruction, moving objects dcteclion and target tracking Additive mmimax similarity function r ' for image A N , x N B N, X N size, w ith d e m e n t s size, with clcmentsn^ and image : "I N ,N , u ụ n u ix [ a „ h „ ) M oving objects detection and tracking 3,1 B ackgrou n d recon stru ction algorithm In this section, \vc have introduced an effective algorithm for b a c k g ro u n d leconslruclion The algorithm takes odd quantity o f the frames o f input video scq u cn cc in w hich movint; objccts arc present and produced background o f the dynamic scenc F ram es tor p ro cessing take out through the set interval Algorithm includes two basic procedures: calculating b in a ry m atrix o f motion detection between neighbours w ork frames and background reconstruction for each o f two frames, rh e constructed images arc classified as input data (work frames) for the follovvmg Iteration o f algorithm Algorithm steps are described as the following: ííxtraclion o f N frames o f input video streams for vector constructio n, w hich includes images o f these work frames: .(2) R Bogush et a! / VNU Journal oj Science Mathematics - Physics 25 (2009) Ị43-Ị5Ỉ 145 N > & N ^ Krnod 2) (3) wheirc L- interval b etw een w o rk frames ( jL G {20 50} w itch guarantee the correct o f background rccoinstruclỉon); k - n u m b er o f image fram es from N T esting / for every step: if / = 0(m od2), w h ore / e {>v - 1, (4) - 2, ,!}, then: 2.1 Form ing the binary m a tn x o f motion detection using coloir channel separately as: I if two im ages s'"and 5*'^ foi each RGB r min(5* i * '') _ 0, / / - " i ' >Tw h ere 7' - IS a threshold to determ in e w hether the intensity value al the point changes; q e {l, ,yV-2} The utilization RG B c h a n n e ls improve the accuracy m oving object localization 2.2 Binary image p ro c e s s in g o f m orphological filters For this purpose we use opening operation: (6) w here V is a structurinu elem ent In opposite ease p / = l ( m o d ) & / > 1, producing the vector w hich includes /intermediate backg rou nd as; 3.1 Create the vcctor w ith elem ents looks as matrix o f m otion detection A/*''''"'* This matrix includes moving objects for fram e o f 5*^' Matrix can be calculated as: (7) 3.2 Form ing a vcctor o f the work background Background is defined as result o f removing each pixel o f m oving objects from fram e o f 5*^' and paste o f pixels o f background from frame o f 5* for this area VVe cxtract the m o v in g objcct from frame using ( ): Steps o f the a lu o n th m are repealed The procedure is term inated after (N-1) steps Background update 5.1 D eleting the first fram e from a vector w and produce cyclic shift for each frame shift to the left on one position 5.2 Extract new fram e from video sequence applying interval L and use this image as a position \)L vector vv 5.3 Steps ^ o f the a lg o rith m are repeated till / =1 l o simplify the description, w e use a group o f schematic diagram s (fig l) Background reconstruction is in practice ju st the starting video processing step in a system that is usually supposed to w o rk in real-tim e Therefore, it is im portant to m ake this step time efficient In figure time expenses are resulted by background reconstruction for iterative algorithm, background information fusion a lg o rith m and Gaussian mixture background model for 23 sequences All 146 R Boí^ush ef aỉ / I'NU Journa! o f Science, Maihemaỉics - Physics 25 (2009) Ỉ43-Ỉ Ỉ experiments are im plem ented on a personal com puter (CPU - A M D Athlon (tm) 61 2200 Mhz, F masks o f niiuion dclcction for BIl- algorithm ỉìinan masks o f mtuion dctcciion for Cỉaussian mixture Fig The motion masks for several sequences 148 R Boị^ỉtsh ct aỉ / VNU Journal o f Science Maihematics - Pỉĩysics 25 (2009) Ỉ43-Ỉ5Ì High quality o f the image o f background providing thanks an optim um choicc o f param eters o f A',Ả, \ L For satisfaction criteria o f «quality /computational complexity)) optim al N to be chosen from