Design Methodology for Face Detection Acceleration

Design Me hodology o Face De ec ion Accele a ion Lau en iu Acasand ei Ins i u o de Mic oelec onica de Se illa IMSE-CNM-CSIC Se illa, Spain lau en [email protected] Angel Ba iga Ins i u o de Mic oelec onica de Se illa/Uni . Se illa IMSE-CNM-CSIC/ Uni . Se illa Se illa, Spain [email p o ec ed] Abs ac —A design me hodology o accele a e he ace de ec ion o embedded sys ems is desc ibed, s a ing om high le el (algo i hm op imiza ion) and ending wi h low le el (so wa e and ha dwa e codesign) by add essing he issues and he design decisions made a each le el based on he pe o mance measu emen s and sys em limi a ions. The implemen ed embedded ace de ec ion sys em consumes e y li le powe compa ed wi h he adi ional PC so wa e implemen a ions while main aining he same de ec ion accu acy. The p oposed ace de ec ion accele a ion me hodology is sui able o eal ime applica ions. Keywo ds— ace de ec ion; embedded sys em; design me hodology; ha dwa e & so wa e codesign I. INTRODUCTION Face de ec ion is an impo an aspec o biome ics[1], ideo su eillance and human compu e [2] in e ac ion. De ec ion sys ems equi e huge compu a ional and memo y esou ces due o he complexi y o de ec ion algo i hms. A so wa e de ec ion ealiza ion implemen ed on a low speed, low esou ce, low powe SoC (Sys em on a Chip) i is no e icien . Ins ead a So wa e-Ha dwa e Codesign app oach can be used o build ha dwa e accele a o s o mos compu a ional consuming pa s o de ec ion algo i hms. Real ime ace de ec ion applica ion equi es a high amoun o mul iplie s and memo y esou ces unning a high speed. Due o high amoun o esou ces and high speed equi emen s a so wa e implemen a ion o he ace de ec ion algo i hm is no easible o a low speed, low esou ce, low powe SoC. Recen ly he e ha e been se e al ha dwa e ealiza ions ha accele a e he ace de ec ion p ocess by pa allelizing he de ec ion algo i hm. Thus some o he p oposals a e o speci ic ha dwa e ealiza ions on ASIC [3] while o he s a e implemen ed on FPGA [4-6]. Using he So wa e-Ha dwa e Codesign app oach we can accele a e he de ec ion p ocess by eplacing, wi h e y as dedica ed ha dwa e, some pa s o he so wa e ha usually consumes la ge amoun o clock cycles and/o memo y du ing he un ime. In [7] se e al pa s o he Viola-Jones de ec ion mechanism ha e been pa allelized in o de o un e icien ly on a GPU a chi ec u e. The pu pose o his communica ion is o desc ibe a design me hodology o accele a ing ace de ec ion applica ions. Taking in o conside a ion ha he OpenCV lib a y comes wi h wo baseline applica ions LBP (Local Bina y Pa e n) and Viola-Jones o ideo de ec ion, we ha e chosen Viola-Jones o de eloping a ace de ec ion applica ion o embedded sys ems. The same design me hodology is common o bo h algo i hms. The LBP and Viola-Jones ha e he same de ec ion mechanism, he only di e ence being in how o encode/decode he ea u es om an image. II. FACE DETECTION TECHNIQUE The ace de ec ion echnique is based on he ace de ec ion amewo k p oposed by Viola-Jones [8]. The p oposed amewo k is capable o p ocessing images ex emely apidly while achie ing high de ec ion a es. The speed o he ace de ec ion amewo k elies on h ee impo an key componen s. Fi s ly, he image is ans o med in o “In eg al Image” which allows he ea u es used by he de ec o o be compu ed e y quickly. Secondly, he used classi ie is simple and e icien which is build using he AdaBoos lea ning algo i hm [9] o selec a small numbe o c i ical isual ea u es om a e y la ge se o po en ial ea u es. And hi dly, he classi ie is o med by combining weak classi ie s in a “cascade” which allows backg ound egions o he image o be quickly disca ded while spending mo e compu a ion on p omising ace-like egions. Viola-Jones echnique is based on explo ing he image by means o a window looking o ea u es. This window is scaled o ind aces o di e en sizes. The sys em a chi ec u e is based on a cascade o de ec o s. The i s s ages consis o simple de ec o s, e y as and low cos , ha allows o elimina e hose windows ha do no con ain aces. In he successi e s ages he complexi y o de ec o s a e inc eased in o de o make a mo e de ailed analysis o ea u es. A ace is de ec ed only i i makes i h ough he en i e cascade. The Haa -like ea u es used by he classi ie consis o ec angula a eas whose p ocessing equi es simple a i hme ical ope a ions. The calcula ion is based on he sum o he pixels o each ec angula egion weighed by a weigh . A all scales, hese ea u es o m he “ aw ma e ial” ha will be used by he de ec o . The se o ec angle ea u es in he image is qui e la ge and o e comple e, so o educe ha numbe applies he AdaBoos lea ning algo i hm [9]. The Viola-Jones classi ie employs AdaBoos a each node in he cascade o lea n a high de ec ion a e a he cos o low ejec ion a e mul i ee (mos ly mul is ump) classi ie a each node o he cascade. To acili a e he p ocessing o he ea u es he ope a ions a e no made on he o iginal image bu on an in eg al image. The e o e he de ec ion algo i hm equi es a p ep ocessing s ep ha calcula es his in eg al image. The in eg a ion o he image consis s o adding o each pixel he alues o he p e ious pixels. III. DESIGN METHODOLOGY The s a ing poin o he design me hodology o he embedded sys em is he OpenCV’s Viola-Jones baseline ace de ec ion applica ion. OpenCV (Open Sou ce Compu e Vision), s a ed by In el in 1999, is a lib a y o p og amming unc ions o eal ime compu e ision [10]. OpenCV is eleased unde a BSD license and hence i is ee o bo h academic and comme cial use. I is w i en in C/C++ and was designed o compu a ional e iciency and wi h a s ong ocus on eal- ime applica ions. The hos a ge o he p oposed ace de ec ion sys em is an embedded en i onmen based on LEON3 AMBA Bus p ocesso . The LEON3 is a syn hesizable VHDL so co e o a 32-bi p ocesso complian wi h he SPARC V8 a chi ec u e [11]. The p ocesso is highly con igu able, and pa icula ly sui able o sys em-on-a-chip (SOC) designs. The ull sou ce code is a ailable unde he GNU GPL license. The p ocesso con ols and execu es he majo i y o so wa e applica ion asks while a speci ic IP (In ellec ual P ope y) module accele a e only hose asks ha equi e a high numbe o clock cycles. The design low is based on ou s ages, as shown in Figu e 1. In he i s s age an adap a ion o he so wa e applica ion o execu e on he embedded sys em has been made. In he nex s age an analysis o he new embedded so wa e applica ion is pe o med, in o de o de ec "bo lenecks" and hose asks ha a e sui able o accele a e h ough ha dwa e implemen a ion. In he hi d phase, as esul o he p e ious analysis, solu ions o op imize and accele a e some o he asks o he ace de ec ion p ocess will be p oposed. The ou h s age consis s in he ha dwa e implemen a ion o hose pa s ha consume a many esou ces and ha e la ge un imes. The idea is o o load o he ha dwa e he unc ions wi h a high deg ee o p ocessing and o pa allelize he execu ion o he de ec ion algo i hm. Wi h his he ace de ec ion p ocess can be d as ically accele a ed. Fig. 1. Design me hodology A each s age o he design p ocess he pe o mance o he ace de ec ion sys em is analyzed, in e ms o speed and de ec ion accu acy. The accu acy o ace de ec ion p ocess is analyzed using ROC (Recei e Ope a ing Cha ac e is ic) cu es. The ROC cu e o a gi en ace de ec o shows i s pe o mance as a ade-o be ween he alse accep ance a e and he ace de ec ion a e by a ying i s disc imina ion c i e ion (e.g. a h eshold pa ame e ). IV. EMBEDDED SOFTWARE FACE DETECTION IMPLEMENTATION Viola and Jones o ganized each boos ed classi ie g oup in o nodes o a ejec ion cascade. Each o he nodes con ains an en i e boos ed cascade o g oups o decision s umps (o ees) ained on he Haa -like ea u es om aces and non aces (o o he objec s he use has chosen o ain on). Typically, he nodes a e o de ed om leas o mos complex so ha compu a ions a e minimized (simple nodes a e ied i s ) when ejec ing easy egions o he image. Typically, he boos ing in each node is uned o ha e a e y high de ec ion a e (a he usual cos o many alse posi i es). When aining on aces, o example, almos all (99.9%) o he aces a e ound [12] bu many (abou 50%) o he non aces a e e oneously “de ec ed” a each node. Bu his is sa is ac o y because using, as an example, 20 nodes will s ill yield a ace de ec ion a e ( h ough he whole cascade) o 0.99920 ≈ 98% wi h a alse posi i e a e o only 0.520 ≈ 0.000001%! Du ing algo i hm execu ion, a sea ch window o di e en sizes is swep o e he o iginal image. In p ac ice, 70–80% o non aces a e ejec ed in he i s wo nodes o he ejec ion cascade, whe e each node uses abou en decision s umps. This quick and ea ly “a en ional ejec ” as ly speeds up ace de ec ion. The Haa -like ea u es a e ained o be applied o a e alua ing ec angula window o 20x20 pixel. Fo o he dimensions o he e alua ing window he Haa -like ea u es mus be scaled co espondingly. The OpenCV so wa e implemen a ion o ace de ec ion consis s o 22 cascade de ec o s, con aining 2135 Haa like ea u es. The i s ask o So wa e-Ha dwa e Codesign was adap ing and op imizing he OpenCV baseline applica ion o an embedded en i onmen . We ha e conside ed ha he majo i y o embedded en i onmen s a e capable o unning C/C++ applica ions wi h o wi hou Ope a ing Sys em (OS) suppo . This means ha he esul ing applica ion code has o be compa ible o bo h C, C++ compile s and in he same ime pla o m independen . Ano he conside a ion made is he ac ha mos o he SoC ha e no loa ing poin suppo . Fo i , he esul ing applica ion uses in ege ope a ions ins ead o loa ing poin ope a ions in o de o p ese e he gene ali y o he applica ion o he embedded sys em wo ld. An impo an momen in his s ep was inding an accep able scaling coe icien o he loa ing poin a iables and da a o in ege a iables and da a. A e ying di e en alues and compa ing he esul ed in ege applica ion wi h he loa ing poin applica ion we ound ha by scaling wi h 20 bi s (p ecision o 20 bi s o he loa ing poin decimals) he in ege and loa ing poin applica ions ob ain Pe o mance and de ec ion accu ac ya nalysis OpenCV So wa e Face De ec ion Applica ion Embedded So wa e Implemen a ion Embedded So wa e Anal y sis Algo i hm Op imiza ion Ha dwa e accele a ion iden ical esul s. Also he loa ing poin squa ed oo unc ion necessa y o calcula e a iance o he e alua ing window was eplaced wi h a as in ege squa ed oo e sion. In he end i was ob ained a ace de ec ion s and-alone applica ion compa ible wi h C/C++, using only in ege ype ope a ions and da a. V. EMBEDDED SOFTWARE FACE DETECTION ANALYSIS The nex s ep in applica ion de elopmen was analyzing di e en modes o de ec ion in o de o ind he un ime bo lenecks and op imize he de ec ion. Fo he embedded a ge we ha e decided o use he de ec ion mode whe e he de ec o (Haa -like ea u e) is scaled and he bigges egions con aining aces a e sea ched wi hin an image. In his mode he de ec ions s a s wi h he bigges e alua ion window and bigges e alua ion s ep and p og essi ely, he window oge he wi h he e alua ion s ep a e dec eased un il a egion con aining a ace is de ec ed. In he case ha a egion wi h ace o mul iple aces is de ec ed, he a en ion o he de ec o concen a es in ha egion. The ained classi ie cascade (Haa -like ea u e) is p o ided by OpenCV in an XML ile o ma . Using his XML o ma in an embedded sys em will p oduce memo y and un ime o e head. In o de o a oid he o e head we ha e de eloped an applica ion ha ecei es a XML ile, in e p e s he da a and sa es i in a simple o ma o a C heade ile. The esul ed embedded applica ion can be compiled wi h he cascade o classi ie o he da a can be ansmi ed du ing he execu ion o he applica ion ia an app op ia e in e ace. In o de o ob ain ele an insigh abou which pa s (o unc ions) o he ace de ec ion p og am a e aking mos o he execu ion ime we enable -gp lag in he Eclipse p ojec compila ion op ions in o de o gene a e p o iling in o ma ion ha can be in e p e ed wi h GNU gp o ool. This ool pe mi s one o lea n whe e he p og am spends i s ime and he unc ion calling ee du ing he execu ion. I can also ell which unc ions a e being called mo e o less han a e expec ed. Table I shows he ob ained esul s. As we can see he unc ion Se Ma Ze o() uses 24.64% o he execu ing ime e en su passing he ime spen applying he cascade Haa like- ea u es (20.16%) o he en i e image. The unc ion Se Ma Ze o() is used o se o ze o all he elemen s o a empo a y ma ix ha ing he same dimension o he image. In his ma ix he op le coo dina es o a de ec ed ace a e lagged wi h alue 1. In his mode he de ec ions s a s wi h he bigges e alua ion window and bigges s ep and p og essi ely, he window and he s ep a e dec eased un il a egion con aining a ace is de ec ed. The de ec ion is done in wo s eps:  Fi s S ep: The image is scanned wi h he e alua ion window by applying only he i s wo Haa -like ea u e s ages in o de o apidly de ec egions con aining po en ial aces. I a egion is ound o ha e a po en ial ace hen he coo dina es a e se o alue one in he empo a y ma ix.  Second s ep: Each po en ial ace (s a ing wi h hei coo dina es) om he empo a y ma ix is e alua ed wi h he emaining Haa -like ea u e s ages. I a ue ace is de ec ed hen he coo dina es, wid h and heigh a e s o ed in a lis . A he end o he second s ep, he empo a y ma ix is se o ze o in p epa a ion o he nex image whe e he e alua ing window has smalle dimensions. We can imp o e he speed by no using he unc ion Se Ma Ze o() a he end o he second s ep and ins ead each ime a e we ha e de ec ed a ace du ing he second s ep and ha ace is s o ed in o a lis , we se o ze o he coo dina es inside he empo a y ma ix. A e applying his change, he de ec ion ime is imp o ed wi h 24.64 %. TABLE I. ANALYSIS OF DETECTION SYSTEM BOTTLENECKS Time % # calls Func ion name 24.64 17 Se Ma Ze o() 20.16 3201 RunHaa Classi ie CascadeEmbedd() 14.81 16 Se ImagesFo Haa Classi ie CascadeEmbedd() 13.39 1 In eg al() 11.35 1 LinkDa aToEmbeddClassi ie Cascade() 4.22 511 HResizeLinea () 3.11 262144 sa u a e_ucha () 2.62 512 VResizeLinea () 1.80 296384 sum_elem_p () 1.10 3201 isq 64() VI. EMBEDDED FACE DETECTION OPTIMIZATION The OpenCV ace de ec ion baseline applica ion implemen s de ec ion in wo dis inc modes: Mode 1: Face de ec ion by scaling he image. In his mode he image is scaled using in e pola ion un il i eaches a p ede ined minimal dimension. Each ime he image is scaled he wo in eg al images (no mal= x and squa ed= 2 x), needed o a iance, a e ecalcula ed o he scaled image. The sea ch window has ixed dimension du ing he de ec ion p ocess. Mode 2: Face de ec ion by scaling he classi ie s. In his mode he in eg al images(no mal= x and squa ed= 2 x), needed o a iance, a e calcula ed only once o he o iginal image bu he Haa -Like ea u es o m he classi ie a e scaled p og essi ely un il hei dimensions a e close o he dimension o he o iginal window. This mode lacks he in e pola o used in mode 1. The sea ch window has a a iable dimension du ing de ec ion p ocess. In bo h de ec ion mode he Haa -like ea u es componen s (weigh s and dimensions) a e scaled p opo ionally wi h he dimensions o sea ch window. Tha means o a sea ch window o dimension WxH, he weigh o each ec angle o ming he Haa -like ea u es a e scaled wi h WxH. In he sea ch window, he sum o each applied Haa -like ea u e is calcula ed using:    3 1I scaled I I Sum Weigh A eaeHaa Fea u (1) A ea ep esen s he sum o all pixels inside a componen and I=1, 2 o 3 ep esen s he numbe o componen s o ha Haa -like ea u e. In o de o de e mine he nex weigh alue o he s age sum, each Haa Fea u eSum is compa ed wi h each no malized h eshold o he espec i e Haa -like ea u e as:         no m J Sum J Weigh J no m J Sum J Weigh J Th eseHaa Fea u i S ageS ageSum Th eseHaa Fea u i S ageS ageSum S ageSum , , 1 2 (2) whe e J=[1…2135] ep esen s he Haa -like ea u e indexes in a s age and Th es J no m=σTh esJ Haa Fea u e (σ is s anda d de ia ion o he windows sea ch a ea). I we do no scale he Haa -like ea u e weigh s and adjus he a iance compu a ion by using he o mula  )( 2 22   xxHW adjus ed  , i esul s ha he numbe o a i hme ic ope a ions (di ision, mul iplica ion) and memo y accesses a e dec eased subs an ially. This will make he algo i hm pe o m as e due o a educed numbe o ope a ions needed o compu a ion o he adjus ed a iance o he sea ch window [13]. Figu e 2 shows he p oposed op imiza ion o he de ec ion algo i hms in he wo de ec ion modes (scaling he image and scaling he classi ie s). Fig. 2. P oposed ace de ec ion accele a ion algo i hm In o de o compa e he pe o mance o he OpenCV 2.2 baseline ace de ec ion and he accele a ed Viola-Jones algo i hm, and o analyze he in luence o he con igu a ion pa ame e s, bo h implemen a ions ha e been compiled and speed op imized o 64 bi Win7 OS using Visual S udio 2010 P o esional edi ion. The e i ica ion PC has a Pen ium Dual- Co e CPU T4300, wi h L1 cache o 128KB, L2 uni ied cache 1024KB and he OS is Windows 7 Home Edi ion 64 bi s. Bo h implemen a ion (OpenCV’s implemen a ion and accele a ed Viola-Jones e sion) ha e ecei ed he same es VGA (640x480) images and he same con igu a ion pa ame e s. In Table II he con igu a ion ha e di e en scale ac o (s ) and minimum sea ch window dimensions (swd): s =1.1 and swd=20x20 (Con 1); s =1.2 and swd=20x20 (Con 2); s =1.1 and swd=30x30 (Con 3). TABLE II. P ERFORMANCES OF O PEN C V AND ACCELERATED V IOLA - J ONES IMPLEMENTATION FOR DIFFERENT PARAMETERS Mode 1. Img scaling Mode 2. Haa scaling OpenCV Baseline Op imized e sion OpenCV Baseline Op imized e sion Con 1 speed 708.8 ms 1.41 FPS 185.7ms 5.38FPS 843.9 ms 1.18FPS 188.1ms 5.3FPS Sea ch windows 602348 599816 697582 631343 Con 2 speed 409.5 ms 2.44FPS 164.8ms 6.06FPS 479.7ms 2.08FPS 140.1ms 7.1FPS Sea ch windows 354321 353935 381474 364635 Con 3 speed 348.1ms 2.87FPS 99.4ms 10FPS 456.7 ms 2.18FPS 130.6ms 7.6FPS Sea ch windows 352718 351184 402519 332917 As he p oposed implemen a ion has kep he con ol mechanism o sea ch windows iden ical wi h he one o m he OpenCV baseline i is di icul o make a compa ison wi h p e ious wo k done in accele a ing de ec ion due o he lack o se up in o ma ion and how many sea ch windows a e e alua ed. The e is one excep ion Cho e al [6] whe e hey use mode1 wi h a scaling ac o o 1.2, minimal sea ch window dimension o 20x20 and he sea ch window is applied wi h a e ical ho izon al s ep o 1. Table III shows he compa ison esul s. TABLE III. P ERFORMANCE OF THE O PEN CV BASELINE AND ACCELERATED V IOLA -J ONES USING THE C HO ET AL .[6] SETUP Implemen a ion Mode1-Image scaling Speed Sea ch Windows OpenCV 974.3 ms 1.02FPS 881484 Accele a e Viola-Jones 451.4 ms 2.21 FPS 880585 As shown in Tables II and III he numbe o sea ch windows depends hea ily on he con igu a ion se up and also o he con ol mechanism. Also measu ing he numbe o sea ched windows pe o med by he de ec ion sys em gi es mo e ealis ic in o ma ion abou he de ec ion pe o mance. The speeds ob ained by he accele a ed Viola-Jones implemen a ion in some con igu a ion a e as e ha any single GPU accele a ion [3], [7] and o he highes numbe o sea ch windows (Table III) i has close pe o mances o [3] and [7]. VII. HARDWARE ACCELERATION OF THE EMBEDDED FACE DETECTION SYSTEM A e a ca e ul analysis o he ace de ec ion applica ion i was ound ha he so wa e bo leneck esided in he huge amoun o memo y ead access, mul iplica ion, and squa ed oo ope a ions done by all he sea ch window e alua ions. In o de o de ec a ace om an image, hund eds o housands o sea ch windows a e e alua ed and his ep esen s he mos ime consuming pa o he applica ion. The e o e i was decided o accele a e he e alua ion o sea ch windows by employing a ha wa e IP module [14]. The p oposed IP module and all in e nal componen s a e clocked by he sys em clock (80 Mhz). In o de o keep a high deg ee o lexibly and sha e he ha dwa e esou ces wi h he es o he LEON3 sys em i was decided o he IMSE_OBJECT_DETECTION IP o ha e wo ope a ing modes: he ee mode in which LEON3 p ocesso can use he IP ha dwa e esou ces o implemen o he unc ionali ies, and he ace de ec ion mode. Figu e 3 shows he block diag am o he IP module. Fig. 3. IMSE_OBJECT_DETECTION IP block diag am As i was p e iously men ioned he IP module implemen s he sea ch window algo i hm. The so wa e applica ion will load he comp essed o m o Haa -like ea u es in o he Sha ed Memo y be o e any de ec ion ope a ion. Be o e s a ing any de ec ion ope a ion he con igu a ion egis e s (scale, x-y coo dina es, image dimension, e c.) mus be con igu ed wi h he desi ed con igu a ion alues. When he s a command is gi en he ace de ec ion p ocedu e is ully con olled by he componen Imse_s age_e alua o _uni (see Figu e 3). This uni is he co e engine o accele a ing ace de ec ion. A he end o he de ec ion he componen signalize i a ace is p esen , he S a us egis e is upda ed wi h he de ec ion esul and an in e up is gene a ed. The e alua o uni has a mul iple s a e machine con ol in o de o deal wi h a iable memo y access la encies. Beside he mul iple s a e machine con ol his uni con ains specialized modules. Haa _ ea u e_ ec _calc module is used o calcula e he a ea o in eg al ec angles using only he co ne s da a. Haa _ ea u e_scale is a pipelined module o Haa -like ea u e scaling and sea ch window add ess compu a ion. Sq 64_a ay_pipe16 is 64 bi pipelined in ege squa e oo uni ha has da a ou pu la ency o 16 clocks. Mul41x33signed is a 41x33 signed mul iplie . The Regis e Bank con ains APB bus accessible egis e s ha a e used o he co e con igu a ion and con ol. The APB Sla e In e ace connec s he IP module o he APB bus and enables he LEON3 p ocesso o access he egis e s om Regis e Bank. The AHB Mas e /Sla e In e ace is a simple DMA in e ace. The IP module has a sha ed memo y based on a dual-po RAM wi h AHB in e ace. The Sha ed Memo y is used by he IP module o s o e he comp essed Haa -like ea u es. When he module wo ks in “F ee Mode” LEON3 can use he Sha ed Memo y as addi ional on chip RAM memo y. VIII. RESULTS The p oposed LEON3 ace de ec ion sys em wo ks wi h images (colo ed o g ey) ha ing a esolu ion smalle han 1024x1024 pixels. I ully uses he OpenCV cascade o classi ie s o on al aces and i can s o e in o he IP Sha ed Memo y app oxima ely 2730 Haa -like classi ie s. I also wo ks wi h a g ea e numbe o Haa -like classi ie s bu he ex a classi ie s mus be s o e in o he p og am memo y and hen loaded in o he Sha ed Memo y a he app op ia e momen . The sys em was implemen ed on a Xilinx XC5VLX50 FPGA. The en i e LEON3 ace de ec ion sys em uses 6,435 slices (up o 89% o he de ice u iliza ion) and 10,962 o lip- lops (up o 38% o he de ice u iliza ion). The es ima ed s a ic powe consump ion (measu ed wi h Xpowe Analyze om Xilinx) o he LEON3 co e is 603 mW. The mos powe consuming componen s a e he DDR2 memo y con olle (216 mW), DVI in e ace (136.06 mW) and he clock gene a o s. The LEON3 p ocesso consumes 32.39 mW and e en hough he IMSE_OBJECT_DETECTION IP uses mo e lip lops and has app oxima ely he same amoun o logic, i s powe consump ion is 6 imes less (5.39 mW) han he LEON3 p ocesso . A. Pe o mance In o de o measu e he de ec ion pe o mances o he LEON3 embedded de ec ion sys em h ee dis inc so wa e implemen a ions o ace de ec ion we e compa ed:  Po ed OpenCV so wa e o embedded sys ems.  So wa e accele a ed e sion o he po ed so wa e.  Ha dwa e + So wa e accele a ed e sion o he po ed so wa e. The measu ed pe o mances me ics a e he execu ion ime and he numbe o sea ched windows pe o med. Fo he i s wo implemen a ions he pe o mances we e measu ed o wo dis inc modes o de ec ion (mode 1 and mode 2). Fo each mode, ou di e en se -up pa ame e s we e used (se up 1 o 4) o he minimum size sea ch window (S) and he scale s ep (s ep): 1) S=30x30, s ep=1.2; 2) S=30x30, s ep=1.1; 3) S=20x20, s ep=1.2; 4) S=20x20, s ep=1.1. F om Figu e 4 i esul s ha he accele a ed ace de ec ion applica ion is 3-4 imes as e han he po ed OpenCV applica ion o bo h modes. Using he ha dwa e accele a ion IP he ace de ec ion is 10-12 imes as e han he po ed ace de ec ion applica ion unning exclusi ely on LEON3 p ocesso co e. Fig. 4. De ec ion imes o h ee dis inc implemen a ions o VGA image. a) Scale Image mode (Mode 1), b) Scale Haa -like ea u es mode (Mode 2) B. De ec ion Accu acy In o de o analyze he de ec ion accu acy an speci ic so wa e has been de eloped. The PC based es bench so wa e con igu es LEON3 based de ec ion sys em and send he es images. Then i ecei es he de ec ion esul s o u he analysis. The es se up is based on 2409 on al ace images om he colo FERET da abase [15]. The accu acy o ace de ec ion can be desc ibed using ecei e ope a ing cha ac e is ic (ROC), which is a cu e widely adop ed in signal-de ec ion heo y. An ROC is essen ially a sca e plo ha shows he ela ionship be ween he alse accep ance a e and he ue accep ance a e. Because he numbe o e alua ed nega i e egions is e y high o de ec ion algo i hms, o analysis, i was adop ed a modi ied e sion o ROC cu e in oduced in FDDB[16] in which he ho izon al axis con ains only alse posi i e o he en i e es image se . Figu e 5 shows he ROC cu es o OpenCV so wa e and he IP module. Bo h ha e e y simila esul s. The e is a small di e ence be ween he wo ROC cu es because he da a esul (i.e de ec ed aces) agg ega ion mechanism is sligh ly di e en in he analysis ools. (a) (b) Fig. 5. ROC cu es: a) OpenCV ace de ec ion so wa e, b) IMSE_OBJECT_DETECTION IP based sys em IX. CONCLUSIONS This communica ion p esen s a design me hodology o ace de ec ion embedded implemen a ion. The s a ing poin was OpenCV lib a y esou ces o ideo ace de ec ion. Fo i some modi ica ions has been make adap ing and op imizing he OpenCV baseline applica ion o an embedded en i onmen . This modi ica ions can be summa ized in changing he loa ing poin ope a ions by in ege ones, analyzing he pe o mance in o de o de ec he sys em bo lenecks, and algo i hmic speed- up by no scaling he Haa -like ea u e weigh s and adjus ing he compu a ion o a iance. Finally, hose asks equi ing a g ea e compu a ion a e implemen ed in ha dwa e, accele a ing he ace de ec ion p ocess, and enabling i s applica ion in embedded sys ems ha equi e eal ime. ACKNOWLEDGEMENT This wo k was suppo ed in pa by Spanish Minis e io de Ciencia y Tecnología unde he P ojec TEC2011-24319, and by Jun a de Andalucía unde he P ojec P08-TIC-03674. Co- inanced by FEDER. REFERENCES [1] K. Suzuki, J. Kobayashi, T. Takeshima; K. Yamada, "De ec ion o unusual acial exp ession o human suppo sys ems," 34 h Annual Con e ence o IEEE Indus ial Elec onics. IECON 2008, pp.3414,3418, 10-13 No . 2008. [2] M. Pasche o, G. Del Vesco o, L. Benucci, A. Rizzi, M. San ello, G. Fabb i, F.M.F. Mascioli, "A eal ime classi ie o emo ion and s ess ecogni ion in a ehicle d i e ," IEEE In e na ional Symposium on Indus ial Elec onics (ISIE), pp.1690,1695, 28-31 May 2012. [3] T. Teocha ides, N. Vijayk ishnam, M.J. I win, “A Pa allel A chi ec u e o Ha dwa e Face De ec ion”, P oc. IEEE Compu e Socie y Annual Symposium Eme ging VLSI Technologies and A chi ec u es, pp. 452- 453, 2006. [4] V. Nai , P.O. Lap ise, J.J. Cla k, “An FPGA-Based People De ec ion Sys em”, EURASIP Jou nal on Applied Signal P ocessing, pp. 1047- 1061, 2005. [5] C. Gao, S.L. Lu, “No el FPGA Based Haa Classi ie Face De ec ion Algo i hm Accele a ion", P oc. In e na ional Con e ence on Field P og ammable Logic and Applica ions, 2008. [6] J.Cho, S. Mi zaei, J. Obe g, R. Kas ne , “FPGA-Based Face De ec ion Sys em Using Haa Classi ie s”, P oc. ACM/SIGDA In e na ional Symposium on Field P og ammable Ga e A ays (FPGA'09), pp. 103- 112, 2009. [7] D. He enb ock, J. Obe g, N.T.N. Thanh, R. Kas ne , S.B. Baden, “Accele a ing Viola-Jones Face De ec ion o FPGA-Le el using GPUs”, P oc. IEEE Annual In e na ional Symposium on Field-P og ammable Cus om Compu ing Machines, 2010. [8] P. Viola, M.J. Jones, “Robus Real-Time Face De ec ion”, In e na ional Jou nal o Compu e Vision, .57 n.2, pp.137-154, May 2004. [9] R.E. Schapi e, Y. F eund, P. Ba le , W.S. Lee, “Boos ing he Ma gin: A New Explana ion o he E ec i eness o Vo ing Me hods”, The Annals o S a is ics, pp. 1651-1686, 1998. [10] OpenCV, link: h p://sou ce o ge.ne /p ojec s/openc lib a y/ [11] LEON3, link: h p:// www.gaisle .com/ [12] G. B adski, A. Kaehle , “Lea ning OpenCV”, O’Reilly Media, pp.506- 516, Sep embe 2008. [13] L. Acasand ei, A. Ba iga: “Accele a ing Viola-Jones Face De ec ion o Embedded and SoC En i onmen s”, Fi h ACM/IEEE In e na ional Con e ence on Dis ibu ed Sma Came as (ICDSC’2011), Ghen , Belgium, Aug. 2011. [14] L. Acasand ei and A. Ba iga: ‘FPGA implemen a ion o an embedded ace de ec ion sys em based on LEON3’, In . Con . on Image P ocessing, Compu e Vision, and Pa e n Recogni ion, Jul. 2012. [15] P.J. Phillips, H. Wechsle , J. Huang, P. Rauss: "The FERET da abase and e alua ion p ocedu e o ace ecogni ion algo i hms," Image and Vision Compu ing J, Vol. 16, No. 5, pp. 295-306, 1998. [16] J. Vidi and E.L. Mille : "FDDB: A Benchma k o Face De ec ion in Uncons ained Se ings", Technical Repo UM-CS-2010-009, Dep . o Compu e Science, Uni e si y o Massachuse s, 2010. a) b)