Professional Documents
Culture Documents
Real time video has become reali<able in recent decade, due to advancements in various technological as*ects such as increase in internet band-idth and *rocessing ca*abilities. !n the recent decade, ,ait anal%sis has gained *rominence in situations like sus*icious activit% detection in *ublic *laces. "hough limited to -alking alone= it ma% be a**lied to other +orms o+ locomotion like running also. !n our *ro*osed method o+ ,ait anal%sis, -e use a de+ormable model based a**roach. Silhouettes are obtained +rom the real time motion o+ sub.ects obtained a+ter segmentation and background subtraction. >sing these silhouettes, signatures are derived as gait +eatures. 2ur research is +ocused on identi+%ing the sub.ect through corres*onding gait *atterns. !n their earlier -ork, 6ang, et.al : ; uses C0 1 rinci*le Com*onent 0nal%sis3 a**lied to time var%ing signals derived +rom silhouettes to reduce dimensionalit% o+ the in*ut. attern classi+ication techni7ues -ere a**lied +or Eigen vectors +or recognition. "he dra-back o+ this method is that it is viede*endant. 0nother method b% Sudee* Sarkar uses baseline algorithm com*rising '? e@*eriments and data set= the algorithm estimates silhouettes b% background subtraction and *er+orms recognition b% its tem*oral correlation : ;. Aimitation o+ this is the *oor silhouettes obtained +rom segmentation. Research re*orted b% Da-son uses C0 based a**roach +or dimensionalit% reduction. !t has a disadvantage that the camera lens direction is al-a%s *er*endicular to the direction o+ the *erson -alking. "here+ore, side vie- or *ro+ile vie- o+ an individual is re7uired. : ; 6e have e@*lored that there is a *ossibilit% to derive +ast signatures based on distance ma**ing +unctions built -ith boundar% in+ormation. Bast a**roaches to build level sets or distance ma**ing +unctional have alread% been *ublished b% the authors earlier. :Sandee* V. M;. !n the *resent -ork, -e have attem*ted gait recognition using boundar% signatures along -ith skeleton signatures that are +ound to be translation invariant and scale invariant.
I.
Introduction
"he *rocess o+ identi+%ing a living being b% manner o+ its locomotion is ,ait recognition and is more s*eci+ic to humans. !t is an unobtrusive method -here sub.ect/s gait under consideration can be ca*tured +rom a distance. !n disguise, the sub.ect ma% a**ear more sus*icious. "he ca*tured images need not be o+ high resolution as recognition is *redominantl% boundar% based. "his makes it advantageous com*ared to other methods, -hich are s*eci+icall% content driven methods. Scienti+ic gait anal%sis is attributed to *ioneering -ork o+ 0ristotle in De Motu 0nimalium 12n the ,ait o+ 0nimals3 and much later in '45), ,iovanni 0l+onso $orelli. 6ith the develo*ment o+ *hotogra*h% and cinematogra*h%, it became *ossible to ca*ture image se7uences that revealed details o+ human and animal locomotion that -ere not noticeable b% -atching the movement -ith the naked e%e. 0lthough much earl% research -as done using +ilm cameras, the -ides*read a**lication o+ gait anal%sis to humans -ith *athological conditions such as cerebral *als%, arkinson8s disease, and neuromuscular disorders, began in the earl% '(9)s -ith the availabilit% o+ video camera s%stems that could *roduce detailed studies o+ individual *atients -ithin realistic cost and time constraints. "he develo*ment o+ treatment regimes, o+ten involving ortho*edic surger%, based on gait anal%sis results, advanced signi+icantl% in the '(5)s. Man% leading ortho*edic hos*itals -orld-ide no- have gait labs that are routinel% used to design treatment *lans and +or +ollo-#u* monitoring. :-iki*edia;
II. Background
"he silhouette gait database -e have obtained is +rom C$!R, C!S!0, China. "he gait +rames re*resent movement o+ *erson +rom right to le+t as indicated in the +igure.
E@*eriment -as conducted on the database com*rising '4 individual video having D) +rames each. Sam*le silhouette +rame and corres*onding ED sur+ace *lot o+ distance ma**ed silhouette embedded level set +unction is sho-n in +ig.
Bigure EC 1a3 1b3 1a3C model curve F 1b3C distance ma**ed level set G 1c3C E D *lot o+ G and corres*onding HAS
1c3
!n the *resent -ork, evolving the model curve is un-arranted, as the silhouette boundar% can directl% be re*resented as the model curve that can be intrinsicall% embedded into the level set +unction G. simultaneousl%, the boundar% also corres*onds to region o+ the ob.ect re*resenting segmented version o+ the image. $oundar% embedded level set +unction inherentl% corres*ond to several length based +eatures. "he +eatures that have been attem*ted in the *resent -ork are around boundar% length n) and skeleton o+ the silhouette. 0bsolute boundar% length is onl% suitable -ith ideal database, -here 7uer% is *art o+ the database. !n *ractice, variations are observed in the 7uer% image and database images in terms o+ translation, scaling, and rotation and %et more com*licated being ske-ing. !n the *resent -ork, to accommodate gait movement, deviating +rom the database, -e have used boundar% +eatures in a normali<ed -a%, -ith res*ect to boundar% length. "he corres*onding normali<ed +eatures re*resent the boundar% signatures as belo-C
6hereC
*i@els on the boundar% n#'C number o+ *i@els -ith unit distance outside the region nO'C number o+ *i@els -ith unit distance inside the region nskel C number o+ *i@els l%ing on the skeletal image a+ter s*urring correction.
,ait +rame 0ctual skeleton S*urred skeleton Bigure DC Silhouette +rame and its corres*onding skeleton
S*urring corres*onds to erosion o*eration o+ a set 0 -ith a circular disk -hich is large enough to +it in that set so that it touches the boundar% o+ set 0 at t-o or more *laces. 6ork is going on to add more signatures that can increase scalabilit% and robustness
Skeleton is obtained +rom distance ma**ed level set +unction using the +ollo-ing conditionC Skeleton, S J K@ L @ J Gma@M nskel corres*onds to number o+ *oints on skeleton a+ter s*urring correction. Pere NO#' corres*onds to normali<ed boundar% signature o+ unit length outside and unit length inside the actual boundar%.
'. ?. E. Q.
$oundar% is detected using Cann% edge detector. Distance ma* level set +unction is com*uted using boundar% in+ormation. Skeletal in+ormation is e@tracted. >sing boundar% and skeletal in+ormation +rom level set +unction signatures N#', NO', Nskel are com*uted.
Scaling to a certain e@tent is ensured to be invariant b% normali<ation. "o increase the matching e++icienc%, additional signature is used -ith normali<ed skeleton length o+ silhouette. "hese combinations o+ t-o $oundar% and one Skeletal signature are used in the *resent -ork +or gait recognition. "hese three signatures are com*uted +or all the 5)) images, %ielding a matri@ o+ 5))@E +eature matri@ stored in signature database.
Stage ? com*rises signature com*utation +or the 7uer% image +ollo-ed b% matching to %ield to* best matches. "he metric used +or com*arison is Euclidean metric, though other metrics are also available. "he ob.ective is to obtain signatures -hich are ma@imall% invariant to scaling, rotation, translation, etc. "his is obtained +or each o+ the +rames that are in the database and it is stored in the +orm o+ a matri@, corres*onding to +eature matri@. Stage ? involves similar *rocedure -here the 7uer% image and the all the three signatures are com*uted and is matched -ith the database obtained in *revious section. "he ob.ective is to obtain the signatures -hich are ma@imall% invariant to scaling, rotation, translation, etc.
+or the height. "he $all o*erator used is a ?@? dilator. 2ur e@*erimentation %ielded the correct match -ithin to* ') retrieved images. Po-ever, the *er+ormance deteriorated +or erosion, still retaining the correct match -ithin to* QR. "he result +urther deteriorated +or increased distortion +or more than ?@? mor*hological o*erator and <oom in 1T '.?D3 and <oom out 1T ).9D3. "hus our method has e@hibited invariance to translation and limited invariance to scaling +ig.
VI. Conclusion
>sing the database o+ silhouettes, -e have *resented three signatures -hich are invariant to translation and scaling. "hese signatures are derived using boundar% and skeletal +eatures. 6hen a 7uer% image is encountered, its three signatures are com*uted and the best match is +ound b% calculating its similarit% measure -ith each image in the database. Results are +ound to be satis+actor% and the *reliminar% +indings are encouraging. "he search is also com*utationall% e++icient. "he *resent -ork is no- e@tended to e@*lore additional signatures -hich can render rotation invariance. 6e are also -orking to identi+% uni7ue signatures that e@hibit a *attern similarit% bet-een di++erent gait videos. !t is observed to a limited e@tent, N#', e@hibits similarit% bet-een sub.ects -hich re*resent Unormal gaitV. 0n% sus*icious gait video is +ound to deviate +rom this *attern
Acknowledg$ent %eferences
"he tem*late -ill number citations consecutivel% -ithin brackets :';. "he sentence *unctuation +ollo-s the bracket :?;. Re+er sim*l% to the re+erence number, as in :E;Wdo not use URe+. :E;V or Ure+erence :E;V e@ce*t at the beginning o+ a sentenceC URe+erence :E; -as the +irst ...V Number +ootnotes se*aratel% in su*erscri*ts. lace the actual +ootnote at the bottom o+ the column in -hich it -as
cited. Do not *ut +ootnotes in the re+erence list. >se letters +or table +ootnotes. >nless there are si@ authors or more give all authors/ names= do not use Uet al.V. a*ers that have not been *ublished, even i+ the% have been submitted +or *ublication, should be cited as Uun*ublishedV :Q;. a*ers that have been acce*ted +or *ublication should be cited as Uin *ressV :D;. Ca*itali<e onl% the +irst -ord in a *a*er title, e@ce*t +or *ro*er nouns and element s%mbols. Bor *a*ers *ublished in translation .ournals, *lease give the English citation +irst, +ollo-ed b% the original +oreign# language citation :4;.
:'; ,. Eason, $. Noble, and !.N. Sneddon, U2n certain integrals o+ :?; :E;
:9;
Ai*schit<#Pankel t%*e involving *roducts o+ $essel +unctions,V hil. "rans. Ro%. Soc. Aondon, vol. 0?Q9, **. D?(#DD', 0*ril '(DD. 1references3 X. Clerk Ma@-ell, 0 "reatise on Electricit% and Magnetism, Erd ed., vol. ?. 2@+ordC Clarendon, '5(?, **.45#9E. !.S. Xacobs and C. . $ean, UBine *articles, thin +ilms and e@change anisotro*%,V in Magnetism, vol. !!!, ,.". Rado and P. Suhl, Eds. NeYorkC 0cademic, '(4E, **. ?9'#ED). K. Elissa, U"itle o+ *a*er i+ kno-n,V un*ublished. R. Nicole, U"itle o+ *a*er -ith onl% +irst -ord ca*itali<ed,V X. Name Stand. 0bbrev., in *ress. Y. Yoro<u, M. Pirano, K. 2ka, and Y. "aga-a, UElectron s*ectrosco*% studies on magneto#o*tical media and *lastic substrate inter+ace,V !EEE "ransl. X. Magn. Xa*an, vol. ?, **. 9Q)#9Q', 0ugust '(59 :Digests (th 0nnual Con+. Magnetics Xa*an, *. E)', '(5?;. M. Young, "he "echnical 6riter/s Pandbook. Mill Valle%, C0C >niversit% Science, '(5(.