Geometric Invariants forFacial Feature Tracking with3DTOFCameras
Martin Haker, Martin Böhme · 2007
Thispaperpresents a verysimple feature-based individually, aswellasonthecombination ofthesetwo nosedetector incombined rangeandamplitude dataobtained typesofdata. An important result isthattheperformance bya3Dtime-of-flight camera. Therobust localization ofimageofthedetector onthecombined rangeandintensity datais attributes, suchasthenose, canbeusedforaccurate objectsubstantially better thanon either typeofdataalone. This tracking. We usegeometric features thatarerelated tothe intrinsic dimensionality ofsurfaces. Tofind anoseintheimage,underlines thepotential of3DTOFcameras formachine vision thefeatures arecomputed perpixel; pixels whosefeature valuesapplications. lie inside acertain bounding boxinfeature space areclassified as Previous workhasalready identified thenoseasanimpor- nosepixels, andallother pixels areclassified asnon-nose pixels. tant facial feature fortracking e.g. in(1)and(2). Intheformer Theextent ofthebounding boxislearned onalabeled training set. Despite itssimplicity this procedure generalizes well, thatis,approach thelocation ofthenoseisdetermined bytemplate abounding boxdetermined foronegroupofsubjects accurately matching, undertheassumption thatthesurface aroundthe detects noses ofother subjects. Theperformance ofthedetector tipofthenoseisaspherical Lambertian surface ofconstant isdemonstrated byrobustly identifying thenoseofaperson in albedo. Thisapproach gives veryrobust results underfixed awiderangeofheadorientations. Animportant result isthatlighting conditions andatafixed distance oftheuserfromthe thecombination ofbothrangeandamplitude datadramatically improves theaccuracy incomparison totheuseofasingle typecameraThelatter approach isbased onageometrical model ofdata. Thisisreflected intheequal error rates (EER)obtainedofthenosethatisfitted totheimagedata. onadatabase ofheadposes. Using only therange data, wedetect We alsoconsider thenoseasbeing verywellsuited for noses withanEER of0.66. Results ontheamplitude dataare headtracking, because thenoseisobviously a distinctive slightly better withanEER of0.42. Thecombination ofboth characteristic ofthehumanface. Intermsofdifferential types ofdatayields asubstantially improved EER of0.03. geometry, thetipofthenoseisthepoint ofmaximal curvature