Robust 3D Head Pose Classification using Wavelets by Mukesh C. Motwani Dr. Frederick C. Harris, Jr.,...

Preview:

Citation preview

Robust 3D Head Pose Robust 3D Head Pose Classification using Classification using WaveletsWavelets

byby

Mukesh C. MotwaniMukesh C. Motwani

Dr. Frederick C. Harris , Jr., Thesis AdvisorDr. Frederick C. Harris , Jr., Thesis Advisor

December 5December 5thth, 2002, 2002

A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Computer Engineering

University of Nevada, Reno

OverviewOverview

Problem DescriptionProblem Description Proposed SolutionProposed Solution

Wavelets and PCA Wavelets and PCA Future WorkFuture Work Questions and AnswersQuestions and Answers DemoDemo

What is Pose Classification What is Pose Classification ??

Potential ApplicationsPotential Applications

Face RecognitionFace Recognition

Potential ApplicationsPotential Applications

Virtual RealityVirtual Reality

Potential ApplicationsPotential Applications

Video ConferencingVideo Conferencing

Potential ApplicationsPotential Applications

RoboticsRobotics

Potential ApplicationsPotential Applications

Driver VigilanceDriver Vigilance

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

IdentityIdentity

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

Face DetectionFace Detection

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

Facial ExpressionFacial Expression

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

Changes in distance from CameraChanges in distance from Camera

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

OcclusionOcclusion

What makes Pose What makes Pose Detection so difficult?Detection so difficult?

Changes in IlluminationChanges in Illumination

Previous WorkPrevious Work

Feature BasedFeature Based Appearance BasedAppearance Based Model BasedModel Based Other Combined ApproachesOther Combined Approaches

Feature Based Pose Feature Based Pose DetectionDetection

Kalman FilterKalman Filter Gabor JetsGabor Jets Elastic Graph BunchingElastic Graph Bunching

Appearance Based Pose Appearance Based Pose DetectionDetection

Template MatchingTemplate Matching Neural Networks / Wavelet Neural Neural Networks / Wavelet Neural

NetworksNetworks SVMSVM PCAPCA

Model Based Pose Model Based Pose DetectionDetection

Skin ColorSkin Color AnthropometryAnthropometry

Other Combined Other Combined ApproachesApproaches

Volumetric Frequency Volumetric Frequency RepresentationRepresentation

Hairline contourHairline contour Optical FlowOptical Flow

PCA Based ApproachPCA Based Approach

““View Based”View Based” ““Parametric”Parametric”

1. Motivation for choosing PCA

2. Sub categories in PCA based approach

Block DiagramBlock Diagram

Training StageTraining Stage

Training Images Training Images

Detection StageDetection Stage Detected Detected

PosePose

Query Query ImageImage

Discrete Wavelet Transform

Principal Component

Analysis

ManifoldPlot

Discrete Wavelet Transform

Principal Component

Analysis

Eigen SpaceProjection

Wavelet TransformWavelet Transform

Filter BanksFilter Banks

Wavelet TransformWavelet Transform

Subband DecompositionSubband Decomposition

Low Pass Output

High Pass Output

LL HL

LH HH

DWT of Lena ImageDWT of Lena Image

Selection of Resolution Selection of Resolution LevelLevelDistribution of energy in LL subband at different levels of discrete wavelet transform

Selection of Resolution Selection of Resolution LevelLevel

Principal component analysis Principal component analysis (PCA)(PCA)1. Mean Image (Enlarged)

2. Compute covariance matrix

3. Compute eigen-vectors and eigen-coefficients

Manifold PlotManifold Plot

ExperimentationExperimentation

Facial ExpressionFacial Expression

ExperimentationExperimentation

Changes in distance from cameraChanges in distance from camera

ExperimentationExperimentation

Changes in IlluminationChanges in Illumination

ExperimentationExperimentation

Identity changeIdentity change

PerformancePerformance

Classification accuracy of 84%Classification accuracy of 84% Real time capability 12-15 fpsReal time capability 12-15 fps

ConclusionConclusion

FasterFaster More RobustMore Robust CheaperCheaper

Future WorkFuture Work

Face DetectionFace Detection Identity IndependentIdentity Independent Undecimated Wavelet TransformUndecimated Wavelet Transform

Questions ? / Answers Questions ? / Answers Real-time DemoReal-time Demo