Temporal Correlation and Probabilistic PredictionBased Face Detection Frameworkin Real Time Environment

Embed Size (px)

Citation preview

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    1/21

    TEMPORAL CORRELATION ANDPROBABILISTIC PREDICTION

    BASED FACE DETECTION

    FRAMEWORK

    IN REAL TIME ENVIRONMENTDepartment of Electronics & Electrical Communication

    EngineeringIndian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    2/21

    TEAM

    Indian Institute of Technology, Kharagpur

    Piyush Mayank, Undergraduate Student, IIT

    Kharagpur

    Sudipta Mukhopadhyay, Assistant Professor, IIT

    Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    3/21

    BASIC FACE DECTECTION FRAMEWORK

    FACE DETECTION VIOLA JONES Face

    Detection Algorithm

    Integral Image Representation

    AdaBoost Algorithm for creating Strong ClassifiersCascading of Classifiers

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    4/21

    IMPROVEMENT IN DETECTION

    ALGORITHM

    Increase Detection Rate

    Reduce Computational Complexity

    Decrease False Detection Rate

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    5/21

    PROPOSED ALGORITHMS

    Indian Institute of Technology, Kharagpur

    Search Area Reduction Using Probabilistic

    Prediction

    Minimum Face Size Optimization

    False Detection Removal Using Facial Features

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    6/21

    SEARCH AREA REDUCTION

    Any Rectangle in an image can be determined

    by two diagonal coordinates:

    Use of Detected Face Rectangles of previousframes

    Window Size (WS) of i previous frames. Lets

    say i=3 Indian Institute of Technology, Kharagpur

    (x1,y1)

    (x2,y2)

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    7/21

    SEARCH AREA REDUCTION

    Indian Institute of Technology, Kharagpur

    (x11,y11)

    (x21,y21)

    (x12,y12)

    (x12,y12)

    (x13,y13)

    (x23,y23)

    x

    y

    x1min

    y1min

    x2max

    y2max

    (x1min, y1min)

    (x2max, y2max)

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    8/21

    SEARCH AREA REDUCTION

    Probability of existence does not terminate

    abruptly outside this area.

    In Gaussian Distribution

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    9/21

    SEARCH AREA REDUCTION

    Indian Institute of Technology, Kharagpur

    x

    y

    (x1min, y1min)

    (x2max, y2max)

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    10/21

    MINIMUM FACE SIZE OPTIMIZATION

    To define the minimum size of a face which will be

    searched in the reduced search area.

    Movement of face in frames bear a temporal correlation

    Use of detected face rectangles (DFR) of previousframes

    Window Size of i frames

    Minimum Face Width (MFW) and Length (MFL)

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    11/21

    WORST CASE SCENARIO

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    12/21

    FALSE DETECTION REMOVAL

    Structural Facial Features, the two eyes, nose and

    mouth are detection after successful detection of a

    face

    The face detection is confirmed only if at least oneof these features are detected otherwise the

    detection is rejected

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    13/21

    IMPLEMENTATION

    Indian Institute of Technology, Kharagpur

    Four Input Sequences:

    Input 1: JEVM392_aux_kaumuskel_CamD.mpg

    (Regulated Input)

    Input 2: Real time Regulated Input from Camera Input 3: Behzad.avi (Unregulated Input)

    Input 4: Real Time Unregulated Input from Camera

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    14/21

    IMPLEMENTATION

    PERFORMANCE METRICS

    Detection Rate (in %)

    Frame Rate Per Second ( in sec-1)

    Worst Case Complexity (in msec)

    COMPARED FRAMEWORKS

    VIOLA JONES FACE DETECTION

    PROPOSED ALGORITHMS WITH WINDOW

    SIZE=3,5,7

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    15/21

    RESULTS

    Viola Jones

    FDF

    FDF with

    WS=3

    FDF with

    WS=5

    FDF with

    WS=7

    Detection Rate (in %) 100 100 100 100

    Frame Rate per second (in sec1 ) 26.15 26.29 26.33 26.32

    Worst Case Complexity (inmsec) 99.278 47.235 47.185 47.211

    Indian Institute of Technology, Kharagpur

    Viola Jones

    FDF

    FDF with

    WS=3

    FDF with

    WS=5

    FDF with

    WS=7

    Detection Rate (in %) 96 100 99.2 99.6

    Frame Rate per second (in sec1 ) 18.375 21.16 22.05 21.78

    Worst Case Complexity (inmsec) 125.065 97.63 78.78 94.293

    INPUT

    1

    INPUT

    2

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    16/21

    RESULTS

    Viola Jones

    FDF

    FDF with

    WS=3

    FDF with

    WS=5

    FDF with

    WS=7

    Detection Rate (in %) 58 67.5 68.2 67

    Frame Rate per second (in sec1 ) 20.03 34.05 36.04 23.2

    Worst Case Complexity (inmsec) 56.962 46.678 46.085 46.556

    Indian Institute of Technology, Kharagpur

    Viola Jones

    FDF

    FDF with

    WS=3

    FDF with

    WS=5

    FDF with

    WS=7

    Detection Rate (in %) 83 95.6 92.4 98.5

    Frame Rate per second (in sec1 ) 18.85 21.96 23.15 19.36

    Worst Case Complexity (inmsec) 172.363 132.23 136.63 165.43

    INPUT

    3

    INPUT

    4

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    17/21

    RESULTS

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    18/21

    RESULTS

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    19/21

    RESULTS

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    20/21

    CONCLUSION

    Striking Analogy Between Real Time Face Detection and

    Video Motion Estimation

    Adaptive Search using Probabilistic Prediction

    Window Size is not increased beyond 7 as it would leadto high increase in computational complexity. Tradeoff

    Between Computational Efficiency and Detection

    Efficiency

    Window Size of 5 was found to be most optimal

    System currently works only for frontal faces. Inclusion

    of Profile/Lateral Faces Desirable but not necessary.

    Indian Institute of Technology, Kharagpur

  • 8/11/2019 Temporal Correlation and Probabilistic Prediction Based Face Detection Framework in Real Time Environment

    21/21

    THANK YOU

    Indian Institute of Technology, Kharagpur