Distributed Face Detection by xiuliliaofz


									  Face Detection in
Distributed Camera
  Sensor Networks
  Rajapaksage Jayampathi
    DM Rasanjalee Himali
   Introduction
         Face Detection
         Distributed Camera Sensor Networks (DCS)

   Current Work
         Viola-Jones object detection framework

   Current Limitations:

   Proposed Work
         Distributed Face Detection Framework
         Implementation
   Future Plan
Face Detection
   A computer technology that
    determines the locations and sizes of
    human faces in arbitrary (digital)

   It detects facial features and ignores
    anything else, such as buildings,
    trees and bodies.

   Is a specific case of object-class

   Many algorithms implement the face-
    detection task as a binary pattern-
    classification task
Face Detection [Contd.]
   In our work we focus specially on the problem of face
    detection in still images.

   The most straightforward variety of this problem is the
    detection of a single face at a known scale and orientation.

   Even this, is a nontrivial problem.

   The most immediate application that comes to mind for face
    detection is as the first step in an automated face
Distributed Smart Cameras (DSCs)
   Are real-time embedded systems that achieve computer
    vision using multiple cameras.

   One of the basic and most important problems of smart camera
    networks is face detection

   A smart camera consists of sensing, processing, and
    communication units which deliver some abstracted data of the
    observed scene.

   They perform a verity of image processing algorithms
     Ex : motion detection, segmentation, tracking, and object
      recognition and delivers color and geometric features,
      segmented objects or high level decisions as output
Distributed Smart Cameras (DSCs)
   The main goal for the cameras is to provide sufficient
    processing power and fast memory for processing the images
    in real time while keeping power consumption low

   DSCs introduce distribution and collaboration to smart

   These cameras use distributed algorithms to perform camera

   Multiple threads of processing may take place on different
    processing nodes in parallel.

   These camera sensors generate more data and make analysis
    difficult in many applications.
Distributed Smart Cameras (DSCs)
   Distributed smart cameras distribute not only
    sensing but also processing

   On the one hand, smart cameras can serve as
    processing nodes that perform some fixed
    preprocessing but still deliver data to a central

   On the other hand, processing may be organized in
    a completely decentralized fashion where the smart
    cameras organize themselves and collaborate
Smart Cameras in Sensor Networks
2. Current Work
Current Work
   There are various solutions to this problem

   Most of which deal with faces at arbitrary
    scales, and assume an upright face

   Most methods use a learning algorithm on a
    training set to begin the detection process.
Robust Real-Time Face Detection
[Viola & Jones ,2004]
   capable of processing images extremely
    rapidly while achieving high detection

   Integral Image:
       Introduces a new image representation
        called an integral image that allows for
        very fast feature evaluation
       The integral image can be computed from
        an image using a few operations per pixel.
       Once computed, any one of these Haar-like
        features can be computed at any scale or
        location in constant time.
Robust Real-Time Face Detection
[Viola & Jones ,2004]
   The integral image at location x, y contains the sum of the pixels
    above and to the left of x, y, inclusive:

   where ii (x, y) is the integral image and i (x, y) is the original image

   Using the following pair of recurrences:

   (where s(x, y) is the cumulative row sum, s(x,−1) = 0, and ii (−1, y) =
    0) the integral image can be computed in one pass over the original
Robust Real-Time Face Detection
[Viola & Jones ,2004] [Contd.]
   Uses a simple and efficient classifier
    that is built by selecting a small
    number of important features from
    a huge library of potential features
    using AdaBoost

   Combine successively more complex
    classifiers in a cascade structure
    which dramatically increases the
    speed of the detector by focusing
    attention on promising regions of the

   The final detector is scanned
    across the image at multiple
    scales and locations. Scaling is
    achieved by scaling the detector
    itself, rather than scaling the image.
Disadvantages of Current Work
   many algorithms are centralized algorithms and are not designed for
    distributed or resource constrained environments.

   There are only a handful of parallel architectures for face detection
    have been proposed in the literature so far.

   None of these take into consideration the multiple views different
    cameras may have due to its relative position in a global 3-D coordinate

   Many current approaches assume up-right faces although few
    algorithms have been devised to address multi-view face problem.

   Viola & Jones[2] approach limits itself to a limited set of features and
    classifiers to reduce computation.
3. Our Approach
Problem Statement
   Our work is an extension of the face detection
    algorithm proposed by Viola-Jones.

   The major distribution of load is contributed by two
         Computation of integral image and the
         Face detection.

   The objective of this distributed face detection
    framework is to achieve higher detection rates at low
    false positive rates by using the power of distributed
    computing in a DSC network.
Our Approach
   The advantage of distributed computing is
    achieved at three stages:
    1.   Integral Image Computation

    2.   Integral Image Distribution and

    3.   Cascade and Feature Distribution
1. Integral Image Computation
   The Integral image calculation can be distributed among cameras by
    formulating the problem as a parallel prefix sum calculation

   Given the original image at a sensor s, it can partition the image at
    subwindows to its neighbors. The most intuitive way will be to do a
    row-wise partitioning.

   The neighbors will calculate the partial-integral image for the given
    subwindow and send back to s. s calculates the global integral

                                    Sj         Sk

                         (a)        (b)        (c)
                                            2.                 3. Calculate Partial integral Image
                                            Distribute image
                                            sub windows to            Partial Integral image         SENSOR B
            Image Sub Window

                                                                      Partial Integral image         SENSOR C

1.                                                                                                   SENSOR D
                                                                      Partial Integral image
Receive          Original Image

            Image Sub Window                4.Gather Partial          Partial Integral image         SENSOR E
                                            Integral Images

          Combined Partial Integral Image

                          5.Calculate global Integral

              Global Integral Image
                                            SENSOR A
                                                                             PARALLEL PREFIX SUM
Integral Image Distribution
   A better approach however is to
    incrementally calculate integral image at
    sensor which received the original image
    while distributing it to different nodes for face

   Each incremental version of integral image
    corresponds to a different scale and
                                                        Detect Faces


                                                                   SENSOR B
                            Distribute integral Image

           Integral Image
                                                        Detect Faces

                SENSOR A

                                                                   SENSOR C

The detection process
for different scales are                                  Detect Faces

completed by
neighborhood sensor
nodes concurrently.                                               SENSOR D
3. Cascade and Feature Distribution
   The Viola Jones algorithm limits itself to a limited set
    of features for faster results.

   However, in a distributed environment like DSC
    network, larger number of feature set can be used
    without compromising the faster face detection.

   Also, the pipeline architecture of cascades can be
    implemented in a distributed environment by
    assigning set of adjacent classifiers in the cascade
    to sensors.
Detect Faces

           SENSOR B
                          F                  F             F

                                  T              T                T
Detect Faces             1               2               3
                      SENSOR X        SENSOR Y         SENSOR Z

                                 Attentional Cascade
           SENSOR C

Detect Faces

           SENSOR D
                                                 MASTER NODE PROCESS
      Wait For Image reception
                  Original Image

         Received Image?                Row-wise partition image                    Unused Neighbor?
                                                           Image Sub-window        Yes


                                                       Partial Integral Image No
                                                                                   Received ALL partial
                                                                                    Integral Images?


                                                                           Calculate Global Integral Image

Yes          Received ALL Face         No                                Row-wise partition Integral image
            Detected Sub-Window
                                              Unused Neighbor

                             No                              Yes
                   Face Detected Sub-window        Integral Image Sub-window
                                           SLAVE NODE PROCESS

   Wait For Image reception
                Partial Image Sub-Window

    Received Partial Image
        Sub Window?


Calculate Partial Integral Image           Received Partial ii?

  Send Partial ii to MASTER
                                             Detect Faces

                                    Send Face detected ii to MASTER
4. Future Work
Future Work
   It is possible to devise a distributed face
    detection algorithm in DSC networks
    incorporates multi-view face detection in DSC

   Use Value of Information theory to detect
    faces reliably.
   Multi-Camera Networks: Principles and Applications, Hamid Aghajan, Andrea Cavallaro,
   Robust Real-time Object Detection, Paul Viola , Michael Jones, 2001
   Parallelized architecture of multiple classifiers for face detection, Bridget B. Jung Uk Cho
    ,IEEE International Conference on Application-specific Systems, Architectures and
    Processors (ASAP) , 2009
   CMUcam3: An Open Programmable Embedded Vision Sensor , Anthony Rowe, Adam
    Goode, Dhiraj Goel, Illah Nourbakhsh, , Carnegie Mellon Robotics Institute Technical
    Report, RI-TR-07-13 May 2007
   Fast Multi-View Face Detection, M. Jones, P. Viola, MERL, TR2003-96, July 2003
   Robust Multi-View Multi-Camera Face Detection inside Smart Rooms Using Spatio-
    Temporal Dynamic Programming, Z. Zhang, G. Potamianos, M. Liu, T. Huang, In
    Proceedings of the International Conference on Automatic Face and Gesture
    Recognition, pp.407-412, 2006
   Robust Real-Time Face Detection, P. Viola and M. Jones, International Journal of
    Computer Vision, vol. 57, no. 2, pp. 137-154, 2004.
   Towards a Real-time and Distributed System for Face Detection, Pose Estimation and
    Face-related Features , J. Nesvadba, A. Hanjalic, P. M. Fonseca1, B. Kroon, H. Celik, E.
    Hendriks, Int. Conf. on Methods and Techniques in Behavioral Research, 2005
   A statistical method for 3D object detection applied to faces and cars, Schneiderman, H.
    and Kanade, T, In International Conference on Computer Vision, 2000
   Neural network-based face detection, Rowley, H., Baluja, S, and Kanade, T, IEEE Patt.
    Anal. Mach. Intell, 1998
    Dual camera system for face detection in unconstrained environments Marchesotti,
    L. Marcenaro, L. Regazzoni, C. DIBE, Genoa Univ., Italy, ICIP, 2003

To top