A neural approach to extract foreground from human movement images

1
A neural approach to extract foreground from
human movement images

• S.Conforto, M.Schmid, A.Neri, T.DAlessio
• Compute Method and Programs in Biomedicine
  82(2006) 73-80
• Che-Wei Sung 2007/12/24

2
Outline

• Introduction
• Materials and methods
• Subtraction techniques
• Neural approach
• Qualitative evaluation of results
• Objective evaluation of results
• Conclusions

3
Introduction

• The capture of human movement is a hot topic for
  surveillance, control and analysis.
• In the framework of human movement analysis often
  consists of separating the moving subject (i.e.
  foreground) from the background by techniques
  based on temporal or spatial.

4
Introduction

• Temporal data can be used in two different ways,
  subtraction and flow, while spatial techniques is
  applying markers on foreground.
• Mixed approaches have been presented, but none
  can be considered as outperforming in general
  terms.
• The work in this paper is to development of a
  markerless capture system for movement analysis
  application by making the ANN learn the
  background.

5
Materials and methods

• The moving subject is detected by analyzing the
  differences between the background scene.
• , corresponding to the background image
• ,represents the generic s-th image frame
  extracted from the video sequence gathering the
  moving subject over the background scene.

6
Materials and methods - Subtraction

• 1. Compute the image difference
• 2. For each row of , calculate the
  vectors of mean value
  , and standard deviation
• 3. Determine the 3D-classification interval
• , if
  a pixel lies inside the domain
• , it is
  classified as background, vice versa as
  foreground.
• 4. Detect the largest connected area that
  considered as actual foreground.

7
Materials and methods - ANN

• Neural network makes use of a Kohonen map,
  composed of (88) neurons

8
Materials and methods - ANN

• In this work, background image is partitioned
  into blocks of (88) pixels, and arranged in a
  mono-dimensional vector composed of
• H (643) 192 components for training data.

9
Materials and methods - ANN

• Assume the image is subdivided into B blocks of
  size (88), the training input vector
  Vbb1,2B and the size of each synaptic weight
  vector is randomly initialized in 0,1,
  where h1,2H

10
Materials and methods - Training

• 1. One input vector Vb is randomly extracted
  from the training set, and feeds the network.
• 2. In each neuron nij, the distance dij,b(k)
  between Vb and (k) is calculated
• 3. The best match neuron nBM(k) is defined as
  the nij whose corresponding vector (k) is at
  the minimum distance from Vb.

11
Materials and methods - Training

• 4. The weight vectors are updated by using
  typical Kohonen neighborhood procedure.
• where

12
Materials and methods - Training

• The training has been considered as complete
  when, for the 98 of training samples, the
  association between each Vb and the corresponding
  best match neuron is not altered

13
Materials and methods - Testing

• 1. undergoes Data Shaping, creating a
  set of vectors .
• 2. For each block, the best match neuron is
  identified by considering the minimum Euclidean
  distance criterion.

14
Materials and methods - Testing

• 3. is used to build up a distance
  matrix, whose elements are rearranged respecting
  the spatial of
• , where each element occupies the
  position of block.
• 4. For each row of distance matrix, the mean
  value
• and the standard deviation
  are calculated.
• Blocks with corresponding distance values
  outside the range are considered
  as foreground.
• 5. A segmentation mask is built up by marking
  pixels with 0 for background, 1 for foreground.

15
Qualitative evaluation of results

• The proposed algorithm have been applied to
  analyze human body movement during three motor
  tasks gait, pitching a ball and standing up from
  a chair.
• The training of Kohonens map has met convergence
  after around 90000 presentations of background
  blocks.

16
Qualitative evaluation of results
17
Qualitative evaluation of results
18
Objective evaluation of results

• quality_indexs 0.3shape_regs
• 0.33temp_stabs
  0.37contrasts
• shape_regs the regularity of segmented object
  shape.
• temp_stabs the stability along the video
  sequence of extracted object.
• contrasts the contrast between the inside and
  the outside of the object evaluated along the
  border.

19
Objective evaluation of results
20
Conclusions

• The work proposes a new unsupervised approach for
  foreground extraction in human movement images
  based on ANN and the presented results
  demonstrate it is suitable.