Crowd Estimation and Monitoring Techniques

1
Crowd Estimation and Monitoring Techniques

• Doug Gray
• 3-9-05

2
Outline

• Introduction
• Detection Based Methods
• Mapping Based Methods
• Lin et al., Perspective Transform
• Cho et al., Neural Based
• Our Method, Scene Invariant Estimation

3
What is Crowd Estimation

• Mapping an image to the number of people it
  contains

4
Techniques Overview

• Detection Based
• Run a detector, count or cluster the output
• Mapping Based
• Extract features and map them to a value
• Hybrid
• Use detector output as a feature

5
Detection Based Methods

• Pros
• Good zero output accuracy
• Cons
• Requires a good detection algorithm
• Not scalable to large crowds

6
Detectors

• Range from simple
• Connected Components
• Skin tone classifier
• To very complex
• AdaBoost
• Hough Transform
• Harr Wavelet Transform
• Etc.

7
Detection Clustering

• An inaccurate detector count may provide
  duplicates
• Biliotti 05 proposes
• a clustering algorithm

8
Clustering Results
ICA Clustering With Hausdorff Distance
Maximum of Cross-Correlation Clustering With Eucli
dean Distance
9
Mapping Methods

• Pros
• Scalable to large crowds
• Cons
• Hard to make scene invariant
• Can be divided into feature extraction, density
  estimation and classification

Example with 1D Feature Vector
10
Feature Extraction

• Basic
• Moving pixels
• Edge Pixels
• Complex
• Texture Analysis
• Spectral Content
• Optical Flow Levels
• Fractal Dimension

11
Basic Features

• Simple summation of
• Intensity Values
• Foreground Pixels
• Edges / Fast Edges

12
Histograms

• Expansion of basic features
• May tell us more about people specifically

13
Linear Area Transform (Davies et al.)

• Attempts to capture head motion through a global
  motion parameter g
• Frequency analysis reveals head motion (2 Hz) is
  detectable

14
Optical Flow Levels

• Ixu Iyu It 0
• Pixel directions may correspond to crowd size an
  direction

15
Radar Plots
16
Fractal Dimensions

• Minkowski Sausages!
• Equivalent to dilation of an edge map
  (morphological operation)

17
Fractal Density

• Fractal dimensionality is related to crowd
  density
• Comparable to other texture analysis methods

18
Density Estimation

• Simple Normalization
• Use y coordinate to weight inputs
• Perspective Transform
• Simple geometric correction
• Back Propagated Area (homography)
• Full correction

Density and ROI Map (exaggerated weights)
19
Lin, et al. s Method

• Learning Based Face detection
• Perspective Transform based on face data
• Clustering and Estimation

20
Bootstrap training

• Do initial training
• Run detector on new data
• Retrain with the false detections

21
Vanishing Point

• Relate detection box to distance from axis
• Use linear regression to find vanishing point

22
Perspective Transform

• Knowing the vanishing point and the focal length,
  we can estimate the angle

23
Equidistant lines

• A segment Pi1Pi2 is mapped into the image with
  simple geometry
• This will later
• be used for
• density estimation

24
Detection Grouping
25
Crowd Size Estimate
26
Artificial Results
Accuracy decreases with visual angle The
opposite effect of normal mapping estimation
methods
Puppets are used to calibrate the system with
varying camera angles
27
Real Results
88
222
28
Cho et al. s Method

• Basic Mapping Method
• 3 Simple Features
• Neural Classifier

29
Details

• Reference Image based Background subtraction

• Features
• Sum of Black Pixels
• Sum of White Pixels
• Sum of Edge Pixels

30
Results

• Random Sampling Results
• Paper focused on training methods

31
Our Method

• Feature Histograms
• Density from Homography Estimate
• Several Fitting Methods

32
Feature Extraction
Input Image
Background Detection
Edge Detection with Blob Masking
Edge Detection
33
Feature Histograms

• Edges quantized by angle
• (8 bins)
• Blobs quantized by pixel count
• (6 bins)

34
Back Propagated Homography

• Estimate the homography w.r.t. the ground plane

• For each pixel in the image
• Draw a box
• Compute the boxs coords. on the ground plane
• Find the area of the box
• Normalize area w.r.t. reference person

35
Density Estimate Results
36
Results!
37
References

• S.-Y. Cho, T. W. S. Chow, and C. T. Leung, A
  neural-based crowd estimation by hybrid global
  learning algorithm, IEEE Trans. Syst. Man,
  Cybern. B, vol. 29, pp. 535-541, 1999.
• S-F. Lin, J-Y. Chen, H-X. Chao, Estimation of
  Number of People in Crowded Scenes Using
  Perspective Transformation, IEEE Tran. Syst.
  Man, Cybern. A, vol. 31, pp. 645-654, 2001
• A. C. Davies, J. H. Yin, and S. A. Velastin,
  Crowd monitoring using image processing,
  Electron. Commun. Eng. J., vol. 7, pp. 37-47,
  1995.
• A. N. Marana, L. da F. Costa, R. A. Lotufo,
  Estimating crowd density with Minkowski fractal
  dimension, in Proc. IEEE Int. Conf. Acoust.,
  Speech, Signal Processing, vol. 6, Phoenix, AZ,
  1999, pp. 3521-3524
• T. Schlogl, B. Wachmann, W. Kropatsch, H.
  Bischof, Evaluation of People Counting Systems,
  Unknown
• D. Biliotti, G. Antonini, J. P. Thiran,
  Multi-layer hierarchical clustering of
  pedestrian trajectories for automatic counting of
  people in video sequences, in Proc. IEEE
  Workshop on Motion and Video Computing, 2005
• S. A. Velastin, J. H. Yin, A. C. Davies, M. A.
  Vicencio-Silva, R. E. Allsop, A. Penn, Automated
  measurment of crowd density and motion using
  image processing, Road Traffic Mon. and Cont.,
  pp. 127-132, 1994
• S. Bouchafa, D. Aubert, S. Bouzar, Crowd motion
  estimation and motionless detection in subway
  corridors by image processing, IEEE 1998, pp.
  332-337
• A. N. Marana, L. F. Costa, R. A. Lotufo, S. A.
  Velastin, On the Efficacy of Texture Analysis
  for Crowd Monitoring, Anais do XI SIBGRAPHI,
  1998 pp. 1-8