Multi-temporal Scene Models for Visual Surveillance - PowerPoint PPT Presentation

Loading...

PPT – Multi-temporal Scene Models for Visual Surveillance PowerPoint presentation | free to download - id: ce122-ZDc1Z



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Multi-temporal Scene Models for Visual Surveillance

Description:

Compare each frame to the scene model. Regions that differ from model are changes ... No moving cars. Small a. Large a. 9/29/2006. Nathan Jacobs. 15. Uses of ... – PowerPoint PPT presentation

Number of Views:33
Avg rating:3.0/5.0
Slides: 19
Provided by: cseW
Learn more at: http://www.cse.wustl.edu
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Multi-temporal Scene Models for Visual Surveillance


1
Multi-temporal Scene Models for Visual
Surveillance
  • Nathan Jacobs
  • 9/29/2006
  • Advisor Robert Pless

2
Visual Surveillance
  • Many cameras. Many images.
  • Few people.

3
More Cameras per Person
  • Reduce human time per camera
  • Dont show every frame to a person
  • Cycle through cameras
  • Poor temporal context
  • Modify the video to speed understanding
  • For example, highlight recent changes

4
Detecting Change
  • Build scene model
  • Compare each frame to the scene model
  • Regions that differ from model are changes

Scene Model
Current Frame
Threshold
5
Scene Models
  • What type of data?
  • Intensity visual, IR
  • Local Texture
  • Motion
  • What types of models at each pixel?
  • Mean average value of data
  • Gaussian mean and variance
  • Mixture Model many Gaussians
  • Common Property all scene models can be updated
  • Usually a moving average

6
A Simple Scene Model
  • Moving average of pixel intensity
  • Scene model S( t )
  • Pixel value I( t )
  • Filter constant a in range 0,1
  • Update rule
  • S( t ) a S( t - 1) ( 1 a ) I( t )
  • Filter constant a determines the amount the model
    can change each frame

Larger a
Small a
7
Slowly Updating Model
Slow-updating Mean Model
Current Frame
Large a
Threshold
Difference Image
Foreground Pixels
8
Quickly Updating Model
Fast-updating Mean Model
Current Frame
Small a
Threshold
Difference Image
Foreground Pixels
9
Which a is best?
  • Problem difficult to choose a when setting up a
    camera
  • Large a adapt too slowly to non-interesting
    scene changes
  • Small a adapt too quickly and incorporate
    interesting objects into model
  • Key Idea
  • In surveillance, interesting events happen at
    many time scales
  • Our approach Instead of choosing one filter
    constant, choose a set of filter constants

10
A Set of Scene Models
Slow update
Fast update
a
Current Frame
.99999
0.9999
0.999
0.99
0.9
By keeping many background models we can defer
the decision. Best filter constant depends on
scene, application, time-of-day, weather, image
location
Threshold
11
One Pixel over Time
Pixel intensity I( t )
12
One Pixel Scene Model Values
Scene models with different constants
S( t ) with large a
S( t ) with small a
13
One Pixel Difference of Scene Models
  • Logarithmically-spaced filter constants have some
    interesting properties
  • i.e., a 1-e-x,1-e-2x,,1-e-nx for small
    positive x

14
Difference of Scene Models
Small a
Large a
Threshold
No moving cars
No bus
15
Uses of Difference of Scene Models
  • Important question How long has an object been
    at a particular location?
  • Illegally stopped vehicle
  • Abandoned bag in airport

16
Stopped Vehicle Detection Video
17
Detecting Abandoned Bags
18
Summary
  • Goal Increase number of cameras per person by
    modifying the video
  • Approach Maintain scene models at different time
    scales
  • Result Tool for determining how long ago an
    object appeared, based only on simple pixel
    statistics (not an explicit object model).
About PowerShow.com