Satellite Imagery-based Adaptive Background Modeling and Shadow Suppression - PowerPoint PPT Presentation

Loading...

PPT – Satellite Imagery-based Adaptive Background Modeling and Shadow Suppression PowerPoint presentation | free to download - id: 23bce1-ZDc1Z



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Satellite Imagery-based Adaptive Background Modeling and Shadow Suppression

Description:

Satellite Imagerybased Adaptive Background Modeling and Shadow Suppression – PowerPoint PPT presentation

Number of Views:34
Avg rating:3.0/5.0
Slides: 50
Provided by: Anup3
Learn more at: http://cvrr.ucsd.edu
Category:

less

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

Title: Satellite Imagery-based Adaptive Background Modeling and Shadow Suppression


1
Satellite Imagery-based Adaptive Background
Modeling and Shadow Suppression
  • Research Presentation for Prelim Exam
  • Anup Doshi

February 16, 2007
Committee Members Professor Mohan Trivedi
(Advisor) Professor Truong Nguyen Professor Nuno
Vasconcelos
2
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

3
Situational Awareness is
  • Having a complete understanding of an environment
  • Acquiring knowledge of what events happen when
  • Is the event a common event?
  • Did a daily event not happen today?
  • Is something abnormal happening?
  • Is the event dangerous?
  • Events that likely can include people and
    vehicles, happening anytime day or night

4
Research Motivations
  • First steps detecting objects
  • Robust, Fast
  • Work in Day/Night
  • Remove Shadows/Highlights
  • Types of cameras
  • Omnidirectional, pan-tilt-zoom (ptz), etc.
  • Occlusions
  • General (intra-class variations, pose)
  • Approach
  • Background removal
  • Foreground object modeling

5
Relevant Literature
  • Hybrid Cone-Cylinder Codebook Model for
    Foreground Detection with Shadow and Highlight
    Suppression
  • A. Doshi and M. M. Trivedi. IEEE Intl Conf. on
    Advanced Video and Signal based Surveillance, Nov
    2006, Australia.
  • Satellite Imagery Based Robust, Adaptive
    Background Models and Shadow Suppression
  • A. Doshi and M.M. Trivedi. Submitted to Signal,
    Image, and Video Processing (SIVP) Journal,
    Special Issue on Multi-sensor Object Detection
    and Tracking, Feb 2007.

6
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

7
Prior Work Background Subtraction
  • Mixture of Gaussians Algorithm Grimson,1999
  • Each pixel modeled as multiple Gaussian
    distributions (3-7)
  • Strengths Robust to illumination changes, fast
    automatic learning
  • Weaknesses Highlights and shadows must be dealt
    with still
  • Codebook algorithm Kim,2005
  • Each pixel modeled as multiple codewords
  • Cylinder representing intensity
  • Strengths Highly robust to lighting variations
  • Weaknesses Very sensitive to many parameters

8
Why Codebook and not Gaussians?
  • Codebook Model
  • Captures structural BG motion over long time
    periods
  • Uses limited memory
  • Adaptively deals with local and global
    illumination changes (shadows)
  • Allows foreground motion during training process
  • Works on compressed video
  • Layered modeling allows various levels of BG

9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
Background x Foreground x Codewords o
14
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

15
Prior Work Moving Cast Shadow Suppression
  • Discard the shadows of (moving) foreground
    objects
  • Comparative review in Prati 2003

16
Prior Work HSV Shadow Suppression
  • Idea a shadow has similar chromaticity but
    lower brightness than that of the same pixel in
    the background image.
  • Use this knowledge in HSV color space to weed out
    the shadows. Cucchiara 2001

17
From Cucchiara et al.
18
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

19
Modified Codebook Model HC3
Codebook Background Model
HSV Shadow Suppression Model
Hybrid Cone-Cylinder Codebook Model
20
Modification 1 Color Spaces
  • In RGB space, intensity estimated as L2-norm of
    (R,G,B) vector
  • green (0,1,0) appears brighter than blue (0,0,1)
  • HSV space encodes luminance in Value component
  • Color information encoded separately in Hue and
    Saturation (chrominance)
  • Choose HSV space to model illumination-invariant
    background model

21
Modification 2 Hybrid Cone-Cylinder Volume
  • Advantages of fixed volume simplicity and
    effectiveness
  • Analyzing pixels of similar color as lighting
    intensity increases
  • Forms a cone in HSV space

22
HC3 Background Segmentation Model
23
HC3 Results Comparisons
Raw Video
Ground Truth
HC3 Model Output
Original Codebook Model Output
24
HC3 Results Comparisons
Raw Video
Ground Truth
HC3 Model Output
Original Codebook Model Output
25
HC3 Results Comparisons
Raw Video
Ground Truth
HC3 Model Output
Original Codebook Model Output
26
HC3 Results Comparisons
  • HC3 Performance vs. Other Algorithms
  • Speed 40fps on 320x240 video
  • Memory Minimal (1-2 codewords on average)

27
HC3 Results Comparisons
Sensitive Detections with Omnidirectional Cameras
Night Scene
28
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

29
Satellite-Imagery-based Models
  • Informative source of illumination conditions
  • Ubiquitous and Robust
  • 24-7 Operation

30
Satellite-Imagery-based Models
  • Inverse relation between reflectivity and
    transmissivity

Liou 1992
31
Sensor Fusion Examining Model Parameters
  • Shadow and BG Models heavily dependent upon
    ambient illumination conditions
  • Idea Use auxiliary information to determine
    illumination
  • Satellite Data easily available

32
Satellite-Imagery-based Models
33
(No Transcript)
34
Satellite Data Analysis
Relative Sunlight Intensity (RSI) Normal
Reflecatance Measured Reflectance
35
Satellite Data Analysis
  • Cloudiness State estimated from RSI using MAP
  • Priors assumed uniform (ML) but could be adjusted
    based on weather history
  • Likelihoods modeled as Gaussians, trained using
    manually labeled sample data

36
Satellite Cloudiness ? HC3 Parameters
  • Intuition
  • Sunny days ? more shadows ? decrease a
  • Partly cloudy days ? quick, large intensity
    increases ? increase ß
  • Cloudy days ? small variations, no shadows ?
    decrease ß, Imax and increase a, Imin.

37
Satellite Cloudiness ? HC3 Parameters
  • Overview automatic model parameter discovery
  • Evaluate the performance of each of N parameter
    sets
  • BG Subtraction over range of conditions from
    sunny to overcast
  • Measure number of foreground pixels per frame,
    for each model
  • Reduce the complexity of comparisons
  • Principal Components Analysis on measurements
  • K-means clustering ? 3 representative clusters of
    models
  • Choose model k closest to the mean of each
    cluster
  • Assign to model k a label from overcast, partly
    cloudy, sunny
  • Given satellite-based parameter corresponding to
    labelk, choose model k.

38
Satellite Cloudiness ? HC3 Parameters
39
Satellite Cloudiness ? HC3 Parameters
40
Satellite Cloudiness ? HC3 Parameters
41
Satellite Cloudiness ? HC3 Parameters
42
Satellite Cloudiness ? HC3 Parameters
43
Satellite-based Analysis
  • Automatically discovered parameters
  • Model switching
  • for optimal
  • performance

44
8am Video Satellite-based Cloudiness Parameter
? Overcast
45
Satellite-based Model Switching
46
12pm Video Satellite-based Cloudiness Parameter
? Partly Cloudy
47
Satellite-based Model Switching
48
3pm Video Satellite-based Cloudiness Parameter
? Sunny
49
Satellite-based Model Switching
50
Satellite-based Model Switching
51
Outline
  • Motivation
  • Relevant Work
  • Codebook-based background subtraction
  • Shadow suppression in HSV color space
  • Contributions and Results
  • HC3 Background Segmentation Model
  • Satellite Imagery-based Parameter Estimation
  • Conclusions and Future Work

52
Conclusions
  • Developed HC3 background segmentation and shadow
    suppression model
  • Fast, effective, and fundamentally coherent
  • Shown to outperform state-of-the-art
  • Proposed satellite data as a new paradigm in
    multi-sensory surveillance
  • Cheap, ubiquitous, informative
  • Demonstrated integration scheme with background
    models
  • Presented evidence of usefulness in updating
    model parameters

53
Satellites, backgrounds, and shadows Ongoing and
Future Work
  • Deeper analysis and comparisons
  • HC3 and other models such as Mixture of
    Gaussians.
  • Further integration of satellite information
  • Smooth/continuous parameter control from
    satellite data
  • Examine satellite potential in other models
  • Determine parameter sensitivity of models in
    various conditions
  • Characterize performance over 24-7 operations.

54
Thank you!
  • Professor Trivedi for support, guidance, and
    motivation
  • Professors Nguyen and Vasconcelos for being on my
    committee
  • CVRR Lab, including Brendan, Erik, Junwen,
    Ramsin, Sangho, Shankar, Shinko, Steve, and
    Tarak for helpful discussions, corrections, and
    comments
About PowerShow.com