Motion Detection and Shadow Suppression in Video Streams

Diplomarbeitspräsentationen der Fakultät für Informatik
FAKULTÄT FÜR
Motion Detection and Shadow
Suppression in Video Streams
!NFORMATIK
Diplomstudium:
Informatik
Technische Universität Wien
Institut für Rechnergestützte Automation
Arbeitsbereich für Mustererkennung und Bildverarbeitung
Betreuer: Univ.-Ass. Dipl.-Ing. Dr. Martin Kampel
Philipp Blauensteiner
Motivation
The IHLS Background Model
Motion detection:
Motion Detection is an important task for a vast number of computer vision applications.
Goal: segment an image in a video stream into (moving) foreground and (static) background
The IHLS colour space
Colours represented in terms of hue, luminance and saturation
Saturation is not normalised (in contrast to similiar colour spaces like HSV)
Assures, that
- saturation of achromatic colours is always low
- saturation is independant of brightness function
Shadows of moving objects:
Shadows cast by moving objects are moving as well and are detected as foreground objects
But: classifying those shadows as belonging to the foreground introduces problems:
object counting / connected component
input image
shape of objects
Saturation weighted hue statistics
Hue information instable for weakly saturated colours
represent hue as Cartesian coordinates vector
weigh vector by corresponding saturation
use Euclidean distance to compare chrominance vectors
without shadow suppression
Background model
- Mean luminance: - Saturation weighted hue vector:
correct segmentation
size / position estimation
invalid estimation
Foreground check
Observed pixel (luminance
, saturation
- Luminance’s standard deviation:
- Mean Euclidean distance:
, hue vector
Shadow check (conducted on all foreground pixels)
- Pixel is darker
- Colour is desaturated
- Hue is similar
, where
correct estimation
) is classified as foreground if
reclassify as shadow
Experiments & Results
Methodology
Receiver Operating Characteristics
Stairway
Ground truth:
0.95
0.9
0.85
Detection Rate
Competitors
State of the art algorithms
- RGB background model + shadow check in HSV
- RGB background model + shadow check in c1c2c3
- RGB background model + shadow check in NRGB
- Background modelling and shadow check in NRGB
0.8
0.75
0.7
0.65
0.6
Performance metrics
- Detection rate - False alarm rate
0.5
0
0.002
0.004
0.006
1
In-depth evaluation
- All parameters for each algorithm
variied within the valid limits
- Comparison via receiver
operating characteristcs
(ROC curves)
The proposed approach:
RGB+NRGB
NRGB+NRGB
RGB+HSV
RGB+C1C2C3
Our Approach
0.55
Books
0.008
0.01
0.012
False Alarm Rate
0.014
0.016
0.018
0.02
0.9
RGB+HSV
Detection Rate
0.8
0.7
0.6
RGB+NRGB
NRGB+NRGB
RGB+HSV
RGB+C1C2C3
Our Approach
0.5
0.4
0
0.01
0.02
0.03
0.04
False Alarm Rate
1
PETS2001
0.05
0.06
0.07
RGB+NRGB
0.95
0.9
Detection Rate
Test Sequences
3 Sequences:
- Stairway:
achromatic scene
31 groundtruth frames
- Books:
colourful scene
25 groundtruth frames
- PETS2001:
standard scene for visual
surveillance algorithms
30 ground truth frames
Segmentation Results
1
0.85
NRGB+NRGB
0.8
RGB+NRGB
NRGB+NRGB
RGB+HSV
RGB+C1C2C3
Our Approach
0.75
0.7
0
0.5
1
1.5
2
2.5
False Alarm Rate
3
3.5
4
−3
x 10
Conclusions:
- Application of the advantagous IHLS colour space and saturation weighted hue statistics to visual surveillance
- In-depth evaluation of the developed algorithm and comparison to state of the art methods
- Evaluation has proven that the proposed approach is a meaningful way for motion detection and shadow suppression
- Algorithm satisfies real-time constraints (PAL-resolution, >25 frames per second)
- Published at IEEE Intl. Conf. on Signal and Image Processing (ICSIP06, India) and El. Letters on Comp. Vision and Image Analysis (ELCVIA vol. 6 no. 3)
Parts of this thesis have been sponsored by the EU Project AVITRACK (AST-CT-3002-502818) and the CABS Project
Contact:
[email protected]