26
I. Gómez-Conde, D. Olivieri, X.A. Vila Sobrino, A. Orosa-Rodríguez (University of Vigo) Salamanca (6-8th April, 2011) Intelligent Video Monitoring for Anomalous Event Detection www.milegroup.net

Intelligent video monitoring

Embed Size (px)

DESCRIPTION

Intelligent video monitoring for anomalous event detection.

Citation preview

Page 1: Intelligent video monitoring

I. Gómez-Conde, D. Olivieri, X.A. Vila Sobrino, A. Orosa-Rodríguez

(University of Vigo)

Salamanca (6-8th April, 2011)

Intelligent Video Monitoring for Anomalous

Event Detection

www.milegroup.net

Page 2: Intelligent video monitoring

• Introduction

• Our approach

o Software algorithms for the tele-assistance for the elderly

oMultiple object tracking techniques

oBehavior detectors based on human body positions

• Experimental Results

• Conclusions

Index

Iván Gómez Conde

Page 3: Intelligent video monitoring

o % people (65 years and over)

o % youth (under 15 years)

o In 2050, % elderly people % youth

o Problems:

Sociologic

Economic

Computer Vision can be used as early warning monitor for anomalous event detection!!!

The aging of the population has increased dramatically.

Current problem

Iván Gómez Conde

Page 4: Intelligent video monitoring

The motivation for this paper is the development of a tele-assistance application.

Detect foreground objects

Track these objects in time

Action Recognition

Motivation

Iván Gómez Conde

Page 5: Intelligent video monitoring

o Image analysis

o Machine learning

o Transate the low level signal to a higher semantic level

o Inference actions and behaviors

Present computer aplications go far beyond the simple security camera of a decade ago and now include:

What is the monitoring?

Iván Gómez Conde

Page 6: Intelligent video monitoring

Method for comparing foreground- background segmentation

Feature vector tracking algorithm

Simple real-time histogram based algorithm for discriminating movements and actions

There are several original contributions proposed by this paper:

Contributions

Iván Gómez Conde

Page 7: Intelligent video monitoring

• C++

• OpenCV (Open Source Computer Vision)

Qt

Octave

Software

Iván Gómez Conde

Page 8: Intelligent video monitoring

System

Iván Gómez Conde

This software is an experimental application. The graphical interface provides maximum information.

Page 9: Intelligent video monitoring

Detecting movement

There are several background subtraction methods. We use two methods:

• Running Average

• Gaussian Mixture Model

Iván Gómez Conde

Page 10: Intelligent video monitoring

Running Average

A = Matrix of accumulated pixels

I = Image

Nf = nº of used frames

α = weighting parameter Є [0,1]

Each point of the background is calculated with the mean of the backgrounds over Nf previous frames.

At(Nf) = (1-α) At-1(Nf) + α It

Iván Gómez Conde

Page 11: Intelligent video monitoring

Running average

Iván Gómez Conde

Page 12: Intelligent video monitoring

Gaussian Mixture Model

This method models each background pixel as a mixture of K Gaussian distributions

o K is tipically from 3 to 5

o Eliminates many of the artefacts that Running Average is unable to treat

Iván Gómez Conde

Page 13: Intelligent video monitoring

Gaussian Mixture Model

Iván Gómez Conde

Page 14: Intelligent video monitoring

Testing Methods (% error)

FN + FP 640∙480

Iván Gómez Conde

• False Negatives (FN): Foreground pixels labeled as background

• False Positives (FP): Background pixels labeled as foreground

% error =

Page 15: Intelligent video monitoring

Finding individual objects

• Foreground objects rectangular “blobs”

detect blob

while (∃ blob) do

apply mask

create color histogram

aproximate with gauss

create feature vector

detect new blob

end while

Iván Gómez Conde

Page 16: Intelligent video monitoring

Feature vector for classification

Feature Vector

Size and coordinates of the blob center

Gaussian fitted values of RGB components

Motion vector

Iván Gómez Conde

Page 17: Intelligent video monitoring

Discrimination objects

Norm difference of red channel

Nor

m d

iffe

ren

ce o

f g

reen

ch

an

nel

Iván Gómez Conde

Page 18: Intelligent video monitoring

Tracking algorithm

Once objects have been separated and characterized by their feature vector, we tracks

Tracking is performed by matching features of the rectangular regions

Iván Gómez Conde

Page 19: Intelligent video monitoring

Tracking algorithm

• Position from t to t+1 (x = xo + vt)

Iván Gómez Conde

Page 20: Intelligent video monitoring

Time chart

Bg-Fg Seg. Blob Detection Normal Video Video with Qt

Frame 1 28.3 ms 168.5 ms 33.2 ms 2.5 ms

Frame 30 847.5 ms 5065.4 ms 997.2 ms 75.82 ms

Frame 361 10198.2 ms 60954.1 ms 12000 ms 912.36 ms

Iván Gómez Conde

Page 21: Intelligent video monitoring

Detecting gestures We have considered a limited domain of events

Discrimination arms gestures

o The mass histogram

o Statistical moments

Iván Gómez Conde

Page 22: Intelligent video monitoring

Detecting actions

Normalized Histogram

Iván Gómez Conde

• Basic body position

o Upright

o Lying down

• The inset image is the histogram normalized to unity

Page 23: Intelligent video monitoring

Discrimination actions

Figure 1 Figure 2 Figure 3

µ 0.54 0.33 0.44

σ 0.21 0.17 0.21

µ3

0.17 3.99 3.12

Iván Gómez Conde

Page 24: Intelligent video monitoring

Conclusions

Our software aplication will allow track people and discriminate basic actions

The system is actually part of a more complete tele-monitoring system

The paper opens many possibilities for future study.

o Using our quantitative comparison to optimize parameters

o Combining feature vector with sequential Monte Carlo methods

Iván Gómez Conde

Page 25: Intelligent video monitoring

Conclusions

The histogram model developed in this paper provides detection for a limited set of actions and events:

Real-time method

Easy to implement

Should have utility in real systems

It is not sufficiently robust

Iván Gómez Conde

Page 26: Intelligent video monitoring

Many thanks for your attention

Iván Gómez Conde