domenico bloisi, luca iocchi, dorothy monekosso, paolo remagnino [email protected]

25
A Novel Segmentation A Novel Segmentation Method for Crowded Method for Crowded Scenes Scenes Domenico Bloisi , Luca Iocchi, Dorothy Monekosso, Paolo Remagnino [email protected]

Post on 21-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

A Novel Segmentation A Novel Segmentation Method for Crowded Method for Crowded

ScenesScenes

Domenico Bloisi, Luca Iocchi,

Dorothy Monekosso, Paolo Remagnino

[email protected]

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 2

Video surveillance tasks

A video surveillance system may accomplish a series of well-defined tasks:• To detect objects of interest

(we may want to detect all the moving cars in a street)[Yoneama et al. 2005, <long list>]

• To track objects of interest(we may want to know the exact number of people standing in a room) [Khan and Shah 2006, <long list>]

• To react to particular events(we may want to send an alarm if an unauthorized person enters a restricted area) [Leo et al. 2005, <long list>]

• …

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 3

PB-KU

Visual modeling of people behaviors and interactions for professional training (PB-KU)

• The method is applied to the training of nurses in the School of Nursing in the Faculty of CISM at Kingston University [ http://www.healthcare.ac.uk ].

• The aim is to detect and track people in order to analyze their behavior

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 4

Summary

• Project Overview

• Segmentation

• Height Image Algorithm

• Examples

• Results

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 5

Features

• Background:– Dynamic background (indoor, crowded)

• Number of objects to track:– Up to 15 people in the scene

• Camera:– Two stereo cameras

• Evaluation method:– Evaluation on a on-site built data-set

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 6

General architecture

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 7

Hardware

Videre DesignSTH-MDCS

Intel Core 2 Duo2,0 GHz CPU

Mac mini

Videre DesignSTH-MDCS

Intel Core 2 Duo2,0 GHz CPU

Mac mini

Wireless connection

Firewireconnection

Firewireconnection

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 8

Segmentation(Detecting objects of interest)

BackgroundEstimate

CurrentFrame

BackgroundModel

ForegroundExtraction

List ofDetected Objects

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 9

Background Modeling

Background Imagecomputed from S

(the image displays only the higher

Gaussian values)

Set S of n images from a camera

Raw images

Artificial image

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 10

Foreground Extraction(Background Subtraction Technique)

THRESHOLD T(based on illumination conditions)

blobs (Binary Large OBjectS)

>

T

current frame

foreground image

background image

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 11

Background Subtraction Problems

Background subtraction is a fast and effective technique, but it presents a series of problems:

How to compute a correct background? [Heikkilä and Silven 1999] [Stauffer and Grimson 1999]

How to manage gradual and sudden illumination changes? [Bloisi et al. 2007]

How to manage high-frequencies background objects (such as artificial light flickering, windows) [Bloisi and Iocchi 2008]

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 12

Proposed Solution

Background Subtraction +Stereo Vision +

Edge Detection +Height Image Algorithm

Advantages:robust and efficient foreground extraction, shadow suppression, 3D information, non-moving object filtering, accurate multiple object segmentation.

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 13

System architecture PB-KU

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 14

Height ImageAlgorithm

t: minimum area for a blob to be considered of interest

A: set of found activity blobs

F: final set of the segmented objects we are searching for

H: the set of height images.

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 15

Height Image

a) Active blobsb) Height imagec) Segmented

image

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 16

Example

Segmented image Ground plane view

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 17

Example

Crowd flowAnalysis

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 18

Evaluation and Metrics

• Evaluation On-site dataset

• Metrics Scene accuracy

A is the average accuracy

n

nnai

ˆ

1

Number of detected people

Number of people in the scene

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 19

Segmentation Results

Segmentation Accuracyon 100 randomly chosen images

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 20

Algorithm Speed

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 21

Conclusions

• Summary of results Accurate segmentation even in

case of 15 people in the scene Real-time computation Ground plane view projection for

crowd flow analysis

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 22

Future Work (1)

Add Radio Frequency Identifiers (RFID) to stereo for helping segmentation and dealing with occlusions

RFID

Identity

LocalizationStereo

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 23

Future Work (2)

Crowd flow analysis based on ground plane projection

• ExampleHow many people are near a bed in event of an emergency?

04/18/23A Novel Segmentation Method for Crowded Scenes – VISAPP 2009

Page 24

References

- A. Yoneama C.H. Yeh, C.C.J. Kuo. Robust Vehicle and Traffic Information Extraction for Highway Surveillance, JASP(2005), No. 14, pp. 2305-2321, 2005. - M. Leo, T. D’Orazio, A. Caroppo, T. Martiriggiano P. Spagnolo. Automatic Monitoring of Forbidden Areas to Prevent Illegal Accesses ICAPR (2), pp. 635-643, 2005. - S. Khan and M. Shah. A multiview approach to tracking people in crowded scenes using a planar homography constraint. In ECCV (4), pp. 133–146, 2006. - Y.T. Tsai, H.C. Shih and C.-L. Huang. Multiple human objects tracking in crowded scenes. In ICPR ’06, pp. 51–54, 2006.- J. Heikkilä, O. Silven. A real-time system for monitoring of cyclists and pedestrians. Proc. 2° IEEE International Workshop on Visual Surveillance, pp. 74-81, 1999.-C. Stauffer, W. Grimson. Adaptive background mixture models for real-time tracking. (CVPR'99), pp.246-252, 1999.- D. Bloisi, L. Iocchi, G.R. Leone, R. Pigliacampo, L. Tombolini, L. Novelli. A Distributed Vision System for Boat Traffic Monitoring in the Venice Grand Canal (VISAPP), pp. 549-556, 2007.- D. Bloisi and L. Iocchi. ARGOS - A Video Surveillance System for Boat Traffic Monitoring in Venice. IJPRAI, 2008.

A Novel Segmentation A Novel Segmentation Method for Crowded Method for Crowded

ScenesScenes

Domenico Bloisi, Luca Iocchi,

Dorothy Monekosso, Paolo Remagnino

[email protected]