SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS



Similar documents
Automatic measurement and detection of GSM interferences

1. BACKGROUND 1-1 Traffic Flow Surveillance

Real-time Particle Filters

Performance Center Overview. Performance Center Overview 1

Principal components of stock market dynamics. Methodology and applications in brief (to be updated ) Andrei Bouzaev, bouzaev@ya.

Appendix D Flexibility Factor/Margin of Choice Desktop Research

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999

Journal Of Business & Economics Research September 2005 Volume 3, Number 9

USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES

COMPUTATION OF CENTILES AND Z-SCORES FOR HEIGHT-FOR-AGE, WEIGHT-FOR-AGE AND BMI-FOR-AGE

CHARGE AND DISCHARGE OF A CAPACITOR

Multi- and Single View Multiperson Tracking for Smart Room Environments

A Natural Feature-Based 3D Object Tracking Method for Wearable Augmented Reality

Multi-camera scheduling for video production

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613.

Acceleration Lab Teacher s Guide

Chapter 8: Regression with Lagged Explanatory Variables

Robust 3D Head Tracking by Online Feature Registration

A Novel Approach to Improve Diverter Performance in Liquid Flow Calibration Facilities

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1

Real-time avatar animation steered by live body motion

Chapter 2 Kinematics in One Dimension

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS

Technical Report / Universität Dortmund, SFB 475 Komplexitätsreduktion in Multivariaten Datenstrukturen, No. 2006,23

Maintaining Multi-Modality through Mixture Tracking

Chapter 2 Problems. 3600s = 25m / s d = s t = 25m / s 0.5s = 12.5m. Δx = x(4) x(0) =12m 0m =12m

Unsupervised approach to color video thresholding

LEASING VERSUSBUYING

Load Prediction Using Hybrid Model for Computational Grid

Hybrid System Design for Singularityless Task Level Robot Controllers *

A Bayesian framework with auxiliary particle filter for GMTI based ground vehicle tracking aided by domain knowledge

A Distributed Multiple-Target Identity Management Algorithm in Sensor Networks

Distributed Online Localization in Sensor Networks Using a Moving Target

Distributing Human Resources among Software Development Projects 1

An Online Learning-based Framework for Tracking

Particle Filtering for Geometric Active Contours with Application to Tracking Moving and Deforming Objects

Improved Techniques for Grid Mapping with Rao-Blackwellized Particle Filters

This is the author s version of a work that was submitted/accepted for publication in the following source:

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR

Random Walk in 1-D. 3 possible paths x vs n. -5 For our random walk, we assume the probabilities p,q do not depend on time (n) - stationary

System Performance Improvement By Server Virtualization

PRACTICES AND ISSUES IN OPERATIONAL RISK MODELING UNDER BASEL II

Direc Manipulaion Inerface and EGN algorithms

µ r of the ferrite amounts to It should be noted that the magnetic length of the + δ

Chapter 1.6 Financial Management

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas

AP Calculus BC 2010 Scoring Guidelines

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE

12. TESTING OF CEMENT PART 1.

Bayesian Filtering with Online Gaussian Process Latent Variable Models

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand

Improving Unreliable Mobile GIS with Swarm-based Particle Filters

GoRA. For more information on genetics and on Rheumatoid Arthritis: Genetics of Rheumatoid Arthritis. Published work referred to in the results:

Price Controls and Banking in Emissions Trading: An Experimental Evaluation

Chapter 7. Response of First-Order RL and RC Circuits

Module 3 Design for Strength. Version 2 ME, IIT Kharagpur

NASDAQ-100 Futures Index SM Methodology

CALCULATION OF OMX TALLINN

Measuring macroeconomic volatility Applications to export revenue data,

Usefulness of the Forward Curve in Forecasting Oil Prices

Towards Intrusion Detection in Wireless Sensor Networks

An Unobtrusive Semantic Health-Monitoring Medium

Real Time Bid Optimization with Smooth Budget Delivery in Online Advertising

Motion Along a Straight Line

Segmentation, Probability of Default and Basel II Capital Measures. for Credit Card Portfolios

Appendix A: Area. 1 Find the radius of a circle that has circumference 12 inches.

Random Scanning Algorithm for Tracking Curves in Binary Image Sequences

Answer, Key Homework 2 David McIntyre Mar 25,

SPEC model selection algorithm for ARCH models: an options pricing evaluation framework

Task is a schedulable entity, i.e., a thread

Form measurement systems from Hommel-Etamic Geometrical tolerancing in practice DKD-K Precision is our business.

Hotel Room Demand Forecasting via Observed Reservation Information

Improving timeliness of industrial short-term statistics using time series analysis

Inductance and Transient Circuits

Strategic Optimization of a Transportation Distribution Network

Trends in TCP/IP Retransmissions and Resets

DDoS Attacks Detection Model and its Application

II.1. Debt reduction and fiscal multipliers. dbt da dpbal da dg. bal

Evolutionary building of stock trading experts in real-time systems

Predicting Stock Market Index Trading Signals Using Neural Networks

Nikkei Stock Average Volatility Index Real-time Version Index Guidebook

Hedging with Forwards and Futures

SHB Gas Oil. Index Rules v1.3 Version as of 1 January 2013

Diane K. Michelson, SAS Institute Inc, Cary, NC Annie Dudley Zangi, SAS Institute Inc, Cary, NC

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C.

GUIDE GOVERNING SMI RISK CONTROL INDICES

UPDATE OF QUARTERLY NATIONAL ACCOUNTS MANUAL: CONCEPTS, DATA SOURCES AND COMPILATION 1 CHAPTER 7. SEASONAL ADJUSTMENT 2

Information Theoretic Evaluation of Change Prediction Models for Large-Scale Software

INTRODUCTION TO FORECASTING

A Bayesian Approach for Personalized Booth Recommendation

Resiliency, the Neglected Dimension of Market Liquidity: Empirical Evidence from the New York Stock Exchange

Sampling Time-Based Sliding Windows in Bounded Space

New Fuzzy Dynamic Evaluation For ERP Benefits

RULE-BASED LUNG REGION SEGMENTATION AND NODULE DETECTION VIA GENETIC ALGORITHM TRAINED TEMPLATE MATCHING

MODEL AND ALGORITHMS FOR THE REAL TIME MANAGEMENT OF RESIDENTIAL ELECTRICITY DEMAND. A. Barbato, G. Carpentieri

Term Structure of Prices of Asian Options

Experimental exploration of decision making in production-inventory system

Individual Health Insurance April 30, 2008 Pages

Transcription:

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS Hao Wu and Qinfen Zheng Cenre for Auomaion Research Dep. of Elecrical and Compuer Engineering Universiy of Maryland, College Park, MD-20742 {wh2003, qinfen}@cfar.umd.edu ABSTRACT In his paper, we presen an algorihm for auomaic performance evaluaion of a video racking sysem ha does no require ground-ruh daa. Such an algorihm can play an imporan role in auomaically deermining when he underlying sysem loses rack and needs reiniializaion. The algorihm is based on measuring appearance similariy and racking uncerainy. Several experimenal resuls on vehicle and human racking are presened. Effeciveness of he evaluaion scheme is assessed by comparisons wih ground ruh. The proposed self evaluaion algorihm has been used in an acousic/video based moving vehicle deecion and racking sysem. 1. INTRODUCTION An objec racking sysem can fail under many circumsances. I could be due o illuminaion changes, pose variaion, occlusion, and oher facors. There is a need for auomaic performance evaluaion. Mos of he exising work on racking performance evaluaion has focused on overall algorihmic performance evaluaion using ground-ruh daa. Their usefulness in real ime deermining racking failure is quie limied. In his paper, we presen a racker self-evaluaion algorihm ha auomaically evaluaes he racking qualiy on-he-run and does no require ground-ruh daa. Online self-evaluaion for keeping rack of sysem performance has been sudied for video based objec segmenaion. In [Erdem, 2004], segmenaion and moion consisency along he objec conour and hisogram similariy are calculaed and used o evaluae he goodness of segmenaion and racking. However, a generic racking algorihm may no segmen he objec from he background and hence, he conour informaion may no be available. We address video racking sysems whose arges are bounded by boxes. The rack assessmen is mainly based on appearance similariy and rajecory smoohness. We reduce he confidence in 1 Prepared hrough collaboraive paricipaion in he Advanced Sensors Consorium sponsored by he U.S. Army Research Laboraory under he Collaboraive Technology Alliance Program, Cooperaive Agreemen DAAD19-01-02-0008. racking when here is ambiguiy in he resul. The uncerainy is assessed hrough monioring several ambiguiy measuremens. The paper is organized as follows: ambiguiy feaure exracion and rack evaluaion crierion are discussed in Secion 2 and 3 respecively; Secion 4 gives several experimenal resuls; finally conclusions are given in Secion 5. 2. FEATURES USED FOR SELF-EVALUATION In a common video racker, he locaion and appearance of he arge is represened hrough a represenaive chip specified by a bounding box in he image frame. Conour based rackers can be modified o fi ino such a framework. Inuiively, one may hink ha he appearance change can be used for evaluaion. However, i is no reliable o judge he racking performance solely based on he appearance of he racking box. Appearance change may be caused by wo facors: (1) objec pose change due o camera and/or objec moion and (2) appearance difference measure no consisen wih subjecive evaluaion. The appearance change doesn necessarily indicae poor racking performance. In addiion, in many cases he bounding box includes some background pixels, which makes he appearance evaluaion difficul. In our experience on video surveillance using saic infrared camera, we have noiced ha when racking fails, he size and locaion of bounding box changes irregularly. Once he racking bounding box locks ono background pixels, i changes randomly due o he similariy of he background cluer. Anoher common cause of racking failure is ha he racking bounding box locks ono background objecs. Our goal is o deec any racking failure soon afer i occurs. The following ambiguiy ess are examined in our self evaluaion algorihm. Tes 1: Trajecory complexiy evaluaion Normally, a moving vehicle will no change is direcion and speed dramaically in a few adjacen frames. Therefore, rapid and frequen change in objec moion rajecory is a sign of racking failure. We measure rajecory complexiy as he raio of he rajecory pah

Repor Documenaion Page Form Approved OMB No. 0704-0188 Public reporing burden for he collecion of informaion is esimaed o average 1 hour per response, including he ime for reviewing insrucions, searching exising daa sources, gahering and mainaining he daa needed, and compleing and reviewing he collecion of informaion. Send commens regarding his burden esimae or any oher aspec of his collecion of informaion, including suggesions for reducing his burden, o Washingon Headquarers Services, Direcorae for Informaion Operaions and Repors, 1215 Jefferson Davis Highway, Suie 1204, Arlingon VA 22202-4302. Respondens should be aware ha nowihsanding any oher provision of law, no person shall be subjec o a penaly for failing o comply wih a collecion of informaion if i does no display a currenly valid OMB conrol number. 1. REPORT DATE 00 DEC 2004 2. REPORT TYPE N/A 3. DATES COVERED - 4. TITLE AND SUBTITLE Self-Evaluaion For Video Tracking Sysems 5a. CONTRACT NUMBER 5b. GRANT NUMBER 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) Cenre for Auomaion Research Dep. of Elecrical and Compuer Engineering Universiy of Maryland, College Park, MD-20742 8. PERFORMING ORGANIZATION REPORT NUMBER 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR S ACRONYM(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release, disribuion unlimied 11. SPONSOR/MONITOR S REPORT NUMBER(S) 13. SUPPLEMENTARY NOTES See also ADM001736, Proceedings for he Army Science Conference (24h) Held on 29 November - 2 December 2004 in Orlando, Florida., The original documen conains color images. 14. ABSTRACT 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF ABSTRACT UU a. REPORT unclassified b. ABSTRACT unclassified c. THIS PAGE unclassified 18. NUMBER OF PAGES 5 19a. NAME OF RESPONSIBLE PERSON Sandard Form 298 (Rev. 8-98) Prescribed by ANSI Sd Z39-18

lengh, L pp 1 2, and end poins disance, D pp 1 2, beween wo racking poins p1 = p( τ ) and p2 = p() as shown in Fig. 1. Normally, he larger he raio is, he more complex he rajecory will be. We define rajecory complexiy indicaor as Lpp 1 2 0 if T1 I () = D 1 pp (1) 1 2 1 oherwise We can furher include rajecory direcion change in rajecory complexiy indicaor. P(-) Fig.1 Illusraion of racking rajecory Tes 2: Moion smoohness evaluaion We noiced ha he rajecory incremen beween wo adjacen frames ofen increases when racking fails. We define moion sep as he displacemen of objec box over wo consecuive frames,. Moion smoohness D p ( 1) p ( ) indicaor is defined as 0 if D T p p 2 ( 1) ( ) I () = 2 (2) 1 oherwise The hreshold T 2 is deermined according o prior knowledge of objec moion. For objec racking from a moving camera, camera ego moion should firs be esimaed and removed from he objec displacemen compuaion. Tes 3: Scale consancy evaluaion In general, for medium o long range surveillance, we expec he scale change o be small. We measure arge scale change as he raio of he area of curren arge bounding box, A, o he area of iniial bounding box, P() A 0. Boh he arge scale change and scale change rae are measured and used in rack evaluaion. We define he scale consancy indicaor as A A T31 U T32 U A0 A 0 0 if I () = 3 da da T33 T (3) U 34 Ad 0 Ad 0 1 oherwise Tes 4: Shape similariy evaluaion Shape is an imporan discriminaor for objecs. When he racking bounding box swiches o a differen objec or o he background, he shape of he bounding box ofen also changes. We use aspec raio, Widh Heigh, of he bounding box o represen objec shape and measure he shape similariy as he raio of bounding box aspec raios. The shape similariy indicaor is defined as W H W H 0 if T41 U T42 I () = W 4 0 H0 W0 H0 (4) 1 oherwise Tes 5: Appearance similariy evaluaion Alhough racking evaluaion should no solely depend on appearance similariy, appearance change ofen resuls in racking failure. Therefore, quanifying he appearance change is sill imporan. We use hree appearance change measures o evaluae he appearance sabiliy. The firs one, D, is pixel by pixel difference I beween he curren objec and he iniial objec; he second one, D, is difference of image inensiy H hisograms beween he curren and iniial objecs as used in 0; he hird one, D M, is he sum of weighed differences beween he curren appearance model and he iniial appearance model. Oher measuremen mehods can also be added. We define he appearance similariy indicaor as I () 5 0 if { D T } { D T } { D T } I 51 H 52 M 53 = U U 1 oherwise 3. EVALUATION CRITERION In ideal siuaion, a good racking should have all he five racking evaluaion indicaors equal o one. In pracical circumsances, some unexpeced facors may rigger one or wo of hese indicaors, while he racking performance is sill good. However if hree or more indicaors have been riggered, we conclude ha he racking performance has deerioraed. We fuse he above five es scores o ge a comprehensive racking performance score. We firs learn he uncerainy decision hresholds for each es using empirical daa and hen compue a weighed sum of he five indicaors 5 5 q () = wi (), w = 1 i i i i= 1 i= 1 (5) (6) In general, he larger he q () is, he beer he racking performance. When q () drops below a hreshold, we conclude ha he racking performance has deerioraed and needs o be re-iniialized. The weigh can be learn from raining daa. In our implemenaion, he

appearance weigh, w 5, is se slighly larger han ohers. In implemenaion, one may re-iniialize he sysem only afer q () is below a hreshold for a specified period of ime. 4. EXPERIMEN RESULTS The proposed algorihm was esed on differen surveillance videos. Fig.2 shows evaluaion resuls on an IR vehicle surveillance sequence. The vehicle firs moved sraigh away from he camera and hen made a lef urn. The resuls show ha he self evaluaion algorihm does give a good indicaion of he racking performance. In Fig. 2(a), when he bounding box does no fi he objec well, he evaluaion score drops. Afer re-iniializaion, he bounding box fis he objec and he evaluaion score rises, as shown in Fig. 2(b). We also compared he self evaluaion resul wih ground ruh (Fig.3). I is shown ha as he disance beween he racked objec locaion and he ground ruh increases, our racking confidence score decrease indicaing deerioraion in racking performance. When inegraed ino a moving vehicle deecion and racking sysem [Sankaranayanan, 2004], he proposed algorihm helps he video surveillance sysem mainaining a good arge rack by re-iniializing he racker whenever he racker performance deerioraes. The racking algorihm used in our experimens is he adapive appearance model based racker developed by Zhou, e al [Zhou, 2004]. 5. CONCLUSIONS In his paper, we presen an algorihm for auomaic performance evaluaion of a video racking sysem ha does no require ground-ruh daa. The algorihm is based on measuring appearance similariy and racking uncerainy. Several experimenal resuls on vehicle and human racking are repored. Effeciveness of he evaluaion scheme is demonsraed by comparisons wih ground ruh. The proposed self evaluaion algorihm has been used in an acousic/video based moving vehicle deecion and racking sysem [Sankaranayanan, 2004]. 6. REFERENCES Erdem, C.E. Sankur, B, Tekalp, A.M., 2004: Performance Measures for Video Objec Segmenaion and Tracking, IEEE Trans. Image Processing, 13:931-951. Sankaranayanan, A.C., e al, 2004: Vehicle Tracking using Acousic and Video Sensors, Proc. 24 h Army Science Conference (o appear). Zhou, S., Chellappa, R., Moghaddam, B., 2004: Visual Tracking and Recogniion Using Appearanceadapive Models in Paricle Filers, IEEE Trans. Image Processing (o appear). Fig. 4 shows he resuls of evaluaion of pedesrian deecion and racking from a color surveillance video. The firs hree images are represenaive frames of he surveillance video wih he racking bounding box superimposed. The corresponding racker evaluaion scores are shown in he boom row of Fig.4. In his example, he bounding box swiches o he background and wanders around a ha posiion aferwards. Our self evaluaion crierion correcly repors he racking failure. Fig.5 shows he resuls of evaluaing a pedesrian racking wih parial occlusion and reappearance. The racked person walks behind a moving car. The racker becomes uncerain while parially occluded by he moving vehicle. The racker regains is confidence/performance afer he human reappears. Our racker evaluaion algorihm correcly scores he even. Fig.6 shows he evaluaion resuls for racking a group of pedesrian wih significan occlusion. As he racked human group is blocked by he moving van, he bounding box swiches o he van and loses he arge. Our self-evaluaion score drops when he racker fails. We expec he confidence score will drop furher if arge rajecory direcion is also incorporaed in he evaluaion measuremens.

(a) (b) Fig.2 Improved video racking wih rack evaluaion and appearance updaing. Also shown are he corresponding evaluaion plos. Fig.3 Comparison of self-evaluaion score and he ground ruh. The red line is he disance beween GPS measuremens and racked arge cener; he green line is he evaluaion scores repored by our algorihm. Fig.4 An example of pedesrian racking. Shown in he op hree rows are represenaive frames wih he racking bounding box superimposed. The corresponding racker evaluaion scores are shown in he boom row. Our self evaluaion crierion correcly repors he racking failure.

Fig.5. An example of racking pedesrian wih parial occlusion. The racked person walks behind a moving car. The racker becomes uncerain while parially occluded by he moving vehicle. The racker regains is performance afer he human is cleared of occlusion. Our racker evaluaion algorihm correcly scores he even. Fig.6. An example of racking a group of pedesrian wih significan occlusion. As he racked human group is occluded by he moving van, he bounding box swiches o he van and lose he arge. Our self-evaluaion score drops when he racker fails.