+ All Categories
Home > Documents > Detecting and Distinguishing Nuances of Anomaly in Machine ...

Detecting and Distinguishing Nuances of Anomaly in Machine ...

Date post: 10-Jan-2022
Category:
Upload: others
View: 2 times
Download: 0 times
Share this document with a friend
45
1 Detecting and Distinguishing Nuances of Anomaly in Machine Perception Systems Josef Kittler Centre for Vision, Speech and Signal Processing Dept. Electronic Engineering, University of Surrey Guildford, UK Coauthors: F Yan, T de Campos, D Windridge, W Christmas, M Osman, A Khan, N Faraji Davar and J Illingworth Support by EPSRC is gratefully acknowledged
Transcript

1

Detecting and Distinguishing Nuances of Anomaly in Machine Perception Systems

Josef Kittler Centre for Vision, Speech and Signal Processing

Dept. Electronic Engineering, University of Surrey Guildford, UK

Coauthors: F Yan, T de Campos, D Windridge, W Christmas, M Osman, A Khan, N Faraji Davar and J Illingworth

Support by EPSRC is gratefully acknowledged

2

Outline

  Background   Introduction to anomaly detection   Prior art   Critique of conventional anomaly

detection   Proposed anomaly detection system   Application to tennis video annotation

  number of players anomaly   out-of-play anomaly

  Conclusions

Vision system design

 Architecture  Vision system modules  Training data   Learning system parameters  System integration  System optimisation

3

Advanced concepts in vision systems design

 High level scene interpretation  Ability to adapt and acquire new

competences  Anomaly detection  Acquisition of domain knowledge  Context classification

4

Tennis video interpretation

5

6

Application to sports video annotation

  Aim is to interpret tennis video from the observed visual events

  The states are:   1st serve   2nd serve   Ace   Rally   Point award

7

Tennis annotation system   15 modules

videosource

on/off switch

data flow

frame

lens correctionde−interlace &

homography

corner tracking/ homographies

constructmosaic

mosaic

shotclassification

shotClass

closeup play

shotdetection

shotBoundary

for ball tracking

ballCandidate

finds candidates

detect eventsin 3D world

threeDCues

reasoningbased on rulesof tennis

highLevel

HMM eventdetector

HMMEDetection

anomalyDetection

detects anomalies inball events

anomalyAnalysis

compares anomalyanalysis with ground truth

tennisanotation

white linefinding

whiteLines

calibrationcamera

projection

ok

proj

serveDetection

detects serve pointfilters action results

separates foregroundfor ball tracking

ballForeground

corner tracking /

homographiesalternate

skipHomogs

ballInterpolation

Interpolate ball trackwhen off top of screen

ballTracking

track ball anddetect ballevents in 2D

creates table ofanomaly vs. noise

anomalyVsNoise

foreground

separatesforeground forplayer tracking

track players

playerTrack

classifieshit/serve/non−hitactions

playerActions

playerCount

in no. of playersdetects anomalies

Low−level Ball trackingMosaicing /projection Player tracking High−level

8

Tennis annotation system

  A system that “understands” tennis   In the sense: video in -> score out   High level reasoning relies on low level

processing   Ball tracking is crucial

shot analysis

court detection

ball tracking

player tracking

events detection HMM

camera score

low level high level

Contextual interpretation

  Objects   Object

configurations

9

o i

o j

ω α ω β

Scene Graph Model Graph

Sensory data

World model

Scene model

  Examples in tennis video   Object dynamics   Global view of static environment   Spatial relationship of objects   Visual event evolution

(grammars, Hidden Markov models )

Static background removal

11

Ball/player event detection

12

13

Event Detection I

Visual event detection

14

• Simplified for ease of training • Error correcting Viterbi algorithm • Point awarding model only • Future extensions to game and set models

Match evolution model

15

The tennis annotation system

Testing on tennis doubles

  How will the system respond to tennis doubles?

  What mechanisms are needed for the system to extend its competence to the task of interpreting tennis doubles?

16

ACASVA Project

  Existing machine perception systems   largely no ability to adapt   no ability to extend competence/transfer

knowledge   ACASVA aims to develop mechanisms for

cognitive bootstrapping, i.e. ability to   adapt detectors   transfer knowledge   adapt interpretation processes   acquire new competences

17

18

Background

  Mechanisms needed for system

competence extension   Anomaly detection   Visual event – anomaly association   Model adaptation   Context classifier

Introduction to anomaly

  Anomaly –   an important notion in human

understanding of the environment   deviation from normal order or rule

  Many synonyms signifying different nuances   rarity, irregularity, incongruence,

abnormality, unexpected event, novelty, innovation, outlier

19

Classical anomaly model

  In science/engineering   prove disprove hypothesis   fault detection   outdated model requires adaptation

  Conventional mathematical model   outlier of a distribution   empirical distribution deviates from

the model distribution

20

x

Prior art in anomaly detection

  Edgeworth (1888)   Hundreds of papers   Many approaches

  statistical, NN, classification, clustering, information theoretic, spectral

  Excellent surveys   Markou&Singh (SP 2003, statistical, neural)   Hodge&Austin (AI Review 2004)   Agyemang&Barker&Alhajj (Int Data Anal 2006)   Chandola&Banerjee&Kumar (ACM Surveys 2010)   Saligrama&Konrad&Vodoin (SPM2010, video)

21

Classical model

22

23

Tennis annotation system   15 modules

videosource

on/off switch

data flow

frame

lens correctionde−interlace &

homography

corner tracking/ homographies

constructmosaic

mosaic

shotclassification

shotClass

closeup play

shotdetection

shotBoundary

for ball tracking

ballCandidate

finds candidates

detect eventsin 3D world

threeDCues

reasoningbased on rulesof tennis

highLevel

HMM eventdetector

HMMEDetection

anomalyDetection

detects anomalies inball events

anomalyAnalysis

compares anomalyanalysis with ground truth

tennisanotation

white linefinding

whiteLines

calibrationcamera

projection

ok

proj

serveDetection

detects serve pointfilters action results

separates foregroundfor ball tracking

ballForeground

corner tracking /

homographiesalternate

skipHomogs

ballInterpolation

Interpolate ball trackwhen off top of screen

ballTracking

track ball anddetect ballevents in 2D

creates table ofanomaly vs. noise

anomalyVsNoise

foreground

separatesforeground forplayer tracking

track players

playerTrack

classifieshit/serve/non−hitactions

playerActions

playerCount

in no. of playersdetects anomalies

Low−level Ball trackingMosaicing /projection Player tracking High−level

Anomaly detection & machine perception

  Multiple models   Discriminative classifiers   Ambiguity of interpretation   Contextual reasoning   Hierarchical representation   Data quality   Model pruning

24

Classical notion of anomaly

25

Distribution outlier

Different aspects of anomaly

26

Different aspects of anomaly

27

-Distribution drift -Novelty detection

Data quality

28

Data quality

29

Incongruence/unexpected event

30

  Magritte’s La duree poignard   Model base pruning

  Computational efficiency   Hierarchical representation

o i

o j

ω α ω β

Scene Graph Model Graph

Conventional anomaly detection

31

Proposed anomaly detection system

32

Incongruence detection

  Detecting differences between observations and expectations (anomaly, rare event, incongruence)

  Basic principle – comparison of outputs of weak and strong classifiers (Ketabdar et al 2007)

  Dirac Project (Burget et al 2008, Weinshall et all [2009-2012])

  Exemplified by out-of-vocabulary word detection   Phoneme recognizer (weak classifier)   HMM speech recognizer (strong, contextual classifier)

33

Proposed anomaly detection system

34

Nuances of anomaly

  No anomaly   Noisy measurement   Unknown object   Corrupted

measurement   Congruent labelling   Unknown structure   Spurious

measurement errors

35

  Unexpected structural component

  Unexpected structural component & structure

  Measurement model drift

Next challenge

 Tennis to badminton

36

#players detection

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8Singles

A03WSA03MSJ09WS

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

moving agents

Doubles

A08WDU06WD

41

# player anomaly detection

0 2 4 6 8 100

10

20

30

40

50

60

70

80

90

100

buffer size (in shots)

dete

cted

ano

mal

ies

(% o

f sho

ts)

Train with A03MS+J09WS

A08WD (doubles)U06WD (doubles)A03WS (singles)

42 0 2 4 6 8 10

0

10

20

30

40

50

60

70

80

90

100

buffer size (in shots)

dete

cted

ano

mal

ies

(% o

f sho

ts)

Train with A03WS+A03MS

A08WD (doubles)U06WD (doubles)J09WS (singles)

0 2 4 6 8 100

10

20

30

40

50

60

70

80

90

100

buffer size (in shots)

dete

cted

ano

mal

ies

(% o

f sho

ts)

Train with A03WS+J09WS

A08WD (doubles)U06WD (doubles)A03MS (singles)

Out-of-play anomaly detection

43

Out-of-play anomaly detection

44

  Decision confidence filter

Anomalous ball event detection

45

  Contextual postprocessing

Bayesian surprise

46

  Measure of incongruence

contextual aposteriori probability

non-contextual aposteriori probability

o i

o j

  Bayesian surprise has undesirable properties

  Modified measure of surprise

  Modified measure of surprise in 2 class

case

47

Out-of-play anomaly

2008 Ground Truth Anomaly Triggering Events (27/176)

0

1

2

3

4

5

6

7

8

9

10

11

48

Bootstrapping system

49

Context detection

Contextual annotation

Anomaly detection

Low level event

detection

Knowledge base

Rule base learning

Conclusions

  Novel anomaly detection system   Incongruence detector   Decision confidence filter   Data quality assessment   Computationally efficient outlier detector

  Applied to anomaly detection in tennis video analysis

  Demonstrated merits of all the mechanisms deployed by the system

50


Recommended