[PPT] - DIVIDE: ADAPTIVE CONTEXT-AWARE QUERY DERIVATION FOR IOT DATA STREAMS PowerPoint Presentation

SLIDE 1

SLIDE 2

DIVIDE: ADAPTIVE CONTEXT-AWARE QUERY DERIVATION FOR IOT DATA STREAMS

Mathias De Brouwer Sensors and Actuators on the Web (SAW) workshop – ISWC 2019 Auckland, New Zealand – 26 October 2019

SLIDE 3

CONTEXT-AWARE MONITORING IN IOT

3

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

SLIDE 4

CONTEXT-AWARE MONITORING IN IOT

4

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

SLIDE 5

CONTEXT-AWARE MONITORING IN IOT

5

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

?

SLIDE 6

CONTEXT-AWARE MONITORING IN IOT

6

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

REQUIRES INTELLIGENT

CONSOLIDATION & ANALYSIS OF HETEROGENEOUS, VOLUMINOUS, HIGH VELOCITY DATA

?

SLIDE 7

CONTEXT-AWARE MONITORING IN IOT

7

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

REQUIRES INTELLIGENT

CONSOLIDATION & ANALYSIS OF HETEROGENEOUS, VOLUMINOUS, HIGH VELOCITY DATA

RDF STREAM PROCESSING

SLIDE 8

EXAMPLE USE CASE – PATIENT BOB

8

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

RDF STREAM PROCESSING

SLIDE 9

EXAMPLE USE CASE – PATIENT BOB

9

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Diagnosis Bob: epilepsy attack

RDF STREAM PROCESSING

SLIDE 10

EXAMPLE USE CASE – PATIENT BOB

10

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Diagnosis Bob: epilepsy attack Epilepsy patients are sensitive to light Threshold: light should not go > 250 lumen

RDF STREAM PROCESSING

SLIDE 11

EXAMPLE USE CASE – PATIENT BOB

11

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Diagnosis Bob: epilepsy attack Epilepsy patients are sensitive to light Threshold: light should not go > 250 lumen Query 1: generate alarm if value measured by sensor A1 is > 250

RDF STREAM PROCESSING

SLIDE 12

EXAMPLE USE CASE – PATIENT BOB

12

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Query 1: generate alarm if value measured by sensor A1 is > 250 Diagnosis Bob: concussion

RDF STREAM PROCESSING

SLIDE 13

EXAMPLE USE CASE – PATIENT BOB

13

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Concussion patients are sensitive to light & sound Thresholds: light should not go > 170 lumen, sound should not go > 30 decibels Query 1: generate alarm if value measured by sensor A1 is > 250 Diagnosis Bob: concussion

RDF STREAM PROCESSING

SLIDE 14

EXAMPLE USE CASE – PATIENT BOB

14

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Concussion patients are sensitive to light & sound Thresholds: light should not go > 170 lumen, sound should not go > 30 decibels Query 1: generate alarm if value measured by sensor A1 is > 250 170 Diagnosis Bob: concussion

RDF STREAM PROCESSING

SLIDE 15

EXAMPLE USE CASE – PATIENT BOB

15

Sound sensor A0 Light sensor A1 Temperature sensor A2 … Localization Door/window sensor Motion sensor Wearable

Electronic Health Record of patients Medical domain knowledge Hospital lay-out Care staff

Concussion patients are sensitive to light & sound Thresholds: light should not go > 170 lumen, sound should not go > 30 decibels Query 1: generate alarm if value measured by sensor A1 is > 250 Query 2: generate alarm if value measured by sensor A0 is > 30 170 Diagnosis Bob: concussion

RDF STREAM PROCESSING

SLIDE 16

HOW TO DEAL WITH CONTEXT CHANGES?

16

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no

SLIDE 17

HOW TO DEAL WITH CONTEXT CHANGES?

17

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no

SLIDE 18

HOW TO DEAL WITH CONTEXT CHANGES?

18

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no

SLIDE 19

HOW TO DEAL WITH CONTEXT CHANGES?

19

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no MULTIPLE QUERIES (1 FOR EACH

RELEVANT SENSOR)

(almost) none manually yes

SLIDE 20

HOW TO DEAL WITH CONTEXT CHANGES?

20

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no MULTIPLE QUERIES (1 FOR EACH

RELEVANT SENSOR)

(almost) none manually yes

SLIDE 21

HOW TO DEAL WITH CONTEXT CHANGES?

21

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no MULTIPLE QUERIES (1 FOR EACH

RELEVANT SENSOR)

(almost) none manually yes

SLIDE 22

HOW TO DEAL WITH CONTEXT CHANGES?

22

Approach Real-time reasoning? How to configure? Manual reconfiguration when context changes? FIXED GENERIC

QUERIES

expressive manually no MULTIPLE QUERIES (1 FOR EACH

RELEVANT SENSOR)

(almost) none manually yes DIVIDE none automatically no

SLIDE 23

GENERAL ISSUE VS. GOAL

(Changed) context

23

SLIDE 24

GENERAL ISSUE VS. GOAL

(Changed) context

24

Queries of interest are automatically derived & (re)configured

SLIDE 25

GENERAL ISSUE VS. GOAL

(Changed) context

25

Queries of interest are automatically derived & (re)configured

By performing reasoning on (changed) context instead of

n the real-time data streams

SLIDE 26

GENERAL ISSUE VS. GOAL

(Changed) context

26

Queries of interest are automatically derived & (re)configured

By performing reasoning on (changed) context instead of

n the real-time data streams

DIVIDE

SLIDE 27

DIVIDE – BUILDING BLOCKS

Logic: Notation3 Logic (N3)

27

SLIDE 28

DIVIDE – BUILDING BLOCKS

Logic: Notation3 Logic (N3) Reasoner: EYE reasoner ➔ OWL 2 RL reasoning ➔ define goal (= which triples EYE should look for evidence) ➔ EYE produces a proof with the goal as the last applied rule

28

SLIDE 29

OVERVIEW OF DIVIDE

29

INSTANTIATED QUERY QUERY DERIVATION PREPROCESSED ONTOLOGY (IN EYE IMAGE) REASONER’S GOAL SENSOR QUERY RULE CONTEXT

New or changed (e.g., for a patient or room) Output: filtering queries to run on RSP engine ONTOLOGY PREPROCESSING

ONTOLOGY

SLIDE 30

ONTOLOGY PREPROCESSING

Goal: speed up query derivation

30

SLIDE 31

ONTOLOGY PREPROCESSING

Goal: speed up query derivation How?

31

SLIDE 32

ONTOLOGY PREPROCESSING

Goal: speed up query derivation How?

32

Create N3 copy

f ontology

SLIDE 33

ONTOLOGY PREPROCESSING

Goal: speed up query derivation How?

33

Create N3 copy

f ontology

Create specialized

ntology-specific

rules from OWL 2 RL rules

SLIDE 34

ONTOLOGY PREPROCESSING

Goal: speed up query derivation How?

34

Create N3 copy

f ontology

Create specialized

ntology-specific

rules from OWL 2 RL rules Create compiled prolog image of EYE (with ontology & specialized rules)

SLIDE 35

DIVIDE – BUILDING BLOCKS

35

Ontology: ACCIO ontology

SLIDE 36

DIVIDE – BUILDING BLOCKS

36

LightIntensityAboveThresholdFault ≡ (hasSymptom some LightIntensityAboveThresholdSymptom) and (madeBySensor some (isSubsystemOf some (hasLocation some (isLocationOf some ( (hasDiagnosis some (hasMedicalSymptom some SensitiveToLight)) and (hasRole some PatientRole))))))

Ontology: ACCIO ontology

SLIDE 37

DIVIDE – BUILDING BLOCKS

37

HandleHighLightInRoomAction ≡ LightIntensityAboveThresholdFault and (madeBySensor some (isSubsystemOf some (hasLocation some (isLocationOf some LightingDevice))))

Ontology: ACCIO ontology

SLIDE 38

INPUTS OF QUERY DERIVATION

38

Reasoner goal: look for an AboveThresholdAction

{?x a :AboveThresholdAction.} => {?x a : AboveThresholdAction.}.

SLIDE 39

INPUTS OF QUERY DERIVATION

39

Reasoner goal: look for an AboveThresholdAction

{?x a :AboveThresholdAction.} => {?x a : AboveThresholdAction.}.

Accio ontology – definitions (previous slides) & individuals:

:Concussion a :Diagnosis; :hasMedicalSymptom :ConcussionSensitivenessToLight . :ConcussionSensitivenessToLight :hasThreshold :ConcussionLightThreshold . :ConcussionLightThreshold :hasDataValue "170"^^xsd:float ; :isThresholdOnProperty [ a :LightIntensity ] .

SLIDE 40

INPUTS OF QUERY DERIVATION

40

Reasoner goal: look for an AboveThresholdAction

:Bob a :Person ; :hasRole [ a :PatientRole ] ; :hasLocation :R101 ; :hasDiagnosis :Concussion . :40-a5-ef-05-a4-a6-A0 a :SoundSensor ; :hasLocation :R101 . :40-a5-ef-05-a4-a6-A1 a :LightSensor ; :hasLocation :R101 . :40-a5-ef-05-a4-a6-A2 a :TemperatureSensor ; :hasLocation :R101 . ... {?x a :AboveThresholdAction.} => {?x a : AboveThresholdAction.}.

Accio ontology – definitions (previous slides) & individuals: Context – part about patient Bob and his room:

:Concussion a :Diagnosis; :hasMedicalSymptom :ConcussionSensitivenessToLight . :ConcussionSensitivenessToLight :hasThreshold :ConcussionLightThreshold . :ConcussionLightThreshold :hasDataValue "170"^^xsd:float ; :isThresholdOnProperty [ a :LightIntensity ] .

SLIDE 41

INPUTS OF QUERY DERIVATION

41

Reasoner goal: look for an AboveThresholdAction

:Bob a :Person ; :hasRole [ a :PatientRole ] ; :hasLocation :R101 ; :hasDiagnosis :Concussion . :40-a5-ef-05-a4-a6-A0 a :SoundSensor ; :hasLocation :R101 . :40-a5-ef-05-a4-a6-A1 a :LightSensor ; :hasLocation :R101 . :40-a5-ef-05-a4-a6-A2 a :TemperatureSensor ; :hasLocation :R101 . ... {?x a :AboveThresholdAction.} => {?x a : AboveThresholdAction.}.

Accio ontology – definitions (previous slides) & individuals: Context – part about patient Bob and his room:

:Concussion a :Diagnosis; :hasMedicalSymptom :ConcussionSensitivenessToLight . :ConcussionSensitivenessToLight :hasThreshold :ConcussionLightThreshold . :ConcussionLightThreshold :hasDataValue "170"^^xsd:float ; :isThresholdOnProperty [ a :LightIntensity ] .

Sensor query rule – ???

SLIDE 42

WHAT IS THIS SENSOR QUERY RULE ?

{ ?p :hasRole [ a :PatientRole ] ; :hasLocation ?l ; :hasDiagnosis [ :hasMedicalSymptom [ :hasThreshold [ :hasDataValue ?th ; :isThresholdOnProperty [ a ?prop ] ] ] ] .

?s a :Sensor ; :observes [ a ?prop ] ; :hasLocation ?l .

} => { _:x a :Query ; :pattern :pattern-1 ; :inputVariables (("?th" ?th) ("?s" ?s) ("?prop" ?prop)) ; _:oo a :observation ; :madeBySensor ?s ; :hasResult [ :hasDataValue _:v ] ; :hasSymptom [ a :ThresholdSymptom ; :forProperty [ a ?prop ] ] . } .

42

SLIDE 43

WHAT IS THIS SENSOR QUERY RULE ?

{ ?p :hasRole [ a :PatientRole ] ; :hasLocation ?l ; :hasDiagnosis [ :hasMedicalSymptom [ :hasThreshold [ :hasDataValue ?th ; :isThresholdOnProperty [ a ?prop ] ] ] ] .

?s a :Sensor ; :observes [ a ?prop ] ; :hasLocation ?l .

} => { _:x a :Query ; :pattern :pattern-1 ; :inputVariables (("?prop" ?prop) ("?th" ?th) ("?s" ?s)) ; _:oo a :observation ; :madeBySensor ?s ; :hasResult [ :hasDataValue _:v ] ; :hasSymptom [ a :ThresholdSymptom ; :forProperty [ a ?prop ] ] . } .

43

Representation

f generic query

SLIDE 44

WHAT IS THIS SENSOR QUERY RULE ?

{ ?p :hasRole [ a :PatientRole ] ; :hasLocation ?l ; :hasDiagnosis [ :hasMedicalSymptom [ :hasThreshold [ :hasDataValue ?th ; :isThresholdOnProperty [ a ?prop ] ] ] ] .

?s a :Sensor ; :observes [ a ?prop ] ; :hasLocation ?l .

} => { _:x a :Query ; :pattern :pattern-1 ; :inputVariables (("?prop" ?prop) ("?th" ?th) ("?s" ?s)) ; _:oo a :observation ; :madeBySensor ?s ; :hasResult [ :hasDataValue _:v ] ; :hasSymptom [ a :ThresholdSymptom ; :forProperty [ a ?prop ] ] . } .

44

Representation

f generic query

:pattern-1 a QueryPattern ; sh:prefixes :prefixes ; sh:construct """ CONSTRUCT { ?o a AboveThresholdAction ; forProperty ?prop . } FROM NAMED WINDOW :win ON <http://idlab.ugent.be/grove> [ RANGE PT1S TUMBLING ] WHERE { WINDOW :win { ?o a Observation ; madeBySensor ?s ; hasResult [ DUL : hasDataValue ?v ] . FILTER (xsd:float(?v) > xsd:float(?th)) } } ORDER BY DESC (?t) LIMIT 1""".

SLIDE 45

WHAT IS THIS SENSOR QUERY RULE ?

{ ?p :hasRole [ a :PatientRole ] ; :hasLocation ?l ; :hasDiagnosis [ :hasMedicalSymptom [ :hasThreshold [ :hasDataValue ?th ; :isThresholdOnProperty [ a ?prop ] ] ] ] .

?s a :Sensor ; :observes [ a ?prop ] ; :hasLocation ?l .

} => { _:x a :Query ; :pattern :pattern-1 ; :inputVariables (("?prop" ?prop) ("?th" ?th) ("?s" ?s)) ; _:oo a :observation ; :madeBySensor ?s ; :hasResult [ :hasDataValue _:v ] ; :hasSymptom [ a :ThresholdSymptom ; :forProperty [ a ?prop ] ] . } .

45

Representation

f generic query

SLIDE 46

HOW ARE THE QUERIES DERIVED ?

Instantiated relevant queries appear in this proof Can be extracted & substituted with additional reasoning steps

46

<#lemma22> a r:Inference; r:gives { _:sk_0 a sd:Query. _:sk_0 sd:pattern ns6:pattern-1. _:sk_0 sd:inputVariables (("?prop" SSNiot:LightIntensity) ("?th" "170.0"^^xsd:float) ("?s" <http://occs.intec.ugent.be/ontology/entity#40-a5-ef-05-a4-a6-A1>)). _:sk_2 a sosa:Observation. _:sk_2 sosa:madeBySensor <http://occs.intec.ugent.be/ontology/entity#40-a5-ef-05-a4-a6-A1>. _:sk_2 sosa:hasResult _:sk_3. _:sk_3 DUL:hasDataValue _:sk_1. _:sk_2 SSNiot:hasSymptom _:sk_4. _:sk_4 a SSNiot:ThresholdSymptom. _:sk_4 ssn:forProperty _:sk_5. _:sk_5 a SSNiot:LightIntensity. }; r:evidence ( <#lemma33> ... <#lemma47> ); r:rule <#lemma48>.

➔

SLIDE 47

DIVIDE PERFORMANCE EVALUATION

(a) Ontology preprocessing (b) Query derivation

Set-up: 1-person hospital room with concussion patient & 10 sensors ➔ output: 2 queries (for light sensor & sound sensor)

SLIDE 48

COMPARISON OF DIVIDE WITH REAL-TIME REASONING APPROACHES

Fair comparison: OWL 2 RL reasoning Filtering the same event type Context: 1-person hospital room with concussion patient & 10 sensors

48

(a) DIVIDE (c) Pipe C-SPARQL & RDFox (b) StreamFox

SLIDE 49

49

COMPARISON OF DIVIDE WITH REAL-TIME REASONING APPROACHES

SLIDE 50

50

COMPARISON OF DIVIDE WITH REAL-TIME REASONING APPROACHES

SLIDE 51

DIVIDE APPROACH ON RASPBERRY PI

Context: 1-person hospital room with concussion patient & 10 sensors

SLIDE 52

SUMMARY

52

SLIDE 53

SUMMARY

53

No manual query construction & configuration

SLIDE 54

SUMMARY

54

No manual query construction & configuration Automatic reconfiguration when context changes (~ reduction of # reasoning steps)

SLIDE 55

SUMMARY

55

No manual query construction & configuration Automatic reconfiguration when context changes (~ reduction of # reasoning steps) Efficient query evaluation

SLIDE 56

SUMMARY

56

No manual query construction & configuration Automatic reconfiguration when context changes (~ reduction of # reasoning steps) Efficient query evaluation No high-end hardware needed

SLIDE 57

ADDITIONAL ADVANTAGES OF DIVIDE

57

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

SLIDE 58

ADDITIONAL ADVANTAGES OF DIVIDE

58

Local autonomy & increased responsiveness

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

SLIDE 59

ADDITIONAL ADVANTAGES OF DIVIDE

59

Only relevant data is sent over network ➔ reduced bandwidth usage & delays

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

Local autonomy & increased responsiveness

SLIDE 60

ADDITIONAL ADVANTAGES OF DIVIDE

60

Only perform more complex back-end reasoning for important events ➔ much more scalable Only relevant data is sent over network ➔ reduced bandwidth usage & delays

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

Local autonomy & increased responsiveness

SLIDE 61

ADDITIONAL ADVANTAGES OF DIVIDE

61

Only perform more complex back-end reasoning for important events ➔ much more scalable Flexible privacy management Only relevant data is sent over network ➔ reduced bandwidth usage & delays

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

Local autonomy & increased responsiveness

SLIDE 62

ADDITIONAL ADVANTAGES OF DIVIDE

62

Only perform more complex back-end reasoning for important events ➔ much more scalable Flexible privacy management Only relevant data is sent over network ➔ reduced bandwidth usage & delays

…

Many IoT devices & sensors

DIVIDE perfectly supports the vision of cascading reasoning & edge computing

No synchronization of context data to local filtering devices Local autonomy & increased responsiveness

SLIDE 63

ADDITIONAL ADVANTAGES OF DIVIDE

63

Only perform more complex back-end reasoning for important events ➔ much more scalable Flexible privacy management Only relevant data is sent over network ➔ reduced bandwidth usage & delays

…

Many IoT devices & sensors

➔ DIVIDE helps in solving the trade-off between reasoning expressivity & data velocity DIVIDE perfectly supports the vision of cascading reasoning & edge computing

No synchronization of context data to local filtering devices Local autonomy & increased responsiveness

SLIDE 64

ALTERNATIVE USE CASE: TRANSPORT

64

how to organize car traffic flow in a big city

TOPIC

SLIDE 65

ALTERNATIVE USE CASE: TRANSPORT

65

how to organize car traffic flow in a big city

TOPIC

real-time car movement

STREAMS

SLIDE 66

ALTERNATIVE USE CASE: TRANSPORT

66

how to organize car traffic flow in a big city

TOPIC

real-time car movement

STREAMS

disruptions, event, time of day/week/year, …

CONTEXT

SLIDE 67

ALTERNATIVE USE CASE: TRANSPORT

67

how to organize car traffic flow in a big city

TOPIC

real-time car movement

STREAMS

disruptions, event, time of day/week/year, …

CONTEXT

traffic lights ~ query filters

QUERIES

SLIDE 68

ALTERNATIVE USE CASE: TRANSPORT

68

how to organize car traffic flow in a big city

TOPIC

real-time car movement

STREAMS

disruptions, event, time of day/week/year, …

CONTEXT

traffic lights ~ query filters

QUERIES

Extra challenging because of dependencies between “queries” Note: this is still an idea – needs to be fully worked out

SLIDE 69

FUTURE RESEARCH STEPS

69

SLIDE 70

FUTURE RESEARCH STEPS

70

Usability: make the configuration easier and accessible to non-experts

SLIDE 71

FUTURE RESEARCH STEPS

71

Usability: make the configuration easier and accessible to non-experts Improve framework around key reasoning steps with EYE

SLIDE 72

FUTURE RESEARCH STEPS

72

Usability: make the configuration easier and accessible to non-experts Improve framework around key reasoning steps with EYE Investigate energy-efficiency on low-end devices

SLIDE 73

FUTURE RESEARCH STEPS

73

Usability: make the configuration easier and accessible to non-experts Improve framework around key reasoning steps with EYE Investigate energy-efficiency on low-end devices Generalize and implement DIVIDE for other use case domains

SLIDE 74

FUTURE RESEARCH STEPS

74

Usability: make the configuration easier and accessible to non-experts Improve framework around key reasoning steps with EYE Investigate energy-efficiency on low-end devices Generalize and implement DIVIDE for other use case domains Extend query instantiation to other query parameters (e.g. window)

SLIDE 75

ir. Mathias De Brouwer

PhD Student

Ghent University – imec IDLab – Internet Technology and Data Science Lab E mrdbrouw.DeBrouwer@UGent.be T +32 9 331 49 38 M +32 484 97 43 15 http://idlab.ugent.be http://idlab.technology

Ghent University @ugent Ghent University