<%BANNER%>

Knowledge Assisted Human Activity Recognition for Improved Accuracy and Programmability

Permanent Link: http://ufdc.ufl.edu/UFE0045038/00001

Material Information

Title: Knowledge Assisted Human Activity Recognition for Improved Accuracy and Programmability
Physical Description: 1 online resource (155 p.)
Language: english
Creator: Kim, Eunju
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2013

Subjects

Subjects / Keywords: activity -- activitymodel -- activityrecognition -- fuzzy-logic -- knowledge -- recognition -- semantics
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Activity recognition (AR) is a key technology for developing human-centric applications in smart environments. However,state-of-the-art AR technology cannot be used for addressing real world problems due to insufficient accuracy and lack of a usable activity recognition programming model.  To address these issues, a new AR approach is developed in this dissertation. AR performance is strongly dependent on the accuracy of the underlying activity model. Therefore, it is essential to examine and develop an activity model that can capture and represent the complex nature of human activities more precisely. To address this issue, we introduce generic activity framework (GAF) and activity semantics. The GAF is a refined hierarchical composition structure of the traditional activity theory. Activity semantics are highly evidential knowledge that can identify activities more accurately in ambiguous situations. We compare our activity model with traditional activity model in terms of attainable recognition certainty. Two new AR algorithms—Multilayer Neural Network (MLNNK) based algorithm and fuzzy logic (FL) based algorithm—have been developed in this dissertation. These algorithms utilize and work in tandem with the developed activity framework and model.  The MLNNK based AR algorithm illustrates the high recognition accuracy of generic activity framework modeling approach. FL based AR algorithm utilizes both activity semantics and generic activity framework. For achieving high accuracy, it is important to identify and mitigate the most debilitating sources of uncertainty. The efficacy of AR systems is usually quantified based on the recognition accuracy at the final step of activity recognition process.This method does not reveal the uncertainty sources that affect overall performance significantly. Therefore, it is necessary to quantify every possible uncertainty source through all activity recognition procedures. To address this issue, metrics and measurement methods for each uncertainty source are developed. AR technology should provide programmable interface to developers to support AR system design change according to new application requirements or AR environment changes. This dissertation classifies developers into three categories:smart space developer, activity model and algorithm developers, and application developers. The hierarchical aspects of our generic activity framework decouple the observation subsystem from the rest of the activity model. We demonstrate the value of this decoupling by experimentally comparing the level of effort needed in making sensor changes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Eunju Kim.
Thesis: Thesis (Ph.D.)--University of Florida, 2013.
Local: Adviser: Helal, Abdelsalam Ali.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2013
System ID: UFE0045038:00001

Permanent Link: http://ufdc.ufl.edu/UFE0045038/00001

Material Information

Title: Knowledge Assisted Human Activity Recognition for Improved Accuracy and Programmability
Physical Description: 1 online resource (155 p.)
Language: english
Creator: Kim, Eunju
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2013

Subjects

Subjects / Keywords: activity -- activitymodel -- activityrecognition -- fuzzy-logic -- knowledge -- recognition -- semantics
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Activity recognition (AR) is a key technology for developing human-centric applications in smart environments. However,state-of-the-art AR technology cannot be used for addressing real world problems due to insufficient accuracy and lack of a usable activity recognition programming model.  To address these issues, a new AR approach is developed in this dissertation. AR performance is strongly dependent on the accuracy of the underlying activity model. Therefore, it is essential to examine and develop an activity model that can capture and represent the complex nature of human activities more precisely. To address this issue, we introduce generic activity framework (GAF) and activity semantics. The GAF is a refined hierarchical composition structure of the traditional activity theory. Activity semantics are highly evidential knowledge that can identify activities more accurately in ambiguous situations. We compare our activity model with traditional activity model in terms of attainable recognition certainty. Two new AR algorithms—Multilayer Neural Network (MLNNK) based algorithm and fuzzy logic (FL) based algorithm—have been developed in this dissertation. These algorithms utilize and work in tandem with the developed activity framework and model.  The MLNNK based AR algorithm illustrates the high recognition accuracy of generic activity framework modeling approach. FL based AR algorithm utilizes both activity semantics and generic activity framework. For achieving high accuracy, it is important to identify and mitigate the most debilitating sources of uncertainty. The efficacy of AR systems is usually quantified based on the recognition accuracy at the final step of activity recognition process.This method does not reveal the uncertainty sources that affect overall performance significantly. Therefore, it is necessary to quantify every possible uncertainty source through all activity recognition procedures. To address this issue, metrics and measurement methods for each uncertainty source are developed. AR technology should provide programmable interface to developers to support AR system design change according to new application requirements or AR environment changes. This dissertation classifies developers into three categories:smart space developer, activity model and algorithm developers, and application developers. The hierarchical aspects of our generic activity framework decouple the observation subsystem from the rest of the activity model. We demonstrate the value of this decoupling by experimentally comparing the level of effort needed in making sensor changes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Eunju Kim.
Thesis: Thesis (Ph.D.)--University of Florida, 2013.
Local: Adviser: Helal, Abdelsalam Ali.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2013
System ID: UFE0045038:00001


This item has the following downloads:


Full Text

PAGE 17

1.1 Overview of Sensor based Activity Recognition

PAGE 18

1.2 Research Motivations

PAGE 20

1.3 Current Challenges 1.3.1 Accuracy Concurrent act ivities

PAGE 21

Interleaved activities. Ambiguity Variety Multiple subjects 1.3.2 Programmability Scalable activity observation system

PAGE 22

Flexible activity model or algorithm. Programmable Interface. 1.4 Research Objectives 1.4.1 Accuracy Enhancement of Activity Recognition 1.4.2 Uncertainty Analysis and Management

PAGE 23

1.4.3 Development of Activity Recognition Programming Model 1.5 Dissertation Organization

PAGE 24

2.1 Activity Theory Origin of Activity Modeling

PAGE 26

2 .2 P robabilistic Activity Models

PAGE 27

2.2.1 Hidden Markov Model (HMM)

PAGE 28

2.2.2 Conditional Random Field (CRF) 2.3 Knowledge based Activity Recognition

PAGE 29

2.4 Performance Metrics and Measurement

PAGE 30

2.4.1 Certainty Factor CF(H, E): MB(H, E): MD(H, E):

PAGE 31

2.4.2 Metrics for Activity Recognition Performance

PAGE 33

2.5 Uncertainty Analysis

PAGE 36

2.6 Activity Recognition Programming Model

PAGE 38

2.7 Summary

PAGE 40

3.1 Assumed Theory and Facts about Human Activities 3.1.1 Problem of Activity Theory Based Assumption

PAGE 41

3.1.2 Problem of Probabilistic Assumption of Activity 3.2 Activity Modeling 3.2.1 Problem of Activity Theory Based Model

PAGE 42

3.2.2 Problem of Probability Based Model

PAGE 44

3.3 Activity Recognition Algorithm

PAGE 46

3.4 Activity Observation System

PAGE 47

3.5 Activity Recognition Programming Model

PAGE 49

4.1 Goal: Increased Accuracy in Activity Model

PAGE 50

4. 2 Generic Activity Framework (GAF)

PAGE 51

Subject Time Location Motive Tool

PAGE 52

Object Context

PAGE 53

Sensors Operation Action

PAGE 54

Activity Meta activity 4.3 Semantics Enriched GAF (SGAF)

PAGE 55

4.3.1 Dominance S emantics Key component. Optional component

PAGE 56

4.3.2 Mutuality S emantics Con current component. Exclusive component Ordinary component 4.3.3 Order S emantics No order.

PAGE 57

Strong order. Weak order. Skip chain order 4.3.4 Effect S emantics T = 0 or T 0.

PAGE 58

T > 0 4.3.5 Activity Life Cycle

PAGE 59

Started. Tentatively Started. Paused. Performing.

PAGE 60

Finished. 4.4 Case Study of Activity Modeling based on SGAF

PAGE 61

4. 5 Validation of SGAF Activity Modeling Approaches 4.5.1 Validation Scenario

PAGE 63

4.5.2 Comparison and Analysis

PAGE 66

4. 6 Summary and Discussion

PAGE 68

5.1 Goal: Increased Activity Recognition Accuracy 5.2 MultiLayer Neural Network based Algorithm 5.2.1 Multi Layer Neural Network for GAF

PAGE 71

5.2.2 MLNNK Training

PAGE 73

5. 3 Validation of MLNNK Based Activity Recognition Algorithm 5.3.1 Experiments

PAGE 74

5. 3 .2 Comparison and Analysis of Experiment Result s

PAGE 77

5. 4 Summary and Discussion

PAGE 79

6.1 Goal: Increased Activity Recognition Accuracy and Tolerance to Uncertainty 6.2 Fuzzy Logic for Activity Recognition

PAGE 80

6.2.1 Extension of Fuzzy Set to Activity Semantic Fuzzy Set

PAGE 81

6.2.2 Extension of Fuzzy Operator to Activity Semantic Fuzzy Operator Fuzzy Dominance Operator

PAGE 82

Fuzzy Mutuality Operator 6.2.2 Extension of Fuzzy Rule to Activity Semantic Fuzzy Rule

PAGE 83

Dominanc e semantic rules Mutuality s emantic rules

PAGE 84

Order s emantic rules Effect s emantic rules 6.3 Applying Activity Semantic Fuzzy Logic to Activity Recognition

PAGE 85

6 .3 .1 Fuzzy Membership Function

PAGE 87

Defuzzification of fuzzy value 6 .3 .2 No Training Weight Computation

PAGE 89

The weight of a group

PAGE 90

6. 3 .3 The Computation of Fuzzy Value of Activities 6.3.4 Enforcing Semantics Using Semantic Fuzzy Operators or Rules 6.3.5 Evaluation of Activity Life Cycle

PAGE 91

6 .4 Fuzzy Logic Based Activity Recognition System Fuzzifier

PAGE 92

Activity Recognition Graph (AR Graph) Semantics Evaluat or Activity Life Cycle Heap (ALC Heap)

PAGE 93

Sensor Knowledge Activity Model Knowledge Fuzzy Logic and Rules

PAGE 94

observe checkStopCondition semanticFuzzify buildGraph computeWeight getNextComponent lookupChildrenComponents performFuzzyDominanceOpertion getMaxTime enforceSemantics (

PAGE 95

lookupChildrenComponents lookupChildrenComponents lookupChildrenComponents removeFromGrap h removeFromGraph removeFromGraph removeFromGraph 6.5 Validation of Fuzzy Logic Based Activity Recognition Algorithm

PAGE 96

6.5.1 Experiment 6.5.2 Comparison and Analysis of Experiment Results

PAGE 99

6.6 Summary and Discussion

PAGE 102

7.1 Goal: Developing More Practical Metrics and Measures of Accuracy

PAGE 103

7.2 End to End Analysis of Various Uncertainty Sources 7.2.1 Uncertainties in H uman A ctivities

PAGE 104

7.2.2 U ncertainties in S ensor T echnology 7.2.3 Uncertainties in AR T echniques

PAGE 106

7. 3 Uncertainty Metrics and Measures 7.3.1 Certainty F actor 7.3.2 Correlation Degree of Sensors and Activities

PAGE 107

7.3.3 Sensor D ata C ertainty 7.3.4 Activity C ertainty 7.3.5 Metrics for AR P erformance

PAGE 108

Accuracy for ARP Recall (Sensitivity) Precision

PAGE 109

Specificity 7.3.6 Uncertainty Impact Comparison

PAGE 110

7.4 Validation of Uncertainty Analysis Approaches 7.4.1 Experiment

PAGE 111

Step 1. Identifying target activities. Step 2. Designing activity models.

PAGE 112

Step 3. Activity scenarios.

PAGE 114

Step 4. Smart space instrumentation.

PAGE 115

Step 5. Activity dataset collection.

PAGE 116

Step 6. Activity recognition system implementation. Step 7. Executing the AR algorithms.

PAGE 117

7.4.2 Measuring and Comparing Uncertainties

PAGE 124

7.4.3 Comparison and Analysis of Uncertainty Impacts

PAGE 125

7.5 S ummary and Discussion

PAGE 128

8.1 Goal: Increasing Activity Recognition Programmability 8. 2 Activity Recognition Programmability

PAGE 129

8. 2 .1 Programmability in Activity Observation Subsystem Layer 8. 2 .2 Programmability in Activity Model and Algorithm Layer

PAGE 131

8. 2 .3 Programmability by T hird P arty in A pplication L ayer 8. 3 Case Study of Programmability Activity Knowledge Schema.

PAGE 132

Support of sensor insertion and Deletion. Activity Verification API Design.

PAGE 133

8. 4 Validation of Activity Recognition Programming Model

PAGE 135

8. 5 Summary and Discussion

PAGE 141

} }