Modified Random Forest Architecture for Thermal Sensor-Based Human Activity Detection and Classification in Smart Spaces
- 발행기관 금오공과대학교 대학원
- 지도교수 도움말 Dong Seong Kim
- 심사위원 신수용, 김동성, 이동현, 이헌철, 이재민
- 발행년도 2021
- 학위수여년월 22.2
- 학위명 박사
- 학과 및 전공 도움말 대학원 IT융복합공학과
- 세부전공 IT 융복합공학과
- 원문페이지 140
- 실제URI http://www.dcollection.net/handler/kumoh/000000016212
- UCI I804:47006-000000016212
- 본문언어 영어
초록/요약 도움말
Nowadays, research works into the dynamic and static human activities on smart spaces abounds. Artificial Intelligence (AI) and low-cost non-privacy invasive ambient sensors have made this ubiquitous. This work presents a state-of-the-art analysis, performance evaluation, and future research direction on the use of thermal sensor for human activity detection in smart spaces. One of the aims of activity recognition (especially that of humans) systems using thermal sensors and AI is the non-privacy invasion while ensuring the safety of persons in smart spaces. On a smart factory shop floor, human activity detection systems are put in place to ensure the safety of persons in such an environment. This system should have the ability to monitor issues like fall detection, a common work-related accident. Moreover, detection and classification models need be accurate to guarantee reliability of the system. In this work, a modified random forest architecture is presented to ensure the accurate and efficient detection of human activity based on dataset collected using thermal sensors. Also, the impact of sensor resolution was investigated. Result showed that the proposed modified random forest outperformed other existing solutions and ensemble learning models with an accuracy of 99.96% on 32x24 thermal sensor resolution. In addition, the proposed model maintained an average accuracy of 93% across all available sensor resolutions. Practical concerns such as integrating the thermal sensor with other sensors for data collection were raised to guide future research direction.
more초록/요약 도움말
오늘날, 스마트 공간에서 역동적이고 정적인 인간의 활동에 대한 연구가 이루어지고 있다. 인공지능(artificial intelligence:AI)과 저비용, 사생활 보호조치가 없는 센서들로 인해 유비쿼터스를 만들었다. 본 연구는 스마트 공간에서 열 센서를 사용한 인체 활동 감지 기법의 최신 분석, 성능 평가 및 향후 연구 방향을 제시한다. 열 센서와 인공지능을 이용한 활동 감지 시스템의 목표 중 하나는 스마트 공간에서 사람의 안전을 보장하면서 사생활을 보호하는 것이다. 스마트 팩토리 작업 현장에서는 이러한 환경에서 사람의 안전을 보장하기 위해 인체 활동 감지 시스템이 설치된다. 이 시스템은 일반적인 업무 관련 사고인 낙상 감지와 같은 문제를 감시할 수 있어야 한다. 또한 시스템의 신뢰성을 보장하기 위하여 감지 및 분류 모델은 정확해야 한다. 본 연구에서는 열 센서를 통해 수집한 데이터세트 기반으로 정확하고 효율적인 인체 활동 감지를 보장하기 위해 Ensemble 학습 아키텍처를 제안한다. 또한 센서 해상도의 영향을 조사했다. 결과는 제안하는 수정된 random forest가 32x34 열 센서 해상도에서 99.96%의 정확도로 기존의 다른 솔루션 및 Ensemble 학습 모델을 능가하는 것으로 나타났다. 또한 제안된 모델은 사용 가능한 모든 센서 해상도에서 평균 93%의 정확도를 유지하였다. 데이터 세트 수집을 위해 열 센서와 다른 센서의 통합과 같은 실용적인 문제를 향후 연구로 제기한다.
more목차 도움말
Contents
[List of Figures] i
[List of Tables] iv
[Abbreviations] v
Chapter I Introduction 1
1.1 Motivation for Research 3
1.2 Research Breakdown and Major Contributions 6
Chapter II Background Information 9
2.1 Thermal Sensors for Human Activity Monitoring 9
2.1.1 Basic Concept of the Thermal Sensor 9
2.1.2 Data Collection using Thermal Sensors 11
2.1.3 Thermal Sensors’ Placement in Factory Shop Floor 11
2.1.4 Handling Thermal Noise of Thermal Sensors’ Placement in Factory Shop Floor 12
2.2 Ensemble Learning and Human Activity Detection in Smart Spaces 13
2.2.1 Ensemble Learning Basics 13
2.2.2 Taxonomy of Ensemble Learning 14
2.2.3 Ensemble Learning: Hybrid Learning 16
2.2.4 Ensemble Learning: Machine Learning 17
2.2.5 Ensemble Learning: Deep Learning 17
2.2.6 Theoretical Challenges of Implementing Ensemble Learning Frameworks 19
2.2.7 Criteria for Selecting the Most Suitable Ensemble Learning Frameworks 20
2.3 Smart Spaces Use Cases and Brief Description 21
2.3.1 Human Activity Monitoring in Smart Homes 21
2.3.2 Human Activity Monitoring in Smart Factories’ Shop Floor 23
2.4 Justification for Thermal Sensing Technology for Human Activity Detection 24
2.5 ML-Based Classifiers or Inducers 28
2.5.1 Support Vector Machine 28
2.5.2 Naïve Bayes 29
2.5.3 K-Nearest Neighbor 30
2.5.4 Decision Tree 30
2.5.5 Artificial Neural Network 30
Chapter III Review of Related Works 32
3.1 Methodology for Review of Literature 32
3.2 Review of Recent Works 33
3.3 State-of-the-Art Approaches and Performance Evaluation 36
3.3.1 Machine Learning Approaches and Result Evaluation 36
3.3.2 Deep Learning Methods and Result Evaluation 37
3.3.3 Ensemble Learning Methods and Result Evaluation 38
3.3.4 Datasets used in Recent Thermal Array Based Activity Detection 39
3.3.5 Testbeds used in Recent Thermal Array Based Activity Detection 41
3.4 Summary of Review of Related Works and Research Gaps 44
Chapter IV Proposed Human Activity System Model and Description 47
4.1 System Model and Research Methodology 47
4.2 Experimental and Testbed set up: Thermal Sensor 48
4.3 Control Experiment and Testbed using RP-LiDAR Sensor 53
4.4 Brief Description of Various Datasets used in this Thesis 54
4.4.1 The Thermal Sensor Datasets collected by NSL 55
4.4.2 The eHomeSeniors Datset 55
4.4.3 The RP-LiDAR Dataset collected by NSL 56
4.5 Feature Engineering using Correlation Coefficient 57
4.6 Data Balancing 58
4.6.1 Under-sampling or Down Sampling 59
4.6.2 Over-sampling or Up Sampling 59
4.6.3 Synthetic Minority Oversampling Technique (SMOTE) vs Adaptive Synthetic Sampling (ADASYN) 60
4.7 Brief Description of Ensemble Learning Candidates 60
4.7.1 Gradient Boost (GB) 61
4.7.2 Extreme Gradient Boost (XGB) 62
4.7.3 AdaBoost (AD) 64
4.7.4 Decision Tree (DT) 64
4.7.5 Random Forest (RF) 65
Chapter V Simulation and Performance Validation for Different Scenarios 67
5.1 Experimental Setup 67
5.1.1 Ensemble Learning Parameters and Proposed Modified Random Forest (RF) 68
5.2 Performance Metrics 68
5.2.1 Minimum Classification Error (MCE) 68
5.2.2 Accuracy 68
5.2.3 Loss 70
5.2.4 Precision 70
5.2.5 Recall 70
5.3 Results & Discussions 71
5.3.1 Eliminating Deep Learning candidates using MATLAB 71
5.3.2 Case Study 1: Performance Evaluation using self-collected LiDAR-based Data 73
5.3.3 Case Study 2: Performance Evaluation using eHomeSenior Dataset-Melexis (32 x24 pixel) 76
5.3.4 Case Study 3: Performance Evaluation using NSL dataset collected with OMRON Sensors (4x4, 4x4H, and 32x32 pixel) 80
5.3.5 Case Study 4: Performance Evaluation using Thermal sensors (16x12, and 32x24 pixel) dataset collected by NSL 83
5.3.6 Case Study 5: Impact of Sensor Resolution on the Performance of the Proposed Random Forest Algorithm 86
5.3.7 Case Study 6: Performance Evaluation with Existing Related Works 87
5.3.8 Summary of Results 90
Chapter VI Conclusions, Practical Concerns and Future Works 91
6.1 Conclusion 91
6.2 Practical Concerns and Future Works 92
[Reference] 95
