센서 기반 넘어짐 동작을 인식하기 위한 딥러닝 모델 아키텍처 설계

다국어 초록 (Multilingual Abstract)

With the popularization of Personal Mobility Vehicles (PMV) such as bicycles, electric scooters, etc., user demand is increasing. PMVs are being used as some substitutes for walking or using public transportation, and Sharing System have made it easier for individuals to access the means without having to own them. In particular, as the demand for delivery has increased significantly in the past few years, the number of drivers for motorcycles, bicycles, and electric scooters as a means of delivery has increased significantly. The global electric scooters market size is expected to grow at a Compound Annual Growth Rate (CAGR) of 7.8% until at least 2030.
However, as the number of users increases, the occurrence of large and small traffic accidents is increasing. In the case of a two-wheeled vehicle accident, the body of the rider is exposed to the impact as it is, and thus the accident makes injury more serious. Also, after the first collision, a secondary collision occurs due to surrounding structures. In the process, a significant number of riders suffered serious head and neck injuries. Reducing injuries from crashes is important to protect the safety of occupants. For this, a technology to recognize and judge the current movement state through real-time information about the change in the rider's posture is needed.
In this paper, I performed performance evaluation for motion detection according to the deep learning algorithm when Inertial Measurement Unit (IMU) sensor data on the rider's movement is given, and through this, I proposed a new improved architecture. A convolution operation was performed on each axis of the acceleration and angular velocity sensors, and Residual block was used to design it. Through this, the characteristics of each axis were analyzed individually, and the accuracy was greatly improved.
In order to build the dataset, I conducted accident experiments using a mannequin. I attached a sensor to the back of the mannequin and collected information on acceleration and angular velocity according to the movement of the rider. I extracted and analyzed acceleration and angle information to find out the motion characteristics, and made datasets for Deep Learning.
In the case of a deep learning model, the architecture is implemented using algorithms such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short Term Memory (LSTM). In order to select an appropriate deep learning technique for the dataset used in this study, I performed performance evaluations on the output according to the change of the algorithm while maintaining hyperparameters such as layers and nodes. For this dataset, the CNN-based method showed the best performance, and analysis along each axis and time axis of the sensor was also performed. Based on these studies, I designed a new improved architecture TAMS (Time Attention for Multi Sensor) using the CNN algorithm. TAMS showed good performance with fewer layers and only 1/5 epoch than the comparison model.
To check the operation of the model trained using the dataset, an experiment was performed through the test bed. The model has been ported to an embedded system based on Raspberry Pi, and when an accident is detected, the operation can be confirmed by deploying the airbag. As a result, the Raspberry Pi module detected the accident in real time and activated the airbag immediately after the first collision, confirming that the accident was determined.
From the results of each experiment, real time motion detection of passengers is possible through deep learning using a single IMU sensor, and it was confirmed that the performance was improved compared to the existing model through TAMS design. In addition, it is expected that data for PVM other than the bicycle data used in this study will be available.

목차 (Table of Contents)

• 1 서 론 1
• 1.1 연구 배경 및 필요성 1
• 1.2 논문 구성 5
• 2 이론적 배경 7
• 2.1 딥러닝 7
• 2.1.1 딥러닝 모델 구성 요소 9
• 2.1.2 딥러닝 모델 기법 20
• 2.1.3 성능지표 27
• 2.2 관성 측정 장치 32
• 2.2.1 관성 측정 장치 특성 32
• 2.2.2 데이터 전처리 40
• 3 적용 대상 및 수행 방법 41
• 3.1 동작 상태 규정 41
• 3.2 데이터 수집 43
• 3.2.1 사고 데이터 수집 44
• 3.2.2 데이터 특성 45
• 3.3 데이터 증강 49
• 4 딥러닝 모델에 따른 특성 실험 52
• 4.1 딥러닝 모델 구조 및 학습 환경 52
• 4.1.1 딥러닝 모델 구조 52
• 4.1.2 딥러닝 모델 개발 환경 62
• 4.2 딥러닝 모델에 따른 학습 결과 및 성능 평가 62
• 4.3 사고 예측 모델의 현장 구현 79
• 5 결론 82
• References 84
• Appendix 90
• DNN 90
• PNN 93
• LSTM 96
• CNN 99
• CNN(15) 102
• CNN(6) 105
• PNN-CNN 108
• CNN-PNN 111
• O-CNN 114
• CNNRE-PNN(15) 117
• CNNRE-PNN(6) 120
• TAMS EPOCH 500 123
• TAMS EPOCH 100 126

참고문헌 (Reference)

1. 자전거 교통사고의 통계분석 및 활용, 김명진, 홍종선, 한국데이터정보과학회지, 21(6), 1081-1090, , 2010