(1) IoT-Based Posture Detection and Feedback System: A Wearable Solution for
Maintaining Proper Posture
User’s back을 evaluate posture 대상으로 표현.
Roll angle이 30도 이상(이상치)로 표현 되었을시 —> device alert
Rolle angle, arc tangent normalization으로 정규화 실시
(2) Reliability of sitting posture between physical therapist videobased evaluation and SMART IMU system using rapid upper limb assessment (RULA)
RULA 랑 wearable sensors 로 assessment tools
A group , the upper extremities: lower arm and wrist
B group , neck, trunk, and leg assessments with scores based on specified angles
Namwongsa S et al. 논문에서 musculoskeletal disorders and ergonomic risks 의 연관성이 밝혀짐
Posture assessment methods
Subjective methods
Body part discomfort scale
The use of questionnaires
Objective methods
Ovako Working Posture Analysis System (OWAS) , time based (general)
Posture of back (4) , arms (3) , legs (7) , weight of load handled (3 categories) / 카메라나 wearable machine으로 측정 절대 불가능선에 있음
(1)의 평가 기준으로 4 digit code가 완성되고, 그것으로 work phase (사람이 뭘 하고 있는지 판단하는 기준표)
Action categories (1~4) 4로 갈수록 harmful
Rapid Upper Limb Assessment (RULA) , event based (sensitive)
Rapid Entire Body Assessment (REBA) , event based (sensitive)
This research compared 3D and 2D angle measurements against a motion-capture camera system. (a small sample size and only males.)
SMART inertial measurement unit (IMU) 3개 기계 머리, 목 뒤, 허리 아래
the right temporal area of the head
the seventh cervical spine
the fifth lumbar spine positions) 에 두고, 그것의 각도 차이로 RULA assessment 비교, MSD risk level을 개인적으로 생성 (neck and trunk joint angle)
결과값:
The results from the SMART IMU system and physical therapist show strong correlations. (문제점, 미래 방향성 —> physical therapist의 의문점 그냥 strong correlations로 보이는건 거의 당연시 되는 결과 예측값인데 이걸 위해서 이 연구를 했다기엔 너무 멍청함) + webcam은 그저 physical theraphist한테 보여주려고 record했다는데 이걸 무슨 기계와 연동을 시킨것도 아님.
근데 subjective methods(webcam + physical therapist) and objective method(RULA assessment with the OWAS) 는 아무 영향도 없었다.
(3) NeckCheck: Predicting Neck Strain using Head Tracker Sensors
Results:
Random Forests (RF) , best performance with score.
Understand the contribution of head tracker features in predicting EMG activity
we analyze feature importance
scores from the RF model. Pitch is the most influential feature (0.549), followed by Roll (0.286) and Yaw (0.165).
Head tilting forward and backward, has the strongest correlation with muscle strain, making it a crucial factor in detecting tech neck.
(4) Comparing Posture Classification: A Human Lying Posture Pressure-Map Dataset
방향성 히스토그램(HOG, Histogram of Oriented Gradient)을 활용한 특징 추출, leave-one-subject-out 교차검증 방식
- focuses on pressure field measurement and the subsequent classification by
means of state-of-the-art software methods in machine learning and artificial neural networks (ANNs)
Smart passive approach
Standard machine learning methods, KNN, SVM, LDA
Directly applying the classifier to an image, X feature extraction
Extracting the features, HOG (Computer vision algorithm) / Taking the direction in every 1 by 1 cell and see the diff. of Color in near cells to extract the outlier of picture. It mostly shows that the important feature of image is the outlier which is the line of character or shape of it.
Pressure image segmentation with Fuzzy C-means (FCM)
애매 모호한 관측치가 많은 데이터셋에는 k-중심 군집보다는 집단에 속할 가능성을 나타내는 fuzzy clustering이 훨씬 상대적으로 적합.
집단에 속할 가능성을 계산하기 위해서 computational cost가 증가.
FCM으로 입력값을 여러 레벨로 나누고, 특정 값으로 전체 데이터셋을 일치 시켜서, 잡음 제거 및 안정화 —> PCA로 차원 축소 —> MLP로 분류
Deep Learning
Small scale CNN
Transfer learning of large scale CNN
SRC , Sparse Representation Classification
Entire training dataset
Data and sized lower the training dataset
The mattress foam and outer sheet feature perfect permeability
Utilize a widely favored material with piezoresistive capabilities
6 diff. segments of metrics (611 or 311)
Use the luminance transformation, utilize the gamma correction to express dark and bright part
Log transformation —> dark place를 강조
Owned Dataset (1280)
4 categories with 16 each subclasses
on the back
on the left side
on the right side
on the stomach
기본의 작은 데이터셋에 60개의 여러가지 문제 해결을 위한 train dataset을 통한 totally 60 배의 dataset을 구성.
LDA, linear —> determine the linear classification models (maximize the supervised learning model) with the normalization factor γ(0~1) to balance the property of data variation and stability
QDA, quadratic —> determine the quadratic classification models
전체 dataset을 the processing period rose up to 20 hours in the occluded option
MCR- SRC method with lasso regularization 99%
83.5% during LOSO validation
HOG + KNN (k = 4), 99% prediction
