Work

Sensor Fusion in Human Activity Recognition and Occupancy Detection

Year: 2020
Type: dissertation
Abstract: This thesis investigates the recognition of human activities from the position of the wrist and the feasibility of occupancy detection based on environmental sensors. In preparation to this thesis, a wrist-worn 10-axis human activity tracker combining a 3-axis accelerometer, a 3-axis gyroscope, a 3-axis magnetometer, and a barometric pressure sensor was developed. Four volunteers participated in collecting a first dataset of 190 measurements by performing ten everyday activities. An approach is introduced by which a person's energy expenditure can be estimated independently of the measurement position. Using the acceleration data obtained at the wrist, various methods for detecting cycling are presented. It is also shown that the barometer allows discriminating activities that involve a vertical displacement, such as climbing stairs or riding an elevator, as well as helps to reduce the false alarm rate of a fall detector. The potential of magnetic field sensing is derived from the signal characteristics of the magnetic interference as experienced during cycling and riding an elevator. It is demonstrated that the examined activity set can be discriminated by fusing the data of either the accelerometer and the barometer or the gyroscope and the barometer. To detect and count occupants, an approach using a fusion of environmental sensors from an indoor air quality measurement system is presented. Environmental sensors, as opposed to motion detectors, are nonintrusive, easy to install, low cost, detect nonmoving occupants, do not have dead spots, and can even infer the number of occupants. For this study, measurements of carbon dioxide, volatile organic compounds, air temperature, and relative air humidity were conducted in four student apartments for a total of 49 days. Features were extracted from the environmental sensor data and subsets selected using correlation-based feature selection. In this thesis, a comparison of the supervised learning models RIPPER, Naı̈ve Bayes, C4.5 decision tree, logistic regression, k-nearest neighbours, and random forest is performed. Furthermore, a method is proposed to greatly reduce time and effort of collecting training data in residential buildings. The results indicate that the predictive power of volatile organic compound sensing is comparable to that of carbon dioxide. With a simple Naı̈ve Bayes classifier, the approach detected occupancy and estimated the number of occupants with an accuracy of 81.1% and 64.7%, respectively. In addition, a gas sensor test setup was developed that is capable of characterizing carbon dioxide and volatile organic compound sensors simultaneously.
Language: English
Open Access status: green
OpenAlex ID: W3135712206
Cites: 182
Cited by: 1
Related to: 10