Master of Engineering Science
Electrical and Computer Engineering
Using smartphones for human activity recognition (HAR) has a wide range of applications including healthcare, daily fitness recording, and anomalous situations alerting. This study focuses on human activity recognition based on smartphone embedded sensors. The proposed human activity recognition system recognizes activities including walking, running, sitting, going upstairs, and going downstairs. Embedded sensors (a tri-axial accelerometer and a gyroscope sensor) are employed for motion data collection. Both time-domain and frequency-domain features are extracted and analyzed. Our experiment results show that time-domain features are good enough to recognize basic human activities. The system is implemented in an Android smartphone platform.
While the focus has been on human activity recognition systems based on a supervised learning approach, an incremental clustering algorithm is investigated. The proposed unsupervised (clustering) activity detection scheme works in an incremental manner, which contains two stages. In the first stage, streamed sensor data will be processed. A single-pass clustering algorithm is used to generate pre-clustered results for the next stage. In the second stage, pre-clustered results will be refined to form the final clusters, which means the clusters are built incrementally by adding one cluster at a time. Experiments on smartphone sensor data of five basic human activities show that the proposed scheme can get comparable results with traditional clustering algorithms but working in a streaming and incremental manner.
In order to develop more accurate activity recognition systems independent of smartphone models, effects of sensor differences across various smartphone models are investigated. We present the impairments of different smartphone embedded sensor models on HAR applications. Outlier removal, interpolation, and filtering in pre-processing stage are proposed as mitigating techniques. Based on datasets collected from four distinct smartphones, the proposed mitigating techniques show positive effects on 10-fold cross validation, device-to-device validation, and leave-one-out validation. Improved performance for smartphone based human activity recognition is observed.
With the efforts of developing human activity recognition systems based on supervised learning approach, investigating a clustering based incremental activity recognition system with its potential applications, and applying techniques for alleviating sensor difference effects, a robust human activity recognition system can be trained in either supervised or unsupervised way and can be adapted to multiple devices with being less dependent on different sensor specifications.
Yin, Xizhe, "Leveraging Smartphone Sensor Data for Human Activity Recognition" (2016). Electronic Thesis and Dissertation Repository. 4292.