Research On Inertial Sensor Data Based Human Daily Activity Recognition

As an emerging research in the field of pattern recognition, inertial-sensor based human ac-tivity recognition has attracted lots of attentions from researchers all over the world. The inertial signal of human activity can be collected by using on-body inertial sensor and transmitted to a remote terminal via wireless network. A major goal of inertial-sensor based researches among a range of health-related areas is to monitor daily activities of patients in long term. Human dai-ly activities may provide additional informatidn to medical staff to accurately diagnose chronic diseases, as well as design the therapeutic plan for the patients. The quality of healthcare pro-vided to the elderly and children could be effectively improved by human activity recognition. Although human daily activity recognition has made a great development, due to the diversity of human daily activities and the complexity of the environment, there are still many difficult problems to be solved.Based on previous work, the main contributions of this research are listed as follows:1. High quality features are essential to improve the classification accuracy of pattern recog-nition. According to our best knowledge, EEMD based features are rarely mentioned previously for activity recognition, which have been utilized to discriminate human daily activities in this research.2. Although sensor-based human activity recognition has attracted lots of attentions, most researches require sensors placed in preset orientation and position on human body. Otherwise, the performance of classification will be degraded. In this research, we propose a game theory based feature selection method to evaluate the features. Relevant and distinguished features robust to the placement of sensors are selected.3. Balanced dataset has been utilized by the previous human activity recognition algorithms to train the classifier. However, imbalanced dataset is ubiquitous in human activity recognition. Though the class imbalance problem exists as a universal phenomenon in human activity recog-nition, few researches have mentioned this problem and solved it. In order to reduce the influence of the imbalance datasets problem, the mixed-kernel based weighted extreme learning machine (MK-WELM) has been proposed in this research. Comparing with ELM and weighted ELM methods, experimental results over different human activity datasets demonstrate the effective-ness of the proposed method.4. Fixed placements of inertial sensors have been utilized by previous human activity recog-nition algorithms to train the classifier. However, the distribution of sensor data is seriously af-fected by the locations of sensors. The performance will be degraded when the model trained on one placement is used in others. In order to tackle this cross-location problem, a fast and robust human activity recognition model called M-RKELM is proposed in this research. Experimental results show that the proposed model can adapt the classifier to new locations quickly and obtain good recognition performance.