Abstract
Activity recognition with triaxial accelerometer embedded in mobile phone is an important research topic in pervasive computing field. The research results can be widely used in many healthcare or data mining applications. Numerous classification algorithms have been applied into the activity recognition tasks. Among these algorithms, ELM (Extreme Learning Machine) shows its advantages in generalization performance and learning speed. But because of the randomly generated hidden layer parameters, ELM classifiers usually produce unstable predictions. To construct a more stable classifier for our mobile-phone based activity recognition task, we designed an ensemble learning algorithm called Average Combining Extreme Learning Machine (ACELM), which integrates several independent ELM classifiers by averaging their outputs. To evaluate the algorithm, we collected raw accelerometer data of five daily activities from mobile phones carried by volunteers, and used them to train and test our classifier. The experiment results show that our algorithm has greatly improved the general performance of ELM in mobile-phone based activity recognition task.