Feature extractionand incremental learning to improve activity recognition on streaming data

In this paper, we propose an approach for an online human daily activity recognition system using motion sensor data. From the sensor readings, the system decides which activity is performed when the values change. It uses the previous measurements to interpret the current ones, without the need to wait for future information. The contributions of this study relies on the presentation of two methods to extract features from the sequence of sensor events, a clustering method to handle missing activity labels in dataset and an incremental learning method to deal with complexity and time spent in training since our system works on streaming data. Our methods are evaluated on publicly available real environment datasets.