HL-HAR: Hierarchical learning based human activity recognition in wearable computing

Abstract

In recent years there have been many successes of recognizing the human activity using the data collected from the wearable sensors. Besides, many of these applications use the data from the smartphone. But it is also a challenge in practice for two reasons. Most method can achieve a high precision in the cost of increasing memory consumption, or asking for complicated data source. In this paper, (1) Utilizing Plus-L Minus-R selection to single out the optimal combination from the feature vector extracted; (2) Introducing a fast classification method named H-ELM to resolve the problem of the highly memory consumption in the process of calculation. The main benefit of this factor is to reduce memory usage and increase recognition accuracy with a brief feature vector so that a wearable device can identify activities all by itself. And the wearable device can recognize the sample activities even if keeping away from cellphone. Our results show that this method leads to that we can recognize object activities with the overall accuracy of 93.7% in a very short period of time on the dataset of Human Activity Recognition Using Smartphones Dataset. The selected 25-dimension feature vector nearly contains all the information and after many times of test, it can achieve very high percentage of accuracy. Moreover, the method enables the learning velocity to outperform the state-of-the-art on the Human Activity Recognition domain.