GPU and Data Stream Learning Approaches for Online Smartphone-based Human Activity Recognition

Abstract

The availability of diverse embedded sensors in modern smartphones has created exciting opportunities for developing context-aware services and applications, such as Human activity recognition (HAR) in healthcare and smart buildings. However, recognizing human activities using smartphones remains a challenging task and requires efficient data processing approaches due to the limited resources of the device. For example, the training process is usually performed offline (on the server or the cloud) but rarely online on the mobile device itself, because traditional batch learning usually needs a large dataset of many users. Therefore, building models using complex multiclass algorithms is generally very time-consuming. In this paper, we have experimented two approaches in order to accelerate the training time. In the first approach, we conducted batch-learning experiments using a GPU platform. Results showed that High Performance Extreme Learning Machine (HPELM) offers the best compromise accuracy/time. Moreover, it achieved better performance on two dynamic activities, outperforming SVM in our previous study. In the second approach, we conducted experiments using online stream learning. Unlike the first approach, experiments were performed using accelerometer data only. We also studied the effects of user/device dependency and feature engineering on the classification performance by comparing five constructed real data streams. Experimental results showed that Hoeffding Adaptive Tree has comparable performance to batch learning, especially for user and device dependent data streams.