Abstract:

Sensor-based human activity recognition is an emerging research field. It is an important application of Internet of Things (IoT), and has  very promising application prospects in health care/recovery, elder/invalid people assistant, smart home/office, etc. Accuracy is one of the most important evaluation standards of activity recognition, and appropriate features and classifiers are important accuracy factors. We first extract a novel feature called correlation feature. By combining the theory of compressive sensing and sparse representation, we propose a multi-task compressive sensing method and use it as the classifier to resolve the problem of activity recognition. Finally, we conduct a large amount of experiments on a set of benchmarks with LeaveOneSubject-Out cross validation. Experimental results show that the extracted feature and the proposed method are effective in improving the accuracy of sensor-based activity recognition. Moreover, compared to the corresponding single task method, the proposed classifier can reduce the execution time by nearly 56%.

Key words: machine learning;Internet of Things;body area network;activity recognition;feature extraction