A Novel Framework for Human Activity Recognition with Time Labelled Real Time Sensor Data

Human activity recognition is an effective approach for identifying the characteristics of historical data. In the past decades, different shallow classifiers and handcrafted features were used to identify the activities from the sensor data. These approaches are configured for offline processing and are not suitable for sequential data. This article proposes an adaptive framework for human activity recognition using a deep learning mechanism. This deep learning approach forms the deep belief network (DBN), which contains a visible layer and hidden layers. The processing of raw sensor data is performed by these layers and the activity is identified at the top most layers. The DBN is tested using the real time environment with the help of mobile devices that contain an accelerometer, a magnetometer, and a gyroscope. The results are analyzed with the metrics of precision, recall, and the F1-score. The results proved that the proposed method has a higher F1_score when compared to the existing approach.