Human Activity Recognition with Ensemble Learning

Human activity recognition (HAR) has become a subject of interest for researchers with the rapid increase in the number of smart buildings and cities in recent years. The common applications are in security including crowd analysis, ambient mood detection, and emergency detection. This study proposes an ensemble learning method for Human Activity Recognition using recurrent and temporal convolutional neural networks. Ensembling match scores of Long Short Time Memory, Gated Recurrent Unit, and Temporal Convolutional Network, first, the training of the models was carried out with the peripheral sensor data, and then the embedding process was carried out on the development board. Among the experimental studies using human activity recognition with the “continuous ambient sensor data set” which includes 30 houses' sensor data. The proposed ensembling methodology achieved average recognition accuracy of 96.52 % in the Python. The highest test accuracy of 99.39% is achieved for ‘House 124’ in both software and hardware (Raspberry Pi). Since the base models are diverse and independent, the prediction error decreases when the ensemble approach is used. This paper proves that for Human Activitiy Recognition problem, the proposed architecture achieves reduced generalization error of the prediction. The overall results obtained by using human activity recognition for continuous ambient sensor data set both in Python and in the development board system showed that the proposed method has achieved a considerable performance improvement when compared to classical deen learning algorithms.