ResNet-based Network for Recognizing Daily and Transitional Activities based on Smartphone Sensors

Abstract

In contemporary wearable computing contexts, sensor-based human activity recognition (HAR) has become a popular research topic. Investigators from the Health Applications Research Institute presented promising discoveries to promote healthcare applications, including fall detection, athletic tracking and reporting, and a monitoring scheme for senior activities in intelligent homes. In these services, ordinary and transitory human actions are captured by smartphones' wearable sensors and analyzed as fundamental and complicated motions. Deep learning techniques demonstrated the usefulness and effectiveness of convolutional neural networks (CNNs) in extracting high-level features embedded in sensor data to develop reliable recognition models. CNN faces deterioration of gradient vanishing issues when networks require deeper convolution layers. To overcome the problem, we developed ResNet, a deep residual network for determining daily and transitory activities. Using a significant standard HAR dataset called the KU-HAR dataset that gathered smartphone sensor data of various human actions, we performed experiments to identify the most appropriate ResNet-based models. Experimental findings indicate that the ResNet-18 has the highest accuracy, at 93.54%. The acquired results surpass prior state-of-the-art models by 3.87% in terms of accuracy.