Modified deep residual network architecture deployed on serverless framework of IoT platform based on human activity recognition application

Abstract In the last few years, human activity recognition (HAR) is a subject undergoing intense study in various contexts such as pattern recognition and human-device interaction. HAR applications come to an aid of Telecare system which is paving the way for doctors and nurses to measure the health status of their patients. Due to the ubiquitous influence of smartphones in an individual’s life, we take embedded smartphone sensors into account as our case study. The proposed method, Modified Deep Residual Network, outperforms the accuracy of Human activity recognition compared with state-of-the-art machine learning techniques which are using Raw signals as their input. we defined new pooling layer called smooth-pooling to leverage the model performance. The accuracy of proposed architecture is evaluated on three common dataset that comprises accelerometer and gyroscope raw data. The results demonstrated the proposed method outperforms accuracy of classification while requiring just raw data with lower parameters compared to other works. Furthermore, The proposed HAR method is deployed in our IoT cloud platform which enables users to create scenarios based on what they are doing at home. Using Function as a Service (FaaS) architecture in this platform solves the scalability issues by running each function in a separate container. The IoT platform prepares an infrastructure for developers who want to integrate their application into the platform and use its functionality along with other IoT platform options.