SENSOR BASED REAL-TIME HUMAN ACTIVITY RECOGNITION IN WIRELESS MULTIMEDIA SENSOR NETWORK

Abstract

Human Activity Recognition (HAR) can be widely used in medicine and military applications. Primarily, it can be used as an assistive technology for the healthcare of older people. In general, HAR can be performed using the data collected from the various type of sensors: body, object and ambient sensors. There are two types of sensors: external and internal. A typical example of applications, which use external sensors, is intelligent home implementations that contain many sensors, such as temperature, pressure, light, ultrasonic sensors, and cameras. In this study, data from internal sensors were used for HAR of activities including walking, walking up or downstairs, sitting, standing, and falling from the collected data of smartphones. In addition, all the HAR models experimented with by Wireless Multimedia Sensor Network (WMSN) with the real-time collected dataset. For the recognition, we use various neural network algorithms as CNN, LSTM and traditional Machine learning classification algorithms such as SVM, KNN, and Random Forest Classifier. In addition, we implemented dimensionality reduction to decrease the number of features, which helped reduce computational time and energy consumption, and transfer learning with different scenarios to increase accuracy, and all these functions are implemented and compared in two WMSN architectures.