Machine Learning Based Real-Time Activity Detection System Design

Identification of human activities is a popular pattern recognition problem. In order to solve this problem, solutions based on machine learning are popularly used. Solutions based on the principle of collecting and processing classified data from one person are often used for non-real-time solutions. In this study, a system design is presented in which real time processing of the received acceleration data is performed using a mobile device, and a hardware three-axis accelerometer and the daily movement of the person is detected through different classification methods. Besides, pre-processing is carried out in the training clusters, enabling the system to respond in real time. The open source WISDM (Wireless Sensor Data Mining) dataset is used for classification in system design. The WISDM data set has a continuous-time data set and a discrete-time version of the data set. In this study, the continuous time data was handled again and some modifications were made to the data set and the discretization process was performed. In this respect, the classification performance for the J48 classification algorithm increased from 85.05% to 89.80%, and the performance in the data set for MLP (Multilayer Perceptron) increased from 84.94% to 93.08%. Furthermore, in the system obtained by using the obtained dataset, real-time usage result is taken as 70% performance. Reasons for the success difference between real time system and data set are discussed and solution proposal is presented.