Human Activity Recognition Improvement on Smartphone Accelerometers Using Cima

Abstract

Human activity recognition (HAR) is a research field that focuses on detecting user activities and has wide applications. However, the problems that need to be solved are real-time constraints and imbalanced datasets due to different activity frequencies. Our research aims to apply classification integrated moving averages (CIMA) to HAR by evaluating its performance regarding real-time constraints and imbalanced datasets. We achieved the smartphone accelerometer dataset from Kaggle, which consists of several activities: walking, jogging, climbing, and descending stairs. We develop a general CIMA windowing algorithm with hyperparameters J and W. We benchmark CIMA with two state-of-the-art HAR methods: distributed online activity recognition system (DOLARS) and convolutional neural network (CNN). We conducted some imbalance and model size analysis. The test results show that, with J = 10 and W = 240, CIMA performs better than DOLARS and CIMA with recall, precision, and f1-score of 0.996, 0.993, and 0.994. We also prove that CIMA, assisted by quantization, has the smallest model size compared to the CNN and DOLARS model sizes. Finally, we demonstrate that CIMA performs well for imbalanced datasets, where CIMA’s recall on upstairs and downstairs activities is better than DOLARS and CNN, with values of 1.00 and 0.98, respectively. Key Words: classification integrated moving average, human activity recognition, smartphone, accelerometer, imbalanced dataset