2nd Place Award (Emteq Activity Recognition Challenge: Caring for Inter-user Dependency)
Kohei Adachi, Md. Shafiqul Islam, Sozo Inoue,
Emteq Activity recognition challenge
Human activity recognition is a popular research area. For the last few decades, researchers are using different types of machine learning algorithms to detect activities from various types of data such as motion capture data, accelerometer data, gyroscope data etc. The researchers have used various machine-learning algorithms to detect simple as well as complex human activities. In this paper, we are focusing on “Emteq activity recognition challenge” where the task is to identify 8 types of human activities. The challenge dataset contains accelerometer, gyroscope and magnetometer data. For this purpose, we have extracted important features from the sensor data and used Random Forest classifier to detect activities. We have used one- person leave out cross validation for our proposed model. The F1 score for our proposed model on one-person leave out cross validation is around 47%. Between persons test that is one person’s training data and another person’s test data, the F1 score is around 30%.