Nurse care activity recognition challenge: a comparative verification of multiple preprocessing approaches

Abstract

Although activity recognition has been studied considerably for the last two decades, it is still not so easy to handle complicated activity classes in a specific domain. The 2nd Nurse Care Activity Recognition Challenge Using Lab and Field Data aims to explore a part of those complicated activities by focusing on the nurse caring. Our team, "UCLab", found that the main problem in the challenge is the imbalance and unevenness of the dataset, each of which often happens in real-field data. Considering the problem, we approached the challenge using a Random Forest-based method with multiple preprocessing to classify 12 activity modes. Our approach consists of the following steps: We first preprocessed the acceleration data to obtain uniformly sampled signals. Then we extracted acceleration data with respect to each row of the given label data and extracted feature values. We adopted Random Forest for classification and performed post-processing to the predicted data obtained from the classifier. As a result, we obtained 51.5% accuracy with the trial-based evaluation.