Detecting Human Activities Based on a Multimodal Sensor Data Set Using a Bidirectional Long Short-Term Memory Model: A Case Study

Abstract

Human falls are one of the leading causes of fatal unintentional injuries worldwide. Falls result in a direct financial cost to health systems, and indirectly, to society’s productivity. Unsurprisingly, human fall detection and prevention is a major focus of health research. In this chapter, we present and evaluate several bidirectional long short-term memory (Bi-LSTM) models using a data set provided by the Challenge UP competition. The main goal of this study is to detect 12 human daily activities (six daily human activities, five falls, and one post-fall activity) derived from multi-modal data sources - wearable sensors, ambient sensors, and vision devices. Our proposed Bi-LSTM model leverages data from accelerometer and gyroscope sensors located at the ankle, right pocket, belt, and neck of the subject. We utilize a grid search technique to evaluate variations of the Bi-LSTM model and identify a configuration that presents the best results. The best Bi-LSTM model achieved good results for precision and f1-score, 43.30 and 38.50%, respectively.