Real-time Detection of Human Falls in Progress: Machine Learning Approach
Gürsel Serpen, Rakibul Hasan Khan · Procedia Computer Science · 2018
The research reported in this study investigates the use of machine learning algorithms that receive real time sensory input from three-axis accelerometers and gyroscopes strategically placed on a human body to detect a fall-in-progress early enough and prior to the impact to facilitate inflation of a body airbag. A publicly available dataset that entails 30+ subjects and 4+ fall types including fall forward, backward, left and right was employed in the study. A set of 12 attributes were defined and provided as inputs to machine learning classifiers. Two machine learning classifiers were trained and tested with the dataset through three-fold cross validation to determine the feasibility of the proposed approach. Specifically, Random Forest and Support Vector Machine classification algorithms were employed through ensemble-in-time and their performances were profiled. Both classifiers offered promising performances as they were able to detect entry into the so-called the “free fall phase” for most combinations of the test subjects and fall types considered. In most cases the algorithms were able to deliver a fall detection decision with at least 100 milliseconds remaining before the impact, which is sufficient to inflate a body airbag.