Fall and Normal Activity Classification via Multiple Wearable Sensors

A fall detection and classification system is crucial for reducing the severe consequences of falls, which account for the leading cause of accidents on construction sites. Wearable sensors are one of the technologies used to detect falls. Although much academic work has been dedicated to the study of this class of systems, little attention has been paid to the evaluation of simpler algorithms prior to training on complex ones. This study utilizes the open-source UP Fall Detection Dataset and proposes that effective data processing and simpler baseline models give better results for fall-direction classification. Several data-processing techniques like windowing and filtering are used prior to using simpler baseline models like Neural Network (NN), K-Nearest Neighbor (kNN), Support Vector Machine (SVM), Naïve Bayes (NB) and Discriminant Analysis (DA) Classifiers. It is also investigated how to minimize multisensor cost while achieving acceptable detection accuracy. Based on this robustness analysis, fine kNN and wide NN yield 99.5% accuracy for all five wearable sensors. In comparison, using the best of these sensors (belt and pocket) results in 99% accuracy, with accuracy of all 11 individual activities exceeding 93%. The findings of this study bode well for the development of real-world fall-prediction systems as they enable accurate fall direction identification.