Outlier based Human Fall Detection using One-Class Classification

Abstract

Objective: This work aims to develop a human fall detection method that is trained using data of routine movement of people only collected from accelerometer sensor to stay away from irregular fall detection model. This work also aims to analyze the effect of calculated features on the fall detection model. Methodology: In the proposed method, The fall detection model is built using one-class classification. At first, data of accelerometer sensor in three directions has been used to detect the fall events. Five feature vectors i.e. resultant, variance, standard deviation, Root Mean square, and Euclidean Norm have been calculated. These four features along with one class SVM have been used to build fall detection model in ways (1) Using only resultant features only (2) Using all calculated features. This model was trained using only activities of daily living (ADL) and tested on both daily living activities and fall activities. Findings: It is found that when the model was built using all calculated features, the sensitivity was 100% and specificity was 94.92% when mobile is in the pocket and; sensitivity was 100% and specificity was 93.61% when mobile is in handbag which is better than when the model was built using resultant features only. We have also studied the effect of individual features on this fall detection model and it is found that variance played a very important role to classify fall activities and ADL activities. Novelty -The proposed fall detection method is built using one-class classification so that the proposed model will be reliable to detect falls in real life. In the proposed work, the effect of features on the fall detection model is also analyzed.