Human Gait Patterns Classification based on MEMS Data using Unsupervised and Supervised Learning Algorithms

With the proliferation of smartphones and wearable devices having Micro-Electro-Mechanical Systems (MEMS) sensors built in, data samples of linear acceleration and angular velocity can be collected almost anytime anywhere. These motion data can be used to identify various types of human motions and to detect the anomaly of individuals movements. This work presents attempts to use the unsupervised Affinity Propagation (AP) clustering algorithm and the supervised Support Vector Machine (SVM) classification algorithm to identify four types of human gait motions: walking, jogging, climbing upstairs and downstairs. Features of three-dimensional linear acceleration that can enable the algorithms to identify these motion types correctly were selected by analyzing the variation of the feature values among different motion types. Efficacy of Affinity Propagation (AP), Linear and Non-linear Support Vector Machine (SVM) algorithms were also studied by comparing their ratios of correct, false positive, false negative and F1 score classification. This preliminary study demonstrated Linear SVM achieved the best performance, followed by Affinity Propagation. Quite surprisingly, Non-linear SVM appeared to be inferior to the other two algorithms.