Using trajectory features for upper limb action recognition

Abstract

There is growing interest in using low-cost wearable sensors to model limb movement in applications such as stroke rehabilitation and physiotherapy. This paper presents an algorithm for the detection and classification of arm motion in time series collected by wearable inertial sensors. High level arm trajectory features are obtained from raw sensor data using a sensor orientation tracking algorithm and an arm model. The features are then used in a clustering-based classifier. In the classifier training stage, features are clustered using the k-means algorithm, and a histogram of “key poses” is generated from the clustering as a template for each class. In the recognition stage, new data are segmented and matched to the templates. Experiments on human subjects show that by using trajectory features in the proposed approach, we can achieve higher accuracy than a range of benchmark non-temporal classifiers.