Feature Selection and Activity Recognition System Using a Single Triaxial Accelerometer

• Published in: IEEE transactions on biomedical engineering Vol. 61; no. 6; pp. 1780 - 1786
• Main Authors: Gupta, Piyush, Dallas, Tim
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.06.2014; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Acceleration; Accelerometer; Accelerometers; Accelerometry - instrumentation; Accuracy; Activities of Daily Living - classification; activity recognition; Adult; Algorithms; Bayes Theorem; Classification; detrended fluctuation analysis (DFA); error estimates; feature selection; Female; Fluctuations; Humans; k -nearest neighbor ( k -NN) classifier; leave-one-person-out (LOO) error; Legged locomotion; Male; Monitoring; Monitoring, Ambulatory - instrumentation; Monitoring, Ambulatory - methods; Motor Activity - physiology; Naïve Bayes classifier; Neural Networks (Computer); Patients; Position (location); Recognition; Relief-F algorithm; Searching; sequential forward floating search (SFFS) wrapper algorithm; Time series analysis; Young Adult
• ISSN: 0018-9294; 1558-2531; 1558-2531
• DOI: 10.1109/TBME.2014.2307069

Activity recognition is required in various applications such as medical monitoring and rehabilitation. Previously developed activity recognition systems utilizing triaxial accelerometers have provided mixed results, with subject-to-subject variability. This paper presents an accurate activity recognition system utilizing a body worn wireless accelerometer, to be used in the real-life application of patient monitoring. The algorithm utilizes data from a single, waist-mounted triaxial accelerometer to classify gait events into six daily living activities and transitional events. The accelerometer can be worn at any location around the circumference of the waist, thereby reducing user training. Feature selection is performed using Relief-F and sequential forward floating search (SFFS) from a range of previously published features, as well as new features introduced in this paper. Relevant and robust features that are insensitive to the positioning of accelerometer around the waist are selected. SFFS selected almost half the number of features in comparison to Relief-F and provided higher accuracy than Relief-F. Activity classification is performed using Naïve Bayes and k-nearest neighbor (k-NN) and the results are compared. Activity recognition results on seven subjects with leave-one-person-out error estimates show an overall accuracy of about 98% for both the classifiers. Accuracy for each of the individual activity is also more than 95%.