Feature Selection and Activity Recognition System Using a Single Triaxial Accelerometer (original) (raw)

Activity recognition is required in various applica-4 tions such as medical monitoring and rehabilitation. Previously 5 developed activity recognition systems utilizing triaxial accelerom-6 eters have provided mixed results, with subject-to-subject variabil-7 ity. This paper presents an accurate activity recognition system 8 utilizing a body worn wireless accelerometer, to be used in the real-9 life application of patient monitoring. The algorithm utilizes data 10 from a single, waist-mounted triaxial accelerometer to classify gait 11 events into six daily living activities and transitional events. The 12 accelerometer can be worn at any location around the circumfer-13 ence of the waist, thereby reducing user training. Feature selection 14 is performed using Relief-F and sequential forward floating search 15 (SFFS) from a range of previously published features, as well as 16 new features, are introduced in this paper. Relevant and robust fea-17 tures that are insensitive to the positioning of accelerometer around 18 the waist are selected. SFFS selected almost half the number of fea-19 tures in comparison to Relief-F and provided higher accuracy than 20 Relief-F. Activity classification is performed using Naïve Bayes and 21 k-nearest neighbor (k-NN) and the results are compared. Activity 22 recognition results on seven subjects with leave-one-person-out er-23 ror estimates show an overall accuracy of about 98% for both the 24 classifiers. Accuracy for each of the individual activity is also more 25 than 95%.