Optimal Sensor Configuration for Activity Recognition during Whole-body Exercises

Advances in wearable devices with inertial measurement units (IMUs) for the detection of different motor activities and monitoring training tasks have important applications in tele-rehabilitation. These technologies can play an effective role in improving the quality of life for people with progressive movement disorders such as Parkinson's disease (PD). Considering cost, simplicity, and practicality, a smaller and more efficient number of IMUs that can accurately recognize the type of movement is preferable. The purpose of the current study was to design an affordable and accurate wearable device with IMUs to detect thirty four different motor activities in a customized training program called LSVT-BIG, which is usually used for people with PD. Nine neurologically healthy individuals performed all 34 tasks. The collected data were processed in windows of 2.5 seconds. Eight features in time and frequency domains and discrete wavelet transforms were calculated. Dimension reduction was performed using the PCA algorithm. NM, RBF, SVM, and k-NN<;span style="font-size: 11.25px;"> <;/span>classifiers were then trained and used to recognize the activity. A genetic algorithm was utilized to decide which sensors and signals took part in the classification to produce the best accuracy. Our results showed that the four sensors installed on the left shank, right thigh, left forearm, and right arm provided the optimal number and arrangement to achieve a precision of 94.3% and sensitivity of 93.4% using NM classification. Also, an adaptation algorithm was utilized in order to maintain the quality of recognition for new users.