Improving the activity recognition using GMAF and transfer learning in post-stroke rehabilitation assessment

An important part of developing a performantassessment algorithm for post-stroke rehabilitation is to achievea high-precision activity recognition. Convolutional Neural Net-works (CNN) are known to give very accurate results, howeverthey require the data to be of a specific structure that differsfrom the sequential time-series format typically collected fromwearable sensors. In this paper, we describe models to improvethe activity recognition using the CNN classifier. At first bymodifying the Gramian angular field algorithm by encoding allthe sensors’ channels from a single time window into a single2D image allows to map the maximum activity characteristics.Feeding the resulting images to a simple 1D CNN classifierimproves the accuracy of the test data from 94% for thetraditional segmentation approach to 97.06%. Subsequently, weconvert the 2D images into the RGB format and use a 2D CNNclassifier. This results in increasing the test data accuracy to97.52%. Finally, we employ transfer learning with the popularVGG16 model to the RGB images, which yields to improvingthe accuracy further more to reach 98.53%.