Performance of accelerometer-based gait classification using different signal representations

Inertial sensor technology has been used in diverse applications to assess impaired postural and activity patterns as well as to track the rehabilitation success of patients during the treatment. The approach used for the gait assessment has to be highly reliable and efficient, and this in turn depends on either the relevance of extracted gait features or the information content of the representation type of gait signals. A technically straightforward approach to assess the gait quality is to use gait signals and skip the feature extraction step. This work investigates the performance of gait classification based on Eigensteps and k-Nearest Neighbors and provides a comparison between accuracy rates achieved by different representations of gait signals. The classification accuracy of the Eigensteps method based on time-vector, continuous wavelet transform, and Fourier synchrosqueezed transform of the gait signals are compared. According to the analysis results, the Eigensteps method using Fourier synchrosqueezed transform representation delivers the best performance in terms of the accuracy rate (96.4%), and that with a smaller number of data sets. Therefore, the performance of the applied scheme highly depends on the proper representation of recorded gait signals. In addition, the effect of signal representation and machine learning method on accuracy performance is investigated using Naive Bayes as the classifier in the Eigensteps approach. The evaluation results show that the accuracy can be improved by almost 3.2% using the Naive Bayes method and the continuous wavelet transform representation, but at the cost of computational expense at training time. Accordingly, the method of gait classification based on Eigensteps and k-nearest neighbors can be used as a simple and efficient strategy for the extraction of relevant signal components (features). Better results in terms of accuracy can be achieved by using Naive Bayes learning and extracted features.