AERed: An Autoencoder-Decoder Dimensionality Reduction Method for Wearable-Based Human Activity Recognition

Human activity recognition (HAR) is the process of identifying a person's activities accurately. It is possible by placing sensors on the subject's body and obtaining data from a variety of high-dimensional physiological signals. Recently, sensors such as an accelerometer and gyroscopes have been incorporated directly into wearable devices, making activity identification fairly straightforward. High dimensionality makes it necessary to use an optimization method that might reduce the number of features used in the dataset while taking less time to make activity recognition successful in wearable devices with limited battery life. In this study, we propose the autoencoder reduction method (AERed), a dimensionality reduction (DR) technique based on the symmetric design of a typical autoencoder (AE). With the new structure, there are fewer weights that must be tweaked, which lowers the computational cost. This work makes use of a public-domain dataset from the UCI repository. The dataset's features were reduced from 561 to 256 using the AERed method. The reduced features are classified using a random Bayesian filter support vector machine classifier and reported 95.95% F1-score. This method also consumed a lot less time than baseline methods at the DR stage. The proposed method is validated by performing parameter sensitivity analysis, complexity analysis, and visualization performance analysis.