Performance Evaluation of Hybrid Intelligent Models for Human Activity Recognition

Information on human activities is critical for the general wellbeing of people and health management. In recent times, machine learning has gained popularity in human activity recognition modelling. However, the machine learning algorithms suffer from the curse of dimensionality due to the several recorded features in the data collected and utilized in the human activity recognition. To overcome the drawback, dimensionality reduction is utilized to reduce the dimensionality of the data and extract the most relevant features. However, in machine learning modeling, the type of dimensionality reduction method utilized is a critical factor that affects the machine learning performance. This study evaluates and compares the impact of four dimensionality reduction methods namely, principal component analysis (PCA), Kernel PCA (KPCA), Nonlinear PCA (NLPCA), and Independent Component Analysis (ICA) in human activity recognition using machine learning. First, the discrete wavelet transform (DWT) was applied as a preprocessing step to decompose the time series data into different frequency bands. Then, the PCA, KPCA, NLPCA and ICA were applied to each frequency band. Finally, Support Vector Machine (SVM) as the classification technique was selected and used to classify the activities. The performance of each method was evaluated using a benchmark dataset: UCI-HAR Postural Transitions dataset. The dataset consists of 561 features extracted from smartphone accelerometers and gyroscope signals from smartphones worn by human subjects. The results show that the DWT-ICA-SVM method achieved the highest classification accuracy, precision, recall, and F1-score results on both datasets, followed by DWT-NLPCA-SVM, DWT-KPCA-SVM, and DWT-PCA-SVM. The performance of the foregoing hybrid models with the traditional PCA-SVM method was compared, and the results show the hybrid models superior performance. The paper also investigated the effect of different decomposition levels of the DWT on the classification performance and found that the decomposition level of 4 or 5 provided the best performance for both datasets. Overall, the results suggest that the hybrid dimensionality reduction methods, particularly DWT-ICA-SVM, can effectively reduce the dimensionality of the data and improve the classification accuracy for human activity recognition tasks. The findings of this study are useful for researchers and practitioners working on activity recognition systems using wearable sensors.