Hybrid Deep Learning Approaches for Human Activity Recognition and Postural Transitions Using Mobile Device Sensors

Human activities recognition (HAR) has grown in popularity over the last decade by using technologies such as mobile phones, smartphones, and video cameras. In fact, this growth has been propelled by using artificial intelligence-based techniques to improve the rate of recognition. The aim of the paper is the development of a robust model to recognize twelve human activities data across individuals aged 19–48 years. We have utilized several machine and deep learning techniques including Naive Bayes (NB), stochastic gradient descent (SGD), XGBoost, convolutional neural network (CNN), CatBoost, LGB classifier, K-nearest neighbor (KNN), long short-term memory (LSTM), CNN-GRU, gated recurrent unit (GRU), and CNN-LSTM. Three static postures and three dynamic actions are investigated, including various transitions. The methodology begins with preprocessing the dataset for cleaning, user counts and frequency analysis. The t-stochastic neighbour embedding approach is incorporated for further exploration and visualization with high dimensions. We employ Principal component analysis (PCA) to compute several groupings based on different human activities in three dimensions. Among other machine learning models, SGD exhibits the highest accuracy at 88.6% with a loss rate of 0.79. On the deep learning front, GRU model achieved the maximum accuracy of 84.50% and the lowest loss score of 0.49. These findings demonstrate the efficacy of both conventional machine learning and deep learning techniques in the recognition of HAR, offering promising directions for future study and practical implementation.