Performances of Human Activity Recognition Classification Using LSTM and 1D-2D CNN

The Human Activity Recognition (HAR) System endeavors to proficiently categorize input data acquired from sensors affixed to the human body, meticulously measuring movement and assigning it to specific activity categories. While traditional techniques have long struggled to capture the intricacies of HAR, especially in instances with subtle microactivity sequences, the integration of machine learning within the realm of Artificial Intelligence has proven itself adept at handling such complex data-related challenges. This article considers the utilization of machine learning techniques to enhance human activities by examining datasets for HAR classification on realtime applications. A carefully curated dataset is chosen as the primary input for the processes of training, testing, and evaluation, facilitated by sophisticated models that include Long Short-Term Memory (LSTM), 1D Convolution Neural Network (CNN) for time series data, 1D CNN for the frequency domain, and 2D Image CNN models. These models are meticulously scrutinized, with their performance metrics meticulously compared to determine the most suitable HAR model for the task at hand. Through comprehensive simulations, the results affirm the supremacy of the 1D CNN frequency domain model in terms of performance with 93.28% accuracy, meanwhile the LSTM model showing the best results in term of the computation cost with the lowest time/step of 0.505ms/step. This outcome underscores the potential of advanced machine learning techniques in revolutionizing the field of HAR, particularly when confronted with nuanced and intricate movement sequences.