An Efficient and Optimized CNN-LSTM Framework for Complex Human Activity Recognition System Using Surface EMG Physiological Sensors and Feature Engineering
Repuri Mohan Vamsi, Neha Adapa, Dinesh Yelamanchili, Nurul Amin Choudhury, Badal Soni · 2024
Human Activity Recognition is an interdisciplinary domain that advances algorithms and frameworks that automatically discern and interpret human activities using sensor data. Recognising complex human activities is a laborious task due to the necessity to capture prolonged dependencies and derive effective features from unprocessed sensor data. This paper introduces an efficient and optimized custom hybrid deep learning (CNN-LSTM) framework for recognising complex human activities using raw physiological sensor data. The proposed framework leverages physiological surface Electromyography sensors and employs refined feature learning techniques for enhanced performance. An innovative approach has been formulated involving the utilization of convolutional neural networks (CNN) and subsequent Long-short-term memory (LSTM) layers to handle spatial and temporal dependencies. A robust data pre-processing pipeline has been developed to make the raw sensor data usable for classification by identifying optimal feature sets using a wrapper-based feature selection method. The proposed model achieves the highest performance accuracy of 87.18% and average performance accuracy of 87% in optimized computational time than the baseline and the benchmark models. Further, the training and validation loss were also optimized compared to the baseline model.