A Framework for Human Action Recognition Using Control Features Fusion and Weighted Entropy Based Feature Selection

Human behaviour research, human-computer interface, and unavoidable processing all now heavily examine human activity recognition (HAR). Recently, deep learning (DL)-based algorithms have been successfully used to predict various human movements using time-series data from cell phones and wearable sensors. Although DL-based techniques fared exceptionally well at activity detection, they still have trouble handling time series data. Time-series data still has a few problems, such as difficulty in feature extraction, highly unbalanced data, etc. Additionally, the majority of HAR techniques include manual feature orchestration. Bidirectional long short-term memory (BiLSTM) and convolutional neural network (CNN) are combined in this article. The proposed multibranch CNN-BiLSTM network performs redid feature extraction from the unprocessed sensor input. The model is prepared for learning area characteristics as well as long-term conditions in successive data thanks to the use of CNN and BiLSTM. In this study, we will fine-tune the weight of Convolutional Neural Network (CNN) using a novel self-improved Seagull optimisation algorithm (SI-SOA) technique in order to increase the human activity categorization accuracy. The proposed SI-SOA approach will be conceptual amalgamation of the standard Seagull optimization algorithm (SOA).