Anil Kumar Chengali

The automatic prediction of daily human activities like walking, running, cooking, and office work is called Human Activity Recognition (HAR). The medical industry can greatly benefit from it, especially those working with the elderly, personal health care aides, those keeping patient records for reference in the future, etc. A HAR system can take (a) video or still images of people doing things, or (b) data showing the human body's motions as they do those things gathered from sensors in smart devices (accelerometers, gyroscopes, etc.), smart homes, eldercare, and the Internet of Things (IoT). The suggested HAR applications heavily rely on the latest developments in AI approaches, such as optimisation algorithms from Deep Learning (DL) and Swarm Intelligence (SI). Here, we use open-source data from wearable sensors to construct a reliable HAR system that combines DL and SI. A method for light feature extraction called Residual Bidirectional Long Short-Term Memory (Res-BiLSTM) has been developed. Based on the Marine Predator Algorithm (MPA), we presented novel feature selection approaches to choose the best collection of features. Using three publicly available HAR datasets from the UCI machine learning repository, we assess the performance of the suggested model. We evaluate the suggested model against different DL architectures that have recently been suggested as solutions to the HAR problem. The proposed model surpasses other state-of-the-art approaches in terms of accuracy 96.92%, precision of 95.45%, recall of 94.07%, and F1 score of 96.15% on all three datasets. The suggested approach outperforms several reported results in robustness and activity detection. As well as adapting activity aspects, it has fewer parameters and improved accuracy.