Wearable Sensor-Based Activity Recognition over Statistical Features Selection and MLP Approach

Supervising and examining the physical activities of different individuals by using wireless sensors is beneficial in favor of recognizing postural positioning and moving actions in the everyday ambiance. Accelerometers have been exercised in intelligent environments to identify human activities for many years. Using these accelerometer sensors and human activity recognition systems, people going through many sicknesses and illnesses can be proficiently scrutinized, and thus remedial medication can be dispensed in a well-timed approach. In this research paper, a naïve, and intricate technique founded on certain time domain features on the way to classify the physical activities is suggested. Feature extraction is a momentous step in HAR, and in this study, six statistical features were obtained from the preprocessed data. To further optimize the feature set, the Binary Grey Wolf Optimizer (GWO) was integrated. Further, a Multilayer Perceptron (MLP) was utilized for activity recognition to distinguish between the six activity classes. MLP has the advantage of finding non-linear patterns in data, the algorithm possesses powerful features of reducing overfitting in the data and good generality. It also offers resistance to noises usually experienced in the sensor data. Experimental results evaluated that the proposed framework achieved 87.14% recognition accuracy. This validates that the suggested framework can be used to observe physical activities and discriminate between healthy and unhealthy patients.