GRADIENT TUNED DENSE TRANSFER LEARNING FOR DISABLED PERSON MOVEMENT ISSUE DETECTION WITH LONG FREQUENCY RFID

Abstract

Disabled individuals movement monitoring is an important part in healthcare and assistive technology for guaranteeing their safety,
independence, and overall well-being. The main aim is to detect abnormal movements, recognize daily activities, and identify fall events. There
are diverse wireless sensing technologies such as Wi-Fi, Radio Frequency Identification (RFID), and Bluetooth (ZigBee) are used to monitor the
movements of disabled individuals by detecting disturbances in electromagnetic waves. Among them, RFID is an automatic, non-contact
technology aimed to detect the movement of individuals based on radio frequency tags. Conventional deep learning approaches often addressing
the movement’s detection of disabled individuals, but, the accuracy faced major challenging issues. This paper proposes a novel model called
Gradient Tuned Dense Transfer Learning (GTDTL) model is developed. The developed GTDTL model employs deep transfer learning model
for accurate movement’s detection of disabled individuals with lesser time consumption. The overall structural design of transfer learning model
consists of two phase’s construction namely pre-trained and new model. In the beginning, transfer learning model constructs the pre-trained deep
learning model called DenseCNN model with many layers, including an input layer, hidden layers and an output layer. Initially, number of RFID
data samples is collected from the dataset and it given to the input layer. Consequently, data pre-processing is carried to handle missing data and
outlier’s removal. Followed by, the more pertinent feature selection process is carried out. Finally, activity recognition is performed with the
selected features. As a final phase of transfer learning, the fine tuning process is performed to optimize the error by employing stochastic sampling
squirrel search algorithm. As a final point, accurate and time efficient activity recognition results are obtained at the output layer with high
accuracy and minimal time consumption using RFID data samples. Experimental assessment of proposed GTDTL model is conducted using
various assessment metrics such as accuracy, precision, sensitivity, F1-score, specificity, and recognition time. The quantitatively analyzed results
expose that the proposed GTDTL attains higher accuracy in recognition with minimal time consumption as well as lesser error compared to
traditional deep learning methods.