ResNet-based Deep Neural Network using Transfer Learning for Animal Activity Recognition

Animal activity recognition (AAR) is a challenging and complex task for monitoring animal behavior to determine their health conditions and physical responses. To address this requirement, AAR research aims to develop an internet of things (IoT) platform that anticipates animal behavior based primarily on sensor data collected from wearable devices connected to animals. Because of its superior effectiveness on multi-dimensional sensor data, deep learning (DL) currently dominates sensor-based AAR applications. Nevertheless, DL techniques require significantly more training time than other learning-based strategies. This study employs transfer learning (TL) to generalize the DL model to overcome the issue of training time, indicating significant promise for future research on animal activity identification. Furthermore, we apply the TL to allow the reusability of a pre-trained DL model on entirely new data without the time-consuming operations of model training and data labeling. We conducted experiments on a publicly available AAR benchmark dataset (SAR) utilizing deep residual networks with automatically learned features. Utilizing source data, our technique achieved an overall accuracy of 96.46% (dataset collected by the first sensor) and 98.40% (dataset collected by the second sensor) based on evaluating findings.