Deep-learning-based Multi-behavior Classification of Animals for Efficient Health and Welfare Monitoring.

With the development of sensor technologies, sensors have become increasingly embedded in various fields, becoming an indispensable part of our daily lives, research, and work. Notably, in ethology, surveillance cameras, a type of optical sensor, are extensively used alongside machine learning to analyze animal behaviors. However, simply feeding vast amounts of sensor data into servers for processing is neither efficient nor sustainable. In line with the prevailing trend towards edge computing, it is becoming increasingly important to process and integrate the captured sensor information directly within the sensor itself. While we have not fully achieved this, the application of deep learning methods to facilitate efficient and rapid processing with low computational demands is a necessary progression. In our study, we used a method for outdoor animal behavior analysis using multi-target classification, taking advantage of the potential efficiency gains provided by deep learning. We focus on a polar bear's behaviors captured by an IoT-enabled surveillance camera in a zoo. The image data are first analyzed by using an object detection model to provide location sequences, movement speed, and coordinates of video frames, representing the animal's state. Using these sensor data, we developed a classification model that accurately classifies multiple behaviors. The detection of these behaviors, including stereotypical behavior, illustrates the potential of our system to comprehensively monitor the animal health status. Our method achieved accurate detection [98.3% average precision (AP) 50] and multi-behavior recognition (accuracy of 89.5%), while maintaining robustness against outdoor noise. [ABSTRACT FROM AUTHOR]