Chelonia mydas detection and image extraction from noisy field recordings

Abstract

Wildlife videography is an essential data collection method for conducting research on animals. The video recording process of an animal like the Chelonia Mydas turtle in its natural habitat requires the setting up of special camera traps or by performing complex camera movement to capture the animal in frame whilst the cameraman maneuvers over uneven terrain while filming. The result is hours of footage that only have the presence of the intended subject in it for seconds whilst the rest is background footage; or noisy and blurry footage that has only several usable frames among thousands of noisy and unusable ones. This presents a problem that deep learning models can help to assist, especially in detecting a wildlife subject and extracting usable data from hours of noise and background footage. This paper proposes the use of machine learning models to detect and extract wildlife images of Chelonia Mydas turtles to help prune through hundreds and thousands of frames from several video footages. Our paper shows that utilizing a custom model with various confidence scores can label and crop out images in noisy field video recordings of Chelonia Mydas turtles with up to 99.89% of output images correctly cropped and labeled.