Frame Elimination for Recordings of Bat Flight Kinematics from a High-Speed Video Array

Bat flight is complex because they have the ability to change the shape of their wings while in flight to perform manoeuvres. To accurately capture the subject's complexity, we have set up a synchronized array of
$50\text{high}$
-speed cameras. This array is capable of capturing over 50,000 images per second. The video array is designed to cover a large imaging volume that allows bats to perform full-flight manoeuvres. As a result, many video frames do not show the entire bat. Hence, this study uses deep learning techniques to develop an accurate frame elimination process. Until the controlled environment is set up, we used simulated data created from the blender. For the frame elimination purpose, images needed to be classified based on the location of the bat, where it could be fully or partially presented within the frame or maybe not presented. We used different deep learning techniques, including the ordinary CNN, pretrained models, and object detection model - YOLOv5, to achieve this target. Among all the models, we noted that YOLOv5 gave the best accurate unwanted frame elimination results, where the mAP0.5 value is 0.99 .