Object Detection in Movies – Case Study

This work analyzed video datasets distinctively characterized by varying properties such as duration, fps, and total frames. These datasets also displayed differences in lighting intensities, image shifting, and rotations. The study utilized the YOLOv4 architecture integrated with the Darknet framework for object detection purposes. This combination efficiently extracted bounding box coordinates, defining the Region of Interest (ROI) of the Primary Display Panel in the plane simulator cockpit. Within this ROI, the research employed the SIFT and FAST algorithms, with the Manhattan distance, to determine corner points. The results revealed the FAST algorithm’s ability to detect more key points quicker than SIFT. On the other hand, the latter one turned out to be slightly better when the average Intersection over Union (IoU) is considered. The study’s outcomes can serve for future analysis of pilots’ behavior.