ANALYSIS OF OPENCV OBJECT TRACKING ALGORITHMS WHEN PROCESSING UAV FRAMES

Abstract

Purpose. Investigation and comparison of OpenCV object tracking algorithms effectiveness when processing frames with typical problems for images from UAV cameras. Conducting experiments to determine advantages and disadvantages, failure tendencies, time performance issues.

Methodology. Experimental videos were recorded in a FPV-drone flight simulator for seven typical problems. Correct behavior (ground truth) was defined manually using Python scripts with a bounding box input. Typical problems were determined: angle change, object position change, presence of identical objects, moving object, image noise, obstacle in front of an object, object size scaling. Python scripts were developed and used to automate calculation of four effectiveness criteria: bounding box match rate, center to center distance, success rate, time spent per frame.

Findings. Investigated effectiveness of Boosting, CSRT, KCF, MedianFlow, MIL, Mosse, TLD algorithms when processing frames with problems typical for images from UAV cameras. Carried out an analysis of failure tendencies, similarities between different algorithms, time performance issues. The experiments results were presented as tables with raw values, charts and human-friendly visualization of each algorithm’s work.

Originality. According to the results of the research, the tendencies of object tracking algorithms to work incorrectly when processing frames with problems typical for images taken from a UAV camera were determined.

Practical value. The results can be used for prioritization of certain object tracking algorithms usage or combination, depending on the field of application.