Can Clouds Improve the Performance of Automated Human Detection in Aerial Images?

Abstract

The objective of this paper is to investigate the role of clouds in the effectiveness of automated human detection in aerial imagery acquired by unmanned aerial vehicles (UAVs). The automated processing is carried out with the nested k-means method applied to images taken in poor visibility caused by low-altitude clouds. Data were acquired during a field experiment carried out in the Izerskie Mountains (southwestern Poland). The fixed-wing UAV took RGB aerial photographs of terrain where persons simulated being lost in the wilderness. The UAV flights were conducted in the morning and around the noon, when clouds reduced clarity of aerial images. Subsequent UAV missions were performed in the afternoon and in the evening, when clouds had no impact on imagery. False hit rates ≥ 50 % correspond to clear imagery (8 of 9 non-cloudy cases). In contrast, images impacted by clouds reveal false hit rates ≤ 40 % (5 of 7 cloudy cases). Sensitivity analysis, carried out on a basis of artificially blurred imagery, confirms that reduced image clarity may improve automated human detection.