Influence of Ultralytics Model Type and Batch Size on People Detection Accuracy in Aerial Photographs

The increasing interest in the use of unmanned aerial vehicles for monitoring and security makes the task of accurate and reliable detection of people in aerial photographs relevant. However, the specific angle, scale, and background of such images create difficulties for computer vision algorithms. This paper is devoted to the study of the configuration of YOLOv11 models for solving the problem of detecting people in images obtained from drones. Despite the obvious focus of the topic, the article focuses contains on an empirical comparison of the performance of different model versions of the models, the influence of the batch size influence, as well as the interaction of architectural features and training conditions. As part of the study, experiments were conducted with various pre-trained versions of YOLOv11 models, and the influence of the «batch size» on the generalization ability of the models was investigated. The results obtained can serve as a basis for implementation in real-time image analysis systems.