Research on machine learning methods for detecting objects in difficult shooting conditions

Abstract

The subject matter of the article is research into machine learning methods for object detection in images and videos under complex urban conditions, particularly under poor lighting, the presence of precipitation, high scene complexity, and limited computational resources. The goal of this research is to identify the most effective deep learning models based on convolutional neural networks for object detection tasks under challenging imaging conditions, considering the practical requirements for accuracy and processing speed. The tasks to be solved are: analysis of object detectors (YOLO v8–11, DETR, SSD, Mask R-CNN, Faster R-CNN, RetinaNet); preparation of a dataset with real weather conditions and pedestrian environments in Ukraine; experimental evaluation of selected detectors using the metrics mAP@0.5, mAP@.5:.95, Recall, Precision, IoU, FPS, and F1-Score; and analysis of the obtained results. The methods used are: convolutional neural networks, automated image annotation, comparative analysis of quality metrics (F1-score, mAP@0.5:.95, Precision, Recall, IoU, FPS), and manual correction of annotations. The following results were obtained: the YOLOv10-m and YOLOv11-m models demonstrated the best quality indicators under conditions of limited visibility and varying lighting. The YOLOv11-m model was the most balanced in terms of accuracy and speed across all tested conditions - snow, rain, and sunshine. YOLOv11-m is recommended as the baseline model for implementation in real-time systems, particularly in intelligent assistants for people with visual impairments. Conclusions: The scientific novelty of the results obtained is as follows: 1) a comprehensive evaluation of modern deep learning architectures for object detection (YOLOv8–v11, Faster R-CNN, SSD, Mask R-CNN, DETR, RetinaNet) was carried out under non-laboratory conditions, including real weather scenarios such as snow, rain, and poor lighting, which are typical for urban environments in Eastern Europe; 2) the software tool for automated model evaluation was developed, allowing simultaneous testing of multiple architectures and visualization of performance metrics (F1-score, mAP@0.5, mAP@.5:.95, IoU, Precision, Recall, FPS) with support for manual annotation correction and comparative model analysis; 3) it was experimentally established that the YOLOv11-m model demonstrates the best balance of accuracy and inference speed across various complex imaging conditions, justifying its recommendation as a baseline model for real-time vision-based assistive systems.