Methods and Models of Intelligent Computer Vision for Identifying and Assessing a Person's Functional State in Conditions of Limited Visibility

Abstract

The study is devoted to improving the accuracy and stability of clustering, segmentation, and identification processes in digital images under conditions that impair visibility, such as fog, low light, and noise. The goal of the research is to integrate adaptive preprocessing methods, advanced descriptor analysis, and deep learning architectures to ensure stable and reliable operation of computer vision systems in challenging environments.

To achieve this goal, a sample of 350 images was formed based on our own photos and standardized COCO and CrowdHuman benchmarks. The experimental phase included modeling degradation (fog, noise, obscuration, low contrast, combined effects) and implementing adaptive preprocessing using complex algorithms such as γ-correction, adaptive histogram equalization with limited contrast, Dehazing, and filtering. The analytical phase used architectures for feature extraction and object localization: U-Net and Mask R-CNN for segmentation, while the YOLOv8 model was deployed for identification. Also, unsupervised learning methods, in particular K-Means and DBSCAN, were integrated for clustering the identified entities. This multi-stage workflow made it possible to evaluate how preprocessing affects the subsequent performance of deep learning models in various image degradation scenarios.

The results of the experiments confirmed a significant improvement in segmentation quality, which reached an intersection over union ratio of 0.95, detection with an mAP increase to 0.95, and clustering with a Silhouette Score increase to 0.79. As a result, the accuracy of human identification increased to 99%. In conclusion, it can be stated that the proposed approach ensures the robustness of the system to image degradation and can be used in real-world video surveillance conditions.