Machine Vision Module for Object Detection in Images and Video Streams

This work addresses the topical issue of developing and applying intelligent computer vision methods for the automated detection and segmentation of protective equipment objects in images and video streams. The aim of the work is to identify objects of a specified type in images captured by a video camera, as well as to develop a detection model capable of effectively identifying and localising these objects under various lighting conditions, scales and perspectives. The article examines the data preparation process, in particular the application of augmentation methods to improve the representativeness of the sample, and performs a comparative analysis of the parameters, performance and results of the SAM and YOLO models. The results of experimental studies are presented, demonstrating the positive impact of increasing the volume and diversity of the dataset on data balance and the generalisation ability of computer vision models. The proposed approaches to training separate models for segmentation and classification tasks have proven their effectiveness in the context of automated image processing. In this work, a specialised dataset of protective equipment items was created to address the image segmentation task. During the model pre-training phase, data augmentation techniques—including mirroring, rotation, scaling and brightness adjustment—were applied to this dataset, which significantly increased the diversity of the training examples. Increasing the size of the dataset ensured a more balanced representation of the data and improved the model’s generalisation ability. The results obtained in this work confirm the feasibility and effectiveness of the authors’ proposed approach to this problem, namely the separate training of models for image segmentation and classification. Prospects for further research include expanding the dataset, optimising the computational complexity of the models, and investigating their application in real-time for video analytics.