Оbject recognition system based on the Yolo model and database formation

Abstract

A system for recognizing objects that are captured in real time on a video camera in a noisy environment that changes to the surrounding conditions has been built. The method of filling the database for mobile military objects was studied. For object recognition, the YOLO v8 neural network is used, which allows you to track moving and identify objects that fall into the video from the video camera. This neural network makes it possible to track objects with a change in scale, during movement with obstacles. It has been analyzed that the recognition of objects is carried out on the basis of contour analysis, comparison with a template and detection and matching of features. Artificial intelligence methods based on YOLO v8 were used to recognize military equipment. Trained for different YOLO models using Adam W, Adam, SGD optimizers and 512x512, 640x640, 1024x1024 px image resolution. Improved object recognition is achieved by analyzing contours, comparing patterns, and comparing entered special points. Different image resolutions and optimizers have shown different effects on model performance, and standard evaluation metrics do not provide the most accurate view. The most effective optimizer is gradient descent (SGD), which has shown the best accuracy for combat vehicle recognition. The gradient is usually considered as the sum of the gradients caused by each training element and is used to adjust the model parameters. As a result of the development of the system, indicators with recognition accuracy (accuracy) of 92%, F1-estimate (F1 score) – 89%, average indicator of accuracy (mAP) – 90% were formed. A method of filling the data set and creating a classifier is proposed. A model of combat vehicle recognition was built. Graphs, results of recognition of moving objects in the Yolo8 x neural network are presented.