Modeling and implementation of the process of determining road objects using the RetinaNet convolutional network apparatus

Abstract

Modeling and implementation of the process of determining road objects using the RetinaNet convolutional network apparatus

Evsina V.A., Shirobokova S.N., Zhzhonov V.A., Evsin V.A.

Incoming article date: 29.05.2022

This article discusses the problems of constructing convolutional neural networks for determining road objects. The general relevance and formulation of the problem of determining road objects is presented. The rationale for the use of artificial neural networks for determining road objects has been formed. The Retinanet network architecture is used as the main architecture of an artificial neural network for determining road objects. The general concept of this architecture and the main subnets are visualized. Error functions for the main subnets of the Retina net network are described. The design description of algorithms for constructing data annotation for training an artificial neural network, as well as algorithms for constructing the neural network architecture of classification, regression and feature pyramid is given. The dynamics of changes in the general error function when determining road objects is determined. The result of training an artificial neural network is presented.

Keywords: convolutional neural networks, classification, regression, convolutional neural networks, deep learning, big data, mathematical modeling, computer science, RetinaNet architecture