CHOOSING A NEURAL NETWORK ARCHITECTURE FOR A VEHICLE AUTOPILOT

Abstract

In the paper the task of choosing a neural network architecture for creating an autopilot is considered. An autopilot was created for a virtual vehicle that can move along a defined route and respond to various traffic lights. The selected architecture, namely a convolutional neural network, has high efficiency in the task of image recognition. Autopilot consists of two convolutional neural networks, one of which recognizes the driving route, the other recognizes traffic light signals. Due to the large amount of noise, the traffic light recognition photos were processed to enhance the red channel and null the green and blue, which helped in the recognition of red and yellow colors. As an environment for training neural networks and testing the performance of the autopilot in general, a two-dimensional game with a top view was created. The autopilot model showed almost 100% accuracy in recognizing the route and traffic lights in the test environment. The positive test result showed that the autopilot can perform control in a simple environment and this gives the opportunity to complicate the operating environment. The proposed autopilot uses only images to navigate in space, which distinguishes it from the other existing autopilots and, in particular, makes it cheaper. The relevance of this work is based on studies of the increase in the number of vehicles and harmful emissions into the atmosphere in the future. In the paper the literary sources are analyzed, the rationale for choosing a neural network architecture is explained, the software implementation is described, the results of testing are shown, and the possible direction of development of this topic is indicated in the conclusions.