Experimental studies of a convolutional neural network for application in the navigation system of a mobile robot

Abstract

Presently, convolutional neural networks grow rapidly in the field of objects recognition and obstacle detection in the image. Recognition is one of the most important problems for development of modern mobile navigation systems. We propose to link up an artificial neural network as an additional or a backup solution for higher safety of these systems while the obstacles are being detected. The purpose of this work is to conduct experimental studies in order to identify the strengths and weaknesses of an artificial neural network for practical application in navigation systems for mobile robots. The largest possible detection accuracy and training time of the artificial neural network have been measured experimentally with the use of two methods. There has been evaluated the effect of the subset size of the epoch influencing the accuracy of detection. The dataset structure has been analyzed and the direct relationship between the accuracy of detection and the surface value of the recognized object was found. The mobile robot utilization and the frequency of processed frames from the video stream during the operation of the artificial neural network have been measured.