An approach to classification and evaluation of geometric parameters of complex surfaces based on rgb-d images

Abstract

When developing control systems for formations of modular robotic systems (RS), the task of classification and evaluation of geometric parameters of complex surfaces is relevant. This approach is based on the use of the Intel RealSense D435 depth camera, which evaluates the distances from the camera lens to the scene sections in metric units, the HRNet pre-trained segmentation neural network model that selects the target area of a complex surface in the image, as well as several analytical algorithms that evaluate key surface parameters and classify complex surfaces representing individual terrain sections based on their geometric characteristics. The approbation and evaluation of the quality of the proposed approach was carried out based on a test data set including 4,500 images. This data set includes images of scenes — rooms containing at least one surface on which the PC is to move. The average accuracy indicators (accuracy, recall, precision) of the classification by angle of inclination and by type of elevation differences for the corresponding data set were {0.74, 0.68, 0.67} and {0.76, 0.68, 0.74}, respectively, while the average proportion of correctly classified surfaces in both classifications was 62.6%. According to the results of the testing, the proposed solution makes it possible to successfully classify various surfaces by the type of angle of inclination and by the type of height differences, as well as to evaluate the geometric parameters of complex surfaces using RGB-D images.