: ..
:
: 114
:
: 2025
: .. // . - 2025. - . 114. - .307-344.
: , , ,
(.): mobile robot localaization, scene recognition, image vectorization, graph theory
: . , ( ). , . , , . , , . , . , KITTI-360 . , , , . .
(.): The paper is devoted to the problem of localization of mobile robots based on visual semantic data. The central element of such a problem is scene recognition task, i.e. searching for a correspondence between observed objects and objects on a semantic map. Paper proposes two methods based on definition of geometric features in the observed scene and searching for them on the map using various graph approaches. The proposed method for determining the relationships between objects, used in both methods, allows taking into account the errors in estimating distances by onboard sensors. In addition to using geometric features, the paper also considers the use of neural network models forming a feature vector based on an image, determining their visual similarity. Visual similarity is used to evaluate and sort the results obtained by the proposed methods based on graph approaches. In addition, the open KITTI-360 dataset was modified to evaluate the accuracy of solving scene recognition problems. Experiments on the resulting dataset demonstrated that the proposed approach, which combines geometric features and visual similarity, significantly increases the accuracy of the considered scene recognition methods. Based on the results of the experiments, some recommendations were formulated for the use of these approaches in practice.
PDF
: 71, : 22, : 7.