Determination of the spatial orientation of objects from their images based on neural networks

Klyachin Vladimir Aleksandrovich, Peretyat’ko Nikita Olegovich

One of the tasks of 3D reconstruction of objects from their images is the problem of their spatial orientation relative to the coordinate system associated with the surveillance camera. A geometrically similar problem belongs to the class of inverse problems of projective geometry. However, the explicit solution of this problem is possible only in some special cases. Nevertheless, we set the task of determining the spatial orientation of buildings and structures from their photograph. The aim of the work is to solve the above problem by numerical methods, based on the presence in the photographs of buildings and structures of characteristic straight-line segments, which are the projection of lines running along the corners of the building. Geometrically, it can be justified that the presence of these segments is sufficient to solve the problem. Methods. Solving the problem is based on the use of a trained neural network. Novelty. Elements of the novelty of the presented solution are the use of image preprocessing for training by a neural network. Pre-processing consists in preliminary application of the Hough transform to select straight-line segments. Results. As shown by test calculations, the application of the proposed approach approximately doubles the accuracy of calculating the angle between the direction of the camera and the side plane of the building. Practical relevance. The practical application of the research results is to create a tool that allows you to automatically determine the relative position of objects on the photograph with their subsequent 3D reconstruction.

Keywords:projective geometry; spatial orientation; neural network.