Problems of modern methods of three-dimensional photogrammetry

Abstract. Technologies of three-dimensional photogrammetry, which is one of the methods for creating computer-generated 3D models of objects, have a wide range of scientific and practical applications in fields such as manufacturing, construction, architecture, geodesy, and medicine. However, the primary challenges of photogrammetric methods are related to their high labor intensity. This work explores the fundamentals of the photogrammetric method for obtaining three-dimensional models of objects, analyzing its key drawbacks and limitations associated with the need to identify key elements across numerous images of an object taken from different angles and then align them accordingly. One of the most effective image comparison methods that can be used in photogrammetric processing to identify key elements in object images is the scale-invariant feature transformation (SIFT) algorithm. This paper analyzes the main stages of implementing this algorithm and provides an overview of several modifications that enhance its performance by eliminating redundant key points and reducing the dimensionality of descriptors used to distinguish each key point from others. Further improvements in performance and reduction of errors in 3D model creation can be achieved by removing frames or images that do not contain common features due to sharp changes in shooting angle or specific object characteristics in the preliminary stage. To accomplish this, the use of a neural network is proposed to analyze the similarity between each pair of sequentially taken images, which are preprocessed into binary form. Removing such images not only saves time by avoiding unnecessary searches for key points on an object’s image but also reduces the likelihood of obtaining erroneous matches between key points on different images of the object.