Reconstruction of a video sequence based on a geometric model using a video descriptor

Abstract

In recent years, the problem of reconstructing visual information has been a hot topic in machine vision and remote sensing using UAVs. The scope of the UAV is quite wide. They can monitor the traffic situation, both citywide and in remote areas, control the fire situation in forests or flood waters in the regions, deliver goods in a short time, and much more. Many different intelligent and widely used subsurface reconstruction methods are used for accident-free driving of UAVs in an urban environment with visual guidance. The main visualization problem in the practical applications described above are shadows from buildings, overlapping objects, sun glare, reflections from the surface, which require high-quality image reconstruction. The problem of classifying a dataset of the underlying surface is of great importance for the correct reconstruction of images. The article proposes an algorithm for reconstructing a video sequence based on a geometric model using a video descriptor to classify video into a static background and moving objects. To study the effectiveness of the new method, a qualitative analysis of the restored underlying surface of the urban environment was carried out. The subject of the study is the existing methods and algorithms for constructing descriptors for image classification, as well as methods for reconstructing dynamic images. The object of the study is a set of test video sequences of the terrain map obtained using a UAV. The result of the study is the development of an algorithm for constructing a global video descriptor for object classification and further reconstruction of the underlying surface based on the descriptor. The novelty of the work is an algorithm that allows you to reconstruct a map of the underlying surface based on the construction of an object classification descriptor. The results of calculating the root-mean-square error for evaluating the proposed method for processing the considered map of the area are shown. When conducting a comparative analysis, it was revealed that almost all methods for recovering a video sequence have disadvantages, which are complicated by a number of reasons: the presence of a non-stationary background: objects located at different distances from the UAV camera can be moving; the difficulty of distinguishing between objects and the background in the case of slow movement of foreground objects; lighting conditions, etc. Therefore, this article proposes a method that allows you to overcome these difficulties and more accurately restore the map of the area.