Search

Search Results

• ESTIMATION OF THE SPATIAL POSITION OF AN ON-BOARD CAMERA BY COMPARING AERIAL IMAGES AND SATELLITE IMAGE DATA

  А.Y. Budko , Т.А. Gaida , Z.А. Ponimash

  2026-02-27

  Abstract ▼

  The article describes a method for estimating the spatial position of an onboard camera of an aircraft. This method involves comparing aerial photographs and georeferenced remote sensing (RSS) data by using a neural network detector to detect stable spatiotemporal reference points in both datasets. This method then solves the well-known Perspective-n-Point (PnP) problem for estimating rotation and translation matrices that minimize the reprojection error based on the correspondences between 3D world points and 2D points of their projections onto the onboard camera matrix. This approach can be used to solve the pressing problem of aircraft localization in the absence of global navigation satellite system signals. Road intersections are selected as stable spatiotemporal reference points that are clearly visible in RSS data and aerial photographs. Other local semantic image patterns characteristic of a particular area may serve as an alternative. Since direct comparison of remote sensing and airborne images is difficult due to significant differences in shooting conditions, the use of robust landmark detectors based on artificial neural network (ANN) algorithms is proposed. To train the robust detector, a mixed dataset was created using satellite and airborne imagery. The mixed dataset was labeled using a 3D Gaussian function normalized to unity with a apex at the intersection center, the graph of which is projected onto a 2D mask of the training set. The parameters of the Gaussian function are calculated based on the radius of the circle enclosing the intersection. Using a normalized 3D Gaussian function with a apex at the geometric center of the intersection projection allows the network to predict the probability of each image pixel belonging to the intersection, with a maximum at the intersection center, which increases positioning accuracy due to more precise georeferencing of the landmark point in the global 3D dataset. A U-Net-type artificial neural network was trained as an intersection detector. A differentiable analog of the Dice metric was used as a training quality metric. AdamW, coupled with a CosineAnnealingLR cosine learning rate planner, was used as an optimizer. The final section of the paper presents the results of comparing satellite data and airborne imagery using the proposed method.

• CONVOLUTIONAL NEURAL NETWORK HYBRID ARCHITECTURE DEVELOPMENT USING SPECTRAL TRANSFORMATIONS

  B. V. Kostrov , S.I. Babaev , А.I. Efimov , V. Y. Tarasova

  2026-02-27

  Abstract ▼

  The hybrid convolutional neural network architecture with combining spectral and spatial layers, as well as new methods of subsampling (WalsPooling) and convolution (ConvWals) are proposed. The developed system is used to geographical proximity assess of images pair based on their visual similarity. A pair of different sensors obtained images visual similarity determination is complicated by different scales and sensor tilt angles shooting conditions. Based on the low-altitude image fragment, a search in the database of underlying surface images is performed. The search is performed in the surrounding area of a given route based on the vector of image features, which is formed on the last layer of the convolutional neural network. The system uses the Siamese architecture, since a pair of images must be submitted to the input. The relevance of this problem stems from the need to ensure UAV navigation in the absence or unreliability of a GPS signal. The approach to data set formation and its preprocessing is also considered. The database search is performed in the surrounding area of the route, which reduces computational costs. The experiments include an analysis of the applicability of the proposed layers (WalsPooling, ConvWals) and a comparison with traditional pooling and convolution methods. The paper also presents a linear approximation method with trainable parameters for reducing the dimensionality of the convolutional layer. The main advantage of the approach is its resistance to changes in the scale and angle of shooting due to a combination of spectral and spatial features. The results demonstrate the applicability of the method for UAV navigation in conditions of loss of GPS signal is lost or unreliable. The experiment demonstrated that using images reconstructed after spectral transformation yields the best neural network convergence and mean square error. The developed architecture demonstrates robustness to geometric and brightness distortions, and its quality metrics (Precision = 0.728, Recall = 0.800, F1 = 0.872) confirm the effectiveness of the approach for visual localization tasks based on images from a surface database.

• SPIKING REPRESENTATIONS COMPARISON FOR LOCALIZATION AND NAVIGATION IN THE KEYFRAME MAP

  I.S. Fomin , V.D. Matveev , А.Е. Arkhipov

  273-284

  2025-10-01

  Abstract ▼

  The task of navigating a mobile robotic platform in a known environment has been efficiently solved for a long time and using a flat passability map, which is built using lidar. Nevertheless, situations regularly arise when, for one reason or another, the platform is not equipped with lidar or other active navigation tools. At the same time, a camera is usually installed on the robotic platform, designed for visual monitoring of the situation by the operator, which can also be used for navigation when moving the robot in a known environment. There are well-known examples of navigation algorithms based on the use of sequences of keyframes, for example, visual SLAM. At the same time, various variants of video images (blurred, masked, etc.) are considered as keyframes. In this paper, a cognitive (non-metric, non-spatial) map of keyframes representing a spiking representation of the observed images is considered as a base for navigation. The possibility of using neuromorphic information control elements developed at the RTC to compare the current spiking representation with all spiking representations of a key sequence is analyzed. It is shown that by such a comparison, the keyframe closest to the current one can be determined, and parameters for the shift of spiking representations can also be selected, which is an analog of localization and navigation for a cognitive map. The description of a software tool for emulating the construction of a map and moving in it for experimental testing of the proposed algorithms is given. Data collection and experimental evaluation of the quality of localization and navigation algorithms have been performed. To do this, we have collected several keyframe maps with different patterns of movement between frames. When determining the position of the frame in the map, the quality was from 70 to 98%, when determining the direction of displacement between frames, the accuracy was from 94 to 97%. The results obtained are assessed as sufficient to solve the tasks assigned to the algorithm.