ASSESSMENT OF THE ACCURACY OF LOCALIZATION OF KEY POINTS OF REFERENCE MARKERS BY A NEURAL NETWORK ALGORITHM WHEN ANALYZING VIDEO FROM ON-BOARD CAMERAS OF AN AIRCRAFT

Abstract

In the context of increasing requirements for the accuracy and reliability of aircraft navigation in the presence of various types of noise in the video stream, the problem of developing algorithms that combine the high performance of neural network methods in complex video shooting conditions with the high accuracy of classical approaches in favorable conditions is highly relevant. Aim. To develop and validate an adapted hybrid algorithm. This algorithm should combine noise-robust neural network detection (based on LND-Net) with subpixel refinement of classical methods (the Canny algorithm) based on local image quality analysis, increasing detection reliability in challenging conditions. Materials and methods. A hybrid approach is proposed in which the primary coordinates of keypoints (marker corners) are determined by a neural network (LND-Net). A bimodality criterion for the brightness histogram is used to assess image quality in the vicinity of each point (a 10  10 pixel region). In the presence of pronounced bimodality, a classical contour is activated: adaptive thresholds based on the Otsu method and a Canny edge detector are used to refine the coordinates. In a noisy environment (unimodal distribution), the coordinates from the neural network are used unchanged. The SIOKS metric is introduced for scale-invariant localization accuracy assessment. Validation was conducted on datasets including reference images and frames with synthetic noise (Gaussian noise, blur, and fog), as well as under conditions of partial marker occlusion. Results. An analysis of the impact of image quality on localization errors (using the SIOKS metric) is conducted. It is shown that the proposed hybrid algorithm maintains accuracy close to that of classical methods under ideal conditions (average SIOKS ~ 0.010–0.013) and ensures robust operation in noisy environments where classical methods fail, relying on a robust neural network component. Conclusion. The developed adaptive algorithm successfully resolves the fundamental tradeoff between the subpixel accuracy of classical computer vision methods and the high noise immunity of neural networks. The proposed histogram bimodality criterion has proven its effectiveness as a fast and reliable trigger for switching processing modes in real time. A validation methodology linking pixel accuracy with local image quality analysis is substantiated, which is critical for the correct operation of onboard video surveillance and aircraft navigation systems.