Image-based positioning systems for unmanned aerial vehicles and image-based positioning methods using these systems.

Abstract

The invention proposes an AI-powered image positioning system and method for autonomous navigation of unmanned aerial vehicles (UAVs) under conditions of complete electronic jamming of GNSS satellite signals. The system consists of: (i) an EO / IR camera with an integrated 3-axis gimbal stabilizer (1) operating day and night; (ii) an NVIDIA Jetson Orin NX embedded computer (2) executing a Convolutional Variational Autoencoder (CVAE) model to encode ground images into 512-dimensional hidden feature vectors and matching them with a FAISS indexed geographic reference vector library; (iii) a software module simulating a virtual GNSS receiver that encapsulates absolute coordinates into a UBX-NAV-PVT message; (iv) a Pixhawk Cube 2.1 flight controller running Ardupilot (3) receiving messages via a secondary GPS port and integrating with the IMU via EKF3.The method operates in a 4-step pipeline: image preprocessing (center cropping, 256×256 RGB) → CVAE encoder encoding to create a z-vector latent → FAISS matching to find geographic location → first-stage EKF3 filtering → UBX-NAV-PVT packaging → second-stage EKF3 in Ardupilot; with a total latency of less than 50ms. The CVAE algorithm was trained offline on 256×256 sub-images cropped according to flight direction from a GeoTIFF map, with a double loss function MSE + β-KLdivergence (β = 0.00025). The entire reference library is stored under 100 MB (a reduction of over 98% compared to storing the original image). The system achieved a real-world position error (RMSE) of 0.34 m under standard conditions, and 2.32 m over a 5.12 km field distance at an altitude of 518.9 m and a speed of 52.52 km / h; significantly exceeding the required threshold of 20 m and surpassing comparable products on the market. The invention is universally compatible with all UAVs using Ardupilot firmware, requiring no modification to the flight controller hardware, and is suitable for defense, infrastructure monitoring, and search and rescue applications.