TDSA-SuperPoint: A Novel SuperPoint Based on Tucker Decomposition With Self-Attention for Scene Matching Navigation

Scene matching navigation (SMN) is a crucial method for unmanned aerial vehicles (UAVs) in environments where the global navigation satellite system (GNSS) is unavailable. Traditional methods often struggle to simultaneously achieve accuracy, computational efficiency, and robustness due to varying conditions of the matching images. This article introduces an accurate, fast, and robust SMN method based on a novel SuperPoint improved by Tucker decomposition with a self-attention (TDSA) network. Utilizing the mechanization results of the inertial measurement unit (IMU) and magnetic compass, the proposed system first completes the georeferencing of real-time sensed images, reducing projection errors. To address the high-computational complexity caused by fully convolutional neural network (FCNN) parameters in SuperPoint encoders and decoders, Tucker-2 decomposition is employed to reduce the channel dimensions of model weights, while a self-attention module enhances key feature point information. A TDSA network with a compact representation is constructed, and a feature extraction model is developed based on this network. Finally, to ensure real-time feature extraction during model deployment, the Tucker convolution kernel and self-attention convolution kernel are fused, reducing the number of forward propagation operations. Experimental results indicate that, compared to SuperPoint, the proposed method reduces the number of parameters by 58%, improves feature extraction efficiency by 18%, and enhances SMN accuracy by 50.45%.