Research On Cross-View Image Matching Algorithm For Rejection Environment Navigation

Cross-view geolocalization matches the same geographic target in images from different viewpoints,which is a key technology for autonomous localization and navigation of UAVs in denied environments.The most challenging aspect in this field is the offset of geographic objects in the image and the scale inconsistency between different viewpoints.Existing methods focus on extracting coarse features from partial images,but ignore the effects of scale and offset on matching accuracy.To solve this problem,this paper proposes two matching algorithms based on neural network.On the experimental dataset,the performance of the two algorithms surpasses the current mainstream algorithms and has strong robustness.The main work of the paper is as follows:1.Briefly describe the UAV image positioning and navigation task of denied environment,propose a UAV image positioning and navigation system and the UAV positioning and navigation scheme.The problems of existing cross-view image matching algorithms are analyzed.2.An image matching algorithm based on CNN multi-scale block and attention mechanism is proposed.The algorithm improves the existing feature map partitioning strategy and proposes an attention mechanism,the attention module discovers the relationship between regions and makes each region focus on different features,making the proposed model more robust to offset and scale changes.The model achieves excellent performance on benchmark dataset,significantly outperforming existing models in accuracy on the task of matching UAV and satellite imagery.Furthermore,compared to existing methods(such as LPN),the model reduces inference time by 30%and achieves the same accuracy.When inference time is nearly the same,the model is10% more accurate.3.An image matching algorithm based on Transformer and semantic guidance is proposed.Different from other existing CNN-based methods,this method proposes a Swin-Transformer-based structure to match UAV and satellite images.A semantic guidance module is proposed to implement contextual information mining and feature alignment in the inference stage,which improves the accuracy of the model under offset and scale changes.The method achieves outstanding performance on various accuracy metrics on benchmark datasets,outperforming existing methods such as LPN by nearly 10%.