Research On Image Matching Algorithms For UAV Visual Navigation

Currently,unmanned aerial vehicle(UAV)navigation and positioning primarily rely on the Global Navigation Satellite System(GNSS).However,GNSS is susceptible to interference and spoofing.UAV visual methods refer to the approach of using onboard cameras for positioning and navigation,which can serve as an effective supplement to GNSS.This method involves matching UAV aerial images with satellite images.This paper focuses on researching image matching algorithms for UAV visual navigation and proposes improvements based on the SuperPoint and SuperGlue networks.The main research content and contributions of this thesis are as follows:1.Addressing the issue of poor scale invariance in the SuperPoint network,we redesigned the encoder structure of the network,developed a feature aggregation module to extract and fuse features from multiple scales,and designed a new pyramid pooling structure called SPPFCSP to further explore multi-scale features.To ensure real-time performance of the network,we introduced the latest lightweight backbone network,MobileOne,despite the increased complexity of the network structure,thereby improving the inference speed of the network.Finally,the algorithm’s performance was quantitatively evaluated on the HPatches dataset,demonstrating a 1.2% improvement in homography estimation metrics compared to the original network.2.Addressing the time-consuming aspect of the attention map neural network in the SuperGlue network,we designed a new message passing mechanism that utilizes singlehead attention and linear projection instead of the original multi-head attention,reducing computational complexity and accelerating the inference process.Additionally,when constructing the complete graph,we designed a local neighborhood feature fusion mechanism to enhance the expressive power of the current feature points.Quantitative validation on the HPatches dataset showed a 1.2% improvement in homography estimation metrics compared to the original network.3.Based on the proposed improved algorithms,a UAVvisual navigation system based on image matching was designed,including modules such as image preprocessing,tasks,and the host computer.Finally,the system was validated through real-world scenarios.