Abstract:

Satellite navigation and positioning is one of the critical modules to ensure unmanned aerial vehicle (UAV) flight safety. When the satellite signal is weak or interfered, failure location will affect or even endanger the normal flight of UAVs. Vision-based methods can be used to locate UAVs with image matching. However, current image matching methods are unable to extract robust features due to the great difference between heterogenous images in time, which results in inadequate accuracy and efficiency. This paper proposes a fast feature matching method, which uses residual network to extract multiscale robust features and accelerates coarse matching of low-resolution feature maps using minimum Euclidean distance. The module above is used on high-resolution feature maps to achieve secondary fine matching. The homography matrix is used to correct matching pairs and achieve UAVs localization. The proposed method improves robustness and efficiency of localization for UAVs. The experimental results with Wuhan Suburb actural dataset show that the average accuracy of the proposed method is 2.86 m, which is 2.1% higher than the current typical matching algorithm. This method has obvious advantages on localization robustness and computing speed, which completes all images matching and localization on Jetson Xavier NX, while the localization frequency is optimal with the frequency of 1 Hz.

Key words: unmanned aerial vehicle (UAV), fast visual localization, heterogenous image matching, residual module, Transformer