Key points are not available for this paper at this time.
Estimating the coordinate positions corresponding to each pixel in the Unmanned Aerial Vehicles (UAV) image is a prerequisite in actual fields. However, GPS devices are vulnerable to interference or attack, making it impossible to obtain accurate position and attitude parameters. Therefore, UAV image geo-localization based on image content has become a research hotspot. This paper proposes a UAV image geo-Localization method in a GPS-denied environment based on deep-learning building semantic feature extraction and point-line-patch (PLP) feature matching, which could correct the longitude and latitude coordinates of each image pixel within the range of several km 2 areas. First, we employ an enhanced edge and multi-scale (EEMS) building extraction network to retain building edges and extract multi-scale buildings. Second, we design the PLP feature using the characteristics of buildings and the distribution of adjacent buildings. Then, we design a semantic matching algorithm based on the PLP feature and optimize the result using the inverse compositional Lucas-Kanade (ICLK) method for the UAV image geo-Iocalization. Experiments show that our proposed method can better locate each UAV image pixel's position, benefiting urban and regional spatio-temporal big data fusion and update.
Xu et al. (2022) studied this question.