A Point–Set–Domain Image Object Matching Method for Airborne Object Localization

Abstract

Image object localization is an important research direction in the development of intelligent autonomous control systems for unmanned aerial vehicles (UAVs). Major challenges remain, such as cross-view images, large-scale deformation, and multitemporal variation. We propose a point–set–domain matching method to locate objects. First, the property constraints of a point, including sparsity, repeatability, and distinguishability,are combined into a keypoint response used to optimize convolutional neural networks, creating keypoint detector and feature descriptor models. With these models, we can improve the performance of point matching and obtain the corresponding keypoint set accurately. This approach solves the cross-view problem. Second, a spatial transformation model of the corresponding keypoint set is obtained using keypoint-constrained diffeomorphism matching, which can align the spatial location of two images and solve the large-scale deformation problem. Third, an approach combining probability statistics with watershed maximally stable extremal regions is proposed to divide the object image and reference image into several subregions, and then the similarity based on diffeomorphism is employed to localize the object in the UAV image, which solves the multitemporal variation problem. The experimental results show that the proposed method can successfully determine the location of the object in the UAV image.