Research On Point Feature Matching Algorithms Of UAV Remote Sensing Images

UAV remote sensing platform is gradually becoming an important data acquisition method for its characteristics such as flexibility, swift response, low cost, simple maintenance etc. Considering the ground coverage of UAV images is limited and the overlap degree is high, which means the entire target area can hardly be obtained within a single image, mosaicking is thus a basic work in the UAV image processing. The matching method based on point-feature is often used to acquire corresponding image points in geometric mosaicking. In this paper, combined with characteristics of UAV images, the feature extraction and matching search method was studied deeply on the purpose of finding and formulating algorithms which can meet different requirements of UAV applications and have more excellent performance in efficiency and effectiveness. The main jobs and the innovations are as follows:1. The background and significance of this paper was introduced. Domestic and overseas development status of UAV and feature matching technology was summarized. Besides, the main research content of this paper was clearly determined.2. The theoretical foundation of feature matching was studied. The definition of image matching, the selection of space geometric transformation models and the general process of feature matching were summarized. Then systematic analyses were carried on the principle of classical feature detection algorithms.3. Feature extraction algorithms of UAV remote sensing images were studied. Based on Scale Space theory, some kinds of point-feature detection and local information description with scale and rotation invariance were analyzed, after which experiments were designed and a comprehensive evaluation system was established to compare the performance of four common feature extraction algorithms under different conditions. Then advantages and disadvantages of AKAZE and ORB algorithm were summarized.4. According to different kinds of UAV applications, performance requirements of feature extraction algorithms were analyzed. Then two improved algorithms were proposed based on shortcomings of existing algorithms.To enhance effectiveness, the binary feature descriptor M-LDB was substituted for SIFT descriptor which has a more stable performance. To improve the precision of feature localization in ORB algorithm, SIFT sub-pixels interpolation technique was reasonable simplified and improved. It allows ORB algorithm to own the function of accurate feature localization. Experiments indicated that improved ORB algorithm keeps efficiency of original ORB algorithm, and has a desired precision of feature localization at the same time.5. An improved search algorithm supplied by navigation data was proposed. Several existing matching search algorithms were firstly analyzed. Based on the principle that Kd-tree algorithm can be used for fast explicit search in low dimension, according to the idea of dimension reduction, combined with the characteristics of UAV images, an improved fast explicit search method supplied by low precision POS data was presented. Experiments showed that the algorithm can keep matching ratio and has a search speedup effect when the size of data sets reaches a certain number. it can perform better with higher dimension and larger size of data sets.