Research On Key Technologies Of Near-real-time Unmanned Aerial Photogrammetry

Timeliness is the soul of emergency response.The unmanned aerial photogram-metry is one of the important means in emergency response.So far,it is possible to quickly obtain a large number of aerial images by using unmanned aerial vehicle(UAV),but it is still difficult to complete an UAV mapping within hours.In the big data era of geospatial information,the resolution of aerial image is expanding,and the amount of aerial image is increasing geometrically.There are some obvious problems in the massive imaging autoprocessing,such as low level of automation,low speed of image update,and low efficiency of data processing.The near-real-time photogrammetry is a key technology to achieve the rapid consumption of massive aerial image and UAV emergency measurement.In order to realize the rapid processing of high-definition aerial image,the paper studies the key problems in unmanned aerial photogrammetry,which has important theoretical and practical significance for the development of aerial photogrammetry.The main contents of this paper include:(1)In the establishment of near-real-time UAV photogrammetry hardware platform: A multi-rotor UAV aerial platform is built,which include multirotor UAV platform,flight control system,differential positioning and other modules.The wireless transmission of aerial images in ordinary commercial camera is achieved by using Wi Fi SD card and the individual transceiver module with 700 MHz communication frequency.A difficult management problem is caused for mass historical images were storage in the different hard disk with different interface in the field of aerial photogrammetry.In the paper,based on a grouping management structure,a low-cost scalable high-capacity image storage management system is designed,which can be used to solve the difficult of management and the storing of massive aerial images.In order to solve the problem of high-speed processing of high-definition aerial images,a FPGA hardware processing platform is designed based on cyclone V FPGA processor,and a high-speed aerial image processing hardware system is built based on computer CPU-FPGA cooperative construction.(2)In the study of high-speed sparse matching of aerial images: The corners are used as the feature points.An efficient corner detection algorithm is proposed based on Ada Boost weak classifier algorithm,which has similar speed with FAST algorithm,higher accuracy and robustness than Harris algorithm.By studying the theory of multi-scale differential Gaussian Do G function,the algorithm of the feature scale estimation and binary vector feature description are proposed in the original image.The new feature description algorithm solves the problem of large memory consumption in feature scale estimation,and achieves stable and fast feature point matching for large affine change images(the complex change of scale,visual angle and brightness).In the parallel high-speed implementation of the algorithm,the most time-consuming step(feature vector matching calculation)in SIFT algorithm is implemented on FPGA.The speed of implementation with CPU-FPGA is 246 times faster than CPU implementation for SIFT algorithm.This paper designs a sparse feature point matching framework of CPU FPGA cooperation.The new corner detection algorithm and feature matching calculation are implemented by CPU FPGA cooperation.The implementation with CPU-FPGA for new algorithm is 160 times faster than CPU implementation for SIFT algorithm,and which is nearly ten times faster than GPU implementation for SIFT.(3)In the research of aerial image high-speed dense matching: the SGM algorithm is implemented by CPU-FPGA collaborative acceleration,and the average speed of FPGA is255 times of the CPU realization in the bottom level of pyramid.A comprehensive judgment strategy about disparity accuracy is adopted,and the disparity hole filling method based on contour line is proposed to improve the later optimization accuracy of dense matching.In the process of disparity hole filling,a new contour line detection algorithm is proposed.A dense matching strategy is proposed,which obtains the initial disparity by local feature matching of sampling points and optimization of sparse disparity image,the end obtains the accurate disparity by SGM algorithm in a fixed small disparity search range.Finally,the high-speed dense matching is realized to high-definition aerial images by CPU and FPGA cooperative.The accuracy of the new algorithm is higher than that of the traditional SGM algorithm.The average speed of new dense matching on CPU-FPGA is 32 times of the original SGM algorithm on CPU and nearly ten times of the pyramid SGM algorithm on CPU.The speed of precise matching by FPGA is about 5 times of GPU realization.At the same time,the new algorithm does not need to guess estimating the disparity search range artificially.This paper introduces a new near-real-time CPU-FPGA aerial image processing strategy and a new sparse and dense matching algorithm.Based on the high-definition aerial image data set,the overall matching performance of CPU-FPGA and the accuracy of 3D reconstruction point cloud are compared with the traditional classical algorithms in detail.The final point cloud accuracy of new algorithm is higher than the traditional algorithm.The CPU-FPGA cooperation photogrammetry algorithm realizes the near-real-time unmanned aerial photogrammetry.The new strategy can finish UAV emergency mapping in a half hour if the survey area is less than 5 square kilometers.