Research On Object Recognition Based On Deep Learning And Sequential Image 3D Reconstruction

In recent years,small unmanned aerial vehicles(UAVs)have a wide range of applications in meteorological exploration,map mapping,traffic control and geological surveys.At the same time,new technologies are emerging that are impacting and changing our lives every day.Deep learning is extremely popular in the application of graphic images.This thesis focuses on the image processing,based on the images taken by the UAV airborne camera,to explore the recognition of objects and the 3D reconstruction of objects.On the one hand,the improved neural network target recognition algorithm is used to improve the recognition accuracy;on the other hand,the 3D reconstruction algorithm based on the improved AKAZE algorithm improves the speed of 3D reconstruction.The specific work of this paper is as follows:1.The R-CNN series and YOLO series algorithms of target recognition based on neural network are studied.YOLO series have higher detection accuracy and speed.According to the requirements of real-time and accuracy of this paper,a simplified version of YOLOv3-Tiny algorithm of YOLO series is adopted.Aiming at the problem that the algorithm can not accurately identify driving targets,this paper first selects KITTI data sets(including driving samples),uses clustering method to select preset candidate boxes suitable for identifying various types of vehicles,changes the size of the input image pixels of the network,and improves the network structure.The experimental results show that the improved algorithm can improve the accuracy of the original algorithm under the condition of real-time traffic target recognition in real environment.2.The traditional multi-view 3D reconstruction algorithm SFM is studied,and the classical SIFT and AKAZE algorithms are discussed.Aiming at the slow speed of multi-matching of feature points extracted by SIFT algorithm,an improved AKAZE algorithm is proposed.The feature points are described by LATCH descriptor and roughly matched by FLANN algorithm.Accelerate the speed of feature extraction and matching,so as to improve the speed of 3D reconstruction.In this paper,the experimental results of the YOLOv3-Tiny algorithm before and after the improvement are compared,and the superiority of the improved algorithm in target recognition is proved.The experimental data is analyzed by comparing the SIFT algorithm with the improved AKAZE algorithm.It shows that the method improves the speed of 3D reconstruction,and obtains a better point cloud through dense reconstruction,which verifies the feasibility of the method.