Research And Realization Of Unmanned Aerial Vehicle Three-Dimensional Reconstruction And Object Recognition System

In recent years,deep learning in the research of graphics and image is extremel y hot,binocular stereoscopic vision and object recognition technology are getting mo re and more attention,at the same time,the microprocessor based on the embedded pl atform has a good performance and cheap,and gradually show its advantages in the fi eld of engineering control.This paper aiming at the problem of UAV in the unknown environment,studies the binocular vision obstacle avoidance method based on the e mbedded platform,mainly discusses the three-dimensional reconstruction of the UAV in the unknown environment and the object recognition problem.This paper has achi eved algorithm design and hardware platform build,and has verified the program’s en gineering value through experiments.Firstly,make use of binocular vision technology for 3D point cloud surface reco nstruction.According to the imaging principle of the camera and the principle of bino cular stereo reconstruction,the stereo matching technique is studied.By using the pic tures taken with binocular cameras,converts in the world coordinates,the image coor dinates and the pixel coordinates based on their relationships getting objects’ three-di mensional information.This paper use the surf algorithm to match the feature pointsuse the median filter algorithm to filter the depth map,then according to three-dimen sional information for three-dimensional point cloud reconstruction.Secondly,carry out the research of convolution neural network in object recognition technology.Convolution neural network is a very representative network structure in depth learning technology,and has achieved great success in image and video processing.In this paper,I study the object recognition system based on the embedded system,and put forward my own object detection framework yolo-enhance based on the yolo detection framework.Yolo-enhance makes some changes in the network structure,it simplifies the network,while the classification of data sets and test data sets mixed together,the use of a hierarchical view of the object classification,with a large number of classified data set data expands the test data set to improve recall and improve fastness and accuracy.Finally,building UAV embedded self-aware environment and image recognition hardware testing platform.In order to verify the feasibility of the design of this paper,the paper uses Matrace 100,Guidance and Manifold as the hardware platform,and uses OpenCV and OpenGL to construct three-dimensional reconstruction.The experimental results show that the system which is based on surf algorithm and median filter has a better result of the three-dimensional point cloud image.At the same time using yolo and improved version of yolo-enhance for image recognition detection and classification,the results of the two comparative analysis verifies that the proposed yolo-enhance in the object recognition accuracy and speed are better than yolo.Based on the platform test analysis,the system designed in this paper can complete the three-dimensional reconstruction of the environment and object recognition.