Research On Improved Method Of Binocular Image Matching Based On YOLO

With the rapid development of image processing technology and deep learning algorithm,computer vision technologies are widely used in production and life.In some engineering fields,ranging in high altitude is required.Manual ranging is very difficult and dangerous,so intelligent video surveillance,like binocular vision ranging system is extensively used.As one of the key technologies of computer vision,binocular image matching is the basic support for the achievement of binocular vision ranging.It can quickly and accurately discriminate object categories,establish a visual relationship with similar objects,and analyze the similarity and consistency of the two objects to obtain more meaningful information to assist computers to recognize and understand objects like human beings.How to overcome the changes of illumination,scale and angle in matching to accurately establish the visual connection between images becomes a major challenge of image matching technology.The binocular ranging system according to the principle of left and right horizontal parallax to achieve the visual ranging of fixed target objects,some traditional matching algorithms cannot recognize the extremely similar feature points between different objects precisely,the matching error will affect the binocular ranging result.Therefore,the binocular image matching technology based on the depth learning YOLO(You Only Look Once)target detection algorithm for defining the feature point matching of binocular images has certain application and research value.Firstly,this paper studies the principle of feature point matching of binocular image,and describe the principles of image feature points extraction of various binocular image matching technologies,including the location determination algorithm and direction determination algorithm..The performance of image matching algorithm based on feature points is verified by system simulation,and the advantages,disadvantages and improvements between them are compared.This paper also introduces the current mainstream deep learning target detection algorithm,and compare the performance of different algorithms.Secondly,data set for the training of YOLO v3 is made.The data set adopts two binocular image data sets that consist of images acquired by KITTI and images self-collected by binocular cameras.Then,the specific targets in images,such as cars,bikes and people are marked by LabelImage.The annotation information will be included in an.xml file containing location and category information for different objects in the binocular image.In order to better analyze the characteristics of different target objects in the binocular image,accelerate the network convergence and better predict the target frame through the candidate frame,this paper uses K-means algorithm to cluster the binocular image dataset used in this paper.Candidate boxes that more closely matches the target aspect ratio in the binocular image is obtainedFor binocular images including smaller targets,the feature information of small target objects is contained in large-scale feature map.In multi-scale prediction,when the shallow network extracts features into the deep network,the feature information loss of small target objects in large-scale feature maps will occur.Improving convolutional neural network model of the YOLO v3,and using more scale predictions when extracting image feature information could help recognize smaller target objects better and decrease the omission ration of small target objects.Finally,this paper studies the binocular image matching method based on deep learning target detection algorithm,in order to solve the problem that the similar feature points of different objects are wrongly matched in binocular image matching,and combined with target detection algorithm such as Fast R-CNN is slow,an improved binocular image matching algorithm combining ORB algorithm and YOLO target detection algorithm is proposed.The image features will be extracted by convolutional neural network.According to obtained target frame position information and the category information of the multi-label classification.The region where feature point matching is performed on binocular images is defined,and the ORB matching of binocular image combined with target detection is realized.The simulation results show that the matching accuracy of the proposed method in single-target,double-target and multi-target binocular images is improved compared with the traditional ORB matching algorithm.What's more,this matching method has feasibility and good robustness.