Research And Application Of Target Location Algorithm Based On Object Detection And Binocular Vision

With the development of modern computer vision technology,many technologies in this field are also applied in all walks of life,making people's work and life more and more automated and intelligent.Some applications such as in the scenario,the intelligent system usually need by way of visual perception environment,identify the two-dimensional image of the object object,you also need to get goals in the actual three-dimensional position in space,and to interact with the target,and the process can be used based on the deep learning object detection technology and positioning technology based on binocular vision.With the increasing maturity of object detection algorithm and binocular stereo vision,it has been widely used even for more complex environment,and is developing towards the direction of high precision and low delay.However,these technologies still have great limitations in practical application.For binocular vision,especially binocular matching,due to the large amount of computation,the computation time is relatively long,so the real-time aspect needs to be improved urgently,and on the other hand,there is also the problem of inaccurate binocular matching.The application and popularization of binocular vision system is limited by these two main problems.In order to improve the real-time performance and accuracy of binocular vision matching algorithm,this thesis designs a real-time positioning algorithm based on object detection and binocular vision,which mainly improves the real-time performance and parallax accuracy of binocular vision matching.The main work of this thesis is as follows:(1)In terms of real-time performance,the target position in the two-dimensional image obtained by the object detection algorithm is combined in this thesis.In the binocular matching stage,only the target area is matched instead of the full image,so that the number of pixels to be matched is greatly reduced.In addition,the feature map with small size in the neural network is used for preliminary matching,the approximate parallax is obtained first,and then the reduced parallax range is used to carry out accurate matching on the original image,which reduces the parallax search range and reduces the matching times of a single pixel by about 2/3.(2)In the aspect of parallax accuracy,the improved uniqueness detection rule can filter some mismatching points and greatly improve the parallax accuracy.Compared with the original algorithm,the mismatching rate drops from 50%?60% to about 20%.Combined with the common characteristics of large window and small window,this thesis uses multiple Windows to match.A matching degree value is set for each window.In the cost aggregation stage,the window with the highest matching degree is selected,and its parallax is taken as the final parallax.Based on the uniqueness detection,the mismatching rate can be further reduced by 3%-5%.