| With the rapid development of automobile autonomous driving technology,the role of intelligent ranging system is more and more important.At present,ranging system can be divided into active and passive according to the working mode.Active ranging,such as trigonometry,requires a laser transmitter,which undoubtedly makes the system structure more complex and increases the size and cost of the equipment.Compared with the active ranging,passive ranging has the advantages of low cost,high efficiency,simple structure,no need to set the radiation source,strong adaptability and so on.Among passive ranging methods,binocular vision technology has the advantages of wide ranging range,high ranging accuracy and high efficiency,and has gradually become the focus of researchers.However,there are still many problems to be solved in the binocular passive ranging system.For example,the design method of the passive ranging optical system used in the vehicle-mounted environment is not clear,and the matching accuracy of binocular image in the weak texture and the occlusion area is degraded greatly.These problems undoubtedly greatly affect the accuracy of binocular passive ranging method.To solve these problems,a set of binocular visible light optical imaging system is designed in this paper.On this basis,a stereo matching network based on attention mechanism is built.The binocular optical system is used to collect high-resolution pavement environment images,and the convolutional neural network is trained to get the model to realize fast and high precision stereo matching.Then,the binocular ranging function is realized by using Open CV.The main research content of this paper includes the following aspects:Firstly,the research background and significance of binocular passive ranging method are described,the research status of the active and passive ranging technology and binocular stereo matching algorithm at home and abroad are reviewed,the principle of binocular ranging and the theories related to deep learning are studied,especially the stereo matching part of binocular stereo vision is studied in detail.This paper introduces the basic concept,classification and implementation methods of deep learning,as well as the basic theory of convolutional neural networks.Then,according to the application scenario of the vehicle binocular passive ranging system,the technical indexes and requirements of the vehicle optical system are demonstrated.CODE V optical design software was used to design,simulate and optimize the vehicle visible light optical imaging system in the visible light band.In order to reduce cost and processing difficulty,the lens surface of the optical system designed in this paper adopts spherical surface,with a horizontal field of view of 11.99°,a vertical field of view of 10.57°,and a diagonal field of view of 15.7°,which can meet the requirements of high resolution binocular image acquisition with a minimum distance of 50 meters in the case of three lanes in the city.This environment is close to the actual scene when the car is driving.Finally,in order to improve the accuracy of the vehicle-mounted binocular passive ranging system,a stereo matching network based on attention mechanism is proposed in this paper,which can perform stereo matching processing on the acquired highresolution binocular images.In order to obtain the connection between each channel and each pixel in the feature map and enable the network to better capture the image context information,the channel attention mechanism module and space attention mechanism module are added to the convolutional neural network feature extraction module to improve the matching accuracy of the three-dimensional matching network based on deep learning in the weak texture,no texture,occlusion area and other illdefined regions.A grouping correlation method is used to construct matching cost bodies to make the generated cost bodies more robust.A cost aggregation module is built by a stacked hourglass network,and the parallax map of three stages is output by a progressive refinement method.The data sets of Scene Flow,KITTI2012 and KITTI2015 are used to train and test the network.The experimental results show that,under the premise of guaranteeing the matching speed,the matching accuracy of this algorithm is greatly improved,especially in the ill-defined areas such as weak texture and occlusion area,which effectively balances the matching accuracy and matching speed.Thus,the accuracy of binocular passive ranging system is improved.Finally,the binocular ranging function is realized by Open CV. |
