Research On Binocular Images Stitching And Distance Measurement Technology

Binocular ranging is one of the hot topics in the field of computer vision.It is widely used in assisted driving and unmanned obstacle avoidance scenarios.With the continuous development of the practical application of this technology,people hope that the perception range of binocular distance measurement will become larger,so that pedestrians and vehicles can be detected earlier by the assisted driving system,and obstacles can be perceived by the robot in a larger range.Traditional binocular ranging technology can only get distance information for part of the binocular image content,making the ranging field of view smaller than the field of view of the collected image.In this thesis,binocular image panoramic ranging is achieved by combining the monocular ranging and image stitching algorithms in the binocular ranging technology.The average relative error rate of panoramic ranging test is 3.05%,which expands the transverse field of view of binocular ranging while ensuring the accuracy.The main contents of this thesis are as follows:(1)Binocular ranging based on improved PSM-NetBinocular ranging is mainly divided into four parts: camera calibration,image correction,stereo matching and depth calculation.This thesis uses the Zhang Zhengyou calibration algorithm to get camera parameters,and uses the Bouguet algorithm to correct the image.In the stereo matching part,to solve the slow convergence problem of PSM-Net model of deep learning stereo matching network,the cross-entropy loss function is added to the loss function part,which improves the convergence speed of the model significantly.Finally,the parallax map obtained from the network model is used for binocular ranging test,and the error of the test is analyzed theoretically.(2)monocular ranging based on YOLOv5The binocular ranging section can only measure the information in the left view.For the right view which can not be used,this thesis obtains the distance information through the monocular ranging.YOLOv5 network is trained by COCO dataset,tested by KITTI-Object dataset,and the position of the lower edge of the target detection frame is used as the grounding line position of the object to achieve geometric relationship method for single-view ranging.The accuracy of single-view ranging is verified by real-world scene testing.(3)Image stitching algorithm based on ANAPIn order to display the ranging results better,this thesis uses AANAP algorithm to generate panoramic images from binocular images.To solve the problem that image stitching algorithms are prone to artifacts in disparity areas,an adaptive clipping algorithm is proposed through experiments,combining image overlap area recognition algorithm and image depth estimation.After preprocessing with this algorithm,the quality of image stitching is greatly improved.