Research And Implementation Of Disparity Ranging Based On Parameter Optimization In Autonomous Driving

At present,the field of autonomous driving at home and abroad continues to develop,and auto companies such as Tesla and NIO have set off a smart car revolution.Research on how to improve the intelligent perception of vehicles has become one of the important research directions of research institutes and companies.Using binocular vision for positioning is one of the mainstream directions in environmental perception technology.However,there are still many unresolved problems in the process of binocular positioning,such as rapid camera calibration.Traditional algorithms cannot be applied to complex road environments.The parameters of the existing neural network model are too large and not suitable for indoor distance measurement.How to improve the speed and accuracy of the binocular camera calibration process in this article? How to reduce the parameters of the model while ensuring the accuracy of the disparity network? How to apply the disparity network trained on the standard data set to indoor distance measurement? This paper proposes new ideas and algorithms based on the above problems.The experimental results show that the parallel calibration algorithm proposed in this paper increases the calibration speed by 23.5% compared with the traditional algorithm while ensuring accuracy.Compared with the PSMNet model,the VHNet model proposed in this paper reduces the model volume by 46.6%,and improves the calculation speed and accuracy,suitable for distance measurement in indoor and outdoor scenes.The main work done in this paper is as follows:(1)In order to improve the accuracy and speed of binocular calibration,a parallel calibration algorithm based on sub-pixels is proposed.By selecting the optimal number of calibration images and adopting a parallel calibration method,the calibration speed of the algorithm is increased by 23.5%,and the reprojection error is 0.858 pix,achieving sub-pixel accuracy.(2)In order to ensure the accuracy of the disparity network while reducing the amount of model parameters,a network model VHNet optimized based on PSMNet is proposed.The model uses a novel feature residual network to obtain image features,and uses SPP network structure,building cost volume and 3D pyramids to train the network model.The experimental results show that the overall performance of VHNet is improved compared with PSMNet,the amount of model parameters is reduced by 46.6%,the operating efficiency is improved by0.33%,and the 3-pixel error is reduced by 7%.(3)In order to verify that the algorithm in this paper is suitable for indoor and outdoor ranging,a set of indoor environments is built for testing.The experimental results show that the algorithm in this paper can be applied to indoor and outdoor scene ranging,and the relationship between the ranging accuracy and the focal plane of the camera is also studied.This topic relies on the research project commissioned by ZTE Corporation-"Key Visual Positioning Technologies and Intellectual Property Layout in Autonomous Driving".