Research Of Lightweight And Efflclent Stereo Matching Algorithm

The binocular vision system can recover the depth information of the captured image through the camera imaging geometric model and the parallax principle through two horizontal cameras,and stereo matching is the key.The purpose of stereo matching is to find the synonymous points between the left and right image pairs,and to obtain the depth information of the object by calculating the parallax between the synonymous points.it is widely used in autopilot,aerospace,AR/VR and other fields,and has great commercial and social value.At present,the end-to-end stereo matching algorithm has achieved impressive performance,but there are still some challenging problems.Many high-precision models have slow reasoning speed and large computing resources,so it is difficult to deploy on embedded devices with limited resources.in addition,models that focus on speed often gain this advantage at the expense of precision.Therefore,it has become a difficult problem that the practical application has to face.For this reason,this paper pays attention to the low accuracy of disparity discontinuous region matching,and the following main work has been done in the fast lightweight stereo matching algorithm.The main results are as follows:(1)the stereo matching model is difficult to take into account in terms of memory footprint,efficiency and accuracy.at present,the mainstream methods use 3D convolution for cost aggregation to obtain higher prediction accuracy,but the model efficiency is low and memory consumption is large.In order to solve this problem,a lightweight stereo matching network based on fusion valence is designed in this paper.the network adopts a two-stage strategy from coarse to fine,and carries out matching cost calculation and cost aggregation operation on low scale.the number of parameters and computation of the model are effectively reduced,and a low-scale fusion valence is designed,which only uses 9 3D convolution for cost aggregation.It further ensures the lightweight and high efficiency of the whole stereo matching network.(2)the stereo matching model has a large matching error in the disparity discontinuous region,and the parallax optimization network based on Warp operation often has a large number of parameters and high computational complexity,so it is difficult to meet the requirements of fast and lightweight models.To solve this problem,a disparity optimization network based on wavelet subband information enhancement is proposed in this paper.the network consists of a disparity adaptive update module based on neighborhood consistency and a disparity thinning module based on wavelet high frequency information.it optimizes the disparity map efficiently with limited parameters and computational burden to improve the results of the stereo matching model in disparity discontinuous regions and other morbid regions.In order to verify the effectiveness of the proposed stereo matching model,experiments are carried out on Sceneflow,KITTI2012 and KITTI2015 data sets.The final experimental results show that the proposed lightweight stereo matching network based on fusion valence combined with adaptive parallax optimization network based on geometric information,the final model is only 1.11m parameters.The error percentage of three pixels with 29ms reasoning time on KITTI2015 test set is only 2.11%,the error percentage of three pixels on KITTI2012 test set is only 2.13%,and the endpoint error on Sceneflow test set is only 0.9 pixels.