Research On Efficient Stereo Matching Algorithm Based On Deep Convolution Network

Depth perception is one of the basic technologies in autonomous robots,VR(Virtual Reality)and AR(Augmented Reality).Compared to those dedicated sensors for depth perception such as Lidar and Radar,the stereo vision system has shown much great application potentials with lower costs and wide ranging range.Accurate and reliable stereo matching algorithm enables the ability of stereo vision system for depth perception.Leveraging on the recent developments in deep learning,stereo matching has been cast as a learning problem,and yields significant gains compared to conventional approaches in terms of both accuracy and speed.Thus,we explored stereo matching algorithms based on deep learning.The main innovations of our work are as follows.We proposed an end-to-end stereo matching network called DEINet,which is efficient yet compact.Proposed DEINet includes four modules: feature extraction tower,cost volume matching,matching cost aggregation network,and disparity regression,which correspond to feature extraction,matching cost calculation,matching cost aggregation,and disparity regression in traditional stereo matching algorithms.The feature extraction tower uses different convolution structures to explicitly extract and fuse low-level detail information and high-level semantic information.An enhancement block consisted of convolution and dilated convolution is proposed,which employed at the upper level of the feature extraction tower for better extraction of high-level semantic information.In cost volume matching module,features are rearranged to form cost volumes,followed by 3D convolution layers for matching costs calculation.Matching costs are further aggregated by matching cost aggregation network,the disparity is obtained from it by the disparity regression module.The design choices of each part is verified by ablation experiments.The EPE of DEINet on the Scne Flow dataset is 0.8,which greatly surpasses competitive algorithms;The 3PE of DEINet on KITTI2012 and KITTI2015 is 2.06% and 2.44% respectively,significantly outperforms competitive algorithms,with the smallest amount of parameters(1.8M)and the shortest inference time(0.24s).In order to further improve the accuracy and robustness of DEINet in ill-posed regions such as foreground occlusions and reflective surfaces,and maintain advantages in parameter amount and speed,we proposed the strong correlation matching cost and the multi-level refinement network.The strong correlation matching cost reduces the dimension of the correlation matching function,eliminates interference from irrelevant features,and decreases calculation consuming in the matching cost calculation process.The multi-level refinement network refines the matching cost by learning the matching cost residuals,thereby obtaining a disparity map with higher precision.Experiments show that after replacing the corresponding module of DEINet with the strong correlation matching cost and the multi-level refinement network,the 3PE on KITTI2012 and KITTI2015 decreased by 6.3% and 4.5% respectively,and the 3PE in the foreground area on the KITTI2015 dataset decreased by 8.79%.Benefit from the efficient design of these two modules,the amount of parameters and calculation consumed is further decreased.