Research On Fast Binocular Stereo Matching Algorithm Based On Distance Groupin

Binocular stereo matching is an important algorithm to perceive the depth information of the surrounding environment.Compared with lidar equipment,the binocular stereo matching algorithm has lower cost and higher accuracy.The traditional binocular stereo matching algorithm cannot solve the problems of non-textured areas,reflections,occlusions and so on.With the rapid development of CNN,binocular stereo matching algorithms have made significant progress.However,existing high-precision binocular stereo matching algorithms based on convolutional neural networks usually require high inference time cost and are difficult to perform real-time inference on resource-constrained embedded devices.In addition,although the existing lightweight binocular stereo matching algorithms based on convolutional neural networks reduce the inference time cost,they also lead to a significant reduction in accuracy.In this paper,the following four parts of work are carried out to address the problem of low prediction accuracy of current lightweighting algorithms.(1)The U-Net feature extractor is easy to lose feature information during the downsampling process,and there are problems such as rough edge prediction and large error in weak texture areas.Therefore,this paper proposes a feature extraction module based on sliced convolution,which enables the model to well extract feature information of the input stereo image.(2)Any Net generates distance correlation graphs for a single channel only for each disparity level,thus losing the problem of correlated feature matching information of multiple channels,so this paper proposes a group-wise L1 distance method to construct a cost volume that can fuse feature matching information of multiple channels.In addition,this paper also proposes a combination of group-wise L1 distance and group-wise correlation methods for cost matching,which can achieve better similarity metrics at a faster speed.(3)To address the problem that it is difficult to regularize the cost volume quickly with a cost aggregation network of high computational complexity,a lightweight cost aggregation network is proposed in this paper.The experimental results show that this lightweight cost aggregation network can effectively improve the speed and accuracy of the model.In addition,this paper also proposes a lightweight cost aggregation network based on the attention mechanism of channels and disparities,which can further improve the disparity accuracy of the model by modeling the interdependence between each feature channel and each disparity separately.(4)Based on the above several simple but efficient modules,this paper follows the design principles of small-scale matching and feature reuse to design two stereo matching algorithms:a lightweight stereo matching algorithm based on grouping distance and a stereo matching algorithm based on channel and disparity attention mechanisms.Experiments on the KITTI2012 and KITTI 2015 datasets show that the proposed models can perform real-time inference on resource-constrained devices and achieve faster speeds than the current state-of-the-art lightweight stereo matching models.With Tensor RT optimization,the model proposed in this paper can process input images faster on NVIDIA Jetson Nano devices while maintaining high accuracy.