Research On Lightweight Stereo Matching Based On Neural Network

Stereo matching is an important algorithm in the field of stereo vision and 3D reconstruction.The stereo matching algorithm generates a disparity map containing scene depth information from a pair of stereo image pairs.The traditional stereo matching algorithm cannot solve the problems such as image occlusion and texture loss.With the development of convolutional neural networks,more and more stereo matching algorithms have achieved better results recently,but it is difficult to meet the real-time requirements.The following three aspects of work accordingly:(1)This paper proposes a stereo matching algorithm model based on lightweight neural network,which follows the design principles of lightweight networks such as small-scale matching and feature reuse,and uses some techniques that are not computationally expensive but can effectively improve network performance.It can achieve real-time computing on the GPU while approaching the accuracy of large-scale networks.(2)This paper proposes a cost aggregation attention mechanism based on 3D convolution,and uses an auxiliary training method,and also proposes a cost aggregation attention mechanism based on correlation layers.These two attention mechanisms can be used in other networks.Experiments show that these two attention mechanisms can effectively improve the accuracy of disparity calculation and reduce the mismatch rate.(3)This paper designs an RTG obstacle detection system based on binocular vision.The system uses stereo matching and object detection algorithms as the core to detect obstacles,and can also complete tasks such as occlusion detection and road line detection.It improves the automation level and safety of port operations.