Research And Application Of Landslide Deformation Recognition Method Based On Binocular Vision

Target recognition is one of the important research directions in the field of computer vision,which is to extract and separate specific target information from a complex environment.The recognition technology based on machine vision has many advantages,such as high safety,high precision,low cost,etc.,so it is also widely used in various fields.In the field of landslide deformation area monitoring and identification,the traditional identification methods mainly focus on the contact method to monitor and identify the landslide deformation area.This method undoubtedly increases the safety risk of construction personnel and reduces construction efficiency.Machine vision recognition technology as a non-contact recognition method effectively avoids exposure of people to dangerous environments.Therefore,the use of non-contact visual recognition technology to identify landslide deformation areas can effectively solve the above problems.With the deepening of subject research,the background information of engineering issues has become increasingly complex,and the current simple and low-dimensional information can no longer meet the needs of in-depth development of the subject.With the development of science and technology today,acquiring the three-dimensional perception of the real three-dimensional world provides more abundant information for the development of the subject.Among them,binocular vision is an extremely important method for obtaining multi-dimensional information.This research topic aims to use non-contact solutions to solve the problem of landslide deformation feature information extraction in complex scenes.To this end,this project uses a common USB dual lens to build a stable and low-cost binocular vision system,and at the same time design a set of landslide deformation area recognition schemes for complex scenes.The key research contents of this paper are as follows:(1)Hardware construction of binocular vision systemAccording to the principle and characteristics of binocular vision three-dimensional reconstruction,the hardware composition of the system is determined,and a stable and fast binocular vision software and hardware system is designed,which provides a feasible scheme for collecting multi-dimensional information of landslides.(2)Acquisition of 3D point cloud of landslide sceneSpecifically,a landslide stereo matching algorithm based on regional growth was selected to accurately establish a one-to-one correspondence between the left and right image pixels;a stereo matching algorithm combining the SURT algorithm with regional growth was proposed,and the NCC was improved by the idea of integral graphs.The algorithm accelerates the stereo matching process of left and right images,and finally completes the generation of 3D landslide data.(3)Identification of landslide deformation areaThis topic chooses to use deep learning to solve the problem of landslide deformation area identification.Improved the deep learning model U-Net full convolutional neural network,which greatly improves the accuracy of the network in the recognition of landslide deformation areas.