| Three-dimensional(3D)visual information is the most intuitive understanding of the external world by intelligent machines.The visual 3D reconstruction technology combines the characteristics of machine vision and image processing,can obtain the spatial information of the target of interest,and achieve efficient and accurate non-contact measurement,which has become a research hotspot in the industry.In industry and agriculture,the realization of dynamic 3D geometric information on the target surface through visual methods is an effective way to accomplish tasks such as real-time target monitoring,quality monitoring,and surface data acquisition.It can meet the needs of production and research and is the principal technical means to achieve automation,intelligence and security.The research of dynamic surface reconstruction technology based on stereo vision can promote industrial intelligence,bring enormous social and economic benefits,and has essential research significance in modern industry and agricultural production.Based on the analysis of the characteristics of the current visual systems and related researches,the 3D dynamic surface reconstruction algorithm in a complex working environment for high precision requirements is studied.In general,for different complex environments,it is necessary to thoroughly analyze the actual measurement conditions,construct a suitable vision system for sampling,and complete the measurement task with appropriate algorithms and strategies.The problems of multi-vision system construction,stereo matching and point cloud post-processing is focused,to explore the key points of the breakthrough of existing visual 3D reconstruction schemes and algorithms.Taking the high measurement accuracy as the essential starting point,a general 3D dynamic surface deformation reconstruction algorithm which can be applied to the dynamic change of shape,continuous change of characteristics and complex working environment is designed.It can further meet the practical application requirements of high-precision 3D dynamic measurement and provide theoretical and technical support for improving the performance of existing visual methods in dynamic 3D reconstruction.The works and innovations of this study are as follows:1)The theory of visual measurement system is introduced,and a multi-vision system is established.The basic principle of the classic monocular vision measurement system and binocular vision measurement system is introduced in detail,and its advantages and disadvantages are analyzed.Based on this,a multi-vision-based model suitable for various measurement environments is constructed,and the solution of the critical parameters of the model are given.In the multi-vision system,each binocular vision system is a relatively independent component,and the visual information is correlated by coordinate transformation,which expands the field of view of the visual system,and can acquire comprehensive dense 3D information,which has strong adaptability,versatility and scalability.The experiment for measurement accuracy of the binocular vision system showed that the measurement error of the system was about-0.070±0.025 mm,and the maximum relative error was less than 0.51%.The experiment for the multi-vision system showed that,the average distance between the point cloud representing the same physical position and the theoretical position after coordinate correlation was about 0.2 mm,and the distance between the measurement sample points representing the same physical point was about 0.3 mm,which indicated that the constructed multi-vision system had high measurement and stitching accuracy.2)The image segmentation method based on semantic segmentation neural network is studied and applied to the stereo matching preprocessing problem to improve the quality and efficiency of matching.The traditional image segmentation method and its limitations are expounded.The basic principles of semantic segmentation neural network are introduced and applied to the segmentation task of multi-image.Before training,a large number of sample samples in the actual scene are collected for tagging and data expansion to obtain a training set,and then the semantic segmentation network is trained to complete the pixel-level classification task.The experiment for evaluating the performance of the model on the test set showed that the forward propagation of a single picture took an average of 0.476 s,the memory usage was about 2.6Gb,the pixel precision was 0.993±0.002,the average pixel precision was 0.993±0.002,and the mean Io U(Intersection Over Union)was 0.976 ±0.005,the FWIo U(Frequency Weighted Intersection Over Union)was 0.980±0.005,which showed a satisfactory segmentation performance.Compared with the traditional image segmentation algorithms,the model has strong versatility and stability.While effectively reducing the computational complexity of stereo matching,it retains the essential information of multiview geometry,especially the edge information,with which the accuracy of the extracted 3D point cloud is guaranteed.3)The high-precision 3D point cloud post-processing algorithm is studied,and the point cloud correction method is innovatively proposed for the requirements of high-precision measurement.Firstly,a variety of filtering methods are used to obtain a compact and smooth point cloud,and then the parameters of the multi-vision model are solved to complete the point cloud stitching.On this basis,the unavoidable error generated from the process of point cloud stitching is analyzed,and the importance and necessity of point cloud correction operation under high precision requirements is emphasized,followed by the design of point cloud correction method.The stitched point cloud is resampled and its normal vector is estimated,and then the combination of density clustering and point cloud registration is utilized to improve the overlap of the standard parts of the two point clouds.These algorithms are used to improve the stitching precision of the existing multi-vision system,and provide technical support with the performance of the current multi-vision system.Static point cloud correction experiments showed that the point cloud correction algorithm reduced the measurement error by more than 50% on different types of complex surfaces.4)Comprehensive dynamic 3D reconstruction experiments were carried out under complex working and sampling condition.The concrete-filled steel tubular(CFST)columns under cyclic repeated load were taken as the object,and the multi-vision system was set to dynamically got them sampled.The semantic segmentation network and high-precision point cloud post-processing algorithm were applied to track the dynamic surfaces visually.The experimental results showed that the average maximum absolute error at each sampling time was 2.54 mm,the average absolute error was 1.27 mm,the average relative error was 0.60%,and the average root mean square error was 2.15 mm.The reconstructed model can describe the complex 3D surface and the process of their deformation pretty well,and meet the requirements of engineering. |
PDF Full Text Request