CN-121685850-B - Ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R

CN121685850BCN 121685850 BCN121685850 BCN 121685850BCN-121685850-B

Abstract

The invention provides a ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which comprises the steps of S1, double-view image acquisition, S2, double-view image preprocessing, S3, DUSt3R algorithm and double-view feature processing, S4, three-layer coordinate mapping model construction and point cloud coordinate conversion, S5, three-dimensional grid model generation and geographic positioning fusion optimization, S6, error correction and performance verification, and S7, digital twin platform integration and application. The invention aims to solve the problems of complex three-dimensional reconstruction process, high cost, poor real-time performance, dependence on manual labor in operation and scheduling, insufficient safety supervision and the like of the traditional ship lock.

Inventors

XIONG XIHUI
HU XIAODONG
ZHU JICHENG
ZHANG HAI
Que Xiuping
XU ZHEN

Assignees

上海海达通信有限公司

Dates

Publication Date: 20260508
Application Date: 20260209

Claims (10)

1. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R is characterized by comprising the following steps of: s1, synchronously acquiring a front view image and a side view image of a navigation ship to form a double-view image pair; S2, sequentially performing noise suppression, reflection suppression and distortion correction on the double-view image pair to obtain a preprocessed double-view image pair; S3, inputting the preprocessed double-view image pair into a trained DUSt R algorithm model based on a Transformer framework, extracting deep features of the double-view image, performing feature matching by adopting a feature matching mechanism, estimating dense point cloud by a triangulation algorithm based on effective matching feature pairs obtained by feature matching; S4, arranging a plurality of ground control points in a lock chamber area, acquiring pixel coordinates, lock chamber coordinates and earth longitude and latitude coordinates of each control point, constructing an image pixel-lock chamber coordinate system mapping model and a lock chamber coordinate system-earth longitude and latitude coordinate system mapping model, and converting the dense point cloud into the dense point cloud under the earth longitude and latitude coordinate system through the two mapping models; S5, denoising and downsampling are carried out on the dense point cloud under the geodetic longitude and latitude coordinate system, a three-dimensional grid model is generated through a curved surface reconstruction algorithm based on the processed point cloud, and the three-dimensional grid model is associated with a geographic coordinate system to achieve geographic positioning fusion, so that a three-dimensional grid model after geographic positioning is obtained; S6, acquiring the actual size of the ship and the actual positioning coordinates of the ship, respectively calculating the size measurement error and the positioning error of the three-dimensional grid model after geographic positioning, and if the error exceeds a corresponding preset error threshold, adjusting related algorithm parameters and repeating the steps S3-S5 until the error meets the requirement, thereby obtaining the three-dimensional grid model after error correction.
2. The dual view and DUSt R based ship lock non-calibrated three-dimensional reconstruction method according to claim 1, further comprising: S7, converting the three-dimensional grid model subjected to error correction into a format supported by a digital twin platform, importing the three-dimensional grid model through a platform interface and linking the three-dimensional grid model with ship lock real-time data, and realizing visual display and business application of the model.
3. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R according to claim 1, wherein in S3, the feature extraction module of the DUSt R algorithm model adopts a multi-layer transducer encoder, each layer of encoder comprises a multi-head self-attention layer and a feedforward neural network layer, the multi-head self-attention layer captures long-distance dependency relations among image pixels in parallel through a plurality of groups of attention heads, and the feedforward neural network layer performs dimension transformation and enhancement on extracted features through a nonlinear activation function.
4. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R according to claim 1, wherein in S3, the feature matching mechanism calculates matching similarity of double-view deep features by adopting cross-correlation attention, effective matching feature pairs are screened through a dynamic matching threshold, the dynamic matching threshold is adaptively adjusted according to illumination change of double-view images and ship attitude difference, and false matching features are eliminated through neighborhood feature consistency verification.
5. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which is characterized in that in S4, an affine mapping model is adopted by the image pixel-lock chamber coordinate system mapping model and the lock chamber coordinate system-earth longitude and latitude coordinate system mapping model, mapping matrix parameters are solved through a least square method, a weight distribution strategy of control point coordinates is introduced in the solving process, higher weight is given to control points of a lock chamber key area, the conversion from lock chamber coordinates to earth longitude and latitude coordinates is realized, the association between the lock chamber coordinates and the earth longitude and latitude coordinates is established, and the point cloud has geographical positioning capability.
6. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which is disclosed in claim 1, is characterized in that in S5, the denoising adopts a statistical filtering algorithm, points exceeding the range of the average distance and the standard deviation of the preset number of neighborhood points of each point are judged to be abnormal points and removed by calculating the average distance and the standard deviation of the preset number of neighborhood points, the downsampling process adopts a voxel grid downsampling algorithm, and the voxel size is dynamically adapted according to the density characteristics of dense point clouds and the precision requirements of a three-dimensional grid model.
7. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which is characterized in that in S3, the triangulation algorithm is based on polar geometric constraint of the double views, a linear equation system is constructed to solve three-dimensional space coordinates through projection coordinates of homonymous feature points in two camera coordinate systems, a robust estimation strategy is introduced in the solving process, and the influence of matching errors on a three-dimensional point cloud coordinate solving result is restrained.
8. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R according to claim 1, wherein in S6, the related algorithm parameters comprise the number of layers of a feature extraction network of a DUSt R algorithm model, a feature matching threshold, a robust estimation coefficient of a triangularization algorithm, a control point weight of a coordinate mapping model and a resolution parameter of a curved surface reconstruction algorithm, and parameter adjustment is directionally optimized according to distribution features of a dimension measurement error and a positioning error until the error meets a preset requirement.
9. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which is characterized in that in S5, a poisson curved surface reconstruction algorithm is adopted in the curved surface reconstruction algorithm, in the reconstruction process, a point cloud normal vector is estimated through a neighborhood point fitting plane and unified in the normal vector direction, then a hidden function is constructed based on the point cloud and the normal vector, a triangular grid model is generated by extracting an isosurface of a preset hidden function value, and the smoothness parameter of the hidden function is adaptively set according to the density degree of the point cloud.
10. The ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which is disclosed by claim 1, is characterized in that in S2, a Gaussian filter algorithm is adopted for noise suppression, the size of a Gaussian kernel function is dynamically selected according to the image noise intensity, the reflection suppression is used for solving a self-adaptive threshold value to divide a reflection region through an Ojin method, brightness compression processing is carried out on the reflection region, histogram equalization is carried out on a non-reflection region so as to enhance characteristic contrast, a quadratic polynomial distortion model is adopted for distortion correction, a distortion coefficient is obtained through a standard calibration template, and radial distortion of an image is accurately corrected.

Description

Ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R Technical Field The invention relates to the technical field of intelligent water transport and intelligent traffic intersection, in particular to a ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R. Background The water transportation is an important component of a comprehensive transportation system, plays a key role in the transportation of bulk commodities, and the ship lock is used as a core hub of a inland waterway, so that the operation efficiency and the safety level directly influence the channel traffic capacity. The traditional lock three-dimensional reconstruction technology has the remarkable limitations that the complex calibration process of multiple sensors is relied on, the requirements on professional equipment and personnel are high, the implementation cost is high, the reconstruction process is long in time consumption, the service requirement of lock real-time dispatching cannot be met, meanwhile, the traditional lock operation relies on manual experience dispatching, the ship positioning precision is insufficient, the safety supervision is mainly two-dimensional monitoring, the risks of ship overrun, out-of-range and the like are difficult to intuitively identify, and the full-process closed loop control cannot be formed. Along with the policy guidance of digital transformation of traffic infrastructure, the traditional technology has difficulty in adapting to the construction requirements of intelligent ship locks, namely high efficiency, safety and intelligence, and a technology for three-dimensional reconstruction and geographic positioning of ships with low cost, high efficiency and no calibration dependence is needed, so that core support is provided for digital operation of the ship locks. Disclosure of Invention The invention provides a ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R, which aims to solve the problems that the traditional ship lock three-dimensional reconstruction process is complex, the cost is high, the instantaneity is poor, the operation and the dispatching are dependent on manual work, the safety supervision is insufficient and other pain points, and by fusing the technologies of artificial intelligence, computer vision, a geographic information system and the like, the ship high-precision, quick and low-cost three-dimensional reconstruction and accurate geographic positioning are realized, a 'perception-modeling-application-feedback' closed loop system is constructed, the whole element digital expression, the safe and efficient operation management, the refined maintenance and the scientific decision of an intelligent ship lock are supported, the intelligent upgrading of the inland navigation ship lock is promoted, and technical support is provided for the national and intelligent water transportation construction. In order to achieve the above purpose, the invention adopts the following technical scheme: a ship lock calibration-free three-dimensional reconstruction method based on double views and DUSt R comprises the following steps: s1, synchronously acquiring a front view image and a side view image of a navigation ship to form a double-view image pair; S2, sequentially performing noise suppression, reflection suppression and distortion correction on the double-view image pair to obtain a preprocessed double-view image pair; S3, inputting the preprocessed double-view image pair into a trained DUSt R algorithm model based on a Transformer framework, extracting deep features of the double-view image, performing feature matching by adopting a feature matching mechanism, estimating dense point cloud by a triangulation algorithm based on effective matching feature pairs obtained by feature matching; S4, arranging a plurality of ground control points in a lock chamber area, acquiring pixel coordinates, lock chamber coordinates and earth longitude and latitude coordinates of each control point, constructing an image pixel-lock chamber coordinate system mapping model and a lock chamber coordinate system-earth longitude and latitude coordinate system mapping model, and converting the dense point cloud into the dense point cloud under the earth longitude and latitude coordinate system through the two mapping models; S5, denoising and downsampling are carried out on the dense point cloud under the geodetic longitude and latitude coordinate system, a three-dimensional grid model is generated through a curved surface reconstruction algorithm based on the processed point cloud, and the three-dimensional grid model is associated with a geographic coordinate system to achieve geographic positioning fusion, so that a three-dimensional grid model after geographic positioning is obtained; S6, acquiring the actual size of the ship and the actual positioning coordinates of th