A dynamic three-dimensional reconstruction system and method suitable for a narrow and long type of hydraulic experiment water tank

Abstract

The application discloses a kind of dynamic three-dimensional reconstruction system and method suitable for narrow long type hydraulic experiment water tank, it is related to water conservancy engineering hydraulic model experiment observation technical field, the system includes position adjusting mechanism, multiple groups of independently displaceable shooting unit, synchronous control and data processing system, supporting setting global coordinate system calibration module and hydraulic water body scene special adaptation module;The method is based on the above system, sequentially complete pre-experimental deployment and calibration, circumferential rotation-multiview synchronous image acquisition, single time three-dimensional scene reconstruction, experimental whole process dynamic reconstruction.The application solves the inherent defects such as visual angle blind area, insufficient precision, poor adaptability of traditional two-dimensional observation technology, can realize the axial cumulative error of 50m within long water tank full area reconstruction≤2mm, water scene and no water scene reconstruction precision deviation≤3%, hardware cost is reduced by more than 60% under the same acquisition density.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic engineering model experimental observation technology, specifically involving a dynamic three-dimensional reconstruction system and method adapted to narrow and long hydraulic experimental flumes. Background Technology

[0002] In the field of hydraulic engineering model experiments, narrow and long hydraulic experimental flumes are the core equipment for conducting various hydraulic mechanism experiments such as dam break evolution, dam failure, sediment transport, and water-sediment two-phase flow. Its core structural feature is that the cross-sectional size is much smaller than the axial length, and the whole is long and narrow. Currently, the mainstream experimental observation solution in this field is two-dimensional high-speed camera observation technology. As a conventional observation method in the industry, this technology has many inherent defects. First, in terms of perspective, fixed-position shooting has blind spots and occlusion problems, making it impossible to fully capture the complete spatial information of the experimental area. Second, in terms of accuracy, this technology can only acquire planar projection information and cannot extract depth data and three-dimensional spatial parameters of the experimental scene. This leads to systematic errors in the estimation of key experimental indicators such as breach flow and sediment transport, and the observation accuracy is difficult to meet experimental requirements. In addition, in experimental scenarios where multiphase media coexist, two-dimensional images also have problems such as information overlap and blurred media boundaries, making it impossible to intuitively distinguish the spatial distribution and dynamic evolution process of multiphase media. Finally, in terms of adaptability, it is difficult to adjust the camera position for two-dimensional shooting for experiments of different lengths and types of water tanks, making it difficult to quickly adapt to the observation area requirements of different experiments and resulting in poor deployment flexibility.

[0003] In existing technologies, general multi-view 3D reconstruction technology has been maturely applied in industrial inspection and static mapping scenarios. However, this technology differs significantly from the experimental scenarios of hydraulic flumes, and cannot directly adapt to the dynamic observation needs of flume experiments. Furthermore, existing 3D reconstruction schemes for hydraulic flumes suffer from two major targeted defects, becoming key bottlenecks restricting the technology's practical application. Firstly, the core contradiction is prominent. Related schemes either employ a single-section fixed multi-camera dense deployment method, which is not only extremely costly in terms of hardware but also difficult to adjust camera positions, making it inflexible for flumes of different lengths; or they use a single-camera rotating scanning method, which is only suitable for static scenes and cannot meet the rigid requirement of "simultaneous multi-view synchronous acquisition" in high-speed transient hydraulic experiments. Secondly, multi-unit axial deployment schemes have accuracy risks. Existing multi-unit axial deployment flume 3D reconstruction schemes lack a standardized global coordinate system calibration scheme, which easily leads to the problem of axial stitching errors accumulating and amplifying with the flume length, ultimately resulting in insufficient reconstruction accuracy across the entire area of ​​long flumes. This defect becomes a fatal obstacle to the practical application of such schemes.

[0004] Currently, there is no mature 3D reconstruction observation solution in this field that is suitable for narrow and long hydraulic flume experiments, can be flexibly deployed, and can balance high-speed dynamic synchronous acquisition with low hardware cost, leaving a significant technological gap. This invention addresses the inherent defects of the aforementioned 2D observation technology and the core bottlenecks of existing 3D reconstruction solutions by proposing an integrated solution that balances "high-speed dynamic synchronous acquisition, low hardware cost, flexible adaptation across the entire area, and high-precision reconstruction." This solution fills a technological gap in the field and is not a simple patchwork of existing technologies; it can effectively meet the dynamic observation needs of various narrow and long hydraulic experimental flumes. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a dynamic 3D reconstruction system and method adapted to narrow and long hydraulic engineering experimental tanks. First, it addresses the four inherent defects of traditional 2D high-speed camera observation technology in applications to narrow and long hydraulic engineering experimental tanks: blind spots, insufficient accuracy, information overlap, and poor adaptability, thus breaking through the limitations of traditional observation techniques. Second, it resolves the core contradiction in existing 3D reconstruction schemes for hydraulic engineering tanks, which cannot simultaneously achieve "multi-view synchronous acquisition in high-speed transient experiments" and "low-hardware-cost, full-view dense acquisition," filling the technological gap in rotational 3D reconstruction in dynamic hydraulic engineering scenarios. Simultaneously, it also solves the problems of inconsistent global coordinate systems and axial stitching errors when multiple shooting units are deployed in an axial array. This system addresses the issues of cumulative differential error and insufficient accuracy in full-area reconstruction of long flumes, breaking through the accuracy bottleneck of long flume observation. Furthermore, considering the unique characteristics of hydraulic engineering experiments, it resolves the problems of feature matching failure and poor 3D reconstruction accuracy caused by water reflection, refraction, and transparent flow interference, improving the system's adaptability in water-rich environments. Finally, it achieves dynamic 3D visualization reconstruction and quantitative analysis of the entire process and entire area of ​​narrow and long hydraulic engineering experimental flumes, flexibly adapting to the needs of hydraulic engineering experiments of different sizes and types. It also solves the problems of existing 3D reconstruction systems, such as the inability of the gimbal to perform circumferential stepping and rotating linkage shooting, high cost of full-view acquisition, and insufficient dynamic adaptability, further improving the system's observation accuracy and deployment flexibility.

[0006] The above objectives are achieved through the following technical solutions:

[0007] This invention first provides a dynamic 3D reconstruction system adapted to a narrow and long hydraulic engineering experimental tank. The system includes a position adjustment mechanism, multiple independently movable shooting units, and a synchronization control and data processing system. The position adjustment mechanism is arranged around the narrow and long hydraulic engineering experimental tank, extending in the same direction as the tank's axis, serving as a shared displacement reference for the entire system. Multiple independently sliding movable locking support components are installed on the position adjustment mechanism. Each component can slide freely along the axis of the position adjustment mechanism and be locked in a preset position. Multiple shooting units are arranged in an array along the tank's axis, with adjacent shooting units having an overlap area of ​​≥60%. Each shooting unit is connected to the position adjustment mechanism via a movable locking support component. Each shooting unit includes a shooting support frame and an image acquisition component. The synchronization control and data processing system is communicatively connected to the adjustable gimbal, drive actuator, and camera of all shooting units, enabling synchronized gimbal rotation, locking, and camera shooting, as well as 3D reconstruction processing based on the acquired images.

[0008] Furthermore, the shooting support frame is arranged circumferentially around the water tank, and a circumferential track is provided on the frame. The shooting support frame adopts a closed-loop ring form, a C-type / U-type non-closed-loop open form, or a multi-column structure. The movable locking support component adopts bolt locking, clamp locking, magnetic locking, or pneumatic locking methods.

[0009] Furthermore, the image acquisition component includes N≥3 adjustable gimbals, a drive actuator, and cameras. The number of cameras in a single shooting unit for high-speed transient experiments is ≥5, and the number of cameras in a single shooting unit for long-term conventional experiments is ≥3. The adjustable gimbals are evenly arranged along the circumferential track of the shooting support frame. Each adjustable gimbal is equipped with one camera. The drive actuator is connected to the adjustable gimbal for driving all adjustable gimbals and cameras to synchronously perform step-by-step rotational motion along the circumferential track. After the adjustable gimbal rotates to a preset shooting angle, the locking angle deviation is ≤±0.1°. All cameras are equipped with linear polarizing filters. High-speed transient experiments use high-speed cameras with a frame rate ≥1000fps and a resolution ≥1920×1080, while long-term experiments use cameras with a frame rate of 50-200fps and a resolution ≥1280×720.

[0010] Furthermore, the hardware synchronization trigger delay of the synchronization control and data processing system is ≤1ms, and the software synchronization trigger delay is ≤10ms. It can trigger camera shooting after the gimbal locks and stabilizes with a delay of ≤500μs. The multi-view stereo matching 3D reconstruction algorithm adopted by the synchronization control and data processing system includes incremental SfM algorithm, bundle adjustment algorithm, outlier filtering algorithm and dense reconstruction algorithm.

[0011] Furthermore, the system also includes multiple sets of global coordinate system calibration modules for shooting units, equipped with global calibration targets. Through target placement, coordinate transformation, and registration, the global coordinates of multiple shooting units can be unified, achieving an axial cumulative error of ≤2mm for the reconstruction of the entire area of ​​a water tank within 50m. The system also includes a dedicated adaptation module for hydraulic engineering and water body scenes, including camera polarizing filter configuration, water refraction correction model, and image preprocessing methods for reflection suppression, which can ensure that the reconstruction accuracy in water scenes deviates from that in waterless scenes by ≤3%.

[0012] This invention also provides a dynamic three-dimensional reconstruction method adapted to narrow and long hydraulic experimental tanks, based on the aforementioned dynamic three-dimensional reconstruction system, comprising the following steps:

[0013] Step 1: Pre-experiment deployment and calibration. According to the experimental plan and the size of the water tank, deploy the corresponding number of shooting units, adjust and fix the position of the shooting units, complete the calibration of the global coordinate system of multiple shooting units, the circumferential rotation positioning calibration of the adjustable gimbal, and the calibration of the camera's internal and external parameters and spatial pose. At the same time, complete the calibration of the camera polarizing filter adjustment and the water refraction correction model parameters.

[0014] Step 2: Circumferential rotation - multi-view synchronous image acquisition. After the experiment starts, the synchronous control and data processing system executes a cyclic acquisition process of synchronous stepping rotation - synchronous locking - simultaneous full-camera synchronous shooting - synchronous stepping rotation again, to complete multiple rounds of synchronous acquisition of the full-view range of a single cross section of the water tank, and completes the full-view synchronous image acquisition of multiple moments in the entire experimental process according to the preset time sequence.

[0015] Step 3: Single-moment 3D scene reconstruction. For the multi-view synchronous images acquired at the same moment, the following steps are performed in sequence: coarse-grained noise reduction pre-screening, water scene-specific fine pre-processing, feature matching optimization, water refraction geometry correction, sparse reconstruction and fluid temporal geometric continuity filtering, dense reconstruction, and multi-round data cross-calibration to restore the 3D model of the experimental scene in the tank at that moment.

[0016] Step 4: Dynamic reconstruction of the entire experimental process. The 3D models at different times during the entire experimental process are stitched and fused according to the acquisition time sequence. The point cloud stitching of multiple shooting units, axial error accumulation suppression and temporal global calibration are completed. Finally, a dynamic 3D evolution video of the experimental scene is generated, realizing the 3D visualization reproduction of the entire experimental process.

[0017] The advantages of this invention compared to the prior art are:

[0018] 1. This invention can directly replace traditional two-dimensional high-speed camera observation methods, effectively solving their inherent defects and compensating for the shortcomings of existing three-dimensional reconstruction schemes, bringing about multifaceted technological improvements. By deploying multi-view cameras around the circumference of the water tank, combined with the circumferential synchronous stepping rotation function of the adjustable gimbal, this invention can achieve 360° full-view coverage of the water tank experimental area, completely solving the problems of viewpoint obstruction and blind spots in traditional two-dimensional shooting, and can comprehensively capture the complete spatial information of the experimental scene. With the help of three-dimensional reconstruction of multi-view synchronous images, this invention can accurately obtain the depth information and three-dimensional spatial parameters of the experimental scene, completely solving the systematic error problem of two-dimensional plane projection, significantly improving the measurement accuracy of key experimental indicators, and providing a guarantee for the accuracy of experimental data.

[0019] 2. In the observation of multiphase media, this invention can intuitively reconstruct the three-dimensional spatial distribution and interface evolution characteristics of multiphase media such as water, soil, and sand, solving the problems of overlapping information and blurred boundaries in two-dimensional images. This provides more intuitive and accurate observational data for experimental mechanism research, helping researchers to delve deeper into the intrinsic laws of hydraulic experiments. Simultaneously, this invention possesses flexible adaptability across all scenarios. The number and installation positions of the imaging units, as well as the configuration of the gimbal and camera, can be flexibly adjusted according to the size of the water tank, the type of experiment, and observation requirements. It can adapt to narrow and long water tanks of any size, as well as various hydraulic mechanism experiments, significantly improving the system's versatility and practicality.

[0020] 3. This invention can also realize the temporal 3D reconstruction of the entire experimental process, stitching together the 3D models at discrete moments into a dynamic evolution video, completely reproducing the 3D evolution process of the experimental phenomena, and providing comprehensive support for the visualization analysis, post-processing, and results display of experimental data. In terms of cost control, the combined structure of a "single shooting unit + circumferentially rotatable gimbal" eliminates the need for a large number of cameras to be deployed circumferentially in the water tank. Only through the synchronous stepping and rotating linkage of the gimbal can dense acquisition of data from all angles be achieved. At the same acquisition density, the number of cameras is only 1 / 4 to 1 / 3 of that of a fixed multi-camera solution, reducing hardware costs by more than 60%, while ensuring the integrity of the acquired data, achieving a balance between low cost and high acquisition quality.

[0021] Furthermore, the synchronous control and data processing system enables synchronized linkage between the gimbal rotation and camera shooting actions. It can flexibly adjust the gimbal rotation angle, shooting sequence, and acquisition interval according to the dynamic characteristics of the experiment, dynamically adapting to the observation needs of different experiments and improving the intelligence level of observation and the targeted nature of data acquisition. Through global coordinate system calibration and point cloud stitching optimization, this invention solves the problem of accumulated axial stitching errors in long water tanks, breaking through the accuracy bottleneck of long water tank observations. For the special scenarios of hydraulic engineering experiments, through polarizing mirror configuration, refraction correction models, and reflection suppression algorithms, this invention effectively solves the problems of water reflection and refraction interference. The reconstruction accuracy deviation between water-containing and waterless scenarios is ≤3%, ensuring the stability of reconstruction accuracy in hydraulic engineering experimental scenarios.

[0022] 4. This invention achieves synergistic effects that cannot be achieved by individual technical features through the organic combination of various technical features, significantly improving the overall performance of the system. Regarding the core acquisition logic, the synergistic effect of "multi-camera synchronous group rotation + simultaneous acquisition at the same time + multi-round data mutual calibration" not only ensures the spatiotemporal synchronization of high-speed dynamic scenes, meeting the requirements of multi-view acquisition at the same time in high-speed transient experiments, but also achieves dense acquisition of all views at low hardware cost. Under the same acquisition density, the number of cameras is only 1 / 4 to 1 / 3 of that of a fixed multi-camera scheme, reducing hardware costs by more than 60%, producing unexpected technical effects that existing technologies cannot achieve.

[0023] 5. In terms of axial observation and accuracy control, the synergistic effect of axial multi-unit array deployment and global coordinate unified calibration is particularly prominent. Using a movable imaging unit alone can only achieve full-view coverage of a single section, which cannot meet the observation needs of the entire axial area of ​​the long flue. Using a shared position adjustment mechanism alone can only provide an axial movement reference, which cannot solve the problems of blind spots and insufficient planar projection accuracy in single-view shooting. Setting up a global coordinate calibration module alone, without the support of multi-unit array deployment, cannot achieve full-area coverage of the long flue. This invention organically combines a shared position adjustment mechanism, movable locking support components, multiple sets of imaging units, and a global coordinate calibration module. By deploying multiple sets of imaging unit arrays, it achieves full-area coverage of the long flue. By using a global coordinate calibration module, it achieves coordinate unification and error suppression. This not only achieves full-area coverage of the long flue and high-precision reconstruction of the core observation area, but also reduces deployment costs through a shared position adjustment mechanism. It solves the core contradiction between full-area coverage of the long flue and error accumulation, and achieves a synergistic effect that cannot be achieved by a single technical feature.

[0024] 6. In terms of scene adaptation, using general 3D reconstruction technology alone cannot directly adapt to the large aspect ratio structure of narrow and long water tanks and dynamic experimental scenarios. Using traditional 2D observation technology alone cannot obtain 3D spatial data. This invention combines mature 3D reconstruction technology with dynamic scenes of hydraulic water tanks, water body anti-interference technology, and global calibration technology through a scene adaptation combination scheme. It is applied to water tank experimental scenarios, filling the gap in the field and realizing cross-border integration and innovative application of technology.

[0025] 7. Regarding system linkage, using an adjustable gimbal alone cannot achieve precise circumferential synchronous rotation positioning and shooting linkage; using a synchronous control and data processing system alone lacks an effective actuator to support its control logic; and using a shooting support frame alone cannot achieve flexible rotation of the image acquisition components. This invention, through the synergistic combination of an "axially movable shooting unit + circumferentially rotatable gimbal + drive actuator + global synchronous control," achieves integrated observation of "axial full-area coverage + circumferential full-view acquisition + dynamic synchronous linkage." This ensures the integrity of the observation range, improves the accuracy and efficiency of data acquisition, and achieves a balance between low hardware cost and high observation performance. Its synergistic effect is far superior to the individual application of any single technical feature. Attached Figure Description

[0026] Figure 1 : Overall layout diagram of the dynamic three-dimensional reconstruction system adapted to narrow and long hydraulic experimental tanks of this invention;

[0027] Figure 2 This invention provides a schematic diagram of the cross-sectional structure of the supporting frame.

[0028] Figure 3 : A schematic diagram of the dynamic three-dimensional reconstruction method adapted to narrow and long hydraulic experimental tanks of the present invention;

[0029] Figure 4 : Schematic diagram of the dynamic 3D reconstruction algorithm based on incremental SfM;

[0030] Figure 5 : Actual photos of the wooden target object in the water tank during the experiment; Figure 5 (a) is the top view of the block, and (b) is the bottom view of the block;

[0031] Figure 6 Rendering of the 3D reconstruction model of the wooden target object; Figure 6 (a) is the top-view reconstruction result of the block, and (b) is the bottom-view reconstruction result of the block.

[0032] Explanation of reference numerals in the attached drawings: 1-Narrow and long hydraulic test tank, 2-Position adjustment mechanism, 3-Movable locking support assembly, 4-Picture unit, 5-Picture support frame, 6-Global calibration target. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] like Figure 1 As shown in this embodiment, a dynamic 3D reconstruction system adapted to a narrow and long hydraulic engineering experimental tank is presented. The system includes a position adjustment mechanism 2, multiple independently movable imaging units 4, and a synchronous control and data processing system. These components cooperate to achieve full-view, full-area dynamic 3D reconstruction of the narrow and long hydraulic engineering experimental tank 1. The position adjustment mechanism is positioned around the narrow and long hydraulic engineering experimental tank, extending in the same direction as the tank's axis. Serving as a shared displacement reference for the entire system, its core function is to provide all imaging units with a moving path and fixed points along the tank's axis, providing fundamental support for the flexible deployment of the imaging units. Multiple independently sliding movable locking support components 3 are installed on the position adjustment mechanism. Each component can slide freely along the axis of the position adjustment mechanism and can be locked in a preset position, thereby achieving independent position adjustment and precise positioning of a single imaging unit. The position adjustment mechanism can take various forms, such as a track system, gantry, fixed pre-embedded bracket or linear robotic arm. The track system can be a single track, double track or multiple track, and the extension direction of the track is consistent with the axial length direction of the water tank, which can adapt to the deployment requirements of different scenarios.

[0035] Multiple imaging units are arrayed along the axis of the water tank, with an overlapping area between adjacent units, and the overlap is no less than 60%. This overlapping area is used to achieve global coordinate registration, forming a flexible multi-unit deployment architecture. Depending on the size of the water tank and the experimental observation requirements, the number of imaging units, their axial spacing, and installation positions can be flexibly adjusted. This ensures both full axial coverage of the water tank and optimized overlap in the core observation area, thus adapting to the experimental needs of narrow and long water tanks of different lengths and types. Each imaging unit is connected to a position adjustment mechanism via a movable locking support assembly, allowing for independent adjustment and fixation along the water tank axis. Each imaging unit includes an imaging support frame 5 and an image acquisition component, which work together to achieve single-section, full-view image acquisition.

[0036] The imaging support frame is arranged circumferentially around the water tank. A circumferential track is installed on the frame, providing mounting points and rotation paths for the image acquisition components distributed circumferentially along the water tank's cross-section, thus achieving full-view coverage of a single cross-section of the water tank. The imaging support frame can adopt a closed-loop ring design, a C-shaped / U-shaped open design, or a multi-column structure. In the closed-loop ring design, the ring surface is perpendicular to the axial length of the water tank, coaxially enclosing the water tank within the inner ring space of the frame. The C-shaped / U-shaped open design has its opening facing one side of the water tank, suitable for scenarios where one side of the water tank is obstructed. In the multi-column structure, the columns are evenly distributed circumferentially around the water tank, also achieving full-view coverage of the water tank's cross-section. The movable locking support component can employ various locking methods such as bolt locking, clamp locking, magnetic locking, or pneumatic locking, flexibly selected based on the shape of the position adjustment mechanism to ensure the stability of the imaging unit and prevent displacement during the experiment from affecting observation accuracy.

[0037] The image acquisition component comprises N≥3 adjustable gimbals, drive actuators, and cameras. For high-speed transient experiments, each shooting unit has at least 5 cameras, while for long-term conventional experiments, it has at least 3. The adjustable gimbals are evenly distributed along the circumferential track of the shooting frame. Each adjustable gimbal carries one camera. Each gimbal is equipped with a drive actuator, which is connected to the adjustable gimbal for synchronous stepping rotation along the circumferential track of the shooting frame. The adjustable gimbal has a synchronous locking function, achieving precise locking after rotating to a preset shooting angle with a locking angle deviation not exceeding ±0.1°, ensuring the stability of the shooting angle. The drive actuator can be a stepper motor, servo motor, rack and pinion transmission mechanism, or synchronous belt transmission mechanism, etc., capable of driving the adjustable gimbal to achieve smooth and precise stepping rotation. The rotation angle can be flexibly set according to observation requirements, with a rotation step angle range of 15°-45° / step, and the step angle in the core observation area not exceeding 30°. All cameras are equipped with linear polarizing filters to suppress water reflection interference. The frame rate and resolution of the cameras are adjusted according to the dynamic characteristics of the experiment. High-speed transient experiments use high-speed cameras with a frame rate ≥1000fps and a resolution ≥1920×1080, while long-term experiments use cameras with a frame rate of 50-200fps and a resolution ≥1280×720.

[0038] The synchronous control and data processing system communicates with the adjustable gimbals, drive actuators, and cameras of all shooting units. Its core linkage logic is as follows: after the adjustable gimbal is driven to synchronously rotate to a preset angle and locks, all cameras are triggered to simultaneously acquire images. The synchronous triggering accuracy of the synchronous control and data processing system meets rigid requirements, with hardware synchronous triggering latency ≤1ms, suitable for high-speed transient experiments; and software synchronous triggering latency ≤10ms, suitable for conventional long-term experiments. On one hand, the system realizes synchronous triggering control of all cameras, image acquisition and storage, and linkage control between the drive actuators and the adjustable gimbals. It can precisely control the synchronous stepping rotation and synchronous locking of the adjustable gimbals, and the synchronous linkage of camera shooting actions. The delay after the gimbal locks and stabilizes is ≤500μs before camera shooting is triggered, avoiding vibration interference. On the other hand, it performs 3D reconstruction processing based on the acquired multi-view synchronous images, ultimately outputting a 3D model of the experimental scene and a dynamic evolution video. The synchronous control and data processing system adopts a multi-view stereo matching 3D reconstruction algorithm, specifically including incremental SfM algorithm, bundle adjustment algorithm, outlier filtering algorithm and dense reconstruction algorithm. Its synchronization triggering method can be hardware synchronization or software synchronization to adapt to the dynamic acquisition requirements of different experiments.

[0039] The system also includes multiple global coordinate system calibration modules for each imaging unit, used to achieve coordinate unification and error suppression across the entire area. This module is equipped with six global calibration targets. Through target placement, coordinate transformation, and registration, the global coordinates of multiple imaging units are unified, effectively suppressing the accumulation of axial stitching errors. Specifically, the global coordinate system calibration module places movable global calibration targets in the overlapping area of ​​adjacent imaging units and places one set of fixed global calibration targets at equidistant axial intervals. It acquires multi-view images of the calibration targets and calculates the transformation matrix between the local and global coordinate systems of each imaging unit. Combined with a global bundle adjustment algorithm, the transformation matrix is ​​optimized, ultimately achieving an axial cumulative error of ≤2mm for the full-area reconstruction of the long water tank.

[0040] The system also includes a dedicated adaptation module for hydraulic engineering and water body scenarios to address issues related to water reflection and refraction interference. This includes camera polarizing filter configuration, a water refraction correction model, and image preprocessing methods for reflection suppression, ensuring the accuracy and stability of reconstruction in water-related scenarios. Specifically, the camera polarizing filter uses a linear polarizing filter, with the polarization angle adjustable within the 0°-90° range based on the water reflection intensity. The water refraction correction model completes parameter calibration based on multi-view acquisition data from a standard calibration block of known size, obtaining key parameters such as refractive index and refraction angle correction coefficients. The image preprocessing method for reflection suppression combines the polarizing filter's acquisition characteristics with image segmentation algorithms, along with Gaussian filtering for noise reduction and histogram equalization for grayscale enhancement, ensuring a signal-to-noise ratio ≥30dB for the preprocessed image.

[0041] The present invention also provides a dynamic three-dimensional reconstruction method for the above-mentioned dynamic three-dimensional reconstruction system adapted to narrow and long hydraulic experimental tanks. This method realizes the dynamic three-dimensional reconstruction of narrow and long hydraulic experimental tanks, which corresponds one-to-one with the above-mentioned system. The specific implementation steps are carried out in a coherent manner to ensure the integrity and operability of the experimental process.

[0042] Step 1: Pre-experiment deployment and calibration. Based on the experimental plan and tank dimensions, deploy the corresponding number of camera units. Adjust and fix the positions of the camera units, complete the global coordinate system calibration of multiple camera units, the circumferential rotation positioning calibration of the adjustable gimbal, and the calibration of camera internal and external parameters and spatial pose. Simultaneously, complete the calibration of the camera polarizer adjustment and the water refraction correction model parameters. The specific implementation process proceeds smoothly, taking into account adaptability to different scenarios and the repeatability of operations. Based on the dimensions of the narrow and long hydraulic engineering experimental tank and the experimental type, select suitable position adjustment mechanisms, camera support frames, and movable locking support components. Determine the number of cameras and rotation parameters in conjunction with the experimental type to ensure a high degree of matching between the deployment plan and experimental requirements.

[0043] Regarding the selection of position adjustment mechanisms, for observations of the entire area of ​​a long flume, a dual-track system is selected, with the track extension direction consistent with the flume axis and arranged in parallel on both sides of the flume as a common displacement reference for the entire system; for high-speed transient observations of a small-scale flume, a single-track system is selected and arranged on one side of the flume to simplify deployment costs; for long-term experiments at a fixed site, a fixed pre-embedded bracket or gantry frame is selected to ensure deployment stability; for flexible observations in complex sites, a linear robotic arm is selected, which can flexibly adjust the spatial position of the imaging unit. The choice of the imaging rig is determined based on the observation requirements. If a 360° observation of the water tank cross-section without blind spots is required, a ring-shaped closed-loop rig is selected, with the ring surface perpendicular to the water tank axis, coaxially enclosing the water tank in the inner ring space, and the circumferential track of the rig completely surrounding the water tank. If there is an obstruction on one side of the water tank, a C-shaped / U-shaped non-closed-loop open rig is selected, with the opening facing the unobstructed side of the water tank, and the circumferential track is laid along the inner side of the opening to ensure full-view coverage of the unobstructed area. For long-term experiments, a multi-column rig is selected, with multiple columns evenly arranged around the circumference of the water tank, and a ring-shaped circumferential track set on the top of the columns, to meet the requirements of long-term stable operation.

[0044] The selection of movable locking support components matches the shape of the position adjustment mechanism. If the position adjustment mechanism is a track system, bolt locking or clamp locking is selected. The component slides with the track and is locked in place by bolts or clamps after sliding to the preset position. If it is a fixed pre-embedded bracket, bolt locking is selected and it is connected to the preset fixed point on the bracket. If it is a linear robotic arm, magnetic locking or pneumatic locking is selected to realize the quick locking and unlocking of the shooting unit and improve deployment efficiency. The deployment of the image acquisition components should be determined based on the type of experiment. For high-speed transient experiments, the number of cameras in a single shooting unit should be ≥5, and for long-term conventional experiments, it should be ≥3. Adjustable gimbals should be evenly distributed on the circumferential track of each shooting unit, with the same number of cameras as the number of cameras. Each adjustable gimbal should be equipped with one high-speed camera. The frame rate and resolution of the cameras should be adjusted according to the dynamic characteristics of the experiment. For high-speed transient experiments, high-speed cameras with a frame rate ≥1000fps and a resolution ≥1920×1080 should be selected, while for long-term experiments, cameras with a frame rate of 50-200fps and a resolution ≥1280×720 should be selected.

[0045] Each camera is equipped with a linear polarizing filter, and the polarization angle is adjusted according to the intensity of water reflection, with an adjustment range of 0°-90° to ensure reflection suppression. Each adjustable gimbal is equipped with a drive actuator. If high gimbal rotation accuracy is required, a servo motor with a synchronous belt drive mechanism is selected. If moderate accuracy is required, a stepper motor with a gear and rack drive mechanism is selected. The drive actuator is rigidly connected to the adjustable gimbal to ensure the smoothness and accuracy of the rotation. The rotation step angle selection range is 15°-45° / step, and the step angle of the core observation area is ≤30°. Regarding the deployment of the synchronous control link, the synchronous control and data processing system establishes communication connections with the drive actuators, adjustable pan-tilt units, and cameras of all shooting units via industrial Ethernet, and deploys a synchronous trigger module. For high-speed transient experiments, a hardware synchronization method is adopted, with a trigger delay of ≤1ms. The hardware synchronization adopts a differential signal wiring method, and the trigger signal timing is such that the trigger signal is sent after the pan-tilt unit is locked and stabilized with a delay of ≤500μs. At the same time, delay calibration is performed to ensure that the trigger deviation of all cameras is ≤100μs. For routine experiments, a software synchronization method is adopted, with a trigger delay of ≤10ms. The software synchronization adopts the TCP / IP communication protocol and uses a network clock synchronization method to ensure that the clock deviation of all devices is ≤1ms, ensuring the synchronous linkage of pan-tilt unit rotation, locking, and camera shooting.

[0046] Hardware protection is equally essential. The camera, gimbal, and drive mechanism all feature waterproof and splash-proof designs with a protection rating of ≥IP65. A waterproof protective cover is added to the camera lens to prevent water splashes from damaging the equipment. When deploying equipment around the water tank, a minimum of 0.5m of operating space is reserved to avoid spatial interference between the equipment and the water tank or experimental setup. Dustproof seals are also installed at the connection between the track system and the gimbal to prevent sediment from entering and affecting equipment operation. The adaptation to the water scene needs to be carried out in stages. First, a polarizing filter is installed and adjusted. The linear polarizing filter is fixed to the front of the camera lens, and the polarization angle is adjusted by rotating the polarizing filter. During the adjustment process, images of the water in the tank are acquired in real time to observe the reflection suppression effect until the water reflection is reduced to a minimum. The optimal polarization angle is recorded and fixed to ensure that the polarization angle remains constant throughout the experiment.

[0047] Subsequently, the parameters of the water refraction correction model were calibrated. Clean water of the required experimental level was injected into the tank, and standard calibration blocks of known size were placed at different depths and positions within the tank. Multi-view images of the calibration blocks were captured using a camera. Based on the image data, the key parameters of the water refraction correction model were calculated and entered into the synchronous control and data processing system for subsequent image refraction correction. Specific parameters for image preprocessing were also set. Noise denoising employed Gaussian filtering with a 3×3 Gaussian kernel size; grayscale enhancement used a histogram equalization algorithm; and reflection suppression employed an image segmentation algorithm based on polarization characteristics. This ensured that the signal-to-noise ratio of the preprocessed image was ≥30dB, providing high-quality image data for subsequent feature matching and 3D reconstruction.

[0048] Pre-experiment calibration work includes positioning calibration of the imaging units, global coordinate system calibration of multiple imaging units, gimbal rotation positioning calibration, and camera parameter calibration to ensure the accuracy of subsequent experiments. During imaging unit positioning calibration, the position of each imaging unit along the axis of the position adjustment mechanism is adjusted by the movable locking support component. The axial spacing between adjacent imaging units is set according to the width of the water tank and the camera field of view to ensure full axial coverage of the water tank and a field of view overlap of ≥60% in the core observation area. After adjustment, the imaging units are locked and fixed. The calibration of the global coordinate system for multiple shooting units must strictly follow the established procedure. Six global calibration targets are selected, using a checkerboard pattern with 8cm sides, the size of which is suitable for the water tank cross-section. The calibration targets are made of waterproof, rigid material with an anti-reflective surface treatment. Next, the calibration targets are deployed. One set of movable global calibration targets is placed in the overlapping area of ​​adjacent shooting units, positioned at the bottom of the water tank, and their position can be adjusted as needed. One set of fixed calibration targets is deployed every 10m along the axis, rigidly connected to the bottom of the water tank with expansion bolts to ensure stable positioning and no displacement. Images of the calibration targets are acquired, and all cameras are controlled synchronously to capture images through a synchronization control and data processing system. The target images are calibrated by acquiring no fewer than 10 images from different perspectives for each target. Coordinate transformation and registration are performed. Based on the calibrated target images, the transformation matrix between the local and global coordinate systems of each imaging unit is calculated. The least squares method is used to perform preliminary calibration of the transformation matrix to ensure that the transformation error is ≤0.1mm. The global bundle adjustment algorithm is used to optimize all transformation matrices in a coordinated manner to suppress coordinate registration errors, and finally achieves an axial cumulative error of ≤2mm for the reconstruction of the entire long water tank area. The calibration results are saved by storing the global coordinate transformation matrix and the local coordinate system parameters of each imaging unit in the synchronous control and data processing system for subsequent point cloud stitching and error suppression.

[0049] During gimbal rotation and positioning calibration, control signals are sent through the synchronous control and data processing system to drive all adjustable gimbals to rotate synchronously along the circumferential track. A rotation step angle is set, and each time a preset angle is reached, all gimbals are controlled to lock synchronously in position. The gimbal position parameters and locking points corresponding to that angle are recorded, completing the circumferential rotation and positioning calibration of all adjustable gimbals. This ensures that the gimbal rotation angles of all shooting units are consistent and synchronized, with a locking angle deviation ≤ ±0.1°. During camera parameter calibration, Zhang's calibration method is used to calibrate the intrinsic and extrinsic parameters of all cameras, obtaining parameters such as the camera intrinsic parameter matrix K and distortion coefficients. The distortion coefficient calibration accuracy is ≤ 0.001. Multi-view checkerboard calibration is used to complete the camera spatial pose calibration, determining the position and attitude of each camera relative to the world coordinate system. The intrinsic and extrinsic parameters and spatial pose information of all cameras, as well as the rotation and positioning parameters of the adjustable gimbals, are recorded and stored, providing a foundation for subsequent 3D reconstruction.

[0050] Step 2: Circumferential Rotation - Multi-view Synchronous Image Acquisition. After the experiment starts, the synchronous control and data processing system executes a cyclic acquisition process of synchronous stepping rotation - synchronous locking - simultaneous full-camera synchronous shooting - synchronous stepping rotation again, completing multiple rounds of synchronous acquisition across the entire field of view of a single section of the water tank, and completing the synchronous image acquisition of multiple moments throughout the entire experimental process according to a preset time sequence. This embodiment supplements typical scene parameters and abnormal working condition handling to ensure the integrity and synchronization of the acquired data, and the specific implementation process takes into account both generality and specificity. After the experiment starts, the synchronous control and data processing system starts the cyclic acquisition process of "synchronous stepping rotation - synchronous locking - simultaneous synchronous shooting - synchronous stepping rotation again" to ensure the synchronization and standardization of the acquisition actions.

[0051] First, the synchronous control and data processing system sends synchronous rotation commands to all drive actuators, driving all adjustable gimbals and their mounted cameras to rotate synchronously along the circumferential track at preset step angles. During rotation, the gimbal position information is fed back in real time, with a position feedback accuracy ≤0.05°. When all gimbals synchronously rotate to the preset shooting angle, the position sensors send a position signal, and the synchronous control and data processing system sends a lock command. All adjustable gimbals immediately lock synchronously to ensure a stable shooting angle, with a locking angle deviation ≤±0.1°. After locking stabilizes, the angle is maintained for 500μs to ensure no vibration interference. Once stabilized, the synchronization control and data processing system sends a synchronization trigger signal to all cameras, controlling all cameras to simultaneously acquire multi-view image data of the water tank experimental area. The hardware synchronization trigger delay is ≤1ms, and the software synchronization trigger delay is ≤10ms. The image data is transmitted to the data processing system in real time and stored in RAW format to ensure that image details are not lost. After a single-angle shot is completed, the synchronization control and data processing system sends an unlock command, and all adjustable gimbals unlock synchronously. The drive actuator continues to drive the gimbals to rotate synchronously to the next preset shooting angle, repeating the above locking and shooting process.

[0052] Once the gimbal completes one full rotation and captures images at all preset angles, the single-moment full-view and multi-view image acquisition of the water tank is complete. According to the experimental progress, the above cyclical acquisition process is repeated according to a preset time sequence to complete the synchronous acquisition of full-view images at multiple moments throughout the experiment, ensuring the capture of key dynamic information. The preset time sequence interval is 0.1-1 seconds, which can be adjusted according to the dynamic characteristics of the experiment. During the acquisition process, abnormal conditions must be handled carefully. The synchronous control and data processing system should detect abnormal conditions such as gimbal rotation jamming, camera trigger failure, and image frame loss in real time. If gimbal rotation jamming is detected, the rotation command should be stopped immediately, an alarm signal should be sent, and a backup gimbal should be activated. Each shooting unit should have one backup gimbal to continue the acquisition process. If camera trigger failure is detected, the trigger signal should be resent immediately. If three consecutive trigger failures occur, the backup camera should be activated to ensure uninterrupted data acquisition. If image frame loss is detected, the time and location of the lost frame should be recorded. After the experiment, data should be interpolated from adjacent moments to ensure data integrity.

[0053] To adapt to different experimental scenarios, this embodiment provides complete implementation parameters for three typical scenarios to ensure the feasibility of the technical solution. In the long-range flume observation scenario, the flume dimensions are 50m in length, 2m in width, and 1.5m in depth. The experiment type is a long-term water and sediment transport experiment. The position adjustment mechanism uses a dual-track system, which is arranged in parallel on both sides of the flume. There are 10 sets of shooting units, with an axial spacing of 5m between adjacent shooting units. The overlap of the field of view in the core observation area is ≥70%. The shooting support frame adopts a closed-loop ring form. Each shooting unit has 3 cameras, a frame rate of 50fps, and a resolution of 1280×720. The rotation step angle is 45° / step. The synchronization method is software synchronization, with a trigger delay of ≤10ms. The acquisition time series interval is 1s, and the acquisition time is 120min. The global calibration target is set up with movable targets in the overlapping area of ​​adjacent shooting units and one set of fixed targets every 10m along the axis, for a total of 5 sets of fixed targets. The water refraction correction model parameters are a refractive index of 1.33 and a refraction angle correction coefficient of 0.98. The reconstruction accuracy requirements are an axial cumulative error of ≤2mm and an accuracy deviation of ≤3% between water-containing and waterless scenes.

[0054] In a small-scale high-speed transient experiment, the water tank measures 8m in length, 0.5m in width, and 0.3m in depth. The experiment type is a high-speed transient experiment on dam-break evolution. A single-track system is used for the position adjustment mechanism, positioned on one side of the water tank. Two camera units are deployed, with an axial spacing of 4m between adjacent units. The overlap of the field of view in the core observation area is ≥80%. The camera support frame uses a C-shaped opening. Each camera unit has 5 cameras, a frame rate of 1000fps, and a resolution of 1920×1080. The rotation step angle is... The capture time is 15° / step; hardware synchronization is used, with a trigger delay of ≤1ms; the acquisition time series interval is 0.1s, and the acquisition duration is 10s; the global calibration target deployment is as follows: one set of movable targets is deployed in the overlapping area of ​​adjacent shooting units, and one set of fixed targets is deployed every 10m along the axis. Since the water tank is 8m long, one set of fixed targets is deployed; the water refraction correction model parameters are a refractive index of 1.33 and a refraction angle correction coefficient of 0.97; the reconstruction accuracy requirement is an axial cumulative error of ≤1mm, and an accuracy deviation of ≤2% between water-containing and waterless scenes.

[0055] In the long-term time series experiment scenario of water and sediment two-phase flow, the water tank dimensions are 20m in length, 1m in cross-sectional width, and 0.8m in depth. The experiment type is a long-term time series evolution experiment of water and sediment two-phase flow. The position adjustment mechanism uses a fixed pre-embedded bracket. The number of shooting units is 5 groups, with an axial spacing of 4m between adjacent shooting units, and the field of view overlap of the core observation area is ≥70%. The shooting support frame uses a multi-column structure. Each shooting unit has 4 cameras, a frame rate of 200fps, and a resolution of 1920×1080. The rotation step angle is 30° / step. The synchronization method is software synchronization, with a trigger delay of ≤10ms. The time series acquisition interval is 0.5s, and the acquisition duration is 60min. The global calibration target deployment is as follows: movable targets are deployed in the overlapping area of ​​adjacent shooting units, and one set of fixed targets is deployed every 10m axially, for a total of 2 sets of fixed targets. The water refraction correction model parameters are a refractive index of 1.34 and a refraction angle correction coefficient of 0.99. The reconstruction accuracy requirement is an axial cumulative error of ≤2mm, and an accuracy deviation of ≤3% between water-containing and waterless scenes.

[0056] Step 3: Single-moment 3D scene reconstruction. For multi-view synchronous images acquired at the same time, a multi-view stereo matching 3D reconstruction algorithm is adopted. The algorithm sequentially performs image coarse-grained noise reduction pre-screening, water scene-specific fine pre-processing, feature matching optimization, water refraction geometry correction, sparse reconstruction and fluid temporal geometric continuity filtering, dense reconstruction and multi-round data cross-consistency calibration to restore a high-precision 3D model of the experimental scene in the tank at that moment.

[0057] Single-moment 3D scene reconstruction first involves coarse-grained image noise reduction pre-screening. Based on the temporal image sequence acquired in step 2, the inter-frame gray-level variance statistics of local pixel regions (such as a 3×3 neighborhood window) in multiple consecutive images are first calculated. Specifically, a local pixel window is defined. Mean of gray variance within for: in, Represents the target pixel A local neighborhood window centered on the target; Represents a local neighborhood window Total number of pixels within; Represents the number of consecutively extracted temporal image frames; The spatial coordinates within the window are The pixels at time grayscale value; This means that the point is at The average grayscale value over time within a continuous frame of an image.

[0058] Based on the above mean gray-level variance Construct a feature mask discriminant for high-light scintillation in water bodies. :when Greater than the preset coarse-grained high-frequency noise cutoff threshold hour, Assign a value of 0; when Less than or equal to hour, The value is assigned to 1. When the mean gray-level variance of a pixel's neighborhood in multiple consecutive frames of images is higher than the preset coarse-grained threshold (i.e., the mask value is 0), the region is determined to be a high-frequency flicker noise region on the water surface, and the subsequent process masks this region, reducing the computational scale. This coarse-grained noise reduction pre-screening step completes the binary distinction between the noise domain and the effective domain of the entire image with low computational overhead. The feature mask generated by this step marks the high-frequency flicker noise region in the image, providing prior information on noise distribution for subsequent fine preprocessing.

[0059] Based on the coarse-grained noise reduction pre-screening and noisy region labeling, the system sequentially performs fine-grained preprocessing operations on the screened and retained image regions to eliminate systematic distortions during imaging and residual interference introduced by the aquatic environment. This provides a high-quality image base for downstream feature extraction and matching. The process includes distortion correction, reflection suppression, and grayscale enhancement. Distortion correction is based on the distortion coefficients obtained from camera calibration, correcting the image pixel coordinates to restore the realistic image scene. Specifically, during distortion correction, the radial distortion coefficient is combined with... With tangential distortion coefficient For physical coordinates Perform the following nonlinear correction: in, and These are the corrected physical coordinates. Indicates the relative radial distance from the uncorrected point to the optical center ( ), Represents the radial distortion coefficient of the camera. This represents the tangential distortion coefficient of the camera.

[0060] Reflection suppression utilizes the characteristics of polarizing mirror acquisition and a dedicated image segmentation algorithm to remove residual highlight areas on the water surface; noise reduction employs a 3×3 Gaussian filter, whose two-dimensional Gaussian kernel weight function is defined as: in, and These represent the horizontal and vertical coordinate offsets of the calculated point relative to the center point within the pixel's neighborhood, respectively. This represents the standard deviation of the spatial Gaussian filter and is used to control the smoothness of the image. It represents the base of the natural logarithm.

[0061] Gray-level enhancement employs a histogram equalization algorithm, which uses a cumulative distribution function to equalize the gray levels of the original image. Mapped to enhanced grayscale levels : in, Represents the total number of possible gray levels in the image; Represents the grayscale index of the original image to be mapped; Represents the first in the original image grayscale value The grayscale probability density function is used. After the above refined preprocessing, the image signal-to-noise ratio is improved to over 30dB, water reflection and light and shadow fluctuation interference are suppressed, and the image quality meets the requirements for high-precision feature extraction. The preprocessed image data is then fed into the feature matching and optimization pipeline.

[0062] The feature matching optimization step is performed on the preprocessed high-quality image. Through scale-space feature extraction, descriptor matching, and outlier removal strategies, the association of corresponding points between images from different viewpoints is established. First, Gaussian difference scale-space feature points are extracted to enhance the feature robustness of regions with smooth grayscale in water bodies. in, A two-dimensional Gaussian difference image generated in the representative scale space; and These represent the scale parameters as follows: and The two-dimensional Gaussian kernel function; This represents the original two-dimensional image that was input.

[0063] Sub-pixel localization is then achieved by offset correction using a ternary second-order Taylor series expansion. in, The sub-pixel-level coordinate offset vector representing the feature point in the interpolation space; Represents the difference scale space function The inverse of the Hessian matrix at the sampling point; Represents the difference scale space function The gradient vector at the sampling point.

[0064] After achieving sub-pixel precise localization of feature points, the system generates a SIFT feature descriptor for each feature point. First, based on the gradient direction distribution of the local neighborhood of the feature point, a gradient direction histogram is calculated. The direction corresponding to the peak of the histogram is taken as the principal direction of the feature point, achieving rotation invariance of the descriptor. Then, with the feature point as the center, its coordinate axes are rotated to the principal direction. A 16×16 pixel neighborhood window is selected around the feature point, and this window is evenly divided into 4×4 sub-regions, each covering 4×4 pixels. Within each sub-region, the cumulative gradient magnitudes in eight directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°) are calculated, and the cumulative values ​​in each direction are Gaussian weighted to enhance the contribution of the neighborhood center. Finally, the gradient statistics in the eight directions of each of the 4×4 sub-regions are concatenated to generate a 4×4×8=128-dimensional SIFT feature descriptor vector for each feature point. The generated SIFT feature descriptor (e.g.) and After that, calculate the Euclidean distance for matching: in, Represents two feature descriptors and The Euclidean distance between them; and They represent feature descriptors respectively. and In the The feature values ​​are in 128 dimensions; 128 represents the total number of vector dimensions of the SIFT feature descriptor.

[0065] Then, Lowe's Ratio Test was introduced to eliminate mismatches in water refraction: in, This represents a target feature descriptor in the reference image. Indicates the target image with The nearest matching descriptor Indicates the target image with The second closest matching descriptor, This represents the L2 norm of the eigenvector.

[0066] Finally, the RANSAC algorithm is used to eliminate exterior points based on the epipolar geometric constraint equations. in, and These represent the homogeneous coordinate vectors of the matching points in the target image and the reference image, respectively; Represents the fundamental matrix between two views; superscript Represents the transpose of a matrix or vector. Ensures feature matching accuracy ≥ 95%. After matching is complete, the pixel coordinates of all feature points are then processed through water refraction geometry correction.

[0067] In the water refraction correction stage, based on the established water-air interface refraction model, corrections are performed according to Snell's Law: in, and These represent the refractive indices of air and water, respectively. and These represent the angle of incidence of light in air and the angle of refraction in water, respectively. As light travels from the surface of the target object inside the water tank to the camera outside the tank, it is refracted and deflected at the water-air interface, causing a deviation between the observed pixel positions of feature points in the image and their actual spatial positions. This step, based on a pre-calibrated water refraction correction model (typical water refractive index = 1.33; air refractive index ≈ 1.0), performs geometric correction to this deviation: First, the pixel coordinates of the matched feature points are back-projected through the camera intrinsic parameter matrix to obtain the normalized incident ray direction on the air side. Based on this, the angle between this ray and the normal to the water-air interface—that is, the air-side angle of incidence—is calculated. Then, based on Snell's law of refraction, the angle of refraction on the water side is solved. The true direction of light propagation in the water medium is determined; then, combined with the spatial position of the camera's optical center and the spatial plane equation of the water-air interface, the true spatial coordinates of the corrected feature points are obtained. The corrected coordinates, along with the camera's intrinsic and extrinsic parameters calibrated in step 1, are sent to the sparse reconstruction module to ensure that the reconstruction accuracy in water scenes deviates from that in waterless scenes by ≤3%.

[0068] After obtaining the multi-view matched feature points and accurate camera pose parameters after refraction correction, the system enters the 3D sparse reconstruction stage. Based on the multi-view geometry principle, the initial 3D coordinates of spatial points are calculated using triangulation intersection, and global refinement is performed through bundle adjustment. Simultaneously, to address the non-physical transient flying point interference unique to water scenes, an anomaly identification and removal mechanism based on the fluid temporal geometric continuity is introduced. Specifically, in the sparse reconstruction stage, a known-constrained projection basis is first established using a pinhole imaging model: in, Scale factor; Represents the homogeneous coordinates of the feature point in the image pixel coordinate system; The intrinsic parameter matrix representing the camera; and These represent the camera's rotation matrix and translation vector, respectively. Represents the homogeneous coordinates of a 3D point in the world coordinate system.

[0069] Then through multiview Figure 3 Angular intersection solution for initial sparse 3D point cloud: in, Representing the Projected homogeneous coordinates of two-dimensional feature points in a view; The camera projection matrix representing the corresponding view; The vector represents the position of the 3D point to be determined in the world coordinate system. This represents the vector cross product operation.

[0070] Then, point cloud coordinates are refined collaboratively using pose-fixed bundle adjustment (BA): in, This indicates a minimization operation. This indicates that a summation operation is performed on all spatial points across all visible viewpoints. Represents a robust kernel function. Point In view Actual observed pixel coordinates This represents the reprojection function that maps points in three-dimensional space to two-dimensional pixels. Represents the camera intrinsic parameter matrix. They represent the first Given rotation and translation parameters for each view This represents the coordinates of the 3D point to be optimized. It represents the square of the L2 norm of a vector (i.e., the square of the Euclidean distance).

[0071] Next, a temporal continuity filtering optimization module is introduced. This module is activated only after at least one prior point cloud has been constructed; if the process is in the first point cloud construction phase, it is not activated. This module constructs a non-intrusive anomaly identification criterion based on the temporal geometric continuity of the fluid. Water flow evolution follows continuum mechanics, while surface high-gloss noise exhibits non-physical high-frequency abrupt changes. The system calculates and outputs the spatial discrete point cloud at the current moment. The reliable point set that was cleaned and anchored in the previous moment is invoked. Construct a KD-Tree topology index. The system queries via cross-frame spatial mapping. Any point to be verified exist Find the absolute nearest neighbor in the matrix and calculate its deformation displacement field. : in, Represents the point to be verified Relative to its absolute nearest neighbor The deformation displacement distance; This represents the Euclidean distance. To adapt to different deformation scales under varying flow field conditions, the system extracts the global mean of the deformation displacement field in real time. with standard deviation The calculation formulas are as follows: as well as: in, This represents the total number of spatial points within the current timeframe. Based on the aforementioned statistical moments, the system constructs an adaptive physical truncation threshold. : in, This represents the anomaly tolerance gain factor, typically with a value of 3.0. Based on this, the geometric distortion discrimination functional is defined. When the indicator function result is 1, meaning the spatial displacement of the point to be verified exceeds the evolutionary statistical limit, the system determines it to be a non-physical transient flying point caused by high-frequency light and shadow interference and performs hard rejection, achieving high-precision dynamic tracking with low computing power overhead.

[0072] After three stages of processing—triangulation, bundle adjustment, and time-series geometric continuity filtering—the system outputs a high-precision sparse point cloud whose spatial coordinates possess physical continuity and consistency. This sparse point cloud serves as the initial geometric prior and is then passed to the dense reconstruction and cross-consistency calibration stages.

[0073] Dense reconstruction is based on the PatchMatch Stereo algorithm, combined with the normalized cross-correlation (NCC) measure: in, Representative space surface The photometric consistency cost function value; The covariance function representing the color or grayscale distribution of two image patch windows; Represents a local neighborhood window image patch in the reference image; This represents the corresponding neighborhood window image patch mapped to the observed image; and These represent the grayscale standard deviations of the image patches in the reference image window and the mapped observation image window, respectively. After the NCC (Network Convergence) process, pixel-level depth maps and normal maps corresponding to each viewpoint are output. After the depth maps are generated, a geometric consistency check is performed, cross-referencing the estimated depth differences of the same spatial point under different viewpoints, eliminating unreliable depth estimates exceeding a preset threshold, and retaining reliable depth values ​​that meet the consistency constraints. The consistency-checked multi-view depth maps are then fused into a unified dense 3D point cloud and output.

[0074] After the dense point cloud is generated, the system triggers multiple rounds of data cross-calibration, eliminating biases through redundancy consistency evaluation from multiple viewpoints to ensure the purity of the spatial geometric benchmark. The system utilizes data acquired at different rotation angles at the same time, as well as shared observation data from adjacent imaging units within overlapping fields of view, to construct a weighted evaluation function based on reprojection consistency. : in, Representing spatial points Evaluation value of reprojection bias under multi-viewpoint redundant observation; Representing the containment point The set of all redundant observation perspectives; Representative perspective The confidence weights (adaptively adjusted based on camera image quality); Representative point From the perspective The actual observed pixel coordinates below; This represents the reprojection mapping function. The system sets the reprojection deviation threshold. , be on point Evaluation value If the point cloud cluster is determined to have local spatial drift or imaging interference, it is directly removed or multi-view weighted correction is performed to obtain a high-precision dense point cloud.

[0075] After dense matching and multiple rounds of cross-validation, the system finally outputs a high-precision 3D dense point cloud model of the experimental scene inside the tank at that moment. Its spatial geometric reference has been verified for consistency of redundant data from multiple perspectives and can be directly used as the standard input for the subsequent full-time dynamic stitching in step 4. At this point, the reconstruction of the 3D scene at a single moment is completed. Step 3 is then executed one by one for each moment of the entire experiment to obtain the 3D model sequence for each moment, and then the process proceeds to step 4.

[0076] Step 4: Dynamic Reconstruction of the Entire Experiment. For the single-frame 3D models at each moment of the entire experiment output in Step 3, the following processing is performed sequentially: First, the global coordinate transformation matrix calibrated before the experiment is called to unify the point clouds reconstructed by each shooting unit to the same global coordinate system, completing the point cloud stitching of multiple shooting units; Second, to address the axial cumulative error easily generated by long-distance stitching of multiple arrays in a narrow and long water tank, a global bundle adjustment joint optimization with an anisotropic penalty term is introduced. A maximum penalty weight is applied along the radial direction of the water tank in the loss function, forcing the translation vector drift to converge to the geometric center line of the water tank, compressing the error release degrees of freedom to a one-dimensional axial space, and completely suppressing axial cumulative drift; Third, the 3D models at different moments are temporally aligned, and an iterative nearest-point algorithm is used to eliminate temporal deviations and ensure temporal continuity based on the overlapping areas of adjacent moments; Finally, the optimized 3D models at different moments are stitched and fused according to the time sequence to generate a dynamic 3D evolution video of the experimental scene, realizing the 3D visualization and reproduction of the entire experimental process.

[0077] Before proceeding with global optimization, the system first unifies the point cloud coordinates of multiple imaging units. Based on the global coordinate transformation matrix calibrated and saved in step 1, the independently reconstructed 3D point clouds of each imaging unit are transformed from their respective local coordinate systems to a unified global coordinate system. This eliminates coordinate reference differences between different imaging units, providing a consistent geometric reference framework for subsequent cross-unit joint optimization and stitching. This step directly reuses the transformation matrix obtained through least squares calibration and global bundle adjustment co-optimization in the pre-experiment calibration stage, eliminating the need for repeated online calculations and ensuring the accuracy and efficiency of coordinate unification.

[0078] In the global joint optimization stage of multi-array splicing, to address the axial cumulative drift easily generated by long water tanks, the system reconstructs the underlying nonlinear optimization objective function and introduces an axial drift suppression step based on the topological prior of narrow and long spaces. The system constructs a joint loss function that includes an anisotropic penalty term. : in, This represents the reprojection error term under Huber robust kernel function constraints; Representing three-dimensional points In the The coordinates of the observed pixels on the camera; Represents the combination of camera intrinsic parameter matrix Rotation matrix With translation vector The projection function; Represents the topological constraint regularization coefficient; The radial drift diagonal weight matrix, specifically designed for narrow and elongated water tanks, is defined as follows: In this matrix, the penalty weight along the axial direction of the water tank (X-axis) is 0, while a very large Laplace penalty weight is assigned in the two radial dimensions perpendicular to the axial direction of the water tank (Y-axis and Z-axis).

[0079] This mechanism forces the camera translation vector to converge radially to near the geometric centerline of the water tank, and forcibly compresses the error release degrees of freedom into a one-dimensional axial space, thus solving the problem of cumulative error divergence in long-distance multi-view stitching.

[0080] After completing global joint optimization, the process proceeds to temporal alignment and video generation. Based on 3D models from different moments throughout the experiment, the models are stitched together according to the acquisition time sequence. For the globally optimized 3D models at each moment, the Iterative Closest Point (ICP) algorithm is used to eliminate deviations caused by temporal acquisition, ensuring temporal continuity, based on the overlapping areas of 3D models from adjacent moments. Finally, using the high-quality dense point cloud after global optimization and temporal alignment as input, Poisson surface reconstruction is performed to generate 3D continuous surface models of experimental scenarios such as water bodies and sediment deposition. A dynamic 3D evolution video with a frame rate of 25fps is then rendered and output, clearly reproducing the dynamic evolution process of the entire experiment, such as dam break evolution, water and sediment transport, and landslide dam failure. The video can be previewed, stored, and exported in real time, providing comprehensive support for the visualization analysis, post-processing, and results presentation of experimental data.

[0081] Furthermore, the synchronous control and data processing system also features post-processing capabilities, enabling quantitative analysis of the reconstructed 3D model and dynamic evolution video. It extracts key experimental parameters such as breach width, depth, sediment transport, and water flow velocity, automatically generating experimental reports and reducing manual data processing workload while improving experimental efficiency. Simultaneously, the system supports data backup and recovery, allowing for dual local and cloud backups of collected image data, 3D models, dynamic videos, and experimental reports, ensuring the security and integrity of experimental data.

[0082] To verify the effectiveness and superiority of this invention, two sets of physical verification experiments were designed, targeting waterless and water-containing scenarios respectively. The reconstruction accuracy, hardware cost, and deployment flexibility of the present invention's solution were compared with existing technical solutions. The experimental data are authentic and reproducible, as detailed below:

[0083] Experiment 1: Verification Experiment of 3D Reconstruction of Wooden Target Object in Waterless Scene. The experimental equipment used was a narrow and long hydraulic experimental tank with a length of 10m, a cross-sectional width of 0.8m, and a depth of 0.5m. A wooden target object with dimensions of 100cm×50cm×30cm was placed at the bottom of the tank. The present invention's scheme deployed two sets of shooting units, with an axial spacing of 5m between adjacent shooting units and a field of view overlap of 75%. Each shooting unit used 3 cameras (frame rate 100fps, resolution 1920×1080), with an adjustable gimbal rotation step angle of 30° / step, and software synchronization was used. The existing technical solution used a dense deployment scheme of 6 cameras (the same camera model as the present invention) fixed on a single cross-section. Experimental results show that the 3D model of the wooden target object reconstructed by the present invention has a size error of ≤0.5mm, a surface texture reproduction degree of ≥98%, and an axial splicing error of ≤0.8mm; the existing technology has a size error of ≤0.6mm, a surface texture reproduction degree of ≥97%, and an axial splicing error of ≤1.2mm; the present invention requires only half the number of cameras as the existing technology, reduces hardware costs by 65%, shortens deployment time by 40%, and allows for flexible adjustment of the shooting unit position to adapt to water tanks of different lengths.

[0084] Experiment 2: Verification Experiment of Two-Phase Flow and Dam-Break Evolution in Dynamic Scenes with Water. A dynamic dam-break scenario was simulated in a 50m long large flume. Under harsh engineering conditions with severe water reflection, refraction deformation, and the high likelihood of cumulative errors in a long flume, the system accurately captured high-frequency evolution data at the moment of the breach, thanks to the adaptive frame-skipping mechanism based on dynamic information entropy introduced in this invention. Through KD-Tree fluid temporal geometric continuity filtering, non-physical "flying point" noise generated by water surface highlights and transparent refraction in the point cloud was eliminated at a rate of 99.2%. Furthermore, combined with joint optimization using anisotropic axial constraints, the axial cumulative error of the full-area reconstruction of the 50m flume was strictly controlled to 1.8mm (better than the standard of ≤2mm). Experimental measurements showed that the reconstruction accuracy deviation between the transient scene with water and the static scene without water was only 2.6% (≤3%). Physical verification fully demonstrates that the reconstruction accuracy of the present invention is no less than that of the existing high-precision dense deployment scheme, while significantly reducing hardware costs. It also has the advantages of dynamic scene adaptation and flexible deployment, which can effectively meet the dynamic three-dimensional observation requirements of narrow and long hydraulic experimental tanks, thus verifying the scientific nature, practicality and superiority of the present invention's technical solution.

Claims

1. A dynamic three-dimensional reconstruction system adapted to narrow and long hydraulic experimental tanks, characterized in that, The system includes a position adjustment mechanism, multiple independently movable shooting units, and a synchronization control and data processing system. The position adjustment mechanism is positioned around a narrow, elongated hydraulic experimental tank, extending in the same direction as the tank's axis, serving as a shared displacement reference for the entire system. Multiple independently sliding movable locking support components are mounted on the position adjustment mechanism; each component can slide freely along the axis of the position adjustment mechanism and be locked in a preset position. Multiple shooting units are arrayed along the tank's axis, with adjacent shooting units having an overlap area of ​​≥60%. Each shooting unit is connected to the position adjustment mechanism via a movable locking support component, and each shooting unit includes a shooting frame and an image acquisition component. The synchronization control and data processing system communicates with the adjustable gimbals, drive actuators, and cameras of all shooting units, enabling synchronized gimbal rotation, locking, and camera shooting, as well as performing 3D reconstruction processing based on the acquired images.

2. The dynamic three-dimensional reconstruction system for adapting to narrow and long hydraulic experimental tanks according to claim 1, characterized in that, The shooting support frame is arranged around the water tank in a circumferential manner, and a circumferential track is set on the frame. The shooting support frame adopts a closed-loop ring form, a C-type / U-type non-closed-loop open form, or a multi-column structure. The movable locking support component adopts bolt locking, clamp locking, magnetic locking, or pneumatic locking methods.

3. The dynamic three-dimensional reconstruction system for adapting to narrow and long hydraulic experimental tanks according to claim 1, characterized in that, The image acquisition component includes N≥3 adjustable gimbals, a drive actuator, and cameras. The number of cameras in a single shooting unit for high-speed transient experiments is ≥5, and the number of cameras in a single shooting unit for long-term conventional experiments is ≥3. The adjustable gimbals are evenly distributed along the circumferential track of the shooting support frame. Each adjustable gimbal is equipped with one camera. The drive actuator is connected to the adjustable gimbal and is used to drive all adjustable gimbals and cameras to synchronously perform stepping rotational motion along the circumferential track. After the adjustable gimbal rotates to the preset shooting angle, the locking angle deviation is ≤±0.1°. All cameras are equipped with linear polarizing filters. High-speed transient experiments use high-speed cameras with a frame rate ≥1000fps and a resolution ≥1920×1080, while long-term experiments use cameras with a frame rate of 50-200fps and a resolution ≥1280×720.

4. The dynamic three-dimensional reconstruction system for adapting to narrow and long hydraulic experimental tanks according to claim 1, characterized in that, The hardware synchronization triggering delay of the synchronization control and data processing system is ≤1ms, and the software synchronization triggering delay is ≤10ms. It can trigger the camera to take pictures after the gimbal is locked and stabilized with a delay of ≤500μs. The multi-view stereo matching 3D reconstruction algorithm adopted by the synchronization control and data processing system includes incremental SfM algorithm, bundle adjustment algorithm, outlier filtering algorithm and dense reconstruction algorithm.

5. The dynamic three-dimensional reconstruction system for adapting to narrow and long hydraulic experimental tanks according to claim 1, characterized in that, The system also includes multiple global coordinate system calibration modules for shooting units, equipped with global calibration targets. Through target placement, coordinate transformation, and registration, the global coordinates of multiple shooting units can be unified, achieving an axial cumulative error of ≤2mm for the reconstruction of the entire area of ​​a water tank within 50m. The system also includes a dedicated adaptation module for water conservancy scenes, including camera polarizing filter configuration, water refraction correction model, and image preprocessing methods for reflection suppression, which can ensure that the reconstruction accuracy in water scenes deviates from that in waterless scenes by ≤3%.

6. A dynamic three-dimensional reconstruction method adapted to narrow and long hydraulic experimental tanks, characterized in that, The implementation based on the dynamic three-dimensional reconstruction system according to any one of claims 1-5 includes the following steps: Step 1: Pre-experiment deployment and calibration. According to the experimental plan and the size of the water tank, deploy the corresponding number of shooting units, adjust and fix the position of the shooting units, complete the calibration of the global coordinate system of multiple shooting units, the circumferential rotation positioning calibration of the adjustable gimbal, and the calibration of the camera's internal and external parameters and spatial pose. At the same time, complete the calibration of the camera polarizing filter adjustment and the water refraction correction model parameters. Step 2: Circumferential rotation - multi-view synchronous image acquisition. After the experiment starts, the synchronous control and data processing system executes a cyclic acquisition process of synchronous stepping rotation - synchronous locking - simultaneous full-camera synchronous shooting - synchronous stepping rotation again, to complete multiple rounds of synchronous acquisition of the full-view range of a single cross section of the water tank, and completes the full-view synchronous image acquisition of multiple moments in the entire experimental process according to the preset time sequence. Step 3: Single-moment 3D scene reconstruction. For the multi-view synchronous images acquired at the same moment, the following steps are performed in sequence: coarse-grained noise reduction pre-screening, water scene-specific fine pre-processing, feature matching optimization, water refraction geometry correction, sparse reconstruction and fluid temporal geometric continuity filtering, dense reconstruction, and multi-round data cross-calibration to restore the 3D model of the experimental scene in the tank at that moment. Step 4: Dynamic reconstruction of the entire experimental process. The 3D models at different times during the entire experimental process are stitched and fused according to the acquisition time sequence. The point cloud stitching of multiple shooting units, axial error accumulation suppression and temporal global calibration are completed. Finally, a dynamic 3D evolution video of the experimental scene is generated, realizing the 3D visualization reproduction of the entire experimental process.

7. The dynamic three-dimensional reconstruction method for adapting to narrow and long hydraulic experimental tanks according to claim 6, characterized in that, The specific process of calibrating the global coordinate system of multiple shooting units in step 1 is as follows: Select waterproof and anti-reflective checkerboard calibration targets, deploy movable global calibration targets in the overlapping area of ​​adjacent shooting units, and deploy one set of fixed global calibration targets every 10m along the axis. Collect synchronous images of the calibration targets from no less than 10 different perspectives, calculate the transformation matrix between the local coordinate system and the global coordinate system of each shooting unit, perform preliminary calibration using the least squares method, optimize the transformation matrix using the global bundle adjustment algorithm, and finally save the calibration results for subsequent point cloud stitching and error suppression.

8. The dynamic three-dimensional reconstruction method for adapting to narrow and long hydraulic experimental tanks according to claim 6, characterized in that, The specific process of the cyclic acquisition process described in step 2 is as follows: the drive actuator drives the adjustable gimbal and camera to rotate synchronously at a preset step angle of 15°-45° / step, with a position feedback accuracy of ≤0.05°; after all gimbals rotate to the preset angle, they lock synchronously, with a locking angle deviation of ≤±0.1°, and remain locked for 500μs after stabilization; all cameras are synchronously triggered to acquire multi-view images at the same time; after a single-angle shot is completed, the gimbal is synchronously unlocked, and the drive continues to rotate to the next preset angle to repeat the acquisition process until a single-section full-view acquisition is completed.

9. The dynamic three-dimensional reconstruction method for adapting to narrow and long hydraulic experimental tanks according to claim 6, characterized in that, Step 3 involves the following steps: coarse-grained noise reduction pre-screening, refined pre-processing specific to the water scene, feature matching optimization, water refraction geometry correction, sparse reconstruction and fluid temporal geometric continuity filtering, dense reconstruction, and multi-round data cross-calibration. The specific process is as follows: Coarse-grained noise reduction pre-screening: First, based on the time-series image sequence acquired by the high-speed camera, the inter-frame gray-level variance statistics of local pixel regions are calculated. When the mean gray-level variance of a pixel's neighborhood is higher than the preset coarse-grained threshold, it is determined to be a high-frequency flicker noise region of the water body and is suppressed in the downstream calculation. Water scene-specific refined preprocessing: After noise reduction pre-screening, distortion correction, noise reduction, grayscale enhancement and polarization reflection suppression are performed sequentially on the image area to ensure that the signal-to-noise ratio of the preprocessed image is ≥30dB; Feature matching optimization: Gaussian difference is used to extract feature points and generate SIFT descriptors. Lowe's Ratio Test criterion is introduced to eliminate mismatches caused by refraction. Outliers are eliminated based on RANSAC algorithm and epipolar geometry constraints to ensure feature matching accuracy ≥95%. Water Refraction Geometric Correction: Based on the law of refraction, physical coordinate compensation is performed on the coordinates of the matched feature points to eliminate reconstruction errors caused by water refraction and ensure that the reconstruction accuracy deviation between water scenes and waterless scenes is ≤3%. Sparse Reconstruction and Fluid Temporal Geometric Continuity Filtering: Sparse point clouds are calculated using multi-view triangulation and refined by bundle adjustment; KD-Tree topological index is constructed to calculate deformation displacement field; points with spatial displacement exceeding the threshold are identified as non-physical flying points and hard-canceled by using an adaptive physical truncation threshold based on global mean and standard deviation. Dense reconstruction and multi-round data cross-calibration: Based on the filtered sparse point cloud, the PatchMatch Stereo algorithm is used for depth and normal estimation; point cloud data with different rotation angles and overlapping fields of view of adjacent shooting units at the same time are used to perform global consistency registration evaluation and cross-validation, and local data with large geometric deviations are eliminated; finally, a high-precision three-dimensional dense point cloud is obtained.

10. The dynamic three-dimensional reconstruction method for adapting to narrow and long hydraulic experimental tanks according to claim 6, characterized in that, The specific process of dynamically reconstructing the entire experimental process described in step 4 is as follows: Initial point cloud stitching: The global coordinate transformation matrix calibrated before the experiment is called to unify the temporal 3D point clouds collected by each imaging unit to the same global coordinate system, and complete the stitching of point clouds from multiple imaging units. Axial drift suppression based on narrow and long spatial topological priors: To address the error accumulation that easily occurs in long-distance stitching of multiple arrays in narrow and long water tanks, a global bundle adjustment method with an anisotropic penalty term is used for joint optimization. A diagonal weight matrix is ​​set in the joint loss function, with a large Laplacian penalty weight on the two radial dimensions perpendicular to the water tank axis and a small penalty weight on the direction along the water tank axis. The translation vector of the camera pose is forced to converge radially to near the geometric center line of the water tank, and the error release degree of freedom is restricted to one-dimensional space along the water tank axis, thereby suppressing the axial cumulative drift of long-distance array stitching. Temporal alignment and video generation: Finally, for the globally optimized 3D models at different times, the iterative nearest point algorithm is used to eliminate temporal deviations based on the overlapping areas of adjacent times. The optimized 3D models are then stitched and fused together according to the time sequence to generate a dynamic 3D evolution video of the experimental scene.

Images