Portable cluster three-dimensional rapid modeling system and method

Abstract

The invention relates to the technical field of modeling systems, and discloses a portable cluster three-dimensional rapid modeling system and method, and the system comprises an integrated case unit which is internally integrated with a main computing node and a sub computing node; the power supply and power management module is used for monitoring the load of each computing node in real time and dynamically adjusting power supply; a KVM switching table in the man-machine interaction module is used for supporting a user to switch among a plurality of computing nodes through a keyboard or a mouse; the scheduling and modeling module is used for performing three-dimensional modeling on the target area according to the image data; and the environment adaptation module is used for carrying out real-time monitoring and intelligent regulation and control on the internal environment temperature and humidity of the integrated case unit. According to the method, rapid generation and result export of the three-dimensional model of the target area are realized through a closed-loop process of data acquisition by the unmanned aerial vehicle, cluster node parallel calculation and local aggregation optimization.

Description

Technical Field

[0001] This invention relates to the field of modeling system technology, and in particular to a portable cluster 3D rapid modeling system and method. Background Technology

[0002] With the rapid development of UAV remote sensing, oblique photogrammetry, and computer vision technologies, 3D modeling based on image data has been widely used in emergency mapping, disaster assessment, urban planning, and military reconnaissance. Traditional 3D modeling systems typically rely on fixed high-performance computing centers or cloud servers, which suffer from problems such as long deployment cycles, poor environmental adaptability, and slow response speeds, making it difficult to meet the urgent needs for rapid, efficient, and portable modeling capabilities in complex scenarios such as field operations and disaster sites.

[0003] Existing portable modeling devices mostly employ a single-machine architecture, which, limited by computing resources and heat dissipation capabilities, struggles to support real-time processing of large-scale image data. While distributed cluster solutions offer strong computing power, they are often bulky, consume high power, have inconvenient human-computer interaction, and lack environmental adaptation mechanisms and task scheduling optimization strategies for mobile work scenarios. Furthermore, during the modeling process, if a computing node fails, traditional systems typically lack dynamic fault tolerance and task redistribution mechanisms, easily leading to modeling interruptions or incomplete results. Therefore, there is an urgent need for a portable 3D modeling system that integrates high computing power, strong robustness, good human-computer interaction, and environmental adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide a portable cluster 3D rapid modeling system and method, which solves the problems that existing portable modeling devices mostly adopt a single-machine architecture, making it difficult to support the real-time processing of large-scale image data; while distributed clusters are often bulky, power-consuming, and have inconvenient human-computer interaction, and lack environmental adaptation mechanisms and task scheduling optimization strategies for mobile operation scenarios.

[0005] This invention provides a portable cluster 3D rapid modeling system, the system comprising: An integrated chassis unit, which integrates several computing nodes, and the computing nodes are interconnected through a backplane. The computing nodes include main computing nodes and sub-computing nodes. The power supply and power management module is used to monitor the load of each computing node in real time and dynamically adjust the power supply. A human-computer interaction module is disposed on the integrated chassis unit. The human-computer interaction module includes a KVM switch, a keyboard, a mouse, and a monitor. The keyboard, mouse, and monitor are connected to each computing node through the KVM switch. The KVM switch is used to support users to switch between several computing nodes using a keyboard or mouse. The scheduling and modeling module is used to create a 3D model of the target area based on image data. An environmental adaptation module is installed inside the integrated chassis unit. The environmental adaptation module includes an air-cooling unit and a heat dissipation unit. The environmental adaptation module is used to monitor and intelligently control the temperature and humidity inside the integrated chassis unit in real time.

[0006] Preferably, the scheduling and modeling module includes: a data receiving unit configured to receive image data of the target area; The task scheduling and partitioning unit is configured to import the image data into the main computing node, and divide the image data into several sub-tasks through the main computing node and distribute them to the sub-computing nodes. The task execution unit is configured to control the sub-computing nodes to execute the sub-tasks in parallel and output the sub-task results; The model building unit is configured to aggregate the results of the subtasks through the main computing node, perform global optimization and texture mapping, and build a three-dimensional model of the target region. The model monitoring unit is configured to switch compute nodes via a KVM switcher to view modeling progress or export results.

[0007] Preferably, the data receiving unit is connected to a drone acquisition device, which is used to acquire image data of the target area. The data receiving unit is provided with several data interfaces, and the drone acquisition device transmits the image data to the main computing node through the data interfaces.

[0008] Preferably, the task scheduling and partitioning unit imports the image data into the main computing node, and divides the image data into several sub-tasks through the main computing node, including: The image data is imported into the main computing node and preprocessed, including image distortion correction, image enhancement, and data indexing. By reading the location corresponding to each image data through the established data index, the spatial range boundary of the target area is generated. The target area is divided into several spatial grids using a grid division method. The number of images in each spatial grid is counted, and the spatial grids are dynamically adjusted based on the number of images. An independent subtask is generated for the image data of each spatial grid. The subtask includes image data, preprocessing parameters, and task execution instructions.

[0009] Preferably, when the task scheduling and partitioning unit distributes subtasks to sub-computing nodes, it specifically does so as follows: Obtain the real-time load and CPU computing power of the sub-computing nodes, and determine the task processing capability of each computing node based on the real-time load and CPU computing power; The number of images for each subtask is determined, and the subtasks are distributed based on the number of images and the task processing capacity.

[0010] Preferably, when the task execution unit controls the sub-computing nodes to execute the sub-tasks in parallel, the task execution instructions include feature extraction, feature matching, sparse reconstruction, and dense point cloud generation.

[0011] Preferably, the task scheduling and partitioning unit is further configured to monitor the task execution data of each sub-computing node in real time when executing sub-tasks, the task execution data including the CPU utilization, GPU utilization, memory usage and task progress of the sub-computing nodes; Based on the task execution data, determine whether the corresponding sub-computing node has failed. If a failure occurs, the unfinished sub-tasks of the sub-computing node will be reassigned.

[0012] Preferably, the model building unit aggregates the results of the subtasks through the main computing node, performs global optimization and texture mapping, and constructs a 3D model of the target region, including: The master computing node receives the subtask results uploaded by each sub-computing node. The subtask results include sparse point clouds, dense point clouds, and image features of each spatial grid. The received subtask results are processed to unify coordinates, transforming sparse and dense point clouds generated in the local coordinate system by different sub-computation nodes into the global coordinate system. Global registration optimization is performed based on the unified sparse point cloud and dense point cloud. The edge of the point cloud of adjacent spatial grids is aligned by the iterative nearest point algorithm to eliminate the stitching gap. Based on the mapping relationship between the preprocessed image data and the sparse and dense point clouds in the global coordinate system, texture coordinates are calculated, and image features are fitted onto the surface of the three-dimensional point cloud to generate a three-dimensional model of the target area.

[0013] Preferably, the model monitoring unit switches computing nodes via a KVM switcher to view modeling progress or export results, including: Switch to the monitoring interface of the master compute node via the KVM switcher to display the task execution status of each sub-compute node in real time. The task execution status includes the number of completed sub-tasks, the progress percentage of the currently running sub-tasks, and the estimated remaining time. It also supports users to view the global 3D model construction process preview on the monitor, and to rotate, scale and translate the preview model using the keyboard and mouse; Once the modeling task is completed, the user issues an export command through the human-computer interaction module. The model monitoring unit packages the 3D model data according to a preset format and exports it to a designated location via an external storage device or network transmission method.

[0014] This invention also discloses a portable 3D rapid cluster modeling method, applied to the aforementioned portable 3D rapid cluster modeling system, the method comprising: The system receives image data of the target area, imports the image data into the main computing node, and divides the image data into several sub-tasks through the main computing node and distributes them to the sub-computing nodes. Control the sub-computing nodes to execute the sub-tasks in parallel and output the sub-task results; The results of the subtasks are aggregated by the main computing node, and global optimization and texture mapping are performed to construct a three-dimensional model of the target region. Use the KVM switcher to switch compute nodes to view modeling progress or export results.

[0015] Compared with existing technologies, the advantages of this invention are that by integrating multiple computing nodes (including master and child nodes) into a single chassis and adopting a backplane interconnect structure, it significantly reduces communication latency and improves data transmission efficiency. Simultaneously, the compact and portable design makes it suitable for non-fixed operating environments such as fieldwork and disaster areas. The task scheduling and partitioning unit in the scheduling and modeling module dynamically generates subtasks based on the spatial distribution of image data (through grid partitioning and image quantity statistics), and intelligently distributes them according to the real-time load and CPU computing power of each sub-computing node, avoiding idle or overloaded computing resources and significantly improving parallel processing efficiency. During task execution, the system monitors the CPU / GPU utilization, memory usage, and task progress of each sub-node in real time. If a node failure is detected, unfinished subtasks can be automatically reassigned to healthy nodes, ensuring uninterrupted modeling tasks and complete and reliable results. By integrating a KVM switch, keyboard, mouse, and monitor, users can seamlessly switch between and access any computing node on a single human-machine interface. The model monitoring unit supports real-time viewing of task status and previewing of the 3D model building process, and allows interactive operations such as rotation and scaling via mouse and keyboard, greatly improving operational convenience and visualization. The model building unit effectively eliminates the stitching gaps between multi-node local reconstruction results through coordinate unification, global registration optimization (such as the iterative nearest point algorithm) and texture mapping, generating a high-precision 3D model with geometric consistency and continuous texture to meet the needs of professional applications. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1This is a functional block diagram of a portable cluster 3D rapid modeling system according to the present invention.

[0018] Figure 2 This is a functional block diagram of the scheduling and modeling module in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 1-2 As shown, the present invention provides a portable cluster 3D rapid modeling system, characterized in that the system comprises: An integrated chassis unit, which integrates several computing nodes, and the computing nodes are interconnected through a backplane. The computing nodes include main computing nodes and sub-computing nodes. The power supply and power management module is used to monitor the load of each computing node in real time and dynamically adjust the power supply. A human-computer interaction module is disposed on the integrated chassis unit. The human-computer interaction module includes a KVM switch, a keyboard, a mouse, and a monitor. The keyboard, mouse, and monitor are connected to each computing node through the KVM switch. The KVM switch is used to support users to switch between several computing nodes using a keyboard or mouse. The scheduling and modeling module is used to create a 3D model of the target area based on image data. An environmental adaptation module is installed inside the integrated chassis unit. The environmental adaptation module includes an air-cooling unit and a heat dissipation unit. The environmental adaptation module is used to monitor and intelligently control the temperature and humidity inside the integrated chassis unit in real time.

[0021] In this embodiment, the integrated chassis unit can be manufactured using high-strength aerospace aluminum (such as 6061-T6), possessing lightweight (total weight ≤35kg), impact resistance (meets MIL-STD-810G vibration standards), and waterproof and dustproof (IP54 rating) characteristics. After the system is powered on, each node automatically obtains a network address (DHCP or static allocation) through the backplane; users can select "independent mode" (complete isolation of each node) or "cluster mode" (enabling the internal network and scheduling platform) via a physical switch on the front panel or software commands. Specifically, the power supply and power management module monitors the load of each computing node in real time and dynamically adjusts the power supply, including: when the load of a single computing node is detected to be below 20%, automatically reducing the power supply of that node to 60%-70% of the baseline value; when the load exceeds 80% and lasts for more than 15 seconds, immediately increasing the power supply to 110% of the baseline value and triggering the heat dissipation unit enhancement mode; at the same time, the module has a built-in intelligent battery management system (BMS), which can seamlessly switch to the built-in lithium battery pack (capacity ≥500Wh) when the external power supply is interrupted, ensuring that the system can work continuously for ≥4 hours in cluster mode and for ≥8 hours in standalone mode. The power supply module also has overvoltage, overcurrent, and short-circuit protection functions. When the input voltage exceeds the AC 180V-240V range or the output current exceeds 1.5 times the rated value of a single node, it automatically cuts off the power supply to the corresponding channel and issues an audible and visual alarm through the human-machine interaction module.

[0022] The environmental adaptation module monitors and intelligently controls the internal temperature and humidity of the integrated chassis unit in real time. This includes: a temperature sensor collecting internal temperature data every 2 seconds; automatically activating the low-speed operation mode of the air-cooling unit when the temperature exceeds 35°C; switching to high-speed operation mode and coordinating with the heat dissipation fins of the cooling unit for forced cooling when the temperature is below 25°C; and automatically reducing the air-cooling unit speed to reduce energy consumption when the temperature is below 25°C. Humidity monitoring uses a high-precision humidity sensor to monitor the relative humidity inside the chassis in real time. When the humidity is above 65%, the built-in dehumidification module is activated, using semiconductor condensation to control the humidity within a suitable range of 40%-60%; if the humidity is below 30%, the humidification component is activated, using ultrasonic atomization technology to maintain humidity balance and prevent static electricity damage to electronic components. Furthermore, the environmental adaptation module has a self-diagnostic function. When the air-cooling unit experiences abnormal speed or the heat dissipation unit efficiency decreases, a fault code is immediately displayed on the human-machine interface module's screen, and a buzzer alarm is triggered to remind the user to perform timely maintenance.

[0023] In some embodiments of this application, the scheduling and modeling module includes: a data receiving unit configured to receive image data of a target area; a task scheduling and partitioning unit configured to import the image data into a main computing node, and divide the image data into several sub-tasks through the main computing node, and distribute them to sub-computing nodes; a task execution unit configured to control the sub-computing nodes to execute the sub-tasks in parallel, and output the sub-task results; a model building unit configured to aggregate the sub-task results through the main computing node, perform global optimization and texture mapping, and build a three-dimensional model of the target area; and a model monitoring unit configured to switch computing nodes through a KVM switch to view the modeling progress or export the results.

[0024] This application achieves rapid generation and export of 3D models of target areas through a closed-loop process of UAV data collection, parallel computing of cluster nodes, and local aggregation optimization.

[0025] In some embodiments of this application, the data receiving unit is connected to a drone acquisition device, which is used to acquire image data of a target area. The data receiving unit is provided with several data interfaces, and the drone acquisition device transmits the image data to the main computing node through the data interfaces.

[0026] Specifically, the drone data collection equipment uses an industrial-grade portable drone equipped with a full-frame camera (≥24 million pixels, sensor size ≥35mm, aperture adjustable from f / 2.8 to f / 11) and an RTK positioning module (positioning accuracy: horizontal ±1cm+1ppm, vertical ±2cm+1ppm), supports manual / automatic flight modes, and has a flight time of ≥30 minutes. The acquisition parameters were set as follows: Flight altitude: adjusted according to the modeling accuracy requirements; the normal altitude for emergency reconnaissance is 50-100m (ground resolution 0.1-0.2m), and the altitude for high-precision modeling is 30-50m (ground resolution 0.05-0.1m); Forward overlap rate: 80%-85% (to ensure sufficient overlap between adjacent images for feature matching); Lateral overlap rate: 70%-75%; Shooting mode: automatic flight path planning mode (such as grid flight, orbital flight), supplemented by manual shooting for the edge of the target area and complex terrain areas to ensure no data blind spots; Image parameters: shutter speed ≥ 1 / 1000s (to avoid image blurring due to drone shake), ISO set to 100-400 (to ensure image signal-to-noise ratio), white balance automatic, storage format JPG+RAW (RAW format is used for subsequent image preprocessing to improve data quality). After the drone finishes its flight, the image data is transmitted to the main computing node via the local USB 3.0 interface (transmission rate ≥ 5Gbps, avoiding reliance on external networks). At the same time, the image data is initially screened (blurred, overexposed, and underexposed images are deleted).

[0027] In some embodiments of this application, the task scheduling and partitioning unit imports the image data into the main computing node and divides the image data into several sub-tasks through the main computing node, including: importing the image data into the main computing node and performing preprocessing, the preprocessing including: image distortion correction, image enhancement and data index establishment; reading the position corresponding to each image data through the established data index to generate the spatial range boundary of the target area; using a grid partitioning method to divide the target area into several spatial grids, counting the number of images in each spatial grid, and dynamically adjusting the spatial grid according to the number of images; An independent subtask is generated for the image data of each spatial grid. The subtask includes image data, preprocessing parameters, and task execution instructions.

[0028] Specifically, this embodiment is the core prerequisite for realizing "cluster parallel computing." The core objective is to rationally divide the massive image dataset and distribute it to different computing nodes for parallel processing, thereby improving the overall modeling efficiency. Considering that the maximum number of cluster nodes is 10 (N≤9), the task division must follow the principles of "load balancing" and "low data correlation" to avoid situations where some nodes are overloaded while others are idle, while ensuring that the data of each subtask is independent, thus reducing the complexity of subsequent result aggregation.

[0029] Specifically, a lightweight distributed scheduling platform is adopted and deployed on the main computing node, which has the functions of automatic task partitioning, node status monitoring, task progress tracking, and automatic fault transfer. After the image data is imported into the main computing node, the scheduling platform first performs preprocessing on the data, including: image distortion correction (based on camera intrinsic parameters, using Zhang Zhengyou calibration method to eliminate the impact of lens distortion on modeling accuracy); image enhancement (historical equalization and gamma correction are performed on low-light and high-contrast images to improve image texture clarity); and data indexing (adding geographic coordinate information (latitude, longitude, and elevation) to each image and generating an image location index table based on RTK positioning data). Specifically, based on a combination of "spatial region partitioning" and "data volume balancing," the image dataset is preferentially divided into N subsets (N being the number of sub-computation nodes, ≤9) according to the spatial location of the target region. This ensures the spatial region corresponding to each subset is continuous, reducing feature matching redundancy between subtasks. Simultaneously, the number of images and data volume in each subset are counted, and the data volume deviation of each subtask is adjusted to ≤10% by adjusting the subset range, ensuring load balancing. The image location index table is read to generate the spatial boundary of the target region. A grid partitioning method is used to evenly divide the target region into N spatial grids. The number of images in each grid is counted; grids with too many images are split, and grids with too few images are merged to ensure a balanced number of images in each grid. An independent task package is generated for each image subset corresponding to each grid, containing image data, preprocessing parameters, and task execution instructions.

[0030] In some embodiments of this application, when the task scheduling and partitioning unit distributes subtasks to sub-computing nodes, it specifically does the following: obtains the real-time load and CPU computing power of the sub-computing nodes, determines the task processing capability of each computing node based on the real-time load and CPU computing power; determines the number of images for each subtask, and distributes the subtasks based on the number of images and the task processing capability.

[0031] Specifically, task packages are distributed to each computing node through the local cluster network. The distribution strategy is "proximity allocation + performance matching" (based on the real-time load rate of the computing node (nodes with a load rate <70% are given priority) and hardware performance (nodes with higher CPU / GPU computing power are allocated tasks with larger data volumes)). After the distribution is completed, a task allocation table is generated and stored on the main computing node.

[0032] In some embodiments of this application, when the task execution unit controls the sub-computing nodes to execute the sub-tasks in parallel, the task execution instructions include feature extraction, feature matching, sparse reconstruction, and dense point cloud generation.

[0033] Specifically, feature extraction is performed using the SIFT (Scale Invariant Feature Transform) algorithm to extract local feature points from the preprocessed image data. The Gaussian difference pyramid has 6 layers, a feature point threshold of 0.03, and a feature point descriptor dimension of 128. To improve efficiency, a GPU-accelerated SIFT implementation is used, with feature extraction time for a single 24-megapixel image ≤ 2 seconds. Feature matching is performed using the FLANN (Fast Nearest Neighbor) algorithm for feature point matching, combined with the RANSAC (Random Sample Consensus) algorithm to remove mismatches. For example, FLANN is first used for coarse matching, with a matching distance threshold of 200; the RANSAC algorithm is then used to filter the fundamental matrix, with 10,000 iterations and an interior point threshold of 1.5 pixels, ultimately controlling the mismatch rate to within 5%. Sparse reconstruction is performed using the Structure for Motion Recovery (SfM) algorithm, calculating the camera pose (position and pose angle) from the matched feature points to generate a sparse point cloud. Initial image pairs are selected, the fundamental and essential matrices are calculated, and the initial camera pose is solved. The Perspective-n-Point (PnP) algorithm is used to iteratively optimize the camera pose. A sparse point cloud is generated based on a triangulation algorithm, and point clouds with reprojection errors > 2 pixels are removed (the reprojection error is ultimately controlled within 1 pixel). Bundle Adjustment (BA) is used to globally optimize the camera pose and sparse point cloud, improving reconstruction accuracy. Dense point cloud generation: A multi-view stereo matching (MVS) based algorithm (such as the PatchMatch Stereo algorithm in COLMAP) is used to generate a dense point cloud based on the sparse point cloud. The image pyramid has 4 layers; the matching window size is 9×9; the depth map fusion weight is 0.8; and the dense point cloud density is 100-200 points / ㎡ (adjusted according to modeling accuracy requirements). With GPU acceleration, a single node can process 50-80 dense point clouds per hour from 24-megapixel images.

[0034] In some embodiments of this application, the task scheduling and partitioning unit is further configured to monitor the task execution data of each sub-computing node in real time when executing sub-tasks. The task execution data includes the CPU utilization, GPU utilization, memory usage, and task progress of the sub-computing node. Based on the task execution data, it is determined whether the corresponding sub-computing node has failed. If a failure occurs, the unfinished sub-tasks of the sub-computing node are reallocated.

[0035] Specifically, the CPU utilization, GPU utilization, memory usage, and task progress of each computing node are monitored in real time (each node reports its progress to the main computing node after completing a sub-task stage (such as feature extraction or sparse reconstruction)); if a node fails (such as power outage or system crash), the unfinished tasks of that node are automatically reassigned to other nodes with lower loads to ensure continuous task execution, and the fault recovery time is ≤3 minutes.

[0036] In some embodiments of this application, the model building unit aggregates the subtask results through a main computing node, performs global optimization and texture mapping, and constructs a 3D model of the target region. This includes: the main computing node receiving subtask results uploaded by each sub-computing node, the subtask results including sparse point clouds, dense point clouds, and image features of each spatial grid; performing coordinate unification processing on the received subtask results, transforming the sparse and dense point clouds generated by different sub-computing nodes in their local coordinate systems to the global coordinate system; performing global registration optimization based on the unified sparse and dense point clouds, aligning the point cloud edges of adjacent spatial grids using an iterative nearest-point algorithm to eliminate stitching gaps; and calculating texture coordinates based on the mapping relationship between the preprocessed image data and the sparse and dense point clouds in the global coordinate system, attaching the image features to the 3D point cloud surface to generate a 3D model of the target region.

[0037] Specifically, after each sub-computing node completes its sub-task, it uploads the sub-task results (including sparse point clouds, dense point clouds, camera intrinsics, camera pose parameters, and image texture data) to the main computing node via the local cluster network. The main computing node's data aggregation module performs a unified format conversion on the results (converting the point cloud data of each node into the PLY format supported by PCL, and the camera parameters into the JSON format), and establishes an intermediate result index library (associating the spatial regions corresponding to the results of each node) to facilitate subsequent global optimization calls.

[0038] The global optimization involves point cloud fusion, employing a spatial coordinate-based point cloud registration algorithm (such as the ICP algorithm (Iterative Closest Point Algorithm)) to fuse the dense point clouds of each node. Using the point cloud of the reference node specified by the main computing node (usually the node with the lowest load and highest processing accuracy during task allocation) as a reference, the transformation matrices (translation and rotation matrices) between the point clouds of other nodes and the reference point cloud are calculated. Based on the transformation matrices, coordinate transformations are performed on the point clouds of each node to achieve point cloud alignment. A voxel filtering algorithm (voxel size 0.05-0.1m) is used to deduplicate and simplify the fused point cloud, reducing data redundancy and improving model fluency.

[0039] Texture mapping is performed using an image projection-based texture mapping algorithm to add realistic textures to the fused dense point cloud. The technical process involves: performing Poisson reconstruction on the fused dense point cloud to generate a 3D mesh model (mesh resolution of 0.1-0.2m to ensure mesh details match the actual terrain / objects); determining the optimal image for each mesh based on camera pose parameters (selecting images with clear textures, a viewpoint perpendicular to the mesh surface, and no occlusion), and sampling texture information from these images; texture stitching: stitching the textures of adjacent meshes, using a feathering algorithm to eliminate texture seams and ensure natural texture transitions; texture compression: using JPEG compression format (compression quality of 85%) to reduce texture data volume and improve model loading speed.

[0040] The main compute node's model output module supports exporting 3D models in two common formats: OBJ format (suitable for viewing local details and editing models, including mesh and texture data during export, with a single model file size ≤10GB); and OSGB format (suitable for large-scale scene visualization and real-time walkthroughs, using a block storage method, supporting LOD (Level of Detail) display, facilitating rapid loading and viewing in emergency situations). The export process supports progress display, with an export speed ≥100MB / s (based on NVMe solid-state drives).

[0041] In some embodiments of this application, the model monitoring unit switches computing nodes via a KVM switch to view modeling progress or export results. This includes: switching to the monitoring interface of the main computing node via the KVM switch to display the task execution status of each sub-computing node in real time. The task execution status includes the number of completed sub-tasks, the progress percentage of the currently running sub-tasks, and the estimated remaining time. Simultaneously, users can view a preview of the global 3D model's construction process on a monitor and perform rotation, scaling, and translation operations on the preview model using the keyboard and mouse. When the modeling task is completed, the user issues an export command through the human-computer interaction module. The model monitoring unit packages the 3D model data according to a preset format and exports it to a designated location via an external storage device or network transmission.

[0042] Specifically, the KVM switcher uses a portable 8-port / 16-port KVM switch (supporting USB 2.0 and HDMI 2.0 interfaces, with 4K@60Hz resolution). Connection methods are as follows: Video signal: The HDMI interface of each node is connected to the input interface of the KVM switch via an HDMI cable, and the output interface of the KVM switch is connected to a portable monitor; USB signal: The USB interface of each node is connected to the USB input interface of the KVM switch via a USB cable, and the USB output interface of the KVM switch is connected to a keyboard and mouse set; Switching method: Supports button switching (KVM switch panel buttons) and hotkey switching (e.g., Ctrl+Alt+F1-F10 corresponding to different nodes), with a switching response time of ≤1 second to ensure smooth operation.

[0043] The scheduling and modeling module provides a visual progress interface on the desktop of each node, displaying the current task stage (data preprocessing, feature extraction, sparse reconstruction, dense point cloud generation, global optimization, texture mapping), completion percentage, remaining time, and node resource usage (CPU, GPU, memory). Users can switch to any node via KVM to view the task progress of that node in real time, or switch to the master node to view the global task progress (summarizing the progress of all nodes). After the model is output, users can export the results in two ways: Local export: Switch to the master node and copy the OBJ / OSGB format model file to an external hard drive via USB 3.0 interface (transfer rate ≥ 5Gbps); Inter-node export: If it is necessary to back up the results on other computing nodes, the model file of the master node can be copied to the target computing node via the local cluster network (transfer rate ≥ 1Gbps). No external network connection is required throughout the process, ensuring data security and environmental adaptability.

[0044] The entire system is deployed locally, with all hardware and software integrated into portable devices. Data transmission occurs only within the local cluster network (a local area network composed of gigabit switches), without relying on external networks such as 4G / 5G or Wi-Fi. The system's required algorithm libraries, drivers, and configuration files are pre-installed on the solid-state drives of each node, allowing it to run directly after startup without needing to download them online, ensuring normal operation even in emergency situations without external network access.

[0045] This invention also discloses a portable 3D rapid cluster modeling method, applied to the aforementioned portable 3D rapid cluster modeling system, the method comprising: The system receives image data of the target area, imports the image data into the main computing node, and divides the image data into several sub-tasks through the main computing node and distributes them to the sub-computing nodes. Control the sub-computing nodes to execute the sub-tasks in parallel and output the sub-task results; The results of the subtasks are aggregated by the main computing node, and global optimization and texture mapping are performed to construct a three-dimensional model of the target region. Use the KVM switcher to switch compute nodes to view modeling progress or export results.

[0046] By breaking down 3D modeling tasks into sub-tasks and processing them in parallel by sub-computing nodes, modeling efficiency is significantly improved, enabling the 3D reconstruction of large-scale target areas to be completed in a short time. The use of a localized cluster network for data transmission and processing avoids dependence on external networks, ensuring data security during transmission and storage, preventing the leakage of sensitive information, and enabling the system to operate stably in complex environments such as emergency sites and remote areas without external network coverage. All hardware and software are integrated into a portable device, combined with a KVM switch to enable convenient control of each computing node and viewing and exporting of results, enhancing the system's portability and operational flexibility, and meeting the needs for rapid, efficient, and secure modeling in scenarios such as emergency rescue, disaster relief, and field exploration.

[0047] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A portable cluster 3D rapid modeling system, characterized in that, The system includes: An integrated chassis unit, which integrates several computing nodes, and the computing nodes are interconnected through a backplane. The computing nodes include main computing nodes and sub-computing nodes. The power supply and power management module is used to monitor the load of each computing node in real time and dynamically adjust the power supply. A human-computer interaction module is disposed on the integrated chassis unit. The human-computer interaction module includes a KVM switch, a keyboard, a mouse, and a monitor. The keyboard, mouse, and monitor are connected to each computing node through the KVM switch. The KVM switch is used to support users to switch between several computing nodes using a keyboard or mouse. The scheduling and modeling module is used to create a 3D model of the target area based on image data. An environmental adaptation module is installed inside the integrated chassis unit. The environmental adaptation module includes an air-cooling unit and a heat dissipation unit. The environmental adaptation module is used to monitor and intelligently control the temperature and humidity inside the integrated chassis unit in real time.

2. The portable cluster 3D rapid modeling system according to claim 1, characterized in that, The scheduling and modeling module includes: a data receiving unit, configured to receive image data of the target area; The task scheduling and partitioning unit is configured to import the image data into the main computing node, and divide the image data into several sub-tasks through the main computing node and distribute them to the sub-computing nodes. The task execution unit is configured to control the sub-computing nodes to execute the sub-tasks in parallel and output the sub-task results; The model building unit is configured to aggregate the results of the subtasks through the main computing node, perform global optimization and texture mapping, and build a three-dimensional model of the target region. The model monitoring unit is configured to switch compute nodes via a KVM switcher to view modeling progress or export results.

3. The portable cluster 3D rapid modeling system according to claim 2, characterized in that, The data receiving unit is connected to the UAV acquisition device, which is used to collect image data of the target area. The data receiving unit is equipped with several data interfaces, and the UAV acquisition device transmits the image data to the main computing node through the data interfaces.

4. The portable cluster 3D rapid modeling system according to claim 2, characterized in that, The task scheduling and partitioning unit imports the image data into the main computing node, and then divides the image data into several sub-tasks through the main computing node, including: The image data is imported into the main computing node and preprocessed, including image distortion correction, image enhancement, and data indexing. By reading the location corresponding to each image data through the established data index, the spatial range boundary of the target area is generated. The target area is divided into several spatial grids using a grid division method. The number of images in each spatial grid is counted, and the spatial grids are dynamically adjusted based on the number of images. An independent subtask is generated for the image data of each spatial grid. The subtask includes image data, preprocessing parameters, and task execution instructions.

5. The portable cluster 3D rapid modeling system according to claim 4, characterized in that, When the task scheduling and partitioning unit distributes subtasks to sub-computing nodes, it does so specifically as follows: Obtain the real-time load and CPU computing power of the sub-computing nodes, and determine the task processing capability of each computing node based on the real-time load and CPU computing power; The number of images for each subtask is determined, and the subtasks are distributed based on the number of images and the task processing capacity.

6. The portable cluster 3D rapid modeling system according to claim 5, characterized in that, When the task execution unit controls the sub-computing nodes to execute the sub-tasks in parallel, the task execution instructions include feature extraction, feature matching, sparse reconstruction, and dense point cloud generation.

7. The portable cluster 3D rapid modeling system according to claim 6, characterized in that, The task scheduling and partitioning unit is also configured to monitor the task execution data of each sub-computing node in real time when executing sub-tasks. The task execution data includes the CPU utilization, GPU utilization, memory usage and task progress of the sub-computing nodes. Based on the task execution data, determine whether the corresponding sub-computing node has failed. If a failure occurs, the unfinished sub-tasks of the sub-computing node will be reassigned.

8. The portable cluster 3D rapid modeling system according to claim 7, characterized in that, The model building unit aggregates the results of the subtasks through the main computing node, performs global optimization and texture mapping, and constructs a 3D model of the target region, including: The master computing node receives the subtask results uploaded by each sub-computing node. The subtask results include sparse point clouds, dense point clouds, and image features of each spatial grid. The received subtask results are processed to unify coordinates, transforming sparse and dense point clouds generated in the local coordinate system by different sub-computation nodes into the global coordinate system. Global registration optimization is performed based on the unified sparse point cloud and dense point cloud. The edge of the point cloud of adjacent spatial grids is aligned by the iterative nearest point algorithm to eliminate the stitching gap. Based on the mapping relationship between the preprocessed image data and the sparse and dense point clouds in the global coordinate system, texture coordinates are calculated, and image features are fitted onto the surface of the three-dimensional point cloud to generate a three-dimensional model of the target area.

9. The portable cluster 3D rapid modeling system according to claim 8, characterized in that, The model monitoring unit switches computing nodes via a KVM switcher to view modeling progress or export results, including: Switch to the monitoring interface of the master compute node via the KVM switcher to display the task execution status of each sub-compute node in real time. The task execution status includes the number of completed sub-tasks, the progress percentage of the currently running sub-tasks, and the estimated remaining time. It also supports users to view the global 3D model construction process preview on the monitor, and to rotate, scale and translate the preview model using the keyboard and mouse; Once the modeling task is completed, the user issues an export command through the human-computer interaction module. The model monitoring unit packages the 3D model data according to a preset format and exports it to a designated location via an external storage device or network transmission method.

10. A portable cluster 3D rapid modeling method, applied to the portable cluster 3D rapid modeling system as described in any one of claims 1-9, characterized in that, The method includes: The system receives image data of the target area, imports the image data into the main computing node, and divides the image data into several sub-tasks through the main computing node and distributes them to the sub-computing nodes. Control the sub-computing nodes to execute the sub-tasks in parallel and output the sub-task results; The results of the subtasks are aggregated by the main computing node, and global optimization and texture mapping are performed to construct a three-dimensional model of the target region. Use the KVM switcher to switch compute nodes to view modeling progress or export results.

Images