A scene-aware based 3D gaussian automated reconstruction scheduling system and method

Abstract

This invention relates to the field of graphics and image processing technology, and discloses a scene-aware 3D Gaussian automated reconstruction scheduling system and method, including a feature perception module, a pipeline scheduling engine, a computation control module, and a monitoring and recovery module. The feature perception module acquires the resolution and displacement span of the image sequence to be reconstructed; the pipeline scheduling engine maps the displacement span to the target reconstruction pipeline through a preset correlation matrix; the computation control module starts the 3D Gaussian splashing computation task according to the parameter constraint set, and establishes the convergence state by limiting the feature point density and suppressing the fluctuation of floating-point accumulation; the monitoring and recovery module acquires the activity parameters of the computation task, and issues a status backtracking instruction when the processor utilization and log writing frequency do not meet the judgment threshold. This invention utilizes computational feature feedback to achieve closed-loop self-healing of the reconstruction task, enhancing the determinism of 3D model generation and the efficiency of automated production.

Description

Technical Field

[0001] This invention relates to a scene-aware 3D Gaussian automated reconstruction scheduling system and method, belonging to the field of graphics and image processing technology. Background Technology

[0002] Current 3D Gaussian splash reconstruction technology is used to construct high-precision digital assets due to its real-time rendering and spatial fidelity. This technology typically uses the structure-of-motion (SOMO) algorithm to parse camera poses from image sequences and generate sparse point clouds, which serve as the initial geometric values ​​for subsequent 3D Gaussian primitive optimization. In industrial-grade applications, the stability of the reconstruction process is affected by the heterogeneity of the image acquisition environment. Because existing reconstruction methods usually require manual parameter setting based on requirements such as indoor close-up or large-scale aerial photography, automated pipelines may experience computational deadlocks or process anomalies due to geometric constraint failures when processing low-overlap or weak-texture images. To address these reliability bottlenecks, linear improvements such as increasing computational resources or improving feature extraction density have been implemented, but this has led to a non-linear increase in computational complexity with the number of images. At the same time, the randomness of the floating-point accumulation order in parallel computing can easily cause numerical convergence instability under complex geometric constraints, resulting in algorithm process suspension, memory overflow, or computational precision loss.

[0003] Current hardware redundancy and feature density enhancement methods are insufficient to balance reconstruction quality and computational cost. Adaptive control schemes at the software level also have shortcomings. For example, Chinese invention patent application CN120472121A discloses a large-scene 3D reconstruction method based on 3D Gaussian sputtering. It uses spatial meshing and monocular depth prior to alleviate the computational pressure of large scenes. However, it is a preset static architecture. The preset computing environment and algorithm convergence process are in an ideal state. In production conditions, the randomness of the floating-point accumulation order of parallel computing and complex geometric constraints cause numerical fluctuations. This scheme lacks low-level perception of process activity and closed-loop self-healing mechanism. When encountering computational deadlock or memory overflow caused by abnormal point cloud distribution, the system cannot spontaneously identify the abnormal state and falls into indefinite suspension, which is difficult to meet the deterministic requirements of large-scale automatic production of 3D assets.

[0004] Therefore, the technical problem to be solved by this invention is how to construct an automated scheduling system with scene perception capability, controlled computation process and process-level fault tolerance based on the visual statistical characteristics and spatial constraint rules of image sequences, so as to solve the problem of pipeline interruption and high execution failure rate caused by environmental uncertainty in the 3D Gaussian reconstruction process. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A scene-aware 3D Gaussian automated reconstruction scheduling system, comprising:

[0006] The feature perception module acquires metadata features of the image sequence to be reconstructed, including image resolution and the displacement span of the image acquisition trajectory in three-dimensional space.

[0007] The pipeline scheduling module is connected to the feature perception module. It maps metadata features to target reconstruction pipelines through a preset association matrix. The target reconstruction pipelines include fine-sculpting pipelines, high-speed pipelines, and aerial photography pipelines.

[0008] The computation control module, connected to the pipeline scheduling module, initiates a 3D Gaussian splash computation task based on the parameter constraint set corresponding to the target reconstructed pipeline. The parameter constraint set includes the upper limit of feature point size, numerical stability constraints, and storage reset parameters. The convergence state of the 3D Gaussian splash computation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the computation process.

[0009] The monitoring and recovery module is connected to the computing control module to obtain the activity parameters of the 3D Gaussian splash computing task. The activity parameters are represented by the processor utilization and the log update frequency. When the processor utilization is lower than 0.1% and the log update frequency is lower than the preset silent threshold, the monitoring and recovery module issues a status backtracking command and updates the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash computing task.

[0010] Preferably, the feature perception module also acquires scene labels of image frames in the image sequence to be reconstructed; when the pipeline scheduling module identifies an attribute conflict between the scene label and the metadata features during the mapping process, it introduces a collaborative decision factor for arbitration and switches the target reconstruction pipeline to the high-speed pipeline; wherein, the collaborative decision factor is obtained based on the consistency between the confidence level of the scene label and the statistical distribution of the metadata features.

[0011] Preferably, the upper limit of the feature point size in the parameter constraint set is determined by the visual contribution scoring mechanism; the calculation control module limits the upper limit of the feature point size through the following sub-steps: Step 31: Calculate the Laplacian response value of the image feature points; Step 32: Retain the feature points in descending order of the Laplacian response value until the upper limit of the feature point size is reached.

[0012] Preferably, the storage reset parameters are used to reconstruct the logical storage path before the 3D Gaussian splash calculation task starts; the calculation control module maps the intermediate cache data and model output data of the 3D Gaussian splash calculation task to the high-speed temporary storage area, and allocates isolated resource handles for different target reconstruction pipelines.

[0013] Preferably, when the feature perception module obtains the displacement span, it extracts the GPS coordinates from the image exchange file format and calculates the Euclidean distance between the start and end image frames in the image sequence; when the Euclidean distance exceeds a preset displacement threshold, the pipeline scheduling module identifies the target reconstruction pipeline as the aerial photography pipeline.

[0014] Preferably, when the target reconstructed pipeline is a precision pipeline, the pipeline scheduling module sets the feature extraction density in the parameter constraint set to a first preset value; when the target reconstructed pipeline is a high-speed pipeline, the feature extraction density is set to a second preset value, and the first preset value is greater than the second preset value.

[0015] Preferably, when scheduling the high-speed pipeline, the computation control module suppresses nondeterministic fluctuations through the following sub-steps: Step 71: Limit the number of parallel threads to 1; Step 72: Perform singular value decomposition operations sequentially by a single thread to eliminate nondeterministic fluctuations in the algebraic operation results under geometric constraints caused by thread competition.

[0016] Preferably, after issuing the status rollback command, the monitoring and recovery module reads the sparse point cloud data backed up by the 3D Gaussian splash calculation task before it stalled as the initial state, lowers the upper limit of the feature point scale according to the degradation strategy, and calls the calculation control module again.

[0017] Preferably, the system also includes a distributed node allocation module, which is used to retrieve idle computing nodes that meet the parameter constraint set in the distributed computing cluster according to the total data volume of the image sequence to be reconstructed, and to distribute the scheduling weight of the 3D Gaussian splash computing task to the selected idle computing nodes.

[0018] A scene-aware 3D Gaussian automated reconstruction scheduling method includes the following steps:

[0019] Step 101: Obtain the metadata features of the image sequence to be reconstructed. The metadata features include the image resolution and the displacement span of the image acquisition trajectory in three-dimensional space.

[0020] Step 102: Map metadata features to target reconstruction pipelines using a preset association matrix. Target reconstruction pipelines include fine-tuning pipeline, high-speed pipeline, and aerial photography pipeline.

[0021] Step 103: Start the 3D Gaussian splashing calculation task according to the parameter constraint set corresponding to the target reconstruction pipeline; wherein, the parameter constraint set includes the upper limit of feature point size, numerical stability constraints and storage reset parameters, and the convergence state of the 3D Gaussian splashing calculation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the calculation process.

[0022] Step 104: Obtain the activity parameters of the 3D Gaussian splash calculation task in real time. The activity parameters are characterized by processor utilization and log update frequency.

[0023] Step 105: When the processor utilization is below 0.1% and the log update frequency is below the preset silent threshold, issue a state backtracking instruction and update the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash calculation task.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. In 3D Gaussian automated reconstruction scheduling, by extracting multimodal metadata features and scene labels from the image sequence to be reconstructed, a heuristic mapping matrix is ​​used to achieve automatic adaptation of the reconstruction pipeline. This fundamentally solves the problem that traditional reconstruction processes rely heavily on manual experience to set parameters. This mechanism can accurately identify key statistical characteristics in the image, such as device type, displacement span, and image gradient. When there is a conflict between the user's explicit labels and metadata, a collaborative decision factor is introduced for arbitration, guiding the system to automatically downgrade to a conservative high-speed mode. This ensures that the reconstruction task can still proceed along the determined technical path under various abnormal sensor data or input error conditions, effectively avoiding the ineffective consumption of computing resources under incorrect parameter settings.

[0026] 2. In the rapid reconstruction mode, hard capping pruning of feature point scale and single-thread numerical stability constraints are introduced. By scoring the visual contribution, core feature points with high Laplacian response values ​​are retained first, and the computational complexity is forcibly controlled from nonlinear growth to a quasi-linear range. Since parallel computation that may cause randomness in the order of floating-point accumulation is abandoned, key algebraic operations such as singular value decomposition are executed sequentially in a single thread. This eliminates the nondeterministic convergence and black-box suspension phenomena that are prone to occur in algorithms under complex geometric constraints from the bottom layer, and improves the determinism of asset generation and industrial-grade delivery efficiency in time-latency sensitive scenarios.

[0027] 3. Construct a real-time monitoring and self-healing closed loop based on computational heartbeats. By dynamically monitoring the synergistic relationship between processor utilization and log update frequency, the system achieves accurate perception of the health status of the reconstruction process. Unlike conventional task heartbeat monitoring, this invention dynamically adjusts the monitoring weights of log updates and resource usage based on the computational characteristics of different pipelines. It can effectively identify computational dead loops or memory overflow risks caused by abnormal distribution of 3D spatial point clouds. Once the process is determined to be in a dead state, the system immediately triggers a state backtracking and parameter adaptive adjustment mechanism to restart the task without manual intervention, thereby enhancing the automation capabilities of the entire large-scale 3D asset production chain. Attached Figure Description

[0028] Figure 1This is a flowchart of the overall process of the 3D Gaussian automated reconstruction scheduling system of the present invention;

[0029] Figure 2 This is a scene-aware reconstruction pipeline decision logic diagram of the present invention;

[0030] Figure 3 This is a diagram showing the environment reset direction and parameter pruning of the rapid reconstruction mode of this invention;

[0031] Figure 4 This is a flowchart of the task monitoring and anomaly self-healing process of the present invention.

[0032] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0034] A scene-aware 3D Gaussian automated reconstruction scheduling system includes:

[0035] The feature perception module acquires metadata features of the image sequence to be reconstructed, including image resolution and the displacement span of the image acquisition trajectory in three-dimensional space.

[0036] The pipeline scheduling module is connected to the feature perception module. It maps metadata features to target reconstruction pipelines through a preset association matrix. The target reconstruction pipelines include fine-sculpting pipelines, high-speed pipelines, and aerial photography pipelines.

[0037] The computation control module, connected to the pipeline scheduling module, initiates a 3D Gaussian splash computation task based on the parameter constraint set corresponding to the target reconstructed pipeline. The parameter constraint set includes the upper limit of feature point size, numerical stability constraints, and storage reset parameters. The convergence state of the 3D Gaussian splash computation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the computation process.

[0038] The monitoring and recovery module is connected to the computing control module to obtain the activity parameters of the 3D Gaussian splash computing task. The activity parameters are represented by the processor utilization and the log update frequency. When the processor utilization is lower than 0.1% and the log update frequency is lower than the preset silent threshold, the monitoring and recovery module issues a status backtracking command and updates the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash computing task.

[0039] Preferably, the feature perception module also acquires scene labels of image frames in the image sequence to be reconstructed; when the pipeline scheduling module identifies an attribute conflict between the scene label and the metadata features during the mapping process, it introduces a collaborative decision factor for arbitration and switches the target reconstruction pipeline to the high-speed pipeline; wherein, the collaborative decision factor is obtained based on the consistency between the confidence level of the scene label and the statistical distribution of the metadata features.

[0040] Preferably, the upper limit of the feature point size in the parameter constraint set is determined by the visual contribution scoring mechanism; the calculation control module limits the upper limit of the feature point size through the following sub-steps: Step 31: Calculate the Laplacian response value of the image feature points; Step 32: Retain the feature points in descending order of the Laplacian response value until the upper limit of the feature point size is reached.

[0041] Preferably, the storage reset parameters are used to reconstruct the logical storage path before the 3D Gaussian splash calculation task starts; the calculation control module maps the intermediate cache data and model output data of the 3D Gaussian splash calculation task to the high-speed temporary storage area, and allocates isolated resource handles for different target reconstruction pipelines.

[0042] Preferably, when the feature perception module obtains the displacement span, it extracts the GPS coordinates from the image exchange file format and calculates the Euclidean distance between the start and end image frames in the image sequence; when the Euclidean distance exceeds a preset displacement threshold, the pipeline scheduling module identifies the target reconstruction pipeline as the aerial photography pipeline.

[0043] Preferably, when the target reconstructed pipeline is a precision pipeline, the pipeline scheduling module sets the feature extraction density in the parameter constraint set to a first preset value; when the target reconstructed pipeline is a high-speed pipeline, the feature extraction density is set to a second preset value, and the first preset value is greater than the second preset value.

[0044] Preferably, when scheduling the high-speed pipeline, the computation control module suppresses nondeterministic fluctuations through the following sub-steps: Step 71: Limit the number of parallel threads to 1; Step 72: Perform singular value decomposition operations sequentially by a single thread to eliminate nondeterministic fluctuations in the algebraic operation results under geometric constraints caused by thread competition.

[0045] Preferably, after issuing the status rollback command, the monitoring and recovery module reads the sparse point cloud data backed up by the 3D Gaussian splash calculation task before it stalled as the initial state, lowers the upper limit of the feature point scale according to the degradation strategy, and calls the calculation control module again.

[0046] Preferably, the system also includes a distributed node allocation module, which is used to retrieve idle computing nodes that meet the parameter constraint set in the distributed computing cluster according to the total data volume of the image sequence to be reconstructed, and to distribute the scheduling weight of the 3D Gaussian splash computing task to the selected idle computing nodes.

[0047] A scene-aware 3D Gaussian automated reconstruction scheduling method includes the following steps:

[0048] Step 101: Obtain the metadata features of the image sequence to be reconstructed. The metadata features include the image resolution and the displacement span of the image acquisition trajectory in three-dimensional space.

[0049] Step 102: Map metadata features to target reconstruction pipelines using a preset association matrix. Target reconstruction pipelines include fine-tuning pipeline, high-speed pipeline, and aerial photography pipeline.

[0050] Step 103: Start the 3D Gaussian splash calculation task according to the parameter constraint set corresponding to the target reconstruction pipeline; the parameter constraint set includes the upper limit of feature point size, numerical stability constraints and storage reset parameters. The convergence state of the 3D Gaussian splash calculation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the calculation process.

[0051] Step 104: Obtain the activity parameters of the 3D Gaussian splash calculation task in real time. The activity parameters are characterized by processor utilization and log update frequency.

[0052] Step 105: When the processor utilization is below 0.1% and the log update frequency is below the preset silent threshold, issue a state backtracking instruction and update the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash calculation task.

[0053] Example 1: In an application scenario involving 3D modeling of substation equipment, the feature perception module acquires metadata features of the image sequence to be reconstructed and identifies that the displacement span of the image acquisition trajectory in 3D space is less than a preset displacement threshold. Simultaneously, based on the image file format information from the image exchange, the resolution was determined to be around 3840×2160. The pipeline scheduling module mapped the scene to the fine-sculpting pipeline using a preset correlation matrix and loaded the corresponding parameter constraint set to start the 3D Gaussian splash calculation task. During the calculation execution phase, due to the high image overlap and the increasing density of feature points, the system backend detected continuous fluctuations in processor utilization and that memory usage was approaching the hardware's physical limits, causing the calculation process to enter a silent state. The monitoring and recovery module obtained the activity parameters of the 3D Gaussian splash calculation task in real time and collected the processor utilization through the kernel interface. With log update frequency and incorporate it into the system health score. The computational logic is dynamically evaluated; among which, To score the system's health, For processor utilization, For log update frequency, Sampling time, Weighting the log update frequency To calculate resource usage weights; when the monitoring and recovery module detects that the log write rate has dropped below a preset silent threshold, and When the rate drops below 0.1%, the system determines that the current pipeline is trapped in a computational dead loop, issues a state rollback command, reads the sparse point cloud data backed up before the stagnation as the initial state, and dynamically switches the target reconstruction pipeline to the high-speed pipeline according to the degradation strategy. Based on the principles of computing resource release and load balancing, the system uses an exponential decay model to smoothly converge computing power overhead, and the monitoring and recovery module executes the formula. Lower the upper limit of feature point size, among which, This represents the upper limit of the updated feature point size, and its value is a positive integer. This is the upper limit parameter for the size of feature points before the system stagnates; The resource attenuation coefficient is a dimensionless constant with a value range of 0.5 to 0.8. This dimensionless constant range is derived from the optimal engineering fault tolerance domain obtained from multiple rounds of destructive stress tests in the laboratory. The detailed test data shows that if the attenuation coefficient exceeds the upper limit of 0.8, it means that the amount of logical memory space stripped away by each round of abnormal retreat is negligible and insufficient to withstand the memory surge induced by the next frame of more complex images, thus causing the process to slip into the quagmire again with a very high reproducibility rate. Conversely, if the coefficient drops below the lower edge of 0.5, it is equivalent to forcibly performing an excessively violent topology pruning operation, which will directly cause irreversible large-area tearing or structural hole defects in the output model shell due to the excessively sparse spatial connection density. This selected range precisely avoids the two extreme failure states of system stagnation or precision disintegration.

[0054] During the execution of the high-speed pipeline, the computational control module determines the upper limit of the feature point size through a visual contribution scoring mechanism. By calculating the Laplacian response value of image feature points Feature points are retained in descending order until they reach a certain level. The preset size limit; among which, The Laplacian response value is used to characterize the high-frequency detail richness of local regions in an image. The threshold for the total number of feature points extracted; simultaneously, the control module calculates the number of parallel threads. By forcibly limiting the dimensionality reduction to 1 and employing a single-threaded sequential execution of singular value decomposition (SVD) operations, the impact of floating-point accumulation randomness on geometric constraint convergence in a multi-threaded environment is eliminated. This allows the convergence characteristic curve of the model training task to return to a stable range after state backtracking. In actual industrial deployments, this forced dimensionality reduction single-threaded strategy is not a global operation spanning the entire high-speed pipeline, but rather is finely isolated and applied to a few abnormal feature points that exhibit rank deficiency or matrix ill-conditioning, specifically in the singular value solution module used for their preceding geometric pose algebra initialization. After smoothly navigating this fragile step that is highly prone to numerical collapse, the system fully re-enables the ultra-concurrent thread matrix engine of the graphics processor during the subsequent large-scale forward rasterization rendering and backpropagation parameter update phases of Gaussian elements. This eliminates the risk of random computational deadlock while ensuring that the system still maintains the high frame rate throughput characteristic highly matching the "high-speed" aspect. The system operates within the specified latency-sensitive threshold. The system generates 3D digital assets internally. While maintaining the geometric accuracy of equipment components, the processing latency is reduced to 1 / 5 of the traditional process. The scene quantification information provided by the feature perception module provides decision input for the pipeline scheduling module. The real-time intervention of the monitoring and recovery module ensures the execution stability of computing tasks under fluctuating computing resources, transforming redundant computing tasks into a controlled automated production flow and solving the problem of task interruption caused by the uncertainty of the computing environment during 3D reconstruction.

[0055] Example 2: Reconstruction Stability under Limited Computing Resources and Dynamic Interference Environments. The experimental platform was built on a portable computing terminal equipped with a graphics processor with 8GB of video memory and 16GB of system memory. This terminal was used to simulate the mobile edge computing conditions used in industrial field inspections. The image sequence used in the experiment was derived from a publicly available industrial 3D scanning dataset, which included 200 consecutive image frames with a resolution of 3840×2160. To simulate the sensor thermal noise and ambient light fluctuations generated during real acquisition, Gaussian white noise with a signal-to-noise ratio of 20dB was superimposed on the color space of the original images, and a 5% random pixel offset was introduced to simulate the vibration of the UAV platform in a windy environment. Key parameter: upper limit of feature point size. The selection logic is as follows: Identify the factors affecting this parameter as the space complexity of feature point matching. Peak signal-to-noise ratio of the reconstruction results The parameter selection needs to balance image reconstruction accuracy and limited video memory capacity. When the scene complexity reaches a preset threshold and the image overlap exceeds 80%, in order to avoid the risk of video memory overflow during motion reconstruction, It tends to the lower limit of the value range.

[0056] The experiment was divided into four groups of comparative observations: the first group was the sample group of the present invention, which adopted the complete scheme of enabling the scene perception scheduling and monitoring recovery module, and set the upper limit of the feature point scale. The first group is set to 8000; the second group is the control group, which uses the traditional 3D Gaussian splashing method, lacks scene awareness logic, uses a fixed sculpting pipeline, and does not set feature point size constraints; the third group is the partially missing control group, which retains the pipeline scheduling logic but disables the monitoring and recovery module; the fourth group is the out-of-range control group, which will... The value was set to 50,000. After the computation task started, the first group identified resolution and displacement features through the feature perception module, switching from the fine-tuning pipeline to the high-speed pipeline. The system recorded that at 45.3 seconds into the process, the processor utilization was reduced due to high-frequency noise inducing the growth of local feature points. When the activity level drops to 0.08%, the monitoring and recovery module identifies this activity parameter as triggering the IO silencing threshold, and then issues a status rollback command and restarts the calculation; among which, This represents the upper limit of the feature point size. Regarding processor utilization, expressed as a percentage, data records show that the first group of samples from this invention completed the reconstruction task within 15.4 minutes, generating a 3D model with a peak signal-to-noise ratio of 24.83 dB. The detailed structure and texture of the equipment in the image sequence were clear, and noise interference was suppressed. In contrast, the second control group, running under the FineCar pipeline, experienced a non-linear increase in memory usage at 8.2 minutes due to the lack of a feature point pruning step, resulting in memory overflow and process interruption. The third partially missing control group, facing the same computational suspension condition, lacked a monitoring and self-healing mechanism, and the system... In the state below 0.1%, the system remained suspended for more than 2 hours without generating any effective output. Data from the fourth group of out-of-range control groups showed that when the upper limit of the feature point size increased from 8000 to 50000, the reconstruction latency increased from 15.4 minutes to 82.7 minutes, while the peak signal-to-noise ratio only increased from 24.83dB to 25.12dB. The performance growth entered the saturation zone, and the peak memory usage reached 7.6GB, approaching the physical hardware boundary. This comparative data established 8000 as the working range that balances computational stability and resource utilization efficiency.

[0057] To measure the gradient response of this invention to problem intensity, three test gradients were set with noise signal-to-noise ratios of 25dB, 20dB, and 15dB. As the image noise power increased, the frequency of state backtracking triggered by the monitoring and recovery module changed in a stepwise manner, occurring 1, 2, and 5 times respectively. However, under each gradient, a stable reconstruction model was output within the preset time delay, and the reconstruction accuracy deviation remained within 1.2%. This experiment confirmed the synergistic effect between the feature perception module and the monitoring and recovery module. Specifically, the pipeline scheduling module achieved initial constraints on computational complexity by reducing feature point density, while the monitoring and recovery module dynamically captured activity parameters. With log update frequency By addressing computational uncertainties caused by environmental disturbances, the system's automated production efficiency exhibits deterministic performance under real-world operating conditions. Based on the aforementioned data observations, this invention constructs a closed-loop scheduling mechanism based on computational feature feedback, enabling the 3D Gaussian reconstruction system to possess self-healing capabilities in heterogeneous computing environments, thus resolving the technical contradiction between high-precision 3D reconstruction and limited computing resources. The entire experimental process is driven by the system's built-in control logic, achieving a stable transformation from raw noisy image input to high-fidelity 3D asset generation, establishing the implementation value of this technical solution in the field of industrial-grade automated reconstruction.

[0058] Example 3: This example combines Figures 1 to 4 This document describes a scene-aware 3D Gaussian automated reconstruction scheduling system and method, such as... Figure 1 As shown, the overall process flow of the scene-aware 3D Gaussian automated reconstruction scheduling system and method starts from the beginning of execution, sequentially entering the image sequence or video input stage and the scene perception module. The pipeline scheduling engine makes decision branches. There is an abnormal self-healing feedback path between the pipeline scheduling engine and the real-time monitoring closed-loop module. According to the difference in input features, it is divided into three branches: scene A, scene B, and scene C, corresponding to the fine-carving reconstruction pipeline, the high-speed reconstruction pipeline, and the aerial reconstruction pipeline, respectively. Each pipeline converges to the automated execution module and is constrained by the heartbeat monitoring from the real-time monitoring closed-loop module. The motion recovery structure, i.e., SfM sparse reconstruction, 3DGS model training, point cloud model compression and lightweighting operations are executed in sequence, and finally the 3D digital asset is output and the process ends.

[0059] like Figure 2 As shown, the pipeline decision logic of the scene-aware 3D Gaussian automated reconstruction scheduling system determines whether an explicit user label is detected. If yes, the label type is parsed as Fine Carving, Speed, or Aerial Photography. If no, automated perception and recognition are initiated, and the EXIF ​​metadata extraction and pre-visual statistics steps are performed. Initial traffic splitting is performed by determining whether the GPS span is greater than the preset displacement threshold Lmax. If the GPS span is greater than Lmax, the system is set to aerial photography mode. If the GPS span is not greater than Lmax, the system load or resolution is further determined to be lower than 1080P. If the load or resolution meets the conditions, the system is set to speed mode; otherwise, the system is set to fine carving mode. Finally, all branch paths are unified and converged to the target pipeline mode determination step, and the process ends after executing the steps of loading parameter sets and starting the reconstruction pipeline.

[0060] like Figure 3As shown, the execution details of the ultra-fast reconstruction mode are divided into the pre-execution environment reset stage and the in-execution parameter pruning stage. The environment reset stage includes three steps: environment variable scanning, environment variable LD\LIBRARY\PATH reorganization, and building a library compatibility sandbox. The in-execution parameter pruning stage includes three constraints: limiting the upper limit of feature points to 8000, limiting the image size to 2400px, and forcing single-thread numerical stability. Through the above technical means, ultra-fast reconstruction is finally completed.

[0061] like Figure 4 As shown, the monitoring and recovery logic involves starting a new 3DGS task subprocess and starting the Observer after the start. The system enters the periodic sampling phase to obtain CPU load and log timestamps, and determines whether the log stagnation is greater than 300s and whether the CPU load is less than 0.1%. If the combined judgment conditions are met, the zombie process is forcibly terminated, the state is rolled back, the parameters are adaptively adjusted, and rescheduling is triggered. If the judgment conditions are not met, the process is further judged whether it has ended normally. If the process has not ended normally, it returns to the periodic sampling phase. If the process ends normally, the standardized JSON metadata is output and the process ends.

[0062] Example 4: In a factory inspection application scenario involving areas shielded by GPS signals, the displacement span acquired by the feature perception module deviates, causing an attribute conflict between this metadata feature and the user-preset precision-carved pipeline labels. The pipeline scheduling module calculates collaborative decision factors... Initiate arbitration proceedings to extract confidence scores for specific scene labels. And calculate the consistency score of the color histogram statistical distribution of the image sequence in the red-green-blue space. Based on the probability and statistics theory of Bach distance, the distribution overlap characterizes the visual similarity features of a multidimensional signal set. The feature perception module extracts the color channel histogram vector of the image to be reconstructed, calculates the Bach distance coefficient between the actual distribution vector and the histogram vector of the preset standard inspection scene, calculates the reciprocal of the coefficient and normalizes it to the interval of 0 to 1, and outputs the statistical distribution consistency score. In this comparison process, the preset standard inspection scene histogram vector used for anchoring is constructed by the system by pre-reading over 10,000 archived photos of substation equipment covering different weather conditions and time periods throughout the four seasons. These image libraries are uniformly projected into an independent hue-saturation-brightness (HSV) color space. After removing the brightness channel perturbation, full-pixel clustering is performed according to 256 levels of color depth to generate a mean one-dimensional distribution matrix with scene generalization. At the same time, when real-time accessing sensor data from different batches or with varying optical quality, the feature perception pipeline performs baseline correction and alignment on the new sequence before extracting the histogram through adaptive histogram equalization (CLAHE) and linear gain compensation operations, thereby eliminating color gamut shift interference caused by hardware noise and photosensitivity errors. Based on the attenuation law of equipment positioning accuracy, the system reads the three-dimensional position accuracy factor output by the GPS receiver and extracts the reciprocal of the position accuracy factor as the confidence score. The two are then fused using a decision mapping function; where, As a collaborative decision-making factor, The confidence score for the scene label. Scoring is applied to the distribution of metadata features and image content. As the confidence level weight, As the distribution weight, in this embodiment, it is set to... and ;when When the calculation result is lower than the preset decision threshold of 0.5, the system determines that the unreliability of the metadata features is high and forces the target reconstruction pipeline to be corrected to the high-speed pipeline to avoid overloading computing resources caused by starting high-precision reconstruction in the environment of location feature failure.

[0063] Determination threshold duration The determination stems from an offline calibration procedure for a specific computing platform. The process involves starting 50 high-speed pipeline tasks with different feature point densities under standard computing power conditions, recording the maximum log writing interval of the computing process under normal conditions as 120 seconds, and adding a 60-second offset to this baseline value to determine the duration of the judgment threshold. `s` serves as the basis for identifying process suspension; simultaneously, after the 3D Gaussian splash calculation task begins, the computation control module triggers state backup every 500 iterations, serializing and storing the current spherical harmonic function coefficients and three-dimensional spatial position parameters into a logical handle in the high-speed temporary storage area, and assigning a unique timestamp identifier to each backup file; when the monitoring and recovery module detects the system health score through the kernel interface... When the sampling rate is below 10% for three consecutive sampling periods and the current input / output silence duration exceeds 180 seconds, the system triggers a closed-loop self-healing response. This involves resetting the stored parameters to locate the most recent valid checkpoint file, reloading the backed-up parameter set into the processor registers and video memory address space, and achieving breakpoint backtracking of the computation task without restarting the feature extraction step. It's important to clarify that this reload process does not involve directly accessing the underlying hardware physical registers; instead, it is achieved by calling the application-level graphics processing unit (GPU) computing architecture interface (such as CUDA API). Specifically, the system retrieves data from the high-speed temporary storage area... The checkpoint file is deserialized, and the saved spherical harmonic function coefficients and 3D spatial position parameters are remapped to the newly allocated safe memory block of the system. The context object mapping of the calculation process is reconstructed simultaneously to update the application layer's logical execution stack and virtual register state. This allows the algorithm execution state to be legally and safely restored within the software protocol framework. This method solves the calculation stagnation caused by changes in ambient lighting inside the factory. After a 4.2s backtracking reset, the model training task resumes its convergence trend. The final generated 3D equipment asset maintains geometric topological integrity, verifying the logical determinism of this invention in handling abnormal operating conditions and feature conflicts.

[0064] Example 5: In a reconstruction scheduling scenario involving heterogeneous computing nodes and multiple batches of optical sensors, the preset correlation matrix called by the feature perception module is filled with data through an offline calibration program, selecting data containing... A set of standard images with known displacement lengths and resolution gradients is used to measure peak memory usage under different pipeline configurations on a preset graphics processor. Computation time The system uses a search algorithm to determine the critical point that causes the video memory utilization rate to fall below 85%, and writes this critical point into a preset correlation matrix. Based on the principle of discrete data boundary optimization, the system establishes a stable mapping boundary for the multi-dimensional parameter matrix. The calculation control module sets the feature point extraction density as the initial control variable, and gradually decreases the control variable at fixed compensation steps. Within a single test cycle, the system reads the value of the graphics processor's video memory usage register and compares it with the current peak video memory usage. When the difference between the memory capacity and the physical upper limit of 85% is detected, and the polarity of the difference is reversed, the corresponding pipeline configuration parameters at this moment are extracted to form a critical judgment point. For the number of image groups, This represents the video memory usage value. To calculate the computation time, and in order to distribute the above-mentioned calibrated tasks to the distributed computing cluster, the distributed node allocation module periodically and continuously listens to the system heartbeat messages broadcast by each heterogeneous computing node in the network topology through the embedded lightweight message queue telemetry transport protocol (MQTT), strips the messages to obtain the current physical memory net balance and processor load percentage of the target node, and dynamically updates the global allocatable node status hash table.

[0065] During scheduling, the algorithm engine compares the computational load threshold specified by the parameter constraint set and uses a resource-weighted round-robin strategy to select the best idle computing node that meets the standard of sufficient video memory and has the lowest data round-trip latency. It then generates an image source data package containing a Uniform Resource Locator (URL) along with the configured pipeline parameters and pushes it to the target physical machine, completing the transfer of control from the scheduling center to the computing endpoint. When the system faces situations where the floating-point arithmetic precision is inconsistent under different hardware architectures, the computing control module runs a pre-debugging program to align with the benchmark and selects images with high curvature gradients from the image sequence. Using 10 feature points as anchor points, the Euclidean distance deviation of the singular value decomposition output feature vectors is calculated in both single-threaded and multi-threaded modes. In Euclidean distance deviation Exceeding the preset numerical tolerance threshold At that time, the number of parallel threads Locked to 1, establishing the convergence state of the three-dimensional geometric constraints.

[0066] Example 6: In a transmission line inspection scenario involving changes in inspection height, variations in ground resolution cause offsets in displacement span determination. The system runs a pre-deployment calibration program to determine the quantization criteria for a preset correlation matrix. This program is based on a reference height. The following standard sequence containing 100 frames of images is acquired, and the pixel displacement between adjacent frames is calculated. With physical displacement span The transformation coefficients are calculated and written into the feature space of the preset correlation matrix. The calculation control module is then used to determine the upper limit of the feature point size. Distribution optimization is performed by calculating the Laplacian response value through Laplacian operator convolution on the sampled images. The probability density distribution is used, and the response value when the cumulative distribution function reaches 90% is selected as the adaptive pruning threshold. Under the condition that the total number of feature points is limited to 8000, the structural features at the connection between the insulator string and the tower are extracted.

[0067] To address the risk of process suspension in heterogeneous edge computing environments, the system initiates automated logic calibration to establish monitoring and self-healing trigger conditions. This process injects 10% simulated computing latency into the target computing node and records the processor utilization. With log update frequency The steady-state envelope interval is determined, and the system health score is fitted using the least squares method. Weighting coefficients in and ,in, and The weights are respectively the log update frequency weight and the computing resource consumption weight, and satisfy the following conditions: The linear constraint relationship, when the system health score output by the calibration program... When the recognition lead time of the response curve reaches 5 seconds when calculating the stagnation point, the system will maintain the corresponding judgment threshold for a certain duration. The logic gates of the computational control module are embedded in the monitoring and recovery module. Within a specified latency-sensitive threshold, the module completes task backtracking. In aerial photography scenarios involving long-distance facility inspections across regions, the feature perception module extracts the GPS coordinates of the start and end image frames from the image exchange image file format of the image sequence to be reconstructed, calculates the three-dimensional Euclidean distance between them, and obtains the physical displacement span. If the physical displacement span exceeds a preset displacement threshold, the module will proceed accordingly. At that time, the pipeline scheduling module identifies the target reconstructed pipeline as the aerial photography pipeline.

[0068] Because the 3D Gaussian reconstruction process generates a large amount of intermediate cache data, the computation control module reconstructs the logical storage path according to the storage reset parameters. It allocates isolated resource handles to the computation process through the kernel file system interface, and maps the model training data and point cloud cache to the independent address space of the high-speed temporary storage area. This solves the write anomaly caused by global storage path conflict under multi-task parallel operation. At the same time, combined with numerical stability constraints, it performs pixel-by-pixel independent calculation of the spherical harmonic function coefficients of the feature point cloud. It uses memory barrier instructions to ensure the atomicity of singular value decomposition operation in single-threaded mode, so that the system maintains the independence of data storage and the stability of geometric convergence when processing aerial-level image sets.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A scene-aware based 3D Gaussian automated reconstruction scheduling system, characterized in that, include: The feature perception module acquires metadata features of the image sequence to be reconstructed, including image resolution and the displacement span of the image acquisition trajectory in three-dimensional space. The pipeline scheduling module is connected to the feature perception module. It maps metadata features to target reconstruction pipelines through a preset association matrix. The target reconstruction pipelines include fine-sculpting pipelines, high-speed pipelines, and aerial photography pipelines. The computation control module, connected to the pipeline scheduling module, initiates a 3D Gaussian splash computation task based on the parameter constraint set corresponding to the target reconstructed pipeline. The parameter constraint set includes the upper limit of feature point size, numerical stability constraints, and storage reset parameters. The convergence state of the 3D Gaussian splash computation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the computation process. The monitoring and recovery module is connected to the computing control module to obtain the activity parameters of the 3D Gaussian splash computing task. The activity parameters are represented by the processor utilization and the log update frequency. When the processor utilization is lower than 0.1% and the log update frequency is lower than the preset silent threshold, the monitoring and recovery module issues a status backtracking command and updates the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash computing task.

2. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, The feature perception module also acquires scene labels of image frames in the image sequence to be reconstructed; when the pipeline scheduling module identifies an attribute conflict between the scene label and the metadata features during the mapping process, it introduces a collaborative decision factor for arbitration and switches the target reconstruction pipeline to the high-speed pipeline; the collaborative decision factor is obtained based on the consistency between the confidence level of the scene label and the statistical distribution of the metadata features.

3. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, The upper limit of the feature point size in the parameter constraint set is determined by the visual contribution scoring mechanism; the calculation control module limits the upper limit of the feature point size through the following sub-steps: Step 31: Calculate the Laplacian response value of the image feature points; Step 32: Retain the feature points in descending order of the Laplacian response value until the upper limit of the feature point size is reached.

4. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, The storage reset parameters are used to reconstruct the logical storage path before the 3D Gaussian splash calculation task starts; the calculation control module maps the intermediate cache data and model output data of the 3D Gaussian splash calculation task to the high-speed temporary storage area and allocates isolated resource handles for different target reconstruction pipelines.

5. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, When the feature perception module obtains the displacement span, it extracts the GPS coordinates from the image exchange file format and calculates the Euclidean distance between the start and end image frames in the image sequence. When the Euclidean distance exceeds the preset displacement threshold, the pipeline scheduling module identifies the target reconstruction pipeline as the aerial photography pipeline.

6. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, When the target reconstructed pipeline is a precision pipeline, the pipeline scheduling module sets the feature extraction density in the parameter constraint set to a first preset value; when the target reconstructed pipeline is a high-speed pipeline, the feature extraction density is set to a second preset value, and the first preset value is greater than the second preset value.

7. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, When scheduling the high-speed pipeline, the computation control module suppresses nondeterministic fluctuations through the following sub-steps: Step 71: Limit the number of parallel threads to 1; Step 72: Perform singular value decomposition operations sequentially by a single thread to eliminate nondeterministic fluctuations in the algebraic operation results under geometric constraints caused by thread competition.

8. The scene-aware 3D Gaussian automated reconstruction scheduling system of claim 1, wherein, After issuing the status rollback command, the monitoring and recovery module reads the sparse point cloud data backed up by the 3D Gaussian splash calculation task before it stalled as the initial state, and lowers the upper limit of the feature point scale according to the degradation strategy, and then calls the calculation control module again.

9. A scene-aware 3D Gaussian automated reconstruction scheduling system according to claim 1, characterized in that, The system also includes a distributed node allocation module, which is used to search for idle computing nodes that meet the parameter constraint set in the distributed computing cluster based on the total amount of data of the image sequence to be reconstructed, and to distribute the scheduling weight of the 3D Gaussian splash computing task to the selected idle computing nodes.

10. A scene-aware 3D Gaussian automated reconstruction scheduling method, used to implement the scene-aware 3D Gaussian automated reconstruction scheduling system as described in claim 1, characterized in that, Includes the following steps: Step 101: Obtain the metadata features of the image sequence to be reconstructed. The metadata features include the image resolution and the displacement span of the image acquisition trajectory in three-dimensional space. Step 102: Map metadata features to target reconstruction pipelines using a preset association matrix. Target reconstruction pipelines include fine-tuning pipeline, high-speed pipeline, and aerial photography pipeline. Step 103: Start the 3D Gaussian splashing calculation task according to the parameter constraint set corresponding to the target reconstruction pipeline; wherein, the parameter constraint set includes the upper limit of feature point size, numerical stability constraints and storage reset parameters, and the convergence state of the 3D Gaussian splashing calculation task is determined by limiting the feature point density and suppressing the floating-point accumulation fluctuation of the calculation process. Step 104: Obtain the activity parameters of the 3D Gaussian splash calculation task in real time. The activity parameters are characterized by processor utilization and log update frequency. Step 105: When the processor utilization is below 0.1% and the log update frequency is below the preset silent threshold, issue a state backtracking instruction and update the parameter constraint set by lowering the upper limit of the feature point size in order to restart the 3D Gaussian splash calculation task.

Images