Mountain torrent debris flow dynamic and static segmentation and directional monitoring method and system

Abstract

The invention discloses a mountain torrent debris flow dynamic and static segmentation and directional monitoring method and system, and the method comprises the steps: firstly carrying out the dynamic boundary filtering and improved RANSAC plane extraction of an original point cloud, constructing a stable terrain reference, and achieving the self-adaptive updating through a multi-frame fusion and time decay mechanism; then, KDTree is adopted to realize preliminary classification of dynamic and static point clouds, and in combination with DBSCAN clustering and small cluster filtering strategies, sensor noise interferences such as raindrops and boulders are effectively eliminated; for a dynamic fluid point cloud, a curvature sensing non-uniform voxel downsampling method is introduced, fluid edge features are retained, and the data volume is compressed. And a main flow direction is further extracted through PCA principal component analysis, and a natural coordinate system is constructed to adapt to the channel terrain. The system realizes resource optimization and emergency sensitive detection based on a Streamz stream processing structure and an event triggering mechanism. The whole process supports low-power-consumption embedded deployment, has real-time performance, robustness and geometric fidelity, and is suitable for automatic debris flow monitoring and early warning in a complex mountainous area environment.

Description

Technical Field

[0001] The present invention belongs to the cross technical field of geological disaster monitoring and computer vision, and particularly relates to a method and system for static and dynamic segmentation and directional monitoring of mountain torrents and debris flows. Background Art

[0002] Mountain torrents and debris flows are complex water-sediment coupling disasters with strong suddenness and great destructive power. Their movement processes have characteristics such as non-linearity, high density, and high sediment concentration, posing a serious threat to mountain engineering facilities and residential areas. Therefore, accurate and real-time monitoring of them is of great significance for disaster reduction and prevention. Traditional monitoring of mountain torrents and debris flows mainly relies on single-point contact devices such as water level gauges and ultrasonic sensors to achieve early warning by measuring apparent parameters such as flow depth and flow velocity. However, due to the significant non-Newtonian fluid characteristics of debris flows, the entrainment of a large amount of sediment and boulders leads to the complication of rheological behavior, and the applicability of traditional flow rate calculation methods based on empirical formulas is significantly reduced. In addition, it is difficult to deploy contact sensors in complex mountain terrains, and the damage rate is high when encountering extreme flood impacts. Moreover, they cannot obtain three-dimensional morphological parameters and surface flow field distributions of debris flows, resulting in insufficient reliability of the early warning system.

[0003] In recent years, non-contact monitoring means such as Doppler radar and laser water level gauges have been gradually applied, improving the data acquisition efficiency. However, such technologies are still limited to obtaining one-dimensional parameters of a local cross-section and cannot analyze the three-dimensional dynamic morphology of debris flows. Signal attenuation under rain and fog conditions and multipath interference caused by channel bends further limit their environmental adaptability. Existing algorithms are mostly based on the assumption of uniform flow and are difficult to accurately depict non-linear processes such as the accumulation, blockage, and breach of debris flow fronts. The development of high-frequency lidar technology provides a new way for three-dimensional monitoring of mountain torrents and debris flows, realizing non-contact dynamic monitoring by continuously collecting point cloud data of river valley cross-sections. However, the processing of massive point clouds under complex mountain terrains faces severe challenges: First, the amount of original point cloud data is huge (>10GB / hour), containing a large number of noise points in non-target areas, and the calculation load of traditional filtering algorithms is high; second, the accurate segmentation of static terrain and dynamic fluid is a key technical bottleneck, and the misclassification rate of existing methods based on simple distance threshold determination is as high as over 30% in areas with complex riverbed morphologies. Existing point cloud processing methods mainly include three categories: segmentation methods based on geometric features (such as region growing method, plane fitting) are difficult to set parameters in complex terrain environments; classification methods based on machine learning have high accuracy, but the training depends on a large amount of labeled data and the inference calculation complexity is high; change detection methods based on temporal contrast can identify dynamic regions, but they have high requirements for point cloud registration accuracy and are insensitive to small changes.

[0004] In addition, current point cloud processing systems mostly adopt batch processing mode, which has significant latency when processing high-frequency point cloud data and is difficult to meet the millisecond-level response requirements for mountain flood and debris flow monitoring. Batch processing systems consume high resources and are difficult to run in real time on on-site devices facing power supply and communication limitations in uninhabited mountainous areas. The reliability of equipment drops sharply under the impact of strongly destructive debris flows, further increasing the difficulty of real-time monitoring. In the face of the challenge of real-time processing of mountain flood and debris flow point cloud data, existing technologies have not provided efficient and robust streaming processing solutions. Especially in key links such as point cloud filtering, dynamic and static segmentation, and geometric feature extraction, there is still a lack of algorithms and system architectures optimized for complex mountain terrain environments. Therefore, there is an urgent need for a method for dynamic and static point cloud segmentation and directional monitoring of mountain floods and debris flows based on streaming processing, which can efficiently process high-frequency lidar point cloud data, achieve precise segmentation of static terrain and dynamic water bodies, and extract key parameters such as flow direction to provide technical support for debris flow early warning. Summary of the Invention

[0005] In order to solve the technical problems such as poor real-time performance and insufficient accuracy existing in the existing mountain flood and debris flow monitoring technologies, the present invention aims to provide a method and system for dynamic and static segmentation and directional monitoring of mountain floods and debris flows. The method collects point cloud data in real time through a high-frequency lidar, and innovatively adopts a dual-path processing strategy and a streaming computing architecture driven by plane changes, controlling the system response time within 50 ms, significantly improving the real-time performance and accuracy of mountain flood and debris flow monitoring in complex mountainous environments, and providing reliable technical support for disaster early warning.

[0006] In order to solve the technical problems, the technical solution of the present invention is:

[0007] A method for dynamic and static segmentation and directional monitoring of mountain floods and debris flows, the method comprising:

[0008] S1: Perform dynamic axial bounding box AABB filtering on the original point cloud, screen the effective point cloud by adaptively adjusting the boundary range to adapt to the changes in the channel morphology; adopt an improved random sample consensus RANSAC algorithm to extract the riverbed plane, and construct a robust reference point cloud through multi-frame point cloud fusion and a reference dynamic update mechanism;

[0009] S2: Based on the constructed reference point cloud, the preliminary classification of dynamic and static point clouds is carried out through the KDTree (K-dimensional tree). By comparing the minimum distance from a point to the reference point cloud with a preset threshold, dynamic fluids and static terrains are distinguished. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used for clustering to remove measurement errors, raindrops, and boulder sensor noise interference. At the same time, small misclassified point clusters are removed through a small cluster filtering strategy. For the dynamic fluid point cloud, a curvature-aware non-uniform voxel downsampling method is introduced to retain the edge features of the fluid and compress the data volume, and the dynamic point cloud and static point cloud after dynamic and static point cloud classification, clustering optimization, and downsampling processing are output.

[0010] S3: Based on the continuous time series and the downsampled point cloud output in step S2, the main flow direction is extracted through principal component analysis (PCA), and a natural coordinate system is constructed to adapt to the channel terrain. The system uses the Streamz stream processing framework, and through the sliding window mechanism, the planar changes are monitored in real time. When the change rate exceeds the set threshold, the dynamic point cloud segmentation process is triggered for deeper analysis, and finally, the construction of the main flow direction coordinate system and the generation of the event trigger flag are realized.

[0011] Further, in step S1, the dynamic axial bounding box filtering includes:

[0012] The mountain flood and debris flow monitoring area usually presents a channel shape, and there are a large number of non-monitoring targets such as vegetation and overhanging rocks around it. The AABB (Axis-Aligned Bounding Box) algorithm is constructed to adaptively define the effective monitoring area, and the filtering formula is:

[0013]

[0014] where P represents the set of original point cloud data, P valid represents the set of effective point clouds after filtering, p represents any point in the point cloud, p x and p y respectively represent the x coordinate and y coordinate of the point. and represent the minimum and maximum boundary values along the x-axis at time t, and represent the minimum and maximum boundary values along the y-axis at time t. This formula screens the effective point cloud through dynamically adjusted spatial range parameters, and can effectively cope with the channel changes caused by debris flows. The boundary parameters are dynamically adjusted according to the effective point ratio r within the sliding window:

[0015]

[0016] where ΔX max represents the adjustment amount of the maximum boundary along the x-axis, λ represents the boundary adjustment coefficient to control the adjustment amplitude, r thIt represents the effective point ratio threshold, which is the expected proportion of effective points. r represents the proportion of effective points in the current frame, σ represents the adjustment sensitivity parameter that controls the steepness of the response curve, and tanh represents the hyperbolic tangent function. This formula realizes the adaptive adjustment of the boundary range, expanding the boundary when the effective point ratio is low and maintaining stability when it is high.

[0017] Furthermore, in the step S1, an improved RANSAC algorithm is used to extract the riverbed plane, including:

[0018] For the improved RANSAC algorithm, the plane model is defined as π: ax + by + cz + d = 0, and the inlier set optimization formula is:

[0019]

[0020] where S inlier represents the optimal inlier set, S represents the candidate inlier set selected from the effective point cloud P valid p represents any point in the point cloud, π represents the plane model defined as π: ax + by + cz + d = 0, where (a, b, c) is the plane normal vector and d is the plane offset. D(p, π) represents the Euclidean distance from point p to plane π, ∈ represents the distance tolerance parameter, and exp represents the natural exponential function. By introducing Gaussian weights, points closer to the plane contribute more, enhancing the robustness of the algorithm to local outliers and being particularly suitable for situations with local boulders and floating logs interfering in the flash flood environment.

[0021] Define the plane change rate double-factor evaluation index:

[0022]

[0023] where Δπ represents the plane change rate, ω1 represents the weight of the normal vector angle change, and ω2 represents the weight of the plane displacement change. represents the unit normal vector of the current frame plane. represents the unit normal vector of the previous frame plane. represents the dot product of the two normal vectors, arccos represents the inverse cosine function, and d t represents the offset of the current frame plane, and d t-1 represents the offset of the previous frame plane. This index comprehensively considers two factors: the change in the normal vector angle and the plane displacement, enabling the system to distinguish between slight water level fluctuations and sudden debris flows. When Δπ > 0.05, the dynamic segmentation process is triggered, avoiding the system's over-response to noise changes while ensuring sensitivity to important events.

[0024] Furthermore, in the step S1, the multi-frame point cloud fusion and reference dynamic update mechanism includes:

[0025] Constructing a Robust Benchmark through Multi-Frame Fusion:

[0026]

[0027] Among them, P base represents the constructed benchmark point cloud, represents the union operation on the valid point clouds of the previous k frames, represents the valid point cloud of the i-th frame, ↓ voxel represents the voxel downsampling operator, which fuses the valid point clouds of the previous k frames and controls the data volume through voxel downsampling. Voxel downsampling not only reduces the computational burden but also smooths local noise, generating a benchmark point cloud with uniform density and improving the subsequent KDTree search efficiency;

[0028] To adapt to long-term environmental changes, a benchmark dynamic update mechanism based on time decay is designed:

[0029]

[0030] Among them, represents the summation within the time window from time t - m to t, Δπ i represents the plane change rate of the i-th frame, β represents the time decay factor, t - i represents the time difference between the i-th frame and the current frame t, e -β(t-i) represents the time decay weight, Γ represents the benchmark reconstruction trigger threshold; this formula calculates the time-weighted cumulative plane change. When the cumulative amount exceeds the threshold, benchmark reconstruction is triggered, giving higher weight to recent changes.

[0031] Furthermore, in the step S2, the KDTree is used for the preliminary classification of dynamic and static point clouds, including:

[0032] Accurately distinguishing static terrain from dynamic fluids is the core of mountain flood and debris flow monitoring. Using the KDTree spatial index, efficient classification of dynamic and static point clouds is achieved:

[0033]

[0034] Among them, P dynamic represents the set of identified dynamic point clouds, P new represents the currently acquired current frame point cloud, p represents any point in the current frame point cloud, q represents any point in the benchmark point cloud, ‖p - q‖2 represents the Euclidean distance between point p and point q, represents finding the nearest point to point p in the benchmark point cloud, τ d represents the dynamic and static classification distance threshold. The KDTree reduces the query complexity from O(n) to O(log n), enabling real-time analysis of large-scale point clouds; the threshold τ dLet it be the threshold value. Noise less than the sensor resolution is regarded as static, and if it is greater than this value, it is determined as dynamic fluid, balancing sensitivity and stability.

[0035] Furthermore, in the step S2, the DBSCAN algorithm is used for clustering, including:

[0036] Apply the DBSCAN algorithm to cluster and optimize the static point set, removing misclassified points caused by measurement errors and sensor noise. For any point p in the static point set, its ∈-neighborhood is defined as:

[0037] N ∈ (p) = {q ∈ P stati c|||p - q||2 ≤ ∈} (8)

[0038] where, N ∈ (p) represents the ∈-neighborhood of point p, P static represents the static point cloud set, q represents any point in the static point cloud, ‖p - q‖2 represents the Euclidean distance between point p and point q, ∈ represents the neighborhood radius parameter of the DBSCAN algorithm. The key parameters of DBSCAN are the neighborhood radius ∈ and the minimum number of points MinPts. Points are classified into core points, border points, and noise points through density judgment; in the monitoring of mountain torrents and debris flows, the random fluctuations of the point cloud mainly originate from raindrops, tiny suspended substances in the air, and sensor errors, usually manifested as small clusters of sparsely distributed points. By setting the value of MinPts, it is ensured that only sufficiently dense point clusters are recognized as real static regions;

[0039] Introduce the minimum effective cluster point number threshold N min Filter small regions:

[0040]

[0041] where, represents the optimized dynamic point cloud set, P dynamic represents the dynamically classified point cloud set after preliminary classification, P static represents the static point cloud set after preliminary classification, p represents any point in the point cloud, L(p) represents the clustering label of point p, -1 represents a noise point, C L(p) represents the cluster with the label L(p), |C L(p) | represents the number of points in the cluster C L(p) N min represents the minimum effective cluster point number threshold, effectively filtering isolated static points caused by measurement errors, retaining real static regions, and improving the accuracy of dynamic point cloud segmentation.

[0042] Furthermore, in the step S2, a curvature-aware non-uniform voxel downsampling method is introduced, including:

[0043] Considering that the segmented dynamic point clouds are usually unevenly distributed, and directly participating in geometric calculations causes the high-density areas to overly affect the results, a non-uniform downsampling method considering geometric features is constructed:

[0044]

[0045] Among them, represents the representative point of voxel I, that is, the center point, V I represents the set of points contained in voxel I, |V I | represents the number of points in voxel I, p represents any point in the voxel, η represents the curvature weight coefficient, represents the curvature gradient at point p, which is calculated by estimating the principal curvatures of the local neighborhood; this formula introduces a curvature weighting term when calculating the voxel center, enabling the downsampled point cloud to better retain the key features of the water surface edge and mutation area, while significantly reducing the data volume and improving the efficiency of subsequent processing.

[0046] Furthermore, in the step S3, the main flow direction is extracted by PCA, including:

[0047] Determine the main flow direction of the debris flow by principal component analysis and establish a natural coordinate system:

[0048]

[0049] Among them, C represents the 3×3 covariance matrix of the point cloud, n represents the total number of points in the point cloud, p i represents the three-dimensional coordinate vector of the i-th point in the point cloud, μ represents the center point of the point cloud, that is, the average coordinate, (p i -μ)(p i -μ) T represents the vector outer product, V represents the eigenvector matrix, whose column vectors form the principal axis coordinate system, Λ represents the eigenvalue diagonal matrix; the eigenvalue decomposition of the covariance matrix reveals the main change directions of the point cloud data. The eigenvector v1 corresponding to the largest eigenvalue represents the main flow direction, the second largest eigenvalue corresponds to v2 representing the width direction, and the smallest eigenvalue corresponds to v3 representing the depth direction. Through coordinate transformation:

[0050] p′ i =V T (p i -μ) (12)

[0051] Among them, p i ′ represents the coordinate of point p i in the principal axis coordinate system, V represents the eigenvector matrix, whose column vectors are the principal axis directions, V T represents the transpose of the eigenvector matrix, p iDenote the original coordinates of the \(i\)-th point in the point cloud, and \(\mu\) represents the center point of the point cloud; project the point cloud onto the principal axis coordinate system to make the subsequent geometric parameter calculation more in line with the actual shape of the river channel. This adaptive coordinate system is suitable for the complex terrain in mountainous areas and can adapt to the bending changes of the river channel without manual intervention.

[0052] Further, in the step S3, the system adopts the Streamz stream processing framework, including:

[0053] Based on the Streamz-based stream processing framework, construct a directed acyclic graph data flow topology structure:

[0054] y i = f i (f i-1 (...f1(x)...)) (13)

[0055] where \(x\) represents the input data, i.e., the original point cloud, and \(y\) i represents the output of the \(i\)-th processing node, and \(f1, f2,..., f\) i represent a series of transformation functions in the data flow, and \(f\) i (f i-1 (...)) represents the function composition operation; adopt a sliding window mechanism to analyze time series data and continuously monitor the change of plane parameters:

[0056] W t = {x t-w+1 , x t-w+2 ,..., x t} (14)

[0057] where \(W\) t represents the sliding window at time \(t\), \(w\) represents the window size, and \(x\) t represents the data at time \(t\), and \(x\) t-w+1 , x t-w+2 ,..., x t represent the continuous data included in the window; based on the event trigger mechanism of the plane change rate, when the change rate exceeds the preset threshold, activate the dynamic point cloud segmentation process:

[0058]

[0059] where \(E\) t represents the event trigger flag at time \(t\), 1 represents trigger, 0 represents non-trigger, and \(\Delta\pi\) t represents the plane change rate at time \(t\), and \(\theta\) represents the event trigger threshold; the system optimizes resource utilization through a dual-path processing strategy: only perform plane monitoring, i.e., the simple path, under normal conditions, and activate the complete analysis process, i.e., the complex path, when a change is detected.

[0060] A dynamic and static segmentation and directional monitoring system for mountain flood debris flow, the system is applied to any one of the above-mentioned methods, and the system includes:

[0061] Point cloud data preprocessing and reference point cloud construction module: Dynamically filter the original point cloud with an axial bounding box AABB, screen the effective point cloud by adaptively adjusting the boundary range to adapt to the change of the channel morphology; use an improved random sample consensus RANSAC algorithm to extract the riverbed plane, and construct a robust reference point cloud through multi-frame point cloud fusion and reference dynamic update mechanism;

[0062] Dynamic and static point cloud segmentation and feature extraction module: Based on the constructed reference point cloud, initially classify the dynamic and static point clouds through a KD tree (KDTree), and distinguish the dynamic fluid from the static terrain according to the comparison between the minimum distance from the point to the reference point cloud and a preset threshold; use the density-based spatial clustering of applications with noise DBSCAN algorithm for clustering to remove measurement errors, raindrops, and boulder sensor noise interference, and at the same time eliminate misclassified small point clusters through a small cluster filtering strategy; for the dynamic fluid point cloud, introduce a curvature-aware non-uniform voxel downsampling method to retain the fluid edge features, compress the data volume, and output the dynamic point cloud and static point cloud after dynamic and static point cloud classification, clustering optimization, and downsampling processing;

[0063] Principal component orientation and flow architecture module: Based on the continuous time series and the downsampled point cloud output in step S2, extract the main flow direction through principal component analysis (PCA) and construct a natural coordinate system to adapt to the channel terrain; the system uses the Streamz stream processing framework to monitor the plane change in real time through a sliding window mechanism. When the change rate exceeds the set threshold, trigger the dynamic point cloud segmentation process for deeper analysis, and finally realize the construction of the main flow direction coordinate system and the generation of an event trigger flag.

[0064] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a dynamic and static segmentation and directional monitoring method for mountain flood debris flow as described in any one of the above.

[0065] A computer-readable storage medium stores a computer program, and when the program is executed by a processor, it implements a dynamic and static segmentation and directional monitoring method for mountain flood debris flow as described in any one of the above.

[0066] Compared with the prior art, the advantages of the present invention are:

[0067] a. Compared with traditional contact monitoring technologies, the present invention realizes the comprehensive monitoring of the three-dimensional morphology of debris flows through non-contact acquisition and analysis of high-frequency lidar point cloud data, overcomes the defects of easy damage and limited measurement points of contact devices in harsh environments, and greatly improves the reliability and accuracy of the monitoring system.

[0068] b. Compared with existing point cloud processing methods, the dynamic and static point cloud segmentation algorithm proposed by the present invention can accurately distinguish static terrain from dynamic fluids, and can achieve a segmentation accuracy rate of over 90% even in complex riverbed environments, providing a reliable data basis for subsequent geometric parameter calculations.

[0069] c. Compared with traditional batch processing systems, the streaming processing architecture designed by the present invention can achieve millisecond-level real-time analysis of point cloud data, reduce the system response delay from the second level to within 50 ms, and at the same time, through the event trigger mechanism and multi-level processing paths, achieve efficient utilization of computing resources, which is particularly suitable for the long-term monitoring requirements in mountainous environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The overall flowchart of the method for dynamic and static point cloud segmentation and directional monitoring of mountain flood debris flows based on streaming processing of the present invention;

[0071] Figure 2 The schematic diagram of the axial boundary box (AABB) dynamic filtering algorithm of the present invention;

[0072] Figure 3 The schematic diagram of the improved RANSAC plane change monitoring algorithm of the present invention;

[0073] Figure 4 The effect diagram of dynamic and static point cloud segmentation of the present invention;

[0074] Figure 5 The schematic diagram of the PCA principal axis orientation algorithm of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0075] The following describes the specific embodiments of the present invention in conjunction with the embodiments:

[0076] It should be noted that the structures, ratios, sizes, etc. shown in this specification are only used to cooperate with the content disclosed in the specification for those skilled in this technology to understand and read, and are not used to limit the limiting conditions that the present invention can be implemented. Any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed by the present invention.

[0077] Meanwhile, terms such as "upper", "lower", "left", "right", "middle", and "one" cited in this specification are only for the sake of clarity in narration and are not used to limit the scope of implementation of the present invention. Changes or adjustments in their relative relationships shall also be regarded as the scope of implementation of the present invention without substantial changes in technical content.

[0078] Example 1:

[0079] This example provides a method for static and dynamic segmentation and directional monitoring of mountain flood and debris flow based on high-frequency lidar point cloud data. Through a streaming processing architecture and real-time analysis algorithms, precise segmentation of static terrain and dynamic fluids, flow direction and motion feature extraction are achieved, solving the technical problems of limited data dimensions and large processing delays in traditional monitoring means, and providing real-time decision support for early identification and warning of mountain flood and debris flow. In the present invention, "river channel" specifically refers to the narrow gully where mountain debris flow occurs, including static terrain - the gully bed (the basal part of debris flow movement) and dynamically flowing debris flow fluid.

[0080] As Figure 1 shown, a method for static point cloud segmentation and directional monitoring of mountain flood and debris flow based on streaming processing includes data acquisition, preprocessing, dynamic and static segmentation, feature extraction, and an event-driven dual-path processing pipeline; it includes the following steps:

[0081] S1: Preprocessing of point cloud data and construction of reference point cloud

[0082] This example uses high-frequency lidar point cloud data collected by the Illgraben debris flow monitoring station in the Canton of Valais, Switzerland, with a sampling rate of 10 Hz.

[0083] S1.1: Axial Aligned Bounding Box (AABB) filtering

[0084] To improve processing efficiency, spatial filtering of the point cloud data is first performed to determine the effective monitoring area. In the example, the horizontal monitoring range is set as [-2m, 14m], and the vertical monitoring range is [-10m, -1m]. A sliding window trigger mechanism is introduced. When the proportion of valid points in consecutive frames is lower than the preset threshold of 0.85, the boundary parameters are automatically adjusted with a step size of 0.5m. Smooth adjustment is achieved through the hyperbolic tangent function, enabling the boundary to be dynamically adjusted according to the channel morphology changes, avoiding system instability caused by sudden changes. This step effectively reduces the amount of data to be processed, improves the calculation efficiency, and ensures that the monitoring area can accurately focus on the channel change area, avoiding interference from surrounding vegetation and slope noise on the monitoring results.

[0085] S1.2: Improved RANSAC plane segmentation

[0086] An improved RANSAC algorithm is used to extract the riverbed plane model, and the key parameters are set as follows: the number of random sampling points is 3, the number of iterations is 100, and the distance threshold is 0.2 m. The plane change rate is evaluated by a two-factor method, and the calculation formula is:

[0087]

[0088] where is the normal vector of the plane for two adjacent frames, and d1 and d2 are the plane offsets. The maximum value strategy can capture the mutations caused by debris flows more sensitively. When the change rate exceeds the preset threshold of 0.05, the dynamic segmentation process is triggered. The threshold is set based on the measured data analysis of Illgraben, which can effectively distinguish sensor noise from actual flash flood and debris flow events. Through the extraction of the plane model, the riverbed basic terrain and flowing water can be accurately distinguished. The two-factor evaluation mechanism is particularly suitable for capturing the mutation characteristics caused by debris flows, improving the timeliness of disaster warning.

[0089] S1.3: Multi-frame reference point cloud fusion

[0090] To build a stable reference, the first 5 frames of point clouds are selected for merging, and the data scale is controlled by voxel downsampling (voxel size 0.08 m). The cumulative plane change is calculated using a time-weighted method:

[0091]

[0092] where β = 0.2 is the time decay factor and m = 10 is the window size. When the cumulative amount exceeds the threshold of 0.5, the reference reconstruction is triggered. This time-weighted mechanism assigns higher weights to recent changes, can respond more sensitively to sudden events, and at the same time avoids the computational overhead caused by frequent updates. The stable reference established by multi-frame fusion can effectively suppress the interference of single-frame noise. The time-weighted accumulation mechanism not only ensures the sensitivity of the system to changes but also avoids false triggering caused by short-term disturbances such as sensor jitter.

[0093] S2: Dynamic and static point cloud segmentation and feature extraction

[0094] S2.1: KDTree nearest neighbor classification

[0095] Using the pre-constructed KDTree index of the reference point cloud, nearest neighbor queries are performed on the newly acquired point cloud. The formula for dynamic point determination is:

[0096]

[0097] Threshold τ dSet it to 0.02 m, which is comprehensively determined based on the measurement accuracy of lidar and the minimum detectable water level change of debris flow. The KDTree spatial index reduces the complexity of nearest neighbor queries from O(n) to O(log n), enabling real-time processing of high-frequency point cloud data. This step is the core of dynamic and static segmentation. Through an efficient spatial index structure, it achieves millisecond-level determination, solves the performance bottleneck of traditional methods in high-frequency data processing, and lays a foundation for subsequent debris flow feature extraction.

[0098] S2.2: DBSCAN Noise Filtering

[0099] Apply the DBSCAN clustering algorithm to optimize the static point set, with parameter configuration: neighborhood radius eps = 0.3 m, minimum number of samples min_samples = 10. The small area filtering strategy reclassifies clusters with a scale of less than 50 points as dynamic points, effectively removing isolated static points caused by measurement errors and improving the accuracy of dynamic point cloud segmentation. This method is particularly suitable for handling noise interference such as raindrops and small floating logs commonly found in mountain flood and debris flow monitoring. This step effectively filters out non-fluid point clouds such as raindrops, splashes, and floating logs in the air, avoiding these interference elements being misidentified as debris flow features and improving the accuracy of subsequent parameter calculations and system stability.

[0100] S2.3: Adaptive Voxel Downsampling

[0101] Sparsify the segmented dynamic point cloud through voxel downsampling technology, with the basic voxel size set to 0.1 m. To retain key features, a curvature-aware adaptive mechanism is adopted: first, estimate the point cloud curvature through local PCA, and then dynamically adjust the voxel size according to the curvature value. Smaller voxel sizes are used in high-curvature regions (such as the debris flow front), and larger voxel sizes are used in flat regions. The adjustment coefficient is set to 0.8, and the maximum considered curvature is set to 0.3. This strategy not only retains important geometric features but also significantly reduces the data volume. The curvature-aware adaptive mechanism ensures that key features such as the debris flow front are retained while reducing the data volume, avoiding the feature loss problem caused by traditional uniform downsampling and providing effective support for accurate flow state parameter estimation.

[0102] S3: Principal Component Orientation and Streaming Architecture

[0103] S3.1: PCA Main Axis Orientation

[0104] Perform principal component analysis on the dynamic point cloud to determine the main flow direction of mountain flood and debris flow. First, calculate the point cloud covariance matrix:

[0105]

[0106] The covariance matrix is decomposed into three main directions. The eigenvalue decomposition corresponds to the largest eigenvalue, which represents the main flow direction (usually along the river channel), the eigenvector corresponding to the second largest eigenvalue represents the width direction (usually perpendicular to the river channel), and the eigenvector corresponding to the smallest eigenvalue represents the depth direction. This adaptive coordinate system is particularly suitable for complex terrain in mountainous areas and can adapt to changes in river channel curvature without manual intervention. After projecting the point cloud onto the principal axis coordinate system, it is divided into 10 segments along the principal axis direction, which is convenient for subsequent segmented analysis of flow characteristics. The establishment of an adaptive coordinate system solves the problem of determining the flow direction in curved rivers in mountainous areas. There is no need to measure the direction of the river channel in advance. The system can automatically adapt to terrain changes and provide a unified reference framework for subsequent cross-sectional analysis and flow calculation.

[0107] S4: Stream processing architecture and task scheduling

[0108] S4.1: Streamz stream processing framework

[0109] A stream processing framework based on Streamz is constructed to decompose the monitoring task into multiple computing nodes such as data reading, plane detection, and dynamic segmentation to form a processing pipeline. A sliding window mechanism (window size 2, step size 1) is designed to continuously monitor changes in plane parameters, and an event trigger mechanism is implemented based on the plane change rate: when the change rate exceeds 0.05, the complete analysis process is activated, otherwise only simple plane monitoring is performed. A dual-path processing strategy is used to optimize resource utilization and realize real-time analysis of high-frequency point cloud data. The event-triggered dual-path processing strategy balances real-time performance and computing resource usage, allowing the system to run at low power consumption under normal water flow conditions, and can quickly switch to full analysis mode when a debris flow event occurs, which is suitable for long-term field deployment monitoring.

[0110] The method of the present invention has been tested and applied on the data of the Illgraben debris flow monitoring station in Switzerland. It processes 10Hz lidar data in real time, can effectively identify and monitor mountain torrents and debris flow events, and provide technical support for disaster warning.

[0111] Figure 2 Schematic diagram of the axial bounding box (AABB) dynamic filtering algorithm of the present invention, the left side is the original point cloud data, and the right side is the effective channel point cloud after filtering by applying the adaptive bounding box parameters; Figure 3 This is a schematic diagram of the improved RANSAC plane change monitoring algorithm of the present invention, showing the plane model fitting process and the calculation method of the plane change rate between consecutive frames; Figure 4 This is the effect diagram of the dynamic and static point cloud segmentation of the present invention, which includes two stages: neighbor distance classification based on KD-Tree and DBSCAN noise filtering. The dark part is the identified dynamic point cloud (mudslide) and the light part is the static terrain; Figure 5This is a schematic diagram of the PCA main axis orientation algorithm of the present invention, showing the method for determining the main flow direction of debris flow and the sectional analysis along the main axis through principal component analysis.

[0112] Embodiment 2:

[0113] This embodiment provides a terminal device, which includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function. The processor described in the embodiments of the present invention can be used for the operation of a static and dynamic point cloud segmentation and orientation monitoring method for mountain flood debris flow based on streaming processing, including the following steps:

[0114] S1: Point cloud data preprocessing and reference point cloud construction

[0115] S1.1: Axial Aligned Bounding Box (AABB) dynamic filtering: Set a spatial parameter vector according to the river channel characteristics, and adjust the boundary parameters through a sliding window trigger mechanism;

[0116] S1.2: Improved RANSAC plane segmentation: Establish a plane model through three random point samplings, and use a dual-factor evaluation of the plane change rate based on the normal vector angle and offset change;

[0117] S1.3: Multi-frame reference point cloud fusion: Merge the initial multi-frame point clouds to construct a reference, apply voxel downsampling and KDTree spatial indexing, and dynamically update the reference based on the cumulative amount of plane changes;

[0118] S2: Static and dynamic point cloud segmentation and feature extraction

[0119] S2.1: KDTree nearest neighbor classification: Use the spatial indexing of the reference point cloud to determine static / dynamic points according to the minimum distance from the point to the reference;

[0120] S2.2: DBSCAN Clustering Optimization: Apply density clustering to filter out isolated noise points in the static point cloud, and set the minimum cluster size threshold to remove small regions;

[0121] S2.3: Voxel Grid Downsampling: Perform dimensionality reduction on the dynamic point cloud to maintain a balanced expression of geometric features;

[0122] S3: Principal Component Orientation and Flow Architecture

[0123] S3.1: PCA Main Axis Orientation: Calculate the covariance matrix of the dynamic point cloud and perform eigen-decomposition to determine the main flow direction of the debris flow and establish a natural coordinate system;

[0124] S3.2: Streamz Flow Processing Topology: Construct a directed acyclic graph data flow structure, use a sliding window mechanism for continuous monitoring, and trigger a dual-path processing strategy based on the plane change rate threshold;

[0125] S3.3: Dask Distributed Computing: Task automatic sharding, work stealing load balancing, and multi-level fault tolerance mechanisms to achieve millisecond-level response.

[0126] Example 3:

[0127] This example provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is the memory device in the terminal device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and, of course, the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. And, in this storage space, there is also stored one or more instructions suitable for being loaded and executed by the processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory.

[0128] One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps in the above example regarding a method for segmenting and directionally monitoring static and dynamic point clouds of mountain torrents and debris flows based on streaming processing; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps:

[0129] S1: Point Cloud Data Preprocessing and Benchmark Point Cloud Construction

[0130] S1.1: Axial Axis-Aligned Bounding Box (AABB) Dynamic Filtering: Set the spatial parameter vector according to the river channel characteristics, and adjust the boundary parameters through the sliding window trigger mechanism;

[0131] S1.2: Improved RANSAC Plane Segmentation: Establish a plane model through three random point samplings, and use the dual factors of the normal vector angle and the offset change to evaluate the plane change rate;

[0132] S1.3: Multi-frame Reference Point Cloud Fusion: Merge the initial multi-frame point clouds to construct a reference, apply voxel downsampling and KDTree spatial indexing, and dynamically update the reference based on the cumulative amount of plane changes;

[0133] S2: Dynamic and Static Point Cloud Segmentation and Feature Extraction

[0134] S2.1: KDTree Nearest Neighbor Classification: Use the spatial indexing of the reference point cloud to determine static / dynamic points according to the minimum distance from the point to the reference;

[0135] S2.2: DBSCAN Clustering Optimization: Apply density clustering to filter out isolated noise points in the static point cloud, and set the minimum cluster size threshold to remove small areas;

[0136] S2.3: Voxel Grid Downsampling: Perform dimensionality reduction processing on the dynamic point cloud to maintain a balanced expression of geometric features;

[0137] S3: Principal Component Orientation and Streaming Architecture

[0138] S3.1: PCA Principal Axis Orientation: Calculate the covariance matrix of the dynamic point cloud and perform eigen decomposition to determine the main flow direction of the debris flow and establish a natural coordinate system;

[0139] S3.2: Streamz Streaming Processing Topology: Construct a directed acyclic graph data flow structure, use the sliding window mechanism for continuous monitoring, and trigger a dual-path processing strategy based on the plane change rate threshold;

[0140] S3.3: Dask Distributed Computing: Task automatic sharding, work stealing load balancing, and multi-level fault tolerance mechanisms to achieve millisecond-level response.

[0141] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code.

[0142] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0143] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0144] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0145] The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the above embodiments. Within the knowledge scope of those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

[0146] Many other changes and modifications can be made without departing from the concept and scope of the present invention. It should be understood that the present invention is not limited to specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A method for static and dynamic segmentation and directional monitoring of mountain torrent debris flows, characterized in that, The method includes: S1: Dynamically filter the original point cloud with an axis-aligned bounding box (AABB), adaptively adjust the boundary range to screen the valid point cloud, and adapt to the changes in the channel morphology; use an improved random sample consensus (RANSAC) algorithm to extract the riverbed plane, and construct a robust reference point cloud through multi-frame point cloud fusion and a reference dynamic update mechanism. S2: Based on the constructed reference point cloud, initially classify the dynamic and static point clouds through a KD-tree (KDTree). Distinguish the dynamic fluid from the static terrain according to the comparison between the minimum distance from the point to the reference point cloud and a preset threshold; use the density-based spatial clustering of applications with noise (DBSCAN) algorithm for clustering to remove measurement errors, raindrops, and boulder sensor noise interference, and at the same time eliminate the misclassified small clusters through a small cluster filtering strategy; for the dynamic fluid point cloud, introduce a curvature-aware non-uniform voxel downsampling method to retain the fluid edge features, compress the data volume, and output the dynamic point cloud and static point cloud after dynamic and static point cloud classification, clustering optimization, and downsampling processing. S3: Based on the continuous time series and the downsampled point cloud output in step S2, extract the main flow direction through principal component analysis (PCA), and construct a natural coordinate system to adapt to the channel terrain; the system uses the Streamz stream processing framework, and monitors the plane changes in real time through a sliding window mechanism. When the change rate exceeds the set threshold, trigger the dynamic point cloud segmentation process for deeper analysis, and finally realize the construction of the main flow direction coordinate system and the generation of an event trigger flag.

2. The static and dynamic segmentation and directional monitoring method for mountain flood and debris flow according to claim 1, characterized in that In step S1, the dynamic axis-aligned bounding box filtering includes: The mountain flood and debris flow monitoring area usually presents a channel morphology, and there are a large number of non-monitoring targets such as vegetation and overhanging rocks around it. Construct an AABB algorithm to adaptively define the effective monitoring area, and the filtering formula is: Among them, P represents the set of original point cloud data, and P valid represents the set of valid point clouds after filtering. p represents any point in the point cloud, and p x and p y respectively represent the x - coordinate and y - coordinate of the point. and represent the minimum and maximum boundary values along the x - axis at time t. and represent the minimum and maximum boundary values along the y - axis at time t. This formula filters valid point clouds through dynamically adjusted spatial range parameters, can effectively cope with the channel changes caused by debris flows, and the boundary parameters are dynamically adjusted according to the proportion r of valid points within the sliding window: Among them, ΔX max represents the adjustment amount of the maximum boundary of the x-axis, λ represents the boundary adjustment coefficient to control the adjustment amplitude, r th represents the effective point ratio threshold, which represents the expected proportion of effective points, r represents the proportion of effective points in the current frame, σ represents the adjustment sensitivity parameter to control the steepness of the response curve, and tanh represents the hyperbolic tangent function; this formula realizes the adaptive adjustment of the boundary range, expanding the boundary when the effective point ratio is low and remaining stable when it is high.

3. A static and dynamic segmentation and directional monitoring method for mountain flood and debris flow according to claim 1, characterized in that In step S1, the improved RANSAC algorithm is used to extract the riverbed plane, including: For the improved RANSAC algorithm, the plane model is defined as π: ax + by + cz + d = 0, and the inlier set optimization formula is: Among them, S inlier represents the optimal inlier set, S represents the candidate inlier set selected from the valid point cloud P valid . p represents any point in the point cloud, π represents the plane model, defined as π: ax + by + cz + d = 0, where (a, b, c) is the plane normal vector and d is the plane offset. D(p, π) represents the Euclidean distance from point p to plane π, ∈ represents the distance tolerance parameter, exp represents the natural exponential function. By introducing Gaussian weights, points closer to the plane contribute more, enhancing the robustness of the algorithm to local outliers, and it is especially suitable for the situation where there are local boulders and floating logs interfering in the flash flood environment; Define a dual-factor evaluation index for the plane change rate: where, Δπ represents the plane change rate, ω1 represents the weight of the normal vector angle change, ω2 represents the weight of the plane displacement change, represents the unit normal vector of the current frame plane, represents the unit normal vector of the previous frame plane, represents the dot product of the two normal vectors, arccos represents the inverse cosine function, d t represents the offset of the current frame plane, d t-1 represents the offset of the previous frame plane. The index comprehensively considers two factors: the change of the normal vector angle and the plane displacement, enabling the system to distinguish between slight water level fluctuations and sudden debris flows. When Δπ > 0.05, the dynamic segmentation process is triggered to avoid the system's over-response to noise changes while ensuring sensitivity to important events.

4. A method for static and dynamic segmentation and directional monitoring of mountain torrents and debris flows according to claim 1, characterized in that, In step S1, the multi-frame point cloud fusion and reference dynamic update mechanism includes: Construct a robust reference through multi-frame fusion: Among them, P base represents the constructed reference point cloud, represents the union operation on the effective point clouds of the previous k frames, represents the effective point cloud of the i-th frame, ↓ voxel represents the voxel downsampling operator, which fuses the effective point clouds of the previous k frames and controls the data volume through voxel downsampling. Voxel downsampling not only reduces the computational burden but also smooths local noise, generates a reference point cloud with uniform density, and improves the subsequent KDTree search efficiency; To adapt to the long-term changes in the environment, design a reference dynamic update mechanism based on time decay: Among them, represents the summation over a time window from time t - m to t, and Δπ i represents the planar change rate of the i-th frame, β represents the time decay factor, t - i represents the time difference between the i-th frame and the current frame t, and e -β(t-i) represents the time decay weight, and Γ represents the reference reconstruction trigger threshold; this formula calculates the time-weighted cumulative planar change, and when the cumulative amount exceeds the threshold, it triggers the reference reconstruction, making the recent changes have a higher weight.

5. A method for static and dynamic segmentation and directional monitoring of mountain flood and debris flow according to claim 1, characterized in that, In step S2, the initial classification of the dynamic and static point clouds through the KDTree includes: Accurately distinguishing the static terrain from the dynamic fluid is the core of mountain flood and debris flow monitoring. Use the KDTree spatial index to achieve efficient dynamic and static point cloud classification: Among them, P dynamic represents the recognized dynamic point cloud set, and P new represents the currently acquired current frame point cloud. p represents any point in the current frame point cloud, q represents any point in the reference point cloud, and ‖p - q‖2 represents the Euclidean distance between point p and point q. represents the nearest point of point p found in the reference point cloud, and τ d represents the dynamic and static classification distance threshold. KDTree reduces the query complexity from O(n) to O(log n) for real-time analysis of large-scale point clouds; the threshold τ d is the threshold. Noise less than the sensor resolution is regarded as static, and values greater than this are determined to be dynamic fluids, balancing sensitivity and stability.

6. A static and dynamic segmentation and directional monitoring method for mountain flood and debris flow according to claim 1, characterized in that, In step S2, the DBSCAN algorithm is used for clustering, including: Apply the DBSCAN algorithm to cluster and optimize the static point set, and remove the misclassified points caused by measurement errors and sensor noise. For any point p in the static point set, its ∈-neighborhood is defined as: N ∈ (p) = {q ∈ P static ||p - q||₂ ≤ ∈} (8) Among them, N ∈ (p) represents the ∈-neighborhood of point p, and P static represents the set of static point clouds, q represents any point in the static point cloud, ‖p - q‖2 represents the Euclidean distance between point p and point q, ∈ represents the neighborhood radius parameter of the DBSCAN algorithm, and the key parameters of DBSCAN are the neighborhood radius ∈ and the minimum number of points MinPts. Points are classified into core points, boundary points, and noise points by density judgment; in the monitoring of mountain floods and debris flows, the random fluctuations of point clouds mainly originate from raindrops, tiny suspended particles in the air, and sensor errors, and usually appear as small clusters of sparsely distributed points. By setting the value of MinPts, it is ensured that only sufficiently dense point clusters are recognized as real static regions; Introduce the minimum effective cluster point number threshold N min Filter small areas: Among them, represents the optimized dynamic point cloud set, P dynamic represents the dynamically point cloud set after preliminary classification, P static represents the static point cloud set after preliminary classification, p represents any point in the point cloud, L(p) represents the clustering label of point p, -1 represents a noise point, C L(p) represents the cluster with the label L(p), |C L(p) | represents the number of points in cluster C L(p) in, N min represents the minimum effective cluster point number threshold, effectively filtering out isolated static points caused by measurement errors, retaining the true static area, and improving the dynamic point cloud segmentation accuracy.

7. A static and dynamic segmentation and directional monitoring method for mountain flood and debris flow according to claim 1, characterized in that In step S2, the introduction of the curvature-aware non-uniform voxel downsampling method includes: Considering that the segmented dynamic point cloud is usually unevenly distributed, and directly participating in geometric calculations causes the high-density area to overly affect the results, construct a non-uniform downsampling method considering geometric features: Among them, represents the representative point of voxel I, i.e., the center point, V I represents the set of points contained in voxel I, |V I | represents the number of points in voxel I, p represents any point in the voxel, η represents the curvature weight coefficient, represents the curvature gradient at point p, which is calculated by estimating the principal curvature of the local neighborhood; this formula introduces a curvature weighting term when calculating the voxel center, enabling the downsampled point cloud to better retain the key features of the water surface edge and mutation area, while significantly reducing the data volume and improving the efficiency of subsequent processing.

8. A static and dynamic segmentation and directional monitoring method for mountain flood and debris flow according to claim 1, characterized in that In the step S3, the main flow direction is extracted by PCA, including: Determining the main flow direction of debris flow through principal component analysis and establishing a natural coordinate system: Among them, C represents the 3×3 covariance matrix of the point cloud, n represents the total number of points in the point cloud, p i represents the three-dimensional coordinate vector of the i-th point in the point cloud, μ represents the center point of the point cloud, i.e., the average coordinate, (p i -μ)(p i -μ) T represents the outer product of vectors, V represents the eigenvector matrix, whose column vectors form the main axis coordinate system, Λ represents the eigenvalue diagonal matrix; the eigenvalue decomposition of the covariance matrix reveals the main change directions of the point cloud data. The eigenvector v1 corresponding to the largest eigenvalue represents the main flow direction, the eigenvector v2 corresponding to the second largest eigenvalue represents the width direction, and the eigenvector v3 corresponding to the smallest eigenvalue represents the depth direction. Through coordinate transformation: p′ i = V T (p i - μ) (12) where p′ i represents the coordinates of point p i in the principal axis coordinate system, V represents the eigenvector matrix, whose column vectors are the principal axis directions, V T represents the transpose of the eigenvector matrix, p i represents the original coordinates of the i-th point in the point cloud, and μ represents the center point of the point cloud; projecting the point cloud onto the principal axis coordinate system makes the subsequent geometric parameter calculation more in line with the actual shape of the river channel. This adaptive coordinate system is suitable for the complex terrain in mountainous areas and can adapt to the bending changes of the river channel without manual intervention.

9. A static and dynamic segmentation and directional monitoring method for mountain torrents and debris flows according to claim 1, characterized in that, In the step S3, the system adopts the Streamz stream processing framework, including: Based on the Streamz-based streaming processing framework, constructing a directed acyclic graph data flow topology: y i = f i (f i-1 (...f1(x)...)) (13) Among them, x represents the input data, i.e., the original point cloud, and y i represents the output of the i-th processing node, and f1, f2,..., f i represent a series of transformation functions in the data stream, and f i (f i-1 (...)) represents the function composition operation; the sliding window mechanism is adopted to analyze the time-series data and continuously monitor the change of the plane parameters: W t = {x t-w+1 , x t-w+2 ,..., x t} (14) Among them, W t represents the sliding window at time t, w represents the window size, and x t represents the data at time t, x t-w+1 , x t-w+2 ,..., x t represent the continuous data contained in the window; based on the event-triggering mechanism of the plane change rate, when the change rate exceeds the preset threshold, the dynamic point cloud segmentation process is activated: Among them, E t represents the event trigger flag at time t, where 1 indicates triggering and 0 indicates non-triggering, and Δπ t represents the plane change rate at time t, and θ represents the event trigger threshold; the system optimizes resource utilization through a dual-path processing strategy: under normal conditions, only plane monitoring is performed, i.e., the simple path, and when a change is detected, the complete analysis process is activated, i.e., the complex path.

10. A static and dynamic segmentation and directional monitoring system for mountain torrents and debris flows, characterized in that, The system is applied to the method described in any one of claims 1-9, and the system includes: Point cloud data preprocessing and reference point cloud construction module: Dynamically filter the original point cloud through an axial bounding box AABB, screen the effective point cloud by adaptively adjusting the boundary range to adapt to the change of the channel morphology; adopt an improved random sample consensus RANSAC algorithm to extract the riverbed plane, and construct a robust reference point cloud through multi-frame point cloud fusion and reference dynamic update mechanism; Dynamic and static point cloud segmentation and feature extraction module: Based on the constructed reference point cloud, initially classify the dynamic and static point clouds through a KDTree, and distinguish the dynamic fluid from the static terrain according to the comparison between the minimum distance from the point to the reference point cloud and a preset threshold; use the density-based spatial clustering of applications with noise DBSCAN algorithm for clustering to remove measurement errors, raindrops, and boulder sensor noise interference, and at the same time eliminate misclassified small clusters through a small cluster filtering strategy; for the dynamic fluid point cloud, introduce a curvature-aware non-uniform voxel downsampling method to retain the fluid edge features, compress the data volume, and output the dynamic point cloud and static point cloud after dynamic and static point cloud classification, clustering optimization, and downsampling processing; Principal component orientation and flow architecture module: Based on the continuous time series and the downsampled point cloud output in step S2, extract the main flow direction through principal component analysis PCA, and construct a natural coordinate system to adapt to the channel terrain; the system adopts the Streamz stream processing framework, and monitors the plane change in real time through a sliding window mechanism. When the change rate exceeds the set threshold, trigger the dynamic point cloud segmentation process for deeper analysis, and finally realize the construction of the main flow direction coordinate system and the generation of an event trigger flag.

Images