Mountain torrent disaster video image recognition and risk assessment method and system

Abstract

The present application relates to the technical field of image recognition, and discloses a mountain torrent disaster video image recognition and risk assessment method and system, which comprises the following steps: acquiring a video frame sequence containing a static reference target, calculating pixel displacement vectors between adjacent frames to construct a motion feature field, converting pixel texture sharpness and displacement vectors to a joint feature space to form a spatiotemporal coupling feature distribution, dividing distribution points into fluid semantic clusters, accompanying rigid body semantic clusters and static background semantic clusters, identifying a solid-liquid two-phase mixed state in a sudden flow medium, calculating a vertical scale representation and a fluid evolution speed according to real-time trajectories of the fluid semantic clusters and a static background space reference, and determining a risk level of a region to be monitored, wherein the present application utilizes the distribution difference between rheological characteristics and background texture to eliminate turbulence interference, ensures the robustness of water level recognition under visual feature chaos, realizes semantic decoupling of solid-liquid two phases in a flood and provides an objective data basis for assessing disaster impact.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for video image recognition and risk assessment of flash flood disasters in mountainous areas. Background Technology

[0002] Current technologies typically utilize digital image processing to extract water body edges and combine them with preset channel cross-sectional geometric parameters to calculate real-time water level height. Mountain torrents are accompanied by intense fluid turbulence, and the torrents contain high concentrations of silt, gravel, and other solid materials. Under these extreme conditions, splashing water, mud adhering to walls, and irregular reflections on the fluid surface cause visual distortion of the liquid-solid physical boundary in the image. Conventional segmentation algorithms based on single-frame grayscale gradients or texture features struggle to distinguish between real fluid edges and instantaneous dynamic noise when faced with monitoring images that are visually chaotic and have low contrast, resulting in significant fluctuations or distortions in the output water level recognition results.

[0003] To suppress image noise, spatial domain filtering or temporal smoothing is commonly used. However, these methods often erase key abrupt changes in water body evolution, creating an irreconcilable technical conflict between recognition accuracy and response sensitivity in monitoring systems. Current solutions lack in-depth analysis of the spatiotemporal features in video sequences, failing to separate the fluid matrix with specific physical laws from the chaotic pixel flow. This results in insufficient understanding of water level evolution trends in feature-deficient environments and a lack of data support for assessing the impact risk caused by solid-liquid two-phase mixtures. For example, Chinese invention patent CN116778395B discloses... A deep learning-based video recognition and monitoring method for flash floods relies on a 3D convolutional neural network to extract global features from video sequences and uses a black-box classification model to determine flash floods. The underlying logic is anchored to the probabilistic statistics of the visual distribution patterns of water flow and ripples. However, in actual mountainous environments with high sediment content and gravel inclusions, the fluid surface exhibits homogeneous characteristics, resulting in the loss of microscopic visual features such as ripples. Due to the lack of an explicit decoupling mechanism for physical semantics, it is difficult to separate the fluid body from the solid accompanying material in complex rheological media with solid-liquid two-phase mixtures. This leads to the fundamental instability of the judgment logic in extreme environments and makes it impossible to achieve quantitative characterization of key physical scales such as water level fluctuations.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a water level identification mechanism that conforms to the laws of fluid motion and realize risk assessment of solid-liquid two-phase mixed state in video streams with missing visual features and drastic fluctuations in dynamic noise. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for video image recognition and risk assessment of flash flood disasters in mountainous areas, comprising the following steps: Step S101: Obtain a video frame sequence of the monitored area scene containing the static reference target in a continuous time series. Step S102: Calculate the pixel-level displacement vector between adjacent frames in the video frame sequence to construct a motion feature field characterizing the dynamic rheological properties of pixels; Step S103: Extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupled feature distribution. Step S104: Perform feature space clustering on the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, phase separation identification is achieved for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. Step S105: Extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. Step S106: Based on the target's vertical scale representation, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster, output the target evolution risk level of the scene to be monitored.

[0006] Preferably, step S102 includes the following sub-steps: step S1021, performing grayscale conversion and Gaussian filtering noise reduction on the video frame sequence; step S1022, using optical flow calculation criteria to extract the horizontal displacement component and vertical displacement component of each pixel between adjacent frames; step S1023, calculating a composite vector based on the horizontal displacement component and the vertical displacement component, and using the composite vector as a pixel-level displacement vector to characterize the dynamic offset intensity and offset direction of the pixel region in the temporal domain.

[0007] Preferably, step S103 includes the following sub-steps: step S1031, calculating the spatial second-order gradient magnitude of each pixel in the video frame sequence as an initial feature quantity characterizing edge sharpness; step S1032, performing normalization processing on the initial feature quantity to obtain the texture sharpness features of each pixel region, wherein the texture sharpness features of the fluid semantic cluster are lower than the texture sharpness features of the accompanying rigid body semantic cluster.

[0008] Preferably, in step S104, clustering stability is improved by constructing a spatiotemporal feedback semantic correction closed loop, including the following sub-steps: Step S1041, for the clustering result of the current frame, the semantic clustering centers of a preset number of previous frames are introduced as prior constraints; Step S1042, the spatiotemporal distance between the current pixel to be classified and each semantic clustering center is calculated. If the spatiotemporal distance exceeds a preset deviation threshold, the current pixel to be classified is determined to be transient motion noise and is removed; Step S1043, the cumulative consistency of pixel-level displacement vectors between adjacent frames is used to correct the categories of pixels with semantic conflicts, so as to ensure the logical consistency between the fluid semantic cluster and the accompanying rigid body semantic cluster in the temporal evolution.

[0009] Preferably, the feature point generation logic in the joint feature space is as follows: an overlay operation is performed based on the product of a preset displacement weighting coefficient and a pixel-level displacement vector, and the product of a preset texture weighting coefficient and a texture sharpness feature.

[0010] Preferably, step S104 for phase identification of a solid-liquid two-phase mixture in a sudden rheological medium includes: marking pixel clusters with displacement vector coherence and texture sharpness features higher than a preset sharpness threshold as displacement-associated rigid body semantic clusters, which correspond to discrete solid particle targets entrained in the fluid; and marking pixel clusters with displacement vector randomness and texture sharpness features lower than a preset sharpness threshold as fluid semantic clusters, which correspond to the fluid body.

[0011] Preferably, the calculation of the target's vertical scale representation in step S105 includes: locking the edge baseline representing the physical boundary of the monitoring scene in the static background semantic cluster; calculating the pixel distance in the vertical direction between the upper boundary line of the fluid semantic cluster and the edge baseline; and converting the pixel distance into physical units based on the preset camera intrinsic parameter matrix and scene depth information. The target vertical scale representation quantity.

[0012] Preferably, step S106 includes the following: extracting the area ratio of the accompanying rigid body semantic cluster in the video frame image as a feature component characterizing the physical destructive force of the sudden rheological medium; constructing a multidimensional risk judgment matrix by combining the transient change rate of the target vertical scale characterization quantity and the fluid evolution velocity; querying the corresponding warning level in the multidimensional risk judgment matrix and outputting a warning instruction characterizing the current monitoring scene status.

[0013] Preferably, after step S101, the method further includes: establishing a scene static background feature library, which includes scene geometric structure features and static texture features in the initial background state; in step S104, by performing differential mapping between the real-time captured video frame sequence and the scene static background feature library, dynamic semantic increments located outside the static background semantic cluster are extracted as the initial processing area for subsequent clustering determination; in step S105, the target vertical scale representation quantity is calculated, which also includes: detecting the occlusion ratio of the static reference in the static background semantic cluster; and performing spatial coordinate compensation on the upper boundary line of the fluid semantic cluster according to the dynamic offset of the occlusion ratio to reduce the target vertical scale representation quantity determination error caused by transient splashing of the medium.

[0014] A video image recognition and risk assessment system for flash flood disasters in mountainous areas, comprising: The data acquisition module is used to acquire video frame sequences of the monitored area scene containing static reference targets in a continuous time series. The feature field construction module is used to calculate the pixel-level displacement vector between adjacent frames in a video frame sequence in order to construct a motion feature field that characterizes the dynamic rheological properties of pixels; The spatiotemporal coupling module is used to extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupling feature distribution; The semantic decoupling module is used to perform feature space clustering and division of the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, it can realize phase separation identification for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. The scale calculation module is used to extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. The risk assessment module is used to output the target evolution risk level of the scene to be monitored based on the target's vertical scale characterization, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster.

[0015] The beneficial effects of this invention are: 1. In the video image recognition of flash flood disasters in mountainous areas, a spatiotemporal feature analysis mechanism resistant to noise interference is constructed by fusing pixel-level temporal displacement vectors with neighborhood gray-level gradient distribution. This mechanism utilizes the distribution difference between rheological features and static background texture in the mapping space to eliminate false edge interference caused by water splashing or mud adhering to the walls in turbulent backgrounds. This makes the system's recognition of the semantic boundary of the water body no longer dependent on the physical boundary of gray-level abrupt changes, ensuring that the water level recognition path still has high robustness and determinism under disaster conditions with extremely chaotic visual features.

[0016] 2. This invention achieves semantic decoupling of the solid-liquid two-phase mixture in floods by establishing a collaborative judgment logic of displacement vector consistency and texture sharpness. This cross-dimensional feature coupling method enables the system to automatically separate solid accompanying objects with constant rigid texture and fluid main body with continuous rheological texture from the chaotic flow pixel clusters. This evolves the traditional single water level monitoring into a physical understanding of the composition properties of fluids, providing a data foundation with objective causal laws for assessing the impact and destructiveness of channel disasters.

[0017] 3. This invention utilizes the inter-frame temporal continuity in video sequences to implement semantic clustering judgment within the feature mapping space, constructing a spatiotemporal feedback semantic correction closed loop. This mechanism filters out incoherent instantaneous motion noise by accumulating and comparing the pixel displacement vectors between adjacent frames, making the judgment results for obstacles or water level evolution in the channel exhibit high stability. It avoids logical breaks caused by missing features in a single frame image or sudden changes in ambient light and shadow, and improves the logical consistency and engineering credibility of risk assessment results in the time dimension. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating the entire process of video image recognition and risk assessment for flash flood disasters in mountainous areas, as described in this invention. Figure 2 This is a logic diagram for the semantic decoupling of solid and liquid phases and the extraction of multidimensional risk features in this invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, an embodiment or embodiment referred to herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. An embodiment appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] This invention is described in detail with reference to the schematic diagrams. When describing the embodiments of this invention, for ease of explanation, the cross-sectional views of the device structure will be partially enlarged without adhering to the general scale. Moreover, the schematic diagrams are only examples and should not limit the scope of protection of this invention. In addition, in actual manufacturing, the three-dimensional spatial dimensions of length, width and depth should be included.

[0023] Furthermore, in the description of this invention, it should be noted that the terms such as "upper," "lower," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or component referred to has a specific orientation, or is constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0024] Unless otherwise explicitly specified and limited, the terms installation, connection, and linking in this invention should be interpreted broadly. For example, they can refer to fixed connection, detachable connection, or integrated connection; similarly, they can refer to mechanical connection, electrical connection, or direct connection, or indirect connection through an intermediate medium, or internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] A method for video image recognition and risk assessment of flash flood disasters in mountainous areas includes the following steps: Step S101: Obtain a video frame sequence of the monitored area scene containing the static reference target in a continuous time series. Step S102: Calculate the pixel-level displacement vector between adjacent frames in the video frame sequence to construct a motion feature field characterizing the dynamic rheological properties of pixels; Step S103: Extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupled feature distribution. Step S104: Perform feature space clustering on the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, phase separation identification is achieved for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. Step S105: Extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. Step S106: Based on the target's vertical scale representation, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster, output the target evolution risk level of the scene to be monitored.

[0026] Preferably, step S102 includes the following sub-steps: step S1021, performing grayscale conversion and Gaussian filtering noise reduction on the video frame sequence; step S1022, using optical flow calculation criteria to extract the horizontal displacement component and vertical displacement component of each pixel between adjacent frames; step S1023, calculating a composite vector based on the horizontal displacement component and the vertical displacement component, and using the composite vector as a pixel-level displacement vector to characterize the dynamic offset intensity and offset direction of the pixel region in the temporal domain.

[0027] Preferably, step S103 includes the following sub-steps: step S1031, calculating the spatial second-order gradient magnitude of each pixel in the video frame sequence as an initial feature quantity characterizing edge sharpness; step S1032, performing normalization processing on the initial feature quantity to obtain the texture sharpness features of each pixel region, wherein the texture sharpness features of the fluid semantic cluster are lower than the texture sharpness features of the accompanying rigid body semantic cluster.

[0028] Preferably, in step S104, clustering stability is improved by constructing a spatiotemporal feedback semantic correction closed loop, including the following sub-steps: Step S1041, for the clustering result of the current frame, the semantic clustering centers of a preset number of previous frames are introduced as prior constraints; Step S1042, the spatiotemporal distance between the current pixel to be classified and each semantic clustering center is calculated. If the spatiotemporal distance exceeds a preset deviation threshold, the current pixel to be classified is determined to be transient motion noise and is removed; Step S1043, the cumulative consistency of pixel-level displacement vectors between adjacent frames is used to correct the categories of pixels with semantic conflicts, so as to ensure the logical consistency between the fluid semantic cluster and the accompanying rigid body semantic cluster in the temporal evolution.

[0029] Preferably, the feature point generation logic in the joint feature space is as follows: an overlay operation is performed based on the product of a preset displacement weighting coefficient and a pixel-level displacement vector, and the product of a preset texture weighting coefficient and a texture sharpness feature.

[0030] Preferably, step S104 for phase identification of a solid-liquid two-phase mixture in a sudden rheological medium includes: marking pixel clusters with displacement vector coherence and texture sharpness features higher than a preset sharpness threshold as displacement-associated rigid body semantic clusters, which correspond to discrete solid particle targets entrained in the fluid; and marking pixel clusters with displacement vector randomness and texture sharpness features lower than a preset sharpness threshold as fluid semantic clusters, which correspond to the fluid body.

[0031] Preferably, the calculation of the target's vertical scale representation in step S105 includes: locking the edge baseline representing the physical boundary of the monitoring scene in the static background semantic cluster; calculating the pixel distance in the vertical direction between the upper boundary line of the fluid semantic cluster and the edge baseline; and converting the pixel distance into physical units based on the preset camera intrinsic parameter matrix and scene depth information. The target vertical scale representation quantity.

[0032] Preferably, step S106 includes the following: extracting the area ratio of the accompanying rigid body semantic cluster in the video frame image as a feature component characterizing the physical destructive force of the sudden rheological medium; constructing a multidimensional risk judgment matrix by combining the transient change rate of the target vertical scale characterization quantity and the fluid evolution velocity; querying the corresponding warning level in the multidimensional risk judgment matrix and outputting a warning instruction characterizing the current monitoring scene status.

[0033] Preferably, after step S101, the method further includes: establishing a scene static background feature library, which includes scene geometric structure features and static texture features in the initial background state; in step S104, by performing differential mapping between the real-time captured video frame sequence and the scene static background feature library, dynamic semantic increments located outside the static background semantic cluster are extracted as the initial processing area for subsequent clustering determination; in step S105, the target vertical scale representation quantity is calculated, which also includes: detecting the occlusion ratio of the static reference in the static background semantic cluster; and performing spatial coordinate compensation on the upper boundary line of the fluid semantic cluster according to the dynamic offset of the occlusion ratio to reduce the target vertical scale representation quantity determination error caused by transient splashing of the medium.

[0034] A video image recognition and risk assessment system for flash flood disasters in mountainous areas includes: The data acquisition module is used to acquire video frame sequences of the monitored area scene containing static reference targets in a continuous time series. The feature field construction module is used to calculate the pixel-level displacement vector between adjacent frames in a video frame sequence in order to construct a motion feature field that characterizes the dynamic rheological properties of pixels; The spatiotemporal coupling module is used to extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupling feature distribution; The semantic decoupling module is used to perform feature space clustering and division of the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, it can realize phase separation identification for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. The scale calculation module is used to extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. The risk assessment module is used to output the target evolution risk level of the scene to be monitored based on the target's vertical scale characterization, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster.

[0035] Example 1: In a field monitoring scenario deployed in a mountainous environment with undulating elevations and narrow gullies, when a sudden torrential downpour triggers a flash flood carrying high concentrations of suspended sediment, as well as irregularly shaped boulders and broken logs, through the video monitoring area, the fluid surface generates unsteady rolling and breaking waves. The resulting random scattered light causes the grayscale boundary between the water body and the riverbank background in the image to physically disappear, leading to insufficient visual feature contrast in the water level identification process. The system acquires a video frame sequence containing static reference targets, calculates the pixel-level displacement vectors between adjacent frames in the video frame sequence to construct a motion feature field characterizing the dynamic rheological properties of pixels, extracts the texture sharpness features of each pixel region in the video frame sequence, and maps them to a joint feature space along with the pixel-level displacement vectors to form a spatiotemporally coupled feature distribution. Within the joint feature space, feature space clustering is performed on the distribution points, identifying those with displacement coherence and texture sharpness features higher than a preset sharpness threshold. The set of pixels is used to lock the semantic cluster of the rigid body with displacement, and features with random displacement vectors and texture sharpness below a preset sharpness threshold are selected. The set of pixels is labeled as a fluid semantic cluster, and the pixel cluster with zero displacement vector is identified as a static background semantic cluster, where a preset sharpness threshold is used. It is a constant determined based on the pixel gradient statistics of the initial background image.

[0036] During this clustering process, the system simultaneously extracts the two-dimensional pixel coordinates of the current pixel to be classified in the original image plane and retrieves the coordinates of the semantic clustering center points preserved in a predetermined number of known valid frames. The system obtains the spatial scale deviation by calculating the geometric distance between the pixel coordinates of the current pixel to be classified and the coordinates of the previous clustering center points. At the same time, it calculates the frame index interval difference between the arrival time of the current video frame and the extraction time of the previous historical frame as the temporal scale deviation. The system introduces a preset constant reference flow velocity mapping constant and multiplies the frame index interval difference by this constant to convert it into a unified equivalent pixel displacement dimension. Finally, the system performs a square root operation on the sum of the squares of the extracted geometric spatial distance and the converted equivalent pixel displacement to obtain a result that covers spatial deformation. Furthermore, a single scalarized spatiotemporal distance value reflecting temporal delay is used. If this spatiotemporal distance is determined to exceed the maximum preset deviation threshold of normal hydrodynamic evolution, the pixel is regarded as transient splash noise deviating from the rheological trajectory and is directly removed. The system locks the edge baseline representing the physical boundary of the monitoring scene in the static background semantic cluster, calculates the pixel distance between the upper boundary line of the fluid semantic cluster and the edge baseline in the vertical direction, and converts the pixel distance into a target vertical scale representation quantity according to the preset camera intrinsic parameter matrix and scene depth information. Combining the area ratio of the accompanying rigid body semantic cluster in the video frame image and the fluid evolution speed, a multi-dimensional risk judgment matrix is ​​constructed, and the risk warning level corresponding to the current hydrodynamic impact intensity and flood peak scale is output.

[0037] Example 2: In a physical simulation ditch test platform with a length of 20m and a width of 1.5m, a variable frequency pump set generates controlled hydrodynamic impact to simulate flash flood conditions in mountainous areas. The image acquisition unit acquires images at a resolution of 1920. A 1080-pixel video frame sequence with a frequency of 30Hz, for the sampling period The settings and factors affecting the values ​​include the characteristic frequencies of the waves on the surface of the monitored fluid. In addition to the computational load of the system processor, to satisfy the sampling theorem and prevent signal aliasing in the motion feature field, the sampling period... With characteristic frequency The reciprocal relationship constraint must be satisfied when the flow rate is between 0.8 and 1.2. Under the condition of / s, the sampling period The timeframe was set to 33.3ms. During the experiment, gravel with a particle size distribution between 50 and 150 mm was added to the fluid, and Gaussian white noise with a signal-to-noise ratio of 20 dB was superimposed onto the video signal. The system determined the preset sharpness threshold by performing pixel gradient statistics on the initial background image. The value was 12.5. The control group was treated with an edge detection method based on the Canny operator for a sediment concentration of 450 kg / m³. In the video frame sequence, due to the reduced grayscale contrast between the water body and the embankment background to 1.12, the output water level pixel position exhibited a random jump of 56 pixels, resulting in a water level calculation deviation of 0.38m. Under the same conditions, the experimental group calculated the average displacement vector magnitude of the fluid pixel region to be 8.4 pixels / frame, with a corresponding texture sharpness feature measurement value of 4.2. Meanwhile, the texture sharpness feature measurement value for the gravel target was 26.8, and its displacement vector direction remained coherent with the fluid evolution direction. Based on the texture sharpness feature being below the preset sharpness threshold, the system... The determination rule locks in the fluid semantic cluster, and the texture sharpness feature is higher than the preset sharpness threshold. Furthermore, the system locks the rigid body semantic cluster with displacement coherence. Based on the vertical pixel distance between the upper boundary line and the edge baseline of the fluid semantic cluster, the system calculates the water level height to be 1.25m. Compared with the manually calibrated baseline value of 1.23m, the absolute deviation is 0.02m.

[0038] For sediment concentrations of 100 kg / 300kg / 550kg / Establish a problem intensity gradient verification system under the working conditions, when the concentration is 100 kg / Increase to 500kg / At that time, the system's accuracy in recognizing fluid semantic clusters remained stable at over 96.8%, even when the sediment concentration exceeded 550 kg / m³. After this performance inflection point, the micro-texture of the fluid surface in the image becomes homogenized, the distinguishability of the texture sharpness feature decreases, and the recognition accuracy drops to 82.5%. Experimental results show that by clustering pixel-level displacement vectors and texture sharpness features in the joint feature space, the system achieves semantic decoupling between the fluid and the accompanying rigid body under the condition of lack of visual contrast. The water level recognition accuracy is no longer limited by the clarity of the static geometric edge, but depends on the statistical significance of the spatiotemporal coupling feature distribution. The system outputs multi-dimensional risk judgment results corresponding to the physical characteristics of the flood.

[0039] Example 3: This example combines Figures 1 to 2 This document describes methods and systems for video image recognition and risk assessment of flash floods in mountainous areas. Figure 1 As shown, step S101 involves acquiring a video frame sequence of the monitored area scene containing a static reference target in a continuous time series. Step S102 involves calculating the pixel-level displacement vector between adjacent frames in the video frame sequence to construct a motion feature field characterizing the dynamic rheological properties of pixels. Step S103 involves extracting the texture sharpness features of each pixel region in the video frame sequence and mapping them to the joint feature space along with the pixel-level displacement vector to form a spatiotemporally coupled feature distribution. Based on this, step S104 involves performing feature space clustering on the distribution points in the joint feature space to achieve phase identification of the solid-liquid two-phase mixture state in the sudden rheological medium. Step S105 involves extracting the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system and calculating the target vertical scale representation quantity and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. Finally, step S106 involves outputting the target evolution risk level of the monitored area scene based on the target vertical scale representation quantity, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster.

[0040] like Figure 2As shown, for the distribution points in the joint feature space, the system performs logical judgment on the pixel-level displacement vectors. If the displacement vector is lower than the preset static threshold, it is identified as a static background semantic cluster. If the displacement vector is random, its texture sharpness feature is determined. When the texture sharpness feature is lower than the preset sharpness threshold, it is identified as a fluid semantic cluster, and the target vertical scale representation and fluid evolution velocity are extracted accordingly. If the displacement vector is coherent and its texture sharpness feature is higher than the preset sharpness threshold, it is identified as a rigid body semantic cluster with accompanying movement, and its area ratio is extracted. The system summarizes the extracted target vertical scale representation, fluid evolution velocity, and area ratio to construct a multi-dimensional risk judgment matrix and calculates and outputs the comprehensive risk feature quantity R. Finally, the system outputs a judgment instruction according to the quantization interval of the comprehensive risk feature quantity R. When it is in the first interval, a risk inspection instruction is issued, and when it is in the second interval, an early warning work order is issued.

[0041] Example 4: In a mountainous runoff area monitoring scenario with fluctuating lighting and churning sediment, random reflections from the fluid surface and local contrast changes caused by sediment movement result in overlapping distributions of displacement and texture features in the joint feature space. The system calculates the dynamic range of the feature sequence within a preset sampling window to calibrate the displacement weighting coefficients. With texture weighting coefficient The system extracts the mean of the displacement components of the motion feature field in the current video frame sequence. and the average contrast of texture sharpness features Displacement weighting coefficient The calculation rule is the average value of displacement components. The ratio of the total feature energy to the texture weighting coefficient. The calculation rule is the contrast mean. The ratio of the total feature energy to the image local feature variance is based on the statistical principle of equivalent physical energy signals. The system sets a preset sampling window, calculates the variance of the displacement component and the variance of the contrast component of all pixels within the window, and performs a summation operation on the two to output the total feature energy. The system then assigns the displacement vector of each pixel to the texture sharpness feature according to the corresponding displacement weighting coefficient. With texture weighting coefficient The proportional superposition operation maps pixel features to a joint feature space, which is based on the local density of points distributed within the space. As a clustering criterion, the neighborhood difference approximation method is used to calculate the local density. The system simultaneously acquires the first point distribution density within the current similarity neighborhood radius and the second point distribution density within the radius of the previous single iteration step size. It calculates the difference between the first and second point distribution densities. When the difference is lower than a predetermined tolerance limit for three consecutive iterations, a local density is determined. The derivative approaches zero when the distance between distribution points is measured. Smaller than the preset similarity neighborhood radius and local density When the derivative approaches zero, the system extracts the boundary centroid coordinates of the fluid semantic cluster. After confirming that the feature distribution density approaches an extreme value and locking the core mathematical pixel set constituting the fluid semantic cluster, the processor needs to explicitly transform this abstract point set into a geometric scale index. The system then retrieves all target pixels enclosed within the fluid semantic cluster, traces their basic row and column coordinates in the original two-dimensional plane of the video, and combines them to generate a binary continuous mask layer representing the current water body. The system performs classical contour detection to scan the closed outer boundary of the mask layer, extracts the image horizontal and vertical axis values ​​of all independent edge pixels that make up this closed boundary line, sums all horizontal axis values ​​and divides them by the total number of edge pixels to obtain the center horizontal coordinate, and sums all vertical axis values ​​and divides them by the total number to obtain the center vertical coordinate. This virtual centroid point, which is finally output by the arithmetic mean of the contour coordinate system, constitutes the aforementioned fluid boundary centroid coordinates, thereby filling the data link gap from feature density stationary points to geometric boundaries.

[0042] After the system identifies the fluid semantic cluster and the accompanying rigid body semantic cluster, it obtains the area ratio of the accompanying rigid body semantic cluster in the video frame image. Based on a fluid dynamics model that directly proportionally measures the impulse of the moving medium to the projected area of ​​the water-facing surface, the system assumes that the rolling stones and broken logs in the channel have a uniform density distribution. It then extracts a preset volume conversion factor and compares it with the area ratio. Perform a product operation to convert the two-dimensional pixel mapping into physical constraints representing the three-dimensional equivalent mass, and combine this with the fluid evolution velocity. Vertical scale representation of the target A multi-dimensional risk assessment matrix is ​​constructed. The system establishes a normalized index table based on historical hydrological data of the monitoring area, including area proportions. Fluid evolution velocity and target vertical scale characterization Convert to by arrive The quantification range is determined, and the comprehensive risk characteristic is calculated based on the product operator. Comprehensive risk characteristics When the system is in the first interval, it sends a risk inspection command to the management terminal, taking into account the comprehensive risk characteristics. When in the second interval, the system will change the monitoring frequency from Hz adjusted to Hz, and generate early warning work orders. This key pre-set volume conversion factor is based on the regional geomorphological offline modeling experiment performed before the equipment was put into operation. Historical sedimentary debris was sampled at the interception section around the monitoring point. Dozens of samples of scour rock and covered vegetation fault zones covering typical grain sizes were collected, and their actual geometric volume was determined using three-dimensional lidar. The mean pixel occlusion area of ​​multi-angle camera images was compared, and linear regression calculation based on large sample data was used to derive the characteristic transformation constant applicable to the soil and rock composition and forest density attributes of this specific area. This value was pre-burned into the system's static storage configuration parameters. Based on this empirical database mapping mechanism, the system can infer compliant parameters for monocular projection to three-dimensional equivalent intensity without initiating real-time 3D reconstruction in the turbid and turbulent flow of sudden flash floods, relying on pre-calibrated coefficients. By adopting a weight dynamic adjustment method based on feature energy distribution and a retrieval method of multi-dimensional risk judgment matrix, the system identifies flash flood risks in mountainous environments with complex solid-liquid mixed states. The extraction of water level data is based on the analysis of joint feature space topological distribution. The early warning output results correspond to the actual kinetic energy impact intensity and material scale in the monitoring scenario, providing data support for flood control decision-making.

[0043] Example 5: In newly deployed image acquisition nodes in mountainous river terrain, the system calibrates the depth mapping relationship by performing multi-point sampling on preset static features with known physical dimensions. The processor acquires the pixel coordinate sequence of the target in the image coordinate system and the measured elevation data of the corresponding position, and calculates the conversion ratio coefficient between pixel displacement and physical length using the least squares algorithm. Conversion ratio coefficient The transformation relationship from the pixel coordinate system to the physical space quantization benchmark was established, and the target vertical scale representation quantity calculated based on the upper boundary line and edge benchmark of the fluid semantic cluster has the comparability of physical quantity dimensions.

[0044] When the system encounters ambient light intensity fluctuations caused by day-night cycles or cloud cover, the processor establishes a dynamic background texture baseline to eliminate interference through feature statistical update logic based on a sliding time window. The system continuously collects data during initial operation. Frame the background image and calculate the gradient distribution magnitude of each pixel region. The processor generates the gradient distribution magnitude through mean calculation. The corresponding static texture fingerprint template has a preset sharpness threshold. Set to 1.5 times the mean global gradient in the static texture fingerprint template, and based on the fundamental statistical law that the environmental background noise follows a Gaussian distribution, the processor synchronously reads... The system calculates the displacement magnitude of the time-series optical flow of the pixel set covered by the static reference in the background image, calculates the standard deviation of the displacement feature sequence, extracts the standard deviation constant and multiplies it by a preset bias parameter, and outputs a preset static threshold to filter out high-frequency, low-amplitude fluctuations caused by natural wind load micro-vibrations. The system executes the time-series benchmark takeover procedure to deal with the full occlusion of the reference object. The system continuously monitors the occlusion status of the two-dimensional pixels of the static reference. When the occlusion ratio reaches the full-scale limit and is maintained for a specified clock cycle, the system freezes the current spatial offset compensation calculation branch and calls the edge baseline pixel coordinates in the most recent unoccluded state within the historical sliding time window as the vertical scale calculation takeover benchmark. The system removes the background signal through differential processing of real-time video frame texture and static texture fingerprint template. The judgment criterion for fluid feature locking is within the statistical envelope under the current illumination conditions.

[0045] Example 6: In the initialization scenario of a newly deployed mountain gully cross-section, in order to enable the system to adapt to the river slope in different geographical locations, the processor records the upper limit value of the water level of the cross-section under historical extreme conditions. Upper limit of flow rate and the upper limit of the rigid body proportion corresponding to the river flow rate. To determine the parameter space normalization benchmark; the processor uses a linear mapping operator to convert the measured physical quantities Convert to quantized components The calculation formula is: This process transforms the target's vertical scale representation, fluid evolution velocity, and accompanying rigid body area ratio into feature vectors with a unified numerical range. The processor queries a multidimensional risk assessment matrix and outputs a risk warning level that matches the current hydrodynamic impact intensity. To measure physical parameters, This is the preset upper limit value of the parameter. The preset lower limit value of the parameter. These are the normalized quantized components; during the clustering process in the joint feature space, the processor calculates the feature points of the pixels to be classified. and adjacent feature points Euclidean distance between To determine semantic attribution, the calculation formula is as follows: The displacement weighting coefficient With texture weighting coefficient Derived from the dynamic range calibration procedure in Example 4; when the Euclidean distance Smaller than the preset similarity neighborhood radius and local density When the derivative approaches zero, the system terminates the clustering iteration and outputs the solid-liquid phase separation identification results. The system outputs a water level identification curve and a risk level warning command that show continuous values ​​under changes in illumination and sediment disturbance. For the Euclidean distance in the joint feature space, The difference between displacement components. The difference between the contrast components. These are displacement weighting coefficients. These are texture weighting coefficients. For local density, in the aforementioned Euclidean distance joint metric model, in the underlying data processing queue, to eliminate... The product term contains The inherent pixel-level physical units cause logical gaps in accumulation. To address this, the system incorporates a static dedimensionalization module before executing the core distance calculation. This module calls a constant virtual transformation matrix in memory with a preset value of 1 and a unit set to frames per pixel. and Immediately after multiplication, the transformation matrix is ​​forcibly multiplied into the expression. Through this step, the displacement deviation feature, which originally carried a specific physical motion dimension, is successfully extracted into a pure variation ratio parameter relative to the reference dynamic range, ensuring its consistency with the texture contrast dimension feature. It exhibits rigorous dimensionless consistency during addition. This built-in calibration design eliminates the scale problem in multidimensional feature space mapping, making the clustering criterion self-consistent in terms of physical laws and mathematical statistics.

[0046] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for video image recognition and risk assessment of flash flood disasters in mountainous areas, characterized in that, Includes the following steps: Step S101: Obtain a video frame sequence of the monitored area scene containing the static reference target in a continuous time series. Step S102: Calculate the pixel-level displacement vector between adjacent frames in the video frame sequence to construct a motion feature field characterizing the dynamic rheological properties of pixels; Step S103: Extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupled feature distribution. Step S104: Perform feature space clustering on the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, phase separation identification is achieved for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. Step S105: Extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. Step S106: Based on the target's vertical scale representation, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster, output the target evolution risk level of the scene to be monitored.

2. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1021, performing grayscale conversion and Gaussian filtering noise reduction on the video frame sequence; Step S1022, using optical flow calculation criteria, extracting the horizontal displacement component and vertical displacement component of each pixel between adjacent frames; Step S1023, calculating a composite vector based on the horizontal displacement component and the vertical displacement component, and using the composite vector as a pixel-level displacement vector to characterize the dynamic offset intensity and offset direction of the pixel region in the temporal domain.

3. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S103 includes the following sub-steps: Step S1031, calculate the spatial second-order gradient magnitude of each pixel in the video frame sequence as the initial feature quantity representing edge sharpness; Step S1032, perform normalization processing on the initial feature quantity to obtain the texture sharpness features of each pixel region, wherein the texture sharpness features of the fluid semantic cluster are lower than the texture sharpness features of the accompanying rigid body semantic cluster.

4. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, In step S104, clustering stability is improved by constructing a spatiotemporal feedback semantic correction closed loop, including the following sub-steps: Step S1041, for the clustering result of the current frame, the semantic clustering centers of the previous preset number of frames are introduced as prior constraints; Step S1042, the spatiotemporal distance between the current pixel to be classified and each semantic clustering center is calculated. If the spatiotemporal distance exceeds the preset deviation threshold, the current pixel to be classified is determined to be transient motion noise and is removed. Step S1043: Using the cumulative consistency of pixel-level displacement vectors between adjacent frames, the pixel categories that have semantic conflicts are corrected to ensure the logical consistency between the fluid semantic cluster and the accompanying rigid body semantic cluster in the temporal evolution.

5. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, The feature point generation logic in the joint feature space is as follows: based on the product of the preset displacement weighting coefficient and the pixel-level displacement vector, and the product of the preset texture weighting coefficient and the texture sharpness feature, a superposition operation is performed.

6. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S104 implements phase separation identification for the solid-liquid two-phase mixture state in the sudden rheological medium, including: marking pixel clusters with displacement vector coherence and texture sharpness features higher than a preset sharpness threshold as displacement-associated rigid body semantic clusters, which correspond to discrete solid particle targets entrained in the fluid; and marking pixel clusters with displacement vector randomness and texture sharpness features lower than a preset sharpness threshold as fluid semantic clusters, which correspond to the fluid body.

7. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S105, calculating the target's vertical scale representation, includes: locking the edge baseline representing the physical boundary of the monitoring scene in the static background semantic cluster; calculating the pixel distance in the vertical direction between the upper boundary line of the fluid semantic cluster and the edge baseline; and converting the pixel distance into physical units based on the preset camera intrinsic parameter matrix and scene depth information. The target vertical scale representation quantity.

8. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S106 includes the following: Extract the area ratio of the semantic cluster of the rigid body with displacement in the video frame image as a feature component characterizing the physical destructive force of the sudden rheological medium; A multidimensional risk assessment matrix is ​​constructed by combining the transient rate of change of the target's vertical scale characteristics with the fluid evolution velocity. Query the corresponding warning level in the multidimensional risk assessment matrix and output the warning command that represents the current monitoring scenario status.

9. The method for video image recognition and risk assessment of flash flood disasters in mountainous areas according to claim 1, characterized in that, Step S101 is followed by: establishing a scene static background feature library, which includes scene geometric features and static texture features in the initial background state; in step S104, by performing differential mapping between the real-time captured video frame sequence and the scene static background feature library, dynamic semantic increments located outside the static background semantic cluster are extracted as the initial processing area for subsequent clustering determination; in step S105, the target vertical scale representation quantity is calculated, which also includes: detecting the occlusion ratio of the static reference in the static background semantic cluster; and performing spatial coordinate compensation on the upper boundary line of the fluid semantic cluster according to the dynamic offset of the occlusion ratio to reduce the target vertical scale representation quantity determination error caused by transient splashing of the medium.

10. A video image recognition and risk assessment system for flash flood disasters in mountainous areas, used to implement the video image recognition and risk assessment method for flash flood disasters in mountainous areas as described in claim 1, characterized in that, include: The data acquisition module is used to acquire video frame sequences of the monitored area scene, including static reference targets, in a continuous time series. The feature field construction module is used to calculate the pixel-level displacement vector between adjacent frames in a video frame sequence in order to construct a motion feature field that characterizes the dynamic rheological properties of pixels. The spatiotemporal coupling module is used to extract the texture sharpness features of each pixel region in the video frame sequence, and map the texture sharpness features and pixel-level displacement vectors to the joint feature space to form a spatiotemporal coupling feature distribution; The semantic decoupling module is used to perform feature space clustering and division of the distribution points in the joint feature space. By identifying fluid semantic clusters with continuous rheological texture features, accompanying rigid body semantic clusters with constant rigid body texture features, and static background semantic clusters with displacement vectors below a preset static threshold, it can realize phase separation identification for the solid-liquid two-phase mixing state in the sudden rheological medium. The static background semantic clusters include static reference markers. The scale calculation module is used to extract the real-time coordinate evolution trajectory of the fluid semantic cluster in the image coordinate system, and calculate the target vertical scale representation and fluid evolution velocity based on the spatial reference coordinates of the static background semantic cluster. The risk assessment module is used to output the target evolution risk level of the scene to be monitored based on the target's vertical scale representation, fluid evolution velocity, and the distribution density of the accompanying rigid body semantic cluster.

Images