A riverway disaster monitoring method based on intelligent video analysis

Abstract

This invention relates to the field of intelligent video analytics, and more particularly to a method for river disaster monitoring based on intelligent video analytics. The method includes: determining whether to perform buffer analysis based on the acquisition stability and interference level of the selected target; during buffer analysis, determining whether to adjust the sample capture method from continuous capture to interval capture based on the feature point change reference value of the selected target; determining whether to adjust the buffering method from vector buffering to original buffering based on the feature prominence of the captured samples to be buffered; and determining whether to trigger a disaster warning or adjust the sample propagation coefficient based on the correlation value of high monitoring demand, based on the buffer pressure coefficient and sample surge coefficient of the video acquisition device. This invention adapts image acquisition and buffering by combining the characteristics of actual floating objects and data generation characteristics, improving data processing efficiency in scenarios with multiple floating objects.

Description

Technical Field

[0001] This invention relates to the field of intelligent video analytics, and in particular to a method for monitoring river disasters based on intelligent video analytics. Background Technology

[0002] Effective monitoring of floating debris in rivers is a crucial aspect of river management. Excessive pollutant discharges from factories or accidents can lead to pollution of rivers by floating debris, causing environmental disasters. Furthermore, large floating objects drifting on the water surface can block water intakes, collide with bridge piers, and accumulate at sluice gates or in narrow river sections, forming obstructions that can damage facilities or even cause secondary disasters such as localized flooding. In practice, this mainly relies on manual patrols along the river or fixed-point monitoring. This method is not only resource-intensive but also fails to meet the growing demands of river management in terms of timeliness and coverage. Therefore, how to achieve effective monitoring of floating debris has become an urgent problem to be solved.

[0003] Chinese Patent Publication No. CN121746970A discloses a method and system for monitoring floating objects on a river surface based on unmanned aerial vehicles (UAVs). The method includes: acquiring riverbank boundary line data of a target area and constructing a riverbank model using the riverbank boundary line data; acquiring lidar point cloud data, visible light image data, and infrared thermal imaging data using UAV patrol based on the riverbank model; fusing the visible light image data and the infrared thermal imaging data to obtain fused image data; and determining floating objects on the river surface using the fused image data and the lidar point cloud data. It is evident that while the above technical solution clearly identifies the appearance characteristics of floating objects through images and accurately obtains key information such as their location and volume using point cloud data, it does not consider the characteristics of actual floating objects, nor does it adaptively acquire and cache images. This can easily lead to problems of high data processing pressure and poor processing efficiency in scenarios with multiple floating objects. Summary of the Invention

[0004] To address this, the present invention provides a river disaster monitoring method based on intelligent video analysis, which overcomes the problems of existing technologies that do not take into account the characteristics of actual floating objects, and that do not adaptively acquire and cache images, which easily leads to high data processing pressure and poor processing efficiency in scenarios with multiple floating objects.

[0005] To achieve the above objectives, the present invention provides a method for river disaster monitoring based on intelligent video analysis, comprising: Whether to perform cache analysis is determined based on the acquisition stability and interference of the selected target. In cache analysis, based on the reference value of the feature point change of the selected target, it is determined whether to change the sample capture method from continuous capture to interval capture. Based on the feature prominence of the captured samples to be cached, determine whether to change the caching method from vector caching to raw caching; Based on the buffer pressure coefficient and sample surge coefficient of the video acquisition device, determine whether to trigger a disaster warning or adjust the sample propagation coefficient based on the correlation value of high monitoring demand; The correlation value for high monitoring demand is related to the number of devices with high monitoring demand, and the devices with high monitoring demand are determined based on the location risk coefficient.

[0006] Furthermore, for acquisition conditions where the acquisition stability is greater than or equal to the preset acquisition stability and the frame selection interference is less than the preset frame selection interference, cache analysis is performed.

[0007] Furthermore, for selected targets whose feature point change reference value is greater than or equal to the preset feature point change reference value, the sample capture method is determined to be continuous capture; For selected targets whose feature point change reference value is less than the preset feature point change reference value, the sample capture method is adjusted to interval capture.

[0008] Furthermore, continuous capture includes: continuously capturing subsequent video frames in time sequence for the current video frame, and uniformly recording the current video frame and the subsequently captured video frames as samples to be cached; The number of video frames subsequently acquired is determined based on the tracking continuity index; The number of video frames and the tracking continuity index are both negatively correlated.

[0009] Furthermore, interval capture includes: taking the time corresponding to the current video frame as the starting point, capturing a video frame once every sampling period, and recording it as a sample to be cached; The sampling period is determined based on the reference value of feature point change; The sampling period is positively correlated with the reference value of feature point change.

[0010] Furthermore, for samples to be cached whose feature prominence is greater than or equal to a preset feature prominence, the caching method is determined to be vector caching; For samples to be cached whose feature prominence is less than the preset feature prominence, the caching method is adjusted to the original caching.

[0011] Furthermore, for the vector caching of the samples to be cached, the following steps are taken: calling the feature extraction model to perform image segmentation on the selected targets in the samples to be cached and convert them into high-dimensional feature vectors.

[0012] Furthermore, for video acquisition devices whose cache pressure coefficient is greater than the preset cache pressure coefficient or whose sample surge coefficient is greater than the preset sample surge coefficient, a disaster warning is issued for the video acquisition device, and the video acquired by the video acquisition device is transmitted to the user terminal.

[0013] Furthermore, for video acquisition devices with a location risk coefficient greater than a preset location risk coefficient, the video acquisition device is determined to be a device with high monitoring demand. For video acquisition devices with a location risk coefficient less than or equal to a preset location risk coefficient, the video acquisition device is determined to be a low monitoring demand device. The location risk coefficient is determined based on the river channel curvature, historical disaster frequency, and water intake distribution density.

[0014] Furthermore, for video acquisition devices with a cache pressure coefficient less than or equal to a preset cache pressure coefficient and a sample surge coefficient less than or equal to a preset sample surge coefficient, the high monitoring demand correlation value corresponding to the video acquisition device is determined. If the correlation value of high monitoring demand corresponding to the video acquisition device is greater than the preset correlation value of high monitoring demand, then the sample propagation coefficient corresponding to the video acquisition device will be increased and adjusted. The increase in the sample propagation coefficient is positively correlated with the correlation value of high monitoring demand.

[0015] Compared with existing technologies, the advantages of this invention are that it reflects the flow stability of the selected target by analyzing the acquisition stability of the selected target, and reflects the interference of image quality on the selection accuracy by using the selection interference degree. Subsequent cache analysis is only performed when the acquisition stability is high enough and the selection interference degree is low enough. This avoids wasting computational and cache resources in scenarios with poor image quality and chaotic target flow, thus improving the overall processing efficiency of the system.

[0016] Furthermore, in the cache analysis, this invention adaptively selects the sample capture method based on the reference value of feature point changes in the selected target. When feature point changes are drastic, continuous capture is used to preserve rich dynamic details; when feature point changes are gradual, interval capture is switched to reduce sampling density. The number of samples in continuous capture is negatively correlated with the tracking continuity index, while the interval size in interval capture is positively correlated with the reference value of feature point changes. This adaptive parameter setting method allows the sample acquisition rate to match the actual movement of the target, thereby improving data processing efficiency.

[0017] Furthermore, in this invention, for samples to be cached, the caching method is adjusted based on feature prominence. Samples with high prominence represent high information gain and are cached using vector caching to segment the target and convert it into a high-dimensional feature vector for compressed storage. Samples with low prominence are highly redundant with existing samples and are switched to the original cache, effectively reducing cache space usage and avoiding the problem of poor data storage performance caused by a single storage method.

[0018] Furthermore, the video acquisition device in this invention is equipped with a sample propagation mechanism to periodically propagate new samples to devices with high monitoring demand within a preset spatial range. When the proportion of devices with high monitoring demand around the device exceeds a threshold, the sample propagation coefficient is adaptively increased, thereby extending the propagation interval and reducing the risk of network congestion. This allows disaster early warning information to dynamically adjust its propagation frequency according to the regional risk density, ensuring information sharing in high-risk areas while avoiding communication load caused by excessive propagation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a river disaster monitoring method for intelligent video analysis according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating how to determine whether to change the sample capture method from continuous capture to interval capture based on the feature point change reference value in an embodiment of the present invention. Figure 3 This is a flowchart illustrating how to determine whether to switch the caching method from vector caching to original caching based on feature prominence, according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating how a video acquisition device is classified as a high-monitoring-demand device or a low-monitoring-demand device based on a location risk coefficient, according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0023] Please see Figure 1 As shown, this is a schematic diagram of a river disaster monitoring method based on intelligent video analysis according to an embodiment of the present invention. The present invention provides a river disaster monitoring method based on intelligent video analysis, comprising: Whether to perform cache analysis is determined based on the acquisition stability and interference of the selected target. In cache analysis, based on the reference value of the feature point change of the selected target, it is determined whether to change the sample capture method from continuous capture to interval capture. Based on the feature prominence of the captured samples to be cached, determine whether to change the caching method from vector caching to raw caching; Based on the buffer pressure coefficient and sample surge coefficient of the video acquisition device, determine whether to trigger a disaster warning or adjust the sample propagation coefficient based on the correlation value of high monitoring demand; The correlation value for high monitoring demand is related to the number of devices with high monitoring demand, and the devices with high monitoring demand are determined based on the location risk coefficient.

[0024] Specifically, at the end of each monitoring cycle, for acquisition conditions where the acquisition stability is greater than or equal to the preset acquisition stability and the frame selection interference is less than the preset frame selection interference, cache analysis is performed.

[0025] In this embodiment of the invention, each video acquisition device is set with an initial image acquisition frequency to control the number of image acquisitions per unit time. Specifically, the initial image acquisition mechanism of the video acquisition device is to acquire the image of the corresponding monitoring area once every five seconds. The target selection is the rectangular box area obtained by selecting floating objects in the image through the YOLO series model.

[0026] Acquisition stability reflects the stability of the flow of selected targets within a unit of time in the monitoring area corresponding to the video acquisition device. Acquisition stability is determined based on the difference between the river flow velocity and the number of selected targets. Acquisition stability and the difference distance between selected targets are negatively correlated. Preferably, when the river flow velocity is greater than or equal to a preset river flow velocity value, the acquisition stability is directly determined to be greater than the preset acquisition stability. For example, the value of acquisition stability is fixed as the preset acquisition stability plus one. Preferably, when the river flow velocity is less than the preset river flow velocity value, the acquisition stability is set as the average value of the difference in the number of selected targets between adjacent acquired images in each acquisition sequence within the monitoring period.

[0027] In this embodiment of the invention, the higher the river flow velocity, the greater the impact on the flow of the selected target. Therefore, when the river flow velocity is greater than the preset river flow velocity, the flow of the selected target is determined to be turbulent, thus ensuring that the acquisition stability is greater than the preset acquisition stability. When the river flow velocity is less than or equal to the preset river flow velocity, the determination is based on the difference in the number of selected targets. Preferably, the flow velocity of different floating objects under different flow velocities can be obtained through simulation. When the flow velocity of the simulated floating objects cannot effectively meet the image monitoring requirements, for example, if the number of images that can be obtained corresponding to 50% of the simulated floating objects is less than the preset number, the corresponding flow velocity is determined to be the failure flow velocity. The minimum value of the failure flow velocity is recorded as the preset river flow velocity. The preset number is set by the user. It can be understood that the larger the preset number, the greater the data analysis accuracy. In this invention, the preset number is 3.

[0028] The bounding box interference degree is used to reflect the degree to which the quality of the acquired image affects the accuracy of bounding box selection. Preferably, the bounding box interference degree is the variance of the grayscale values ​​of each pixel.

[0029] In this embodiment of the invention, the acquisition stability and bounding box interference are used to comprehensively reflect whether the acquired images corresponding to the monitoring period can meet the requirements of buffer analysis. It can be understood that the preset acquisition stability and preset bounding box interference settings are based on the buffer analysis accuracy. The higher the requirement for buffer analysis accuracy, the larger the value of the preset acquisition stability and the smaller the value of the preset bounding box interference. Preferably, through... Under the same conditions, the buffer precision is obtained by separately adjusting the variables for acquisition stability and frame selection interference under the river channel simulation. The minimum values ​​of acquisition stability and frame selection interference that meet the usage requirements are respectively recorded as the preset acquisition stability and preset frame selection interference. The training samples used in the simulation can be historical monitoring videos of the river channel monitoring device when the method described in the embodiments of the present invention is not applied.

[0030] Specifically, for selected targets whose feature point change reference value is greater than or equal to the preset feature point change reference value, the sample capture method is determined to be continuous capture. For selected targets whose feature point change reference value is less than the preset feature point change reference value, the sample capture method is adjusted to interval capture.

[0031] The feature point change reference value reflects the stability of feature point capture corresponding to the selected target. The reference value is the average difference of feature points between adjacent acquired images within the tracking image sequence corresponding to the selected target during the monitoring period. The feature point difference is taken as its absolute value. The tracking image sequence is a sequence of acquired images containing the selected target, continuously tracked and captured using a visual tracking algorithm. The original acquired images are converted to grayscale to obtain a single-channel grayscale image.

[0032] The method for obtaining feature points corresponding to the selected target in a single acquired image is as follows: the acquired image is converted to grayscale to obtain a single-channel grayscale image; adaptive histogram equalization is performed on the area where the selected target is located to enhance local contrast and suppress interference from water surface reflection and uneven lighting; and the ORB algorithm is used to detect feature points in the area where the selected target is located.

[0033] In this embodiment of the invention, a larger feature point change reference value indicates a more drastic change in the features of the selected target between adjacent frames, meaning a faster change in the target's appearance or motion state. In this case, continuous capture can retain richer dynamic details to avoid information loss. Therefore, the sample capture method is determined to be continuous capture rather than interval capture. The greater the user's sensitivity to the accuracy of target change detail capture, the smaller the preset feature point change reference value. Specifically, river monitoring video clips containing different floating object movement speeds and attitude changes are collected, and feature point change reference value thresholds of 3.0, 4.0, 5.0, 6.0, and 7.0 are used respectively to evaluate the combined score of subsequent disaster identification accuracy and cache utilization under continuous and interval capture. Simulation results show that when the threshold is 5.0, the identification accuracy and cache efficiency reach the best balance; therefore, it is set to 5.0.

[0034] Specifically, continuous capture includes: for a sequence of tracked images, continuously selecting a number of images to be acquired, starting with the first image of the sequence, and denoted as samples to be cached; The number of samples selected is determined based on the tracking continuity index; The number of samples screened and the tracking continuity index are both negatively correlated.

[0035] The tracking continuity index is used to reflect the effectiveness of visual tracking of acquired images. The tracking continuity index is the number of acquired images in the tracking image sequence. In this embodiment of the invention, the number of images to be screened is obtained by scaling the number of images acquired in the tracking image sequence through the screening ratio and the transformation value of the tracking continuity index. Preferably, the number of images to be screened is the product of the screening ratio and the transformation value of the tracking continuity index. The transformation value of the tracking continuity index is the product of the reference ratio of the tracking continuity index and the tracking transformation coefficient. The reference ratio of the tracking continuity index is the preset ratio of the tracking continuity index to the tracking continuity index. The tracking conversion coefficient controls the influence of the tracking continuity index on the number of samples to be filtered. The higher the visual tracking effect of the image, the larger the tracking conversion coefficient. The value of the tracking conversion coefficient ranges from (0,1). A larger tracking conversion coefficient means a more sensitive filtering quantity to changes in the tracking continuity index; that is, the higher the tracking continuity index, the more significantly the filtering quantity decreases. In this embodiment, the tracking conversion coefficient is set to 0.6. A larger tracking conversion coefficient indicates a higher requirement for the dynamic balance between tracking performance and sample quantity, meaning the user wants to more aggressively adjust the number of samples to be cached based on changes in the tracking continuity index. Therefore, the greater the user's sensitivity to the balance between visual tracking performance and cache resource consumption, the larger the tracking conversion coefficient.

[0036] In this embodiment of the invention, the screening ratio is a basic proportional coefficient for selecting samples to be cached from the tracked image sequence, and its value ranges from (0,1). The larger the screening ratio, the more samples to be cached are obtained under the same tracking continuity index; the smaller the screening ratio, the fewer samples are obtained. Preferably, the screening ratio is 0.8. The larger the screening ratio, the higher the user's demand for retaining tracking details, that is, the more continuous frames are expected to be collected for subsequent analysis. Therefore, the greater the user's sensitivity to tracking continuity, the larger the screening ratio.

[0037] In this embodiment of the invention, the preset tracking continuity index is a benchmark reference value used to measure the relative level of the actual tracking continuity index. The larger the preset tracking continuity index, the larger the reference ratio under the same actual tracking continuity index, and thus the larger the converted processing value, and the more samples to be cached obtained. In this embodiment of the invention, the preset tracking continuity index is set to 10, which is the minimum number of frames expected in the tracking image sequence. It can be understood that the larger the preset tracking continuity index, the higher the user's benchmark requirement for the length of the tracking sequence, that is, they tend to collect more samples even when the tracking effect is good. Therefore, the greater the user's sensitivity to the bottom line requirement of tracking continuity, the larger the preset tracking continuity index.

[0038] Specifically, interval capture includes: for a tracking image sequence, selecting images to be acquired at intervals starting with the first image of the sequence, and denoting them as samples to be cached; The number of images corresponding to the interval between two adjacent selections is determined based on the feature point change reference value. The number of images at the specified interval is positively correlated with the reference value of feature point changes.

[0039] In this embodiment of the invention, the number of images in the interval is the product of the initial number of intervals and the transformed processing value of the feature point change reference value. The transformed processing value of the feature point change reference value is the product of the reference ratio of the feature point change reference value and the feature transformation coefficient. The reference ratio of the feature point change reference value is the percentage of the preset feature point change reference value to the total feature point change reference value.

[0040] The initial interval number is a base value for the number of images between two adjacent acquisitions in interval capture mode, and its value is greater than or equal to 1. A larger initial interval number results in a larger sampling interval and a sparser number of cached samples when the reference value of feature point changes is stable; a smaller initial interval number results in denser sampling. In this embodiment of the invention, one frame is acquired by default every two frames. It can be understood that the initial interval number reflects the default sampling sparsity when the user does not consider changes in feature points. The greater the user's sensitivity to the balance between cache resource usage and basic sampling density, the larger the initial interval number.

[0041] In this embodiment of the invention, the feature transformation coefficient is used to control the sensitivity of the adjustment of the number of interval images to the reference value of feature point changes, and the value range is (0, 2). The larger the feature transformation coefficient, the larger the number of intervals calculated under the same reference value of feature point changes, and the sparser the sampling; the smaller the feature transformation coefficient, the denser the sampling. In this embodiment of the invention, the feature transformation coefficient is 0.8. It can be understood that the greater the user's sensitivity to the balance between saving cache resources and preserving monitoring details, the larger the feature transformation coefficient, and the more inclined they are to use sparse sampling to reduce cache pressure.

[0042] The preset feature point change reference value is 5.0. This means that the smaller the feature point change reference value relative to the preset feature point change reference value, the more stable and gradual the changes in the selected target features are. In this case, increasing the sampling interval reduces cache usage. Conversely, the larger the feature point change reference value, the more drastic the changes in the target are, requiring denser sampling to capture key changes. The greater the user's sensitivity to the match between feature stability and sampling density, the smaller the preset feature point change reference value should be.

[0043] Specifically, for samples to be cached whose feature prominence is greater than or equal to the preset feature prominence, the caching method is determined to be vector caching. For samples to be cached whose feature prominence is less than the preset feature prominence, the caching method is adjusted to the original caching.

[0044] Feature prominence is used to quantify the uniqueness of the current sample to be cached in the entire set of samples to be cached, that is, the distinguishability of the sample from other samples. The greater the feature prominence, the more unique the features of the sample and the less likely it is to be confused with other samples; the smaller the feature prominence, the more similar the features of the sample are to other samples and the higher the redundancy. For a sample to be cached, its feature prominence and feature vector similarity are negatively correlated. In this embodiment of the invention, the feature vector similarity is the reciprocal of the average cosine similarity of the feature vectors of the sample to be cached and all other samples to be cached.

[0045] In this embodiment of the invention, multiple video clips under different lighting conditions, floating object types, and water surface ripple levels are collected from a river monitoring scenario. Feature extraction is performed on the selected target samples within the video clips, and the feature prominence of each sample is calculated. Switching between vector caching and the original cache is tested using different preset feature prominence thresholds. The cache storage space occupancy rate and the subsequent target recognition accuracy are used as evaluation indicators, and a threshold value is selected to achieve the user's required storage space and recognition accuracy. Simulation results show that when the preset feature prominence is 0.6, the sample proportion of the vector cache is approximately 40%, the storage space is reduced by about 35% compared to the full original cache, and the recognition accuracy decreases by no more than 2%, resulting in the best overall performance.

[0046] Specifically, vector caching for samples to be cached includes: calling a feature extraction model to perform image segmentation on the selected targets in the samples to be cached and convert them into high-dimensional feature vectors.

[0047] In this embodiment of the invention, a pre-trained feature extraction model, such as ResNet or MobileNet, is invoked to segment the target region within the bounding box of the sample to be cached from the original image. This involves cropping image patches according to the coordinates of the bounding rectangle and inputting them into the feature extraction model. After forward propagation, a high-dimensional feature vector of a fixed dimension is output, such as a 512-dimensional or 1024-dimensional floating-point vector. This feature vector, in a compact form, represents the discriminative information of the bounding target, such as texture, shape, and color, and is used for subsequent disaster identification, target matching, and sample propagation coefficient calculation. It can be understood that vector caching converts the original image into a high-dimensional feature vector, significantly reducing storage space while preserving the core features of the target.

[0048] Specifically, for video acquisition devices whose cache pressure coefficient is greater than the preset cache pressure coefficient or whose sample surge coefficient is greater than the preset sample surge coefficient, a disaster warning is issued for the video acquisition device, and the video acquired by the video acquisition device is transmitted to the user terminal.

[0049] Disaster early warning: Send warning reports to users.

[0050] The cache pressure coefficient reflects the data cache pressure of the video acquisition device, and the sample surge coefficient reflects the cache growth of the cached samples corresponding to video acquisition. By covering two risk assessments—system pressure caused by long-term cache accumulation and the risk of cache failure, as well as the environmental anomaly risk reflected by short-term sample surge—effective disaster monitoring can be achieved. In this embodiment of the invention, the cache pressure coefficient is the ratio of the number of samples currently cached by the video acquisition device to the maximum capacity of the cache queue, and the maximum capacity of the cache queue is the maximum number of images that the video acquisition device can cache. In this embodiment of the invention, a higher cache pressure coefficient indicates that the cache queue of the video acquisition device is closer to full load, and the system's processing capacity is approaching saturation. Continuing to increase the number of samples may lead to data loss or processing delays, thus making the need to trigger disaster warnings and promptly upload the acquired videos more urgent. The greater the user's sensitivity to the system's real-time response capability and data integrity, the lower the preset cache pressure coefficient. By collecting cache load data from multiple sets of video acquisition devices in different river scenarios, and conducting several warning tests with different cache pressure coefficient thresholds, using the false alarm rate as the evaluation index (i.e., the proportion of warnings triggered when there are no risk events), simulation results show that when the preset cache pressure coefficient is 0.8, the false alarm rate is less than 5%, resulting in the best overall performance.

[0051] The sample surge coefficient is used to quantify the abnormal increase in the number of newly generated samples to be cached by the video acquisition device within a short period of time. The sample surge coefficient is the ratio of the number of newly generated samples to be cached in the current monitoring period to the average number of newly generated samples to be cached in the most recent monitoring periods prior to this monitoring period. In this embodiment of the invention, the preset sample surge coefficient is 2.0. Data on the change in the number of samples during normal fluctuation periods and before disasters are collected from historical river monitoring data. Early warning tests are conducted using sample surge coefficient thresholds of 1.5, 2.0, 2.5, and 3.0, respectively. The time difference between successful early warnings and the false alarm rate before a disaster occurs are used as evaluation indicators. Simulation results show that when the threshold is 2.0, the average early warning lead time can reach 15 minutes and the false alarm rate is controlled within 8%, resulting in the best overall performance. Therefore, the preset sample surge coefficient is set to 2.0.

[0052] Specifically, for video acquisition devices with a location risk coefficient greater than a preset location risk coefficient, the video acquisition device is identified as a high-monitoring-requirement device. For video acquisition devices with a location risk coefficient less than or equal to a preset location risk coefficient, the video acquisition device is determined to be a low monitoring demand device. The location risk coefficient is determined based on river meander, historical disaster frequency, and water intake distribution density.

[0053] River meandering reflects the degree of curvature of a river section's geometry. Greater meandering indicates faster water flow, a greater likelihood of floating debris accumulating at bends or impacting riverbanks, and a higher probability of disasters. Therefore, it is positively correlated with the location risk coefficient. Specifically, river meandering is calculated by taking a random point on the river within the monitoring area of ​​a video acquisition device, selecting a section of the river between 500 meters upstream and 500 meters downstream, and calculating the ratio of the actual length of this section to the straight-line distance.

[0054] Historical disaster frequency reflects the density of disaster events occurring in a river section over a certain period. A higher frequency indicates a more disaster-prone area and higher risk, thus showing a positive correlation with the location risk coefficient. Specifically, the frequency is calculated by dividing the number of disaster events recorded within the video capture device's field of view since the system's deployment (such as debris blockage, sudden water level rises, and bridge pier collisions) by the number of years to obtain the annual average disaster frequency. This frequency is then normalized, and the highest annual average frequency in the entire river system is used as the upper limit for reference, thus yielding the historical disaster frequency.

[0055] The density of water intakes reflects the concentration of water intake facilities within a river section, such as those for agricultural irrigation, industrial water intake, and drinking water sources. Areas near water intakes are extremely sensitive to floating debris; blockages can directly impact water supply safety. Therefore, higher density equates to a greater risk factor, which is positively correlated with location risk. The search area is defined as 1 kilometer upstream and downstream of the video acquisition device. The total number of water intakes within this search area is counted, and the ratio of this total number to the search area is recorded as the water intake density.

[0056] The location risk coefficient comprehensively reflects the risk level of the video acquisition device. The location risk coefficient is positively correlated with river channel curvature, historical disaster frequency, and water intake density. Preferably, the location risk coefficient is obtained by scaling the river channel curvature, historical disaster frequency, and water intake density separately, followed by dimensionless summation. The scaling process involves multiplying the data value by its corresponding contribution coefficient. For example, the contribution coefficients for river channel curvature, historical disaster frequency, and water intake density are 0.4, 0.3, and 0.3, respectively. It can be understood that the contribution of these three factors to the risk of floating debris in the river is directly proportional to the value of their contribution coefficients. The value of the contribution coefficient reflects the user's emphasis on the three risk factors. If the user believes that a certain factor, such as water intake density, has a greater impact on the disaster, its contribution coefficient can be increased accordingly; conversely, it can be decreased. In this embodiment of the invention, the contribution coefficient is used as the default configuration, which is applicable to most conventional river scenarios.

[0057] In this embodiment of the invention, a higher location risk coefficient indicates a higher degree of curvature in the river section where the video acquisition device is located, a more frequent history of disasters, and a denser concentration of water intakes. This means the river section is more sensitive to and has a higher potential for floating debris hazards, and therefore should be marked as a high-monitoring-demand device. The greater the user's sensitivity to the monitoring intensity of high-risk areas in the river, the lower the preset location risk coefficient, which is 0.5.

[0058] Specifically, for video acquisition devices whose cache pressure coefficient is less than or equal to the preset cache pressure coefficient and whose sample surge coefficient is less than or equal to the preset sample surge coefficient, the high monitoring demand correlation value corresponding to the video acquisition device is determined. If the correlation value of high monitoring demand corresponding to the video acquisition device is greater than the preset correlation value of high monitoring demand, then the sample propagation coefficient corresponding to the video acquisition device will be increased and adjusted. The increase in the sample propagation coefficient is positively correlated with the correlation value of high monitoring demand.

[0059] The high monitoring demand correlation value reflects the density of high-risk river sections around the current device. The larger the high monitoring demand correlation value, the higher the overall disaster risk in the area. Therefore, the sample propagation coefficient needs to be increased and adjusted to accelerate the propagation rate of disaster information between adjacent devices. Preferably, taking the current video acquisition device as the center, the number of devices marked as having high monitoring demand within a preset spatial range centered on the video acquisition device is counted. The high monitoring demand correlation value is the ratio of the number of devices with high monitoring demand within the preset spatial range to the total number of video acquisition devices.

[0060] It is understandable that the more devices with high monitoring demand, the stronger the regional risk coupling, and the more quickly the disaster characteristics collected by a single device need to be disseminated to neighboring devices. The greater the user's sensitivity to the speed of disaster early warning propagation, the smaller the preset high monitoring demand correlation value should be. In this embodiment of the invention, the preset high monitoring demand correlation value is set to 0.5, that is, adjustment is triggered when the proportion of surrounding devices with high monitoring demand exceeds half. This value is determined through simulation: in multiple simulated disaster scenarios, thresholds of 0.3, 0.4, 0.5, and 0.6 are used to evaluate the average delay time for early warning information to propagate from the starting device to adjacent devices. When the threshold is 0.5, the early warning propagation efficiency and the risk of false alarm spread reach the optimal balance, so it is set to 0.5.

[0061] In this embodiment of the invention, each video acquisition device is equipped with a propagation mechanism, that is, after a preset propagation time, the newly added samples within the preset propagation time are propagated to all high monitoring demand devices within a preset spatial range. The preset propagation time = base time × sample propagation coefficient, the initial value of the sample propagation coefficient is 1, the adjusted sample propagation coefficient is the sum of the initial value and the propagation adjustment value, and the propagation adjustment value is the ratio of the high monitoring demand correlation value to the preset high monitoring demand correlation value.

[0062] In this embodiment of the invention, the preset propagation duration is a baseline value for the time interval between the video acquisition device and the surrounding high-monitoring-demand devices to propagate new samples. It can be understood that the longer the preset propagation duration, the sparser the sample propagation and the lower the system communication load, but the real-time performance of disaster warning will decrease accordingly; the shorter the preset propagation duration, the more frequent the propagation and the faster the warning information is updated. Preferably, the preset propagation duration is 30 minutes.

[0063] In this embodiment of the invention, the preset spatial range is the effective geographical area from which the video acquisition device transmits new samples to surrounding devices with high monitoring needs. It is understood that a larger preset spatial range results in a wider coverage of disaster information dissemination and more effective coordination between adjacent devices, but also increases the amount of data transmitted and communication overhead. Conversely, a smaller preset spatial range results in more targeted dissemination, but may miss high-risk areas at the periphery. In this embodiment of the invention, the preset spatial range is 2 kilometers.

[0064] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A river disaster monitoring method based on intelligent video analysis, characterized in that, include: Whether to perform cache analysis is determined based on the acquisition stability and interference of the selected target. In cache analysis, based on the reference value of the feature point change of the selected target, it is determined whether to change the sample capture method from continuous capture to interval capture. Based on the feature prominence of the captured samples to be cached, determine whether to change the caching method from vector caching to raw caching; Based on the buffer pressure coefficient and sample surge coefficient of the video acquisition device, determine whether to trigger a disaster warning or adjust the sample propagation coefficient based on the correlation value of high monitoring demand; The correlation value for high monitoring demand is related to the number of devices with high monitoring demand, and the devices with high monitoring demand are determined based on the location risk coefficient. 2.The river disaster monitoring method based on intelligent video analysis according to claim 1, wherein, For acquisition conditions where the acquisition stability is greater than or equal to the preset acquisition stability and the box selection interference is less than the preset box selection interference, cache analysis is performed. 3.The river disaster monitoring method based on intelligent video analysis according to claim 2, characterized in that, For selected targets whose feature point change reference value is greater than or equal to the preset feature point change reference value, the sample capture method is determined to be continuous capture; For selected targets whose feature point change reference value is less than the preset feature point change reference value, the sample capture method is adjusted to interval capture. 4.The river disaster monitoring method based on intelligent video analysis according to claim 3, wherein, Continuous capture includes: continuously capturing subsequent video frames in time sequence for the current video frame, and recording the current video frame and the subsequently captured video frames as samples to be cached; The number of video frames subsequently acquired is determined based on the tracking continuity index; The number of video frames and the tracking continuity index are both negatively correlated. 5.The river disaster monitoring method based on intelligent video analysis according to claim 4, wherein, Interval capture includes: taking the time corresponding to the current video frame as the starting point, capturing a video frame once every sampling period, and recording it as a sample to be cached; The sampling period is determined based on the reference value of feature point change; The sampling period is positively correlated with the reference value of feature point change. 6.The river disaster monitoring method based on intelligent video analysis according to claim 5, wherein, For samples to be cached whose feature prominence is greater than or equal to the preset feature prominence, the caching method is determined to be vector caching; For samples to be cached whose feature prominence is less than the preset feature prominence, the caching method is adjusted to the original caching.

7. The river disaster monitoring method based on intelligent video analysis according to claim 6, characterized in that, Vector caching for samples to be cached includes: calling a feature extraction model to perform image segmentation on the selected targets in the samples to be cached and convert them into high-dimensional feature vectors. 8.The river disaster monitoring method based on intelligent video analysis according to claim 7, wherein, For video acquisition devices whose cache pressure coefficient is greater than the preset cache pressure coefficient or whose sample surge coefficient is greater than the preset sample surge coefficient, a disaster warning is issued for the video acquisition device, and the video acquired by the video acquisition device is transmitted to the user terminal. 9.The river disaster monitoring method based on intelligent video analysis according to claim 8, wherein, For video acquisition devices with a location risk coefficient greater than the preset location risk coefficient, the video acquisition device is determined to be a device with high monitoring requirements. For video acquisition devices with a location risk coefficient less than or equal to a preset location risk coefficient, the video acquisition device is determined to be a low monitoring demand device. The location risk coefficient is determined based on river meander, historical disaster frequency, and water intake distribution density. 10.The river disaster monitoring method based on intelligent video analysis according to claim 9, wherein, For video acquisition devices whose cache pressure coefficient is less than or equal to the preset cache pressure coefficient and whose sample surge coefficient is less than or equal to the preset sample surge coefficient, determine the high monitoring demand correlation value corresponding to the video acquisition device. If the correlation value of high monitoring demand corresponding to the video acquisition device is greater than the preset correlation value of high monitoring demand, then the sample propagation coefficient corresponding to the video acquisition device will be increased and adjusted. The increase in the sample propagation coefficient is positively correlated with the correlation value of high monitoring demand.

Images