Road low visibility monitoring method and system based on video analysis

Abstract

The invention relates to the technical field of traffic management, and discloses a road low-visibility monitoring method and system based on video analysis, and the method comprises the steps: carrying out the quantification of the degradation features of a reference object region obtained through dynamic recognition in a target road monitoring video, and generating a degradation value sequence of multiple features of the reference object region; and performing spatial-temporal feature extraction on the degradation value sequence to obtain a steady-state feature vector of the reference object region, performing multiple linear regression calculation, performing early warning grading on the road visibility in the target road monitoring video according to the obtained dynamic visibility index of the target road monitoring video, and generating an early warning scheme. Through multi-frame trajectory association and streaming verification, four-dimensional features of a reference object area are synchronously extracted, collinearity features are automatically removed, a high-precision dynamic visibility index is output, the problems of repeated signal calculation and prediction errors are avoided, and the accuracy and stability of road low-visibility monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic management, and in particular to a method and system for monitoring low road visibility based on video analysis. Background Art

[0002] Current technical solutions rely on a single visual feature for visibility assessment and are unable to cope with complex environmental interference. Local feature degradation can cause overall monitoring failure, and single-point feature degradation can cause cascading collapse of the monitoring link. Consequently, they lack temporal dynamic filtering capabilities, resulting in a single feature being unable to simultaneously respond to multi-dimensional degradation signals and unable to maintain the output stability of the assessment link under complex interference.

[0003] The current method treats image features as independent decision variables and uses fixed-weight linear superposition, ignoring the nonlinear correlation between features. When there is strong coupling between features, independent weight allocation causes the same degradation signal to be repeatedly calculated, which will increase the prediction deviation. At the same time, multicollinearity is not effectively identified, which will interfere with the stability of the solution and cannot meet the needs of high-precision road visibility monitoring. Summary of the Invention

[0004] The present invention provides a method and system for monitoring low visibility on roads based on video analysis, the main purpose of which is to solve the problems of fragile cascade and low precision in low visibility monitoring on roads based on video analysis.

[0005] To achieve the above objectives, the present invention provides a method for monitoring low visibility on roads based on video analysis, comprising: S1. Dynamically identifying inherent reference objects of image features in a target road surveillance video to obtain continuous reference object regions in the target road surveillance video; S2. quantifying degradation features of the reference object region to obtain a degradation value sequence of multiple features of the reference object region; S3. Extracting spatiotemporal features from the degradation value sequence to obtain a steady-state feature vector of the reference object region; S4, performing a multivariate linear regression calculation on the steady-state eigenvector to obtain a dynamic visibility index of the target road surveillance video; S5. Performing warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain a warning level of the road visibility.

[0006] In a preferred embodiment, the dynamically identifying inherent reference objects of image features in the target road surveillance video to obtain continuous reference object areas in the target road surveillance video includes: Detecting the target road surveillance video with the reference object as the target to obtain an initial candidate frame set of the target road surveillance video; Performing multi-frame trajectory association on the initial candidate frame set to obtain motion trajectory data of the reference object; The motion trajectory data of the reference object is subjected to streaming verification to obtain a continuous reference object area in the target road monitoring video.

[0007] In a preferred embodiment, when performing degradation feature quantization on the reference object region to obtain a degradation value sequence of multiple features of the reference object region, the process includes: Performing single-frame degradation feature extraction on the reference object region to obtain a single-frame feature vector of the reference object region; Quantizing and iterating the single-frame feature vector to obtain a frame sequence feature set of the reference object region; The frame sequence feature set is temporally integrated to obtain a degradation value sequence of multiple features in the reference object area.

[0008] In a preferred embodiment, the single-frame feature vector includes: Using the gradient amplitude of the grayscale image in the reference object area as the edge sharpness quantization value of the target reference object area; Performing LBP feature extraction on the target area of ​​the reference object to obtain an LBP coding map of the target reference object area; Performing histogram variance statistics on the LBP coding image to obtain a texture contrast quantization value of the target reference object area; extracting a channel mean value of the reference object area after format conversion, and quantizing the channel mean value into a color saturation quantization value; The degradation rate of the visual features of the current frame and the benchmark features in the reference object area is used as the target recognizable distance estimation value of the reference object area.

[0009] In a preferred embodiment, the degradation value sequence is subjected to spatiotemporal feature extraction to obtain the steady-state feature vector of the reference object region, including: Taking the mean of the degradation value sequence as the mean subsequence of the degradation value sequence; The standard deviation of the features in the degradation value sequence under the same sliding window is used as the characteristic fluctuation quantum sequence of the degradation value sequence; The mean subsequence and the characteristic fluctuation subsequence are arranged in time order to obtain a steady-state characteristic vector of the reference object region.

[0010] In a preferred embodiment, when the mean subsequence and the characteristic fluctuation subsequence are arranged in time order to obtain a steady-state characteristic vector, the method includes: The mean subsequence and the characteristic fluctuation subsequence are time-aligned and stored to obtain a time series label data set of degradation characteristics of the reference object area; Performing spatiotemporal feature extraction on the time series label dataset to obtain a preliminary feature vector of the degradation feature of the reference object region; The preliminary feature vectors are sequentially fused to obtain a steady-state feature vector of the reference object region.

[0011] In a preferred embodiment, the steady-state feature vector is subjected to a multivariate linear regression calculation to obtain the dynamic visibility index of the target road monitoring video, including: Performing parameter matrix training on the steady-state eigenvector to obtain a regression coefficient vector of the reference object region; Performing weighted summation on the regression coefficient vector to obtain a visibility fusion base value of the reference object area; An activation conversion is performed on the visibility fusion base value to obtain a dynamic visibility index of the reference object area.

[0012] In a preferred embodiment, the parameter matrix training of the steady-state eigenvector to obtain the regression coefficient vector of the reference object region includes: splicing the steady-state feature vectors based on the types of feature labels to obtain an augmented matrix of the reference object region; extracting the covariance of the augmented matrix, and constructing a multidimensional covariance matrix of the reference object region based on the covariance; The inverted covariance matrix is ​​multiplied element by element by the covariance matrix to obtain a regression coefficient vector of the reference object region.

[0013] In a preferred embodiment, the road visibility in the target road monitoring video is graded based on the dynamic visibility index to obtain the warning level of the road visibility, including: Filtering invalid values ​​of the real-time visibility index to obtain a valid visibility sampling set in the target road monitoring video; Performing warning matching on the effective visibility sampling set to obtain a warning level of the target road monitoring video; Perform risk assessment on the warning level to obtain signal control of the target road monitoring video.

[0014] In order to solve the above problems, the present invention also provides a road low visibility monitoring system based on video analysis, the system comprising: S1, the reference object extraction module is used to dynamically identify the inherent reference objects of the image features in the target road monitoring video to obtain a continuous reference object area in the target road monitoring video; S2, the degradation value sequence generation module is used to quantify the degradation characteristics of the reference object area to obtain a degradation value sequence of multiple features of the reference object area; S3, the steady-state feature vector generation module is used to extract spatiotemporal features of the degradation value sequence to obtain a steady-state feature vector of the reference object region; S4, the dynamic visibility index module, used for performing a multivariate linear regression calculation on the steady-state feature vector to obtain a dynamic visibility index of the target road monitoring video; S5. The warning level generating module is configured to perform warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain the warning level of the road visibility.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention breaks through the limitations of traditional single-dimensional perception, characterizes image features into multiple indicators, and constructs a four-dimensional degradation feature collaborative quantification system. By synchronously fusing four complementary indicators: edge clarity, texture contrast, color saturation, and recognizable distance, it reduces the misjudgment rate due to local feature degradation or single feature failure, avoids single-frame errors, realizes cross-modal feature cross-validation, and improves the stability of the monitoring solution.

[0016] 2. The present invention uses a covariance-driven dynamic regression architecture to analyze the nonlinear correlation between features, dynamically adjust the weights of four types of features, fit the linear relationship between each degenerate feature, automatically eliminate highly linearly correlated features, avoid prediction bias caused by ignoring feature correlation, and maximize feature accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a process flow of a method for monitoring low road visibility based on video analysis provided by one embodiment of the present invention; Figure 2 A functional module diagram of a road low visibility monitoring system based on video analysis provided by one embodiment of the present invention; The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] The embodiment of the present application provides a method for monitoring low visibility on roads based on video analysis. The execution subject of the method for monitoring low visibility on roads based on video analysis includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for monitoring low visibility on roads based on video analysis can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 FIG. 1 is a flow chart of a method for monitoring low road visibility based on video analysis according to an embodiment of the present invention. In this embodiment, the method for generating reference information based on artificial intelligence and smart home includes: In an example of the present invention, the dynamic identification of the inherent reference objects of the image features in the target road monitoring video to obtain the continuous reference object area in the target road monitoring video is specifically used for: Detecting the target road surveillance video with the reference object as the target to obtain an initial candidate frame set of the target road surveillance video; Performing multi-frame trajectory association on the initial candidate frame set to obtain motion trajectory data of the reference object; The motion trajectory data of the reference object is subjected to streaming verification to obtain a continuous reference object area in the target road monitoring video.

[0021] Specifically, based on known reference objects, the approximate location range of possible target objects in each frame of surveillance footage is quickly determined. Initial candidate boxes are automatically drawn around these possible target locations, marking the areas of the footage that require further analysis. Specifically, the reference object position frames captured in video frames at different time points are linked together to confirm the uniqueness of the reference object they represent. By linking the positions of multiple time points, a continuous path of the reference object's movement over time is created, recording continuous motion information such as the direction and speed of the reference object's movement.

[0022] Specifically, the motion route being formed is monitored in real time to determine whether the newly added position points at each moment are consistent with the previous position points, whether the changes in movement direction and speed are natural, and to discover and eliminate obvious wrong position points or route jumps caused by temporary occlusion, image noise or recognition errors, to ensure that the position area of ​​the reference object finally output in each frame of video is connected before and after, without abrupt breaks or jumps.

[0023] Furthermore, it avoids indiscriminate detailed inspection of the entire video frame and focuses subsequent analysis on local areas near the reference object where the target is more likely to appear, providing a set of clear area location information that requires in-depth inspection for the subsequent more refined recognition or tracking steps of the solution.

[0024] Furthermore, the static position information of a single frame is converted into a dynamic movement route across time, providing a basis for the subsequent continuous attention and prediction of the future position of the reference object, helping to distinguish multiple similar reference objects or interference objects that may exist in the picture, and ensuring that only the route of the target reference object is tracked.

[0025] Furthermore, the credibility and stability of the reference object's movement route and position information are enhanced, providing a set of verified, temporally continuous and spatially coherent reference object area information for all subsequent visibility analysis that relies on the reference object's position, enabling the entire solution to eliminate single-frame or short-term recognition errors and image interference, and maintain long-term stable operation.

[0026] In general, only the key local areas in the picture are processed, which greatly reduces the amount of data that needs to be analyzed, making the entire surveillance video analysis process faster. Moreover, it only needs to be performed within the area marked by the initial candidate box, avoiding invalid operations on irrelevant backgrounds, improving the accuracy and speed of recognition, and making it practical to perform automated analysis of long-term continuous surveillance videos over a large area.

[0027] In general, it is necessary to clearly understand the actual route, direction change and speed of the reference object in the monitoring area, and directly obtain key information such as the average speed, instantaneous speed, acceleration, and driving direction of the reference object from the formed moving route. Through the correlation confirmation of multiple consecutive frames, it is possible to effectively eliminate the possible short-term interference or positioning errors in a single frame, making target identification more reliable.

[0028] In general, based on the coherent and stable position of the reference object, more accurate parameters such as the target position, speed, and distance can be obtained, avoiding errors introduced by jitter or sudden changes in the reference object position. At the same time, the reference object area is ensured to be uninterrupted in the time dimension, so that the continuous tracking and analysis of visibility that depends on the reference object will not fail due to the interruption of the reference object information, and the availability of the solution can be maintained in actual complex environments.

[0029] In an embodiment of the present invention, when the degradation feature quantization is performed on the reference object region to obtain a degradation value sequence of multiple features of the reference object region, the method includes: Performing single-frame degradation feature extraction on the reference object region to obtain a single-frame feature vector of the reference object region; Quantizing and iterating the single-frame feature vector to obtain a frame sequence feature set of the reference object region; The frame sequence feature set is temporally integrated to obtain a degradation value sequence of multiple features in the reference object area.

[0030] Specifically, the specific performance characteristics of the reference object area in the current single frame due to various interferences that lead to the degradation of image quality are identified and recorded, and the observed specific performance characteristics are converted into a set of numerical description information that can be understood and processed in subsequent steps. A digital label is generated for the reference object area in each frame, which reflects the image quality status of the area in the current frame.

[0031] Specifically, the picture state features scattered on each frame are collected in the chronological order of their appearance in the video. By comparing and correlating the feature values ​​of the previous and next frames, the pattern or trend of the picture state change over time is found, and a continuous reference area picture state feature record covering the entire video timeline is generated.

[0032] Specifically, the information of multiple features reflecting different aspects of picture quality in each time period is aggregated into a more representative comprehensive value, and the frame-by-frame feature sequence that may contain redundancy or subtle fluctuations is refined into a numerical sequence that can better highlight the main change trends, directly quantifying the overall degree of degradation of the picture quality in the reference area at that moment.

[0033] Furthermore, the output groups of features are vectors of core input data for analyzing how the picture quality of the reference area evolves with the frame, providing a basic basis for subsequent visibility wind judgment of the picture quality changes of the reference area between different frames, and helping subsequent steps to find the key video frame positions where the picture quality changes significantly.

[0034] Furthermore, the picture status of a single frame is expanded to a complete view showing how the picture status evolves over time, providing a core data basis for subsequent visibility analysis of the fluctuation, degradation or recovery of the picture quality of the reference area throughout the video process, ensuring the accurate identification of the specific video frame locations where the picture quality has significantly improved or decreased.

[0035] Furthermore, the multi-dimensional picture status time series is condensed into a one-dimensional timeline of picture quality degradation level, providing a core quantitative basis for subsequent analysis of the overall pattern, key turning points and severity of picture quality degradation over time in the reference area of ​​visibility, and accelerating the identification of specific video time periods where picture quality deteriorates significantly and continuously.

[0036] In general, by comparing the feature vectors of different frames, we can clearly locate the specific video frames where the image quality of the reference area begins to deteriorate or recover, find the turning point of the quality change, and provide a decision-making basis for subsequent analysis steps that need to be dynamically adjusted according to the image quality. It also objectively records and reflects the actual impact of various factors on the reference area in the monitoring image.

[0037] In general, it intuitively presents the changing trend of the picture quality of the reference area from the beginning to the end of the video, clearly points out the exact time point when the key change of the picture status occurs, as well as the time period when the picture quality remains relatively stable and the time period when it fluctuates violently, effectively eliminating misjudgments caused by accidental interference and more accurately reflecting the overall change process of the picture quality.

[0038] In general, an objective and continuous quantitative indicator sequence is provided to measure the actual impact of interference factors on the availability of surveillance images, and the specific time intervals when the image quality begins to decline continuously, reaches the worst state, and begins to recover are determined, so that subsequent analysis and decision-making can be focused on video clips where there are indeed significant problems with image quality, improving efficiency.

[0039] In this example of the present invention, the single-frame feature vector includes: Using the gradient amplitude of the grayscale image in the reference object area as the edge sharpness quantization value of the target reference object area; Performing LBP feature extraction on the target area of ​​the reference object to obtain an LBP coding map of the target reference object area; Performing histogram variance statistics on the LBP coding image to obtain a texture contrast quantization value of the target reference object area; extracting a channel mean value of the reference object area after format conversion, and quantizing the channel mean value into a color saturation quantization value; The degradation rate of the visual features of the current frame and the benchmark features in the reference object area is used as the target recognizable distance estimation value of the reference object area.

[0040] Specifically, the intensity of the brightness change at the edge is converted into a value that directly represents the edge clarity of the area. That is, when the gradient amplitude is larger, the more intense the brightness change at the edge, the clearer and sharper the edge is visually; when the gradient amplitude is smaller, the brightness change at the edge is more gradual, and the edge is visually blurred. Specifically, the brightness change pattern of each pixel neighborhood in the target area is identified and recorded, and the brightness relationship pattern of each pixel neighborhood is converted into a specific digital code, and finally an image composed of these code values ​​is formed, providing a basic data layer describing the microscopic texture structure information of the target area surface for subsequent analysis.

[0041] Specifically, the degree of difference in the number of times different texture patterns appear in the entire area is monitored, and the observed distribution breadth and fluctuation amplitude of the texture patterns are comprehensively converted into a numerical value that reflects the strength of the overall texture contrast in the area.

[0042] Specifically, we focus on the data of a specific channel in the image that represents the purity of the color, find the average level of the vividness component in the entire reference area, and directly set this regional average value as the core value that reflects the overall color saturation of the area.

[0043] Specifically, the size of the gap between the visual performance of the reference object area in the current frame and its optimal state is monitored, and the ratio of the gap is converted into a spatial distance value reflecting the decline in recognition ability. A specific value is output indicating the distance at which the recognition system can effectively recognize the reference object under the current picture quality.

[0044] Furthermore, it provides a measurable basic value for the entire visibility analysis, which is used to directly determine whether the edge contour of the reference area in a single frame is clear and usable. It is also one of the most core input data for subsequent analysis of whether the image quality clarity has degraded and the degree of degradation. It helps provide a basis for comparing the edge clarity differences between different frames or between different areas of the same frame.

[0045] Furthermore, the LBP feature focuses on relative brightness, making it more stable against uniform changes in light intensity and improving the reliability of reference object recognition and matching. This provides core, local pattern-based feature data for subsequent identification, matching, or differentiation of different reference objects, or analysis of texture variations of the same reference object under different conditions.

[0046] Furthermore, this value is one of the important bases for judging whether the surface texture of the reference object is normal, whether it has changed significantly, or whether there is interference, and provides supplementary information for subsequent analysis of image quality, especially the loss of texture information due to blur or noise.

[0047] Furthermore, this value is one of the important bases for judging whether the color of the reference object is normal, whether it has faded significantly, changed color, or is affected by interference. It provides a key, objective quantitative indicator for the subsequent evaluation of the color characteristics of the reference object area, and also provides key information for changes in lighting conditions or degradation of picture quality.

[0048] Furthermore, it directly reflects the effective range of the reference object recognition system under current environmental conditions, and is the most direct basis for judging whether the entire monitoring plan can work normally at the expected distance under the current moment and current picture conditions, providing a decision-making basis based on the critical point of recognition capability for whether subsequent visibility analysis needs to initiate response measures.

[0049] In general, the edge clarity of a reference object is crucial for accurate identification and positioning. This value directly reflects the usability of this critical information. The blurrier the edge, the greater the recognition and positioning error. By comparing this value across different reference objects or areas within the same image, we can quickly identify the area with the worst clarity, providing a feature foundation for subsequent visibility analysis.

[0050] In general, different reference objects have unique local texture distributions, and the LBP coding map can effectively capture and characterize this unique texture fingerprint. At the same time, severe image blur or noise will destroy the originally clear texture structure, resulting in chaotic LBP coding map patterns or abnormal eigenvalue distribution, which can help reflect the degree of degradation of picture quality.

[0051] In general, severe image blur reduces local texture diversity, causing previously rich texture patterns to become homogenized, resulting in a decrease in this value. Strong noise, on the other hand, can introduce spurious texture patterns, potentially causing this value to rise abnormally. Therefore, this value indirectly reflects the impact of image quality degradation on texture information. Higher values ​​indicate a greater variety of texture patterns within an area and a more uneven distribution, resulting in visually richer, more complex, and more contrasting textures. Lower values ​​indicate a more uniform texture pattern and a more even distribution, resulting in visually flatter, more blurred, or more homogeneous textures.

[0052] In general, this quantitative value directly indicates the intensity of the inherent color of the reference object. For reference objects that rely on color for identification and classification, a value that is too low means that color-based judgment may be unreliable, and it is necessary to activate a backup strategy or issue a warning. At the same time, certain errors in the video acquisition or processing process may cause the overall image to appear gray or have color deviations, which will be reflected in the saturation value of the reference object area, indicating that the quality of the image color information has declined.

[0053] In general, converting the actual impact of environmental interference on the target recognition of the monitoring system into an intuitive numerical value with spatial distance meaning ensures that the monitoring solution will not make incorrect judgments beyond its actual effective recognition distance, which helps to dynamically judge specific areas and improve the reliability of monitoring results.

[0054] In an example of the present invention, when performing spatiotemporal feature extraction on the degradation value sequence to obtain the steady-state feature vector of the reference object region, the method includes: Taking the mean of the degradation value sequence as the mean subsequence of the degradation value sequence; The standard deviation of the features in the degradation value sequence under the same sliding window is used as the characteristic fluctuation quantum sequence of the degradation value sequence; The mean subsequence and the characteristic fluctuation subsequence are arranged in time order to obtain a steady-state characteristic vector of the reference object region.

[0055] Specifically, the average size of all degradation values ​​representing the picture quality status in the entire time period is obtained, and this single average value is regarded as a subsequence with only one element, which is used to represent the overall level of the original sequence.

[0056] Specifically, for each window of the degradation value sequence, the severity of the fluctuation of the value in the window is obtained, and the fluctuation amplitude values ​​of each window are arranged in chronological order to form a record line that records the stability changes of the original degradation value sequence in a local time period.

[0057] Specifically, the two core aspects reflecting the picture quality status are brought together into the same time series record, ensuring that the status information of these two dimensions is aligned and correlated at each corresponding time point, generating a time series record that covers the entire video and describes the average degradation value and fluctuation value of the picture quality of the reference area at each moment.

[0058] Furthermore, the core benchmark indicators used to measure the average state of picture quality provide a highly summarized value for the entire visibility analysis solution, representing the overall average level of picture quality degradation in the reference area throughout the entire surveillance video.

[0059] Furthermore, it provides core data for subsequent visibility analysis of the stable evolution of the degree of picture quality degradation during the monitoring process, helping subsequent steps to quickly locate the specific video time period where the picture quality degradation value undergoes drastic and abnormal fluctuations.

[0060] Furthermore, through state pattern recognition, state anomaly detection, and state evolution prediction, integrated input data is provided for the subsequent comprehensive analysis of the complete state evolution of the reference area image quality during the entire surveillance video process, so that when analyzing the image quality state, the relationship and influence between its average level and volatility can be considered at the same time.

[0061] In general, the average value directly reflects the overall impact of environmental interference factors on image quality throughout the entire monitoring period. It serves as a core comprehensive indicator for measuring the overall image quality stability and availability of the monitoring system in a specific time period and environment. A higher average value indicates more severe overall degradation and lower system availability.

[0062] Generally speaking, a high value indicates that the degree of image quality degradation within that window fluctuates significantly, resulting in an unstable and unpredictable state; a low value indicates that the degradation within that window varies gradually, is stable, and is predictable. Significant peaks in the fluctuation quantum sequence often correspond precisely to periods of sudden or intermittent interference events. Identifying periods of significant image quality fluctuations helps focus maintenance resources and response measures during the most unstable periods, improving efficiency.

[0063] In general, the combination of two-dimensional information is no longer a single-dimensional numerical change, but a clear depiction of the movement trajectory of the state in the two-dimensional space of level and stability. Accurate identification of state patterns makes it easier to infer the causes of state changes. Based on the understanding of the complete state evolution pattern rather than just the current value, it can more accurately predict future state trends and make more reasonable operation and maintenance decisions accordingly.

[0064] In an example of the present invention, the step of arranging the mean subsequence and the characteristic fluctuation subsequence in chronological order to obtain a steady-state characteristic vector includes: The mean subsequence and the characteristic fluctuation subsequence are time-aligned and stored to obtain a time series label data set of degradation characteristics of the reference object area; Performing spatiotemporal feature extraction on the time series label dataset to obtain a preliminary feature vector of the degradation feature of the reference object region; The preliminary feature vectors are sequentially fused to obtain a steady-state feature vector of the reference object region.

[0065] Specifically, the two values ​​representing the average degradation value and the fluctuation amplitude value of the picture quality are ensured to be strictly corresponding at each specific time point, and a complete record containing two-dimensional status information is generated for each segment on the monitoring video timeline, which is convenient for subsequent search, analysis and use.

[0066] Specifically, from the time-arranged status records, we find the potential patterns of how the picture quality status changes over time, analyze the relationship between the status records of adjacent time periods, and transform and package the identified time evolution patterns and spatial correlation clues into characteristic values ​​that condense the core key characteristics of the state evolution.

[0067] Specifically, multiple features reflecting different aspects of state evolution are collected and integrated to generate a compact feature set that can comprehensively summarize the essential characteristics of the state evolution of the reference area, focusing on the long-term stable evolution pattern of the state and weakening minor instantaneous details.

[0068] Furthermore, a two-dimensional historical record of the image quality status of the reference area covering the entire plan is created for the entire plan. The application scenarios based on the historical status data are marked, retrieved, associated or predicted, and the structured, time-stamped label data is provided. It is the core input data source for any subsequent complex status analysis, pattern mining, model training or effect backtracking involving the time dimension.

[0069] Furthermore, the original complex time-series state data is refined into a set of features that can better reflect the core characteristics of state changes, thereby reducing the dimension and increasing the efficiency of subsequent complex analysis. At the same time, it provides the core input data for subsequent intelligent analysis such as automatic state pattern recognition, state evolution prediction, and state anomaly detection.

[0070] Furthermore, the steady-state feature vector is the highest-level input data for the subsequent complex state understanding, classification, comparison and decision-making of the scheme, generating a final and most representative feature set for the entire scheme about the core pattern of the long-term evolution of the picture quality state of the reference area.

[0071] In general, the system automatically discovers typical patterns of how picture quality status changes over time, correlates and analyzes changes in picture quality status with externally recorded events on a timeline, identifies causal relationships, and forms a reference area status knowledge base that can be preserved and reused for a long time. This provides data support for understanding the impact of different environmental conditions, different equipment statuses, and different time periods on picture quality, and continuously improves the intelligence level and operation and maintenance efficiency of the solution.

[0072] In general, based on feature vectors, the algorithm can automatically classify video clips into "stable and good period", "gradual deterioration period", "sudden interference and turbulence period", etc. By using the historical evolution pattern reflected by the feature vector, the model can predict the possible state of picture quality in the short or medium term in the future. When the feature vector corresponding to the newly generated state data deviates significantly from the historical pattern, it can automatically trigger an alarm. The extracted spatiotemporal features help to conduct more accurate time correlation analysis with external events and infer the root cause of the state change.

[0073] In general, a concentrated identity is generated for each reference area that reflects the core habits and patterns of how its image quality status changes over a long period of time. With this unified and refined feature set, the long-term state stability patterns of different monitoring points can be directly compared, problem points or excellent points can be identified, and environmental changes can be evaluated.

[0074] In an example of the present invention, the steady-state feature vector is subjected to a multivariate linear regression calculation to obtain the dynamic visibility index of the target road monitoring video, including: Performing parameter matrix training on the steady-state eigenvector to obtain a regression coefficient vector of the reference object region; Performing weighted summation on the regression coefficient vector to obtain a visibility fusion base value of the reference object area; An activation conversion is performed on the visibility fusion base value to obtain a dynamic visibility index of the reference object area.

[0075] Specifically, the specific contribution degree and influence mode of each core state characteristic in the steady-state eigenvector to the target effect of actual concern are determined, and a set of mathematical rules for predicting the target effect value based on the steady-state eigenvector is established.

[0076] Specifically, the independent impact of various state characteristics on the final visibility are integrated according to their respective weight ratios to calculate a single value that no longer reflects the state of the picture quality itself but directly represents the actual available visual perception ability in that state, providing the core quantitative basis for the final decision.

[0077] Specifically, the original numerical distribution of the fusion base value is changed to make it fall within a range that is more intuitive or easier to use and to enhance the discrimination near the key threshold. The fusion base value is nonlinearly mapped so that the generated index is more in line with the real law of system performance changes.

[0078] Furthermore, the generated regression coefficient vector provides the core calculation basis for subsequent intelligent prediction of the entire plan, translating the abstract multi-dimensional picture quality state evolution pattern into specific and operational effect indicators, making subsequent state-based decision-making more scientific and data-supported.

[0079] Furthermore, the comprehensive visibility performance of the reference area in the current or predicted state is quantified, providing an objective and comparable single measurement standard for the visibility performance of different reference areas and at different time points. It is the most critical input threshold for the subsequent automatic triggering of alarms, adjustment of system parameters, allocation of maintenance resources or activation of emergency plans.

[0080] Furthermore, the fusion base value is converted into an index value that is easy to compare within a standard range, which facilitates the connection between systems and the comparison of historical data. Through nonlinear conversion, the difference in numerical changes in low-visibility areas is amplified, making the system more sensitive to slight changes in threat monitoring effectiveness.

[0081] In general, the positive weight of each state characteristic in maintaining a positive effect and the absolute value of the negative weight that harms the effect are clearly defined. This directly links abstract state patterns with specific business value.

[0082] In general, the Baseline Value is used to measure the effectiveness of the entire monitoring solution at a specific location and time period. The higher the Baseline Value, the better the solution's performance at that location and time period. The Baseline Value can automatically adjust the sensitivity or confidence threshold of the video analysis algorithm based on the Baseline Value. When the Baseline Value drops to a dangerous level due to sudden interference, the pre-set emergency response process is automatically initiated.

[0083] In general, as a standard output interface, it is convenient to integrate with other systems and transmit a unified visibility risk signal. In low visibility ranges and high-risk areas, a slight decrease in the fusion base value will be converted into a more significant exponential decrease, triggering alarms more sensitively and seizing disposal time. In high visibility ranges, that is, safe areas, the index changes relatively smoothly, avoiding unnecessary frequent alarms.

[0084] In an example of the present invention, the parameter matrix training of the steady-state feature vector to obtain the regression coefficient vector of the reference object region includes: splicing the steady-state feature vectors based on the types of feature labels to obtain an augmented matrix of the reference object region; extracting the covariance of the augmented matrix, and constructing a multidimensional covariance matrix of the reference object region based on the covariance; The inverted covariance matrix is ​​multiplied element by element by the covariance matrix to obtain a regression coefficient vector of the reference object region.

[0085] Specifically, the eigenvalues ​​in the originally flat steady-state eigenvector are classified according to the state attributes represented by the features, and a dimension of feature category is explicitly added to the data structure, so that each eigenvalue not only has its own numerical value, but also has a clear semantic category affiliation.

[0086] Specifically, the changing trend of state features of any two different categories in the augmented matrix in historical data is accurately measured to capture the overall fluctuation pattern, and a mathematical matrix is ​​constructed that centrally stores the synergistic change intensity and direction information between all feature categories.

[0087] Specifically, the information about how features change together and how they change independently after excluding the influence of other features is mathematically fused to obtain a set of regression weight coefficients for the net contribution of each feature to the target effect under the premise of considering the interdependence of all other features.

[0088] Furthermore, a complete core state feature record library is created for each reference area, which is semantically classified according to state attributes, providing a core data structure for subsequent semantic category-based feature analysis, comparison, retrieval and model input, so that different subsequent analysis modules can accurately extract their target specific category feature sets.

[0089] Furthermore, a key data structure describing the inherent coordinated change rules among various core state features of the reference area is generated for the entire plan, which is the basis for the subsequent multi-feature joint analysis, state pattern anomaly detection, data generation simulation and advanced probabilistic inference of the plan.

[0090] Furthermore, the generated regression coefficient vector clarifies the complex interdependencies between state features, avoiding prediction bias or model instability caused by ignoring these dependencies. It is the core mathematical basis for the entire scheme to conduct high-precision, robust and intelligent predictions.

[0091] In general, the augmented matrix introduces the dimension of feature category labels, upgrading the steady-state feature vector describing the core pattern of the reference object area state from a one-dimensional numerical list to a feature file with a clear semantic structure, solving the problems of poor feature interpretability and chaotic management, and providing a strong basic data structure support for refined analysis, interpretable modeling, modular system design, and efficient knowledge management, laying the foundation for solution evolution.

[0092] In general, the multidimensional covariance matrix solves the problem of viewing a single state feature in isolation, and represents the complex network relationship of the coordinated evolution of various core state features in the reference area. It is the core data structure for understanding the internal coupling mechanism of the system, realizing accurate anomaly detection, simulating real state evolution, evaluating risk propagation paths, and building robust advanced models.

[0093] In summary, the regression coefficient vector, derived by element-by-element multiplication of the covariance matrix and its inverse, is the mathematical core of dependency-aware, high-precision, and robust prediction. It addresses the interference caused by feature multicollinearity, reveals the unique net contribution of features to the target effect, and improves the accuracy, stability, and interpretability of the prediction model. This makes intelligent predictions and decisions based on this coefficient more reliable in complex and changing traffic monitoring environments, and is a key technical support for the solution's implementation of advanced intelligent operations and maintenance.

[0094] In an example of the present invention, the step of performing warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain the warning level of the road visibility includes: Filtering invalid values ​​of the real-time visibility index to obtain a valid visibility sampling set in the target road monitoring video; Performing warning matching on the effective visibility sampling set to obtain a warning level of the target road monitoring video; Perform risk assessment on the warning level to obtain signal control of the target road monitoring video.

[0095] Specifically, it detects and marks unreliable values ​​in the real-time visibility index stream that clearly exceed a reasonable range, undergo drastic changes compared with previous and subsequent values, or are generated under specific conditions. It removes the identified suspicious values ​​from the data stream to prevent them from entering subsequent processing links, and generates a time series record that only contains verified valid index values ​​that reflect the actual visibility status.

[0096] Specifically, based on preset rules, it is determined which predefined risk level range the current effective visibility index value belongs to, and a standardized risk level identification is produced for the current monitoring scene, which clearly indicates the safety level of the current road monitoring environment in terms of visual perception capabilities.

[0097] Specifically, the warning level is converted into operational road traffic management instructions. Based on the preset rule base, the optimal signal control plan that matches the current warning level is formulated or selected to generate a control instruction set that can directly drive downstream traffic signal control equipment.

[0098] Furthermore, it prevents invalid or abnormal data from interfering with the normal operation of downstream modules, avoids causing misjudgments, false alarms or unstable states, and ensures that the final output visibility status reports, alarms and decision-making recommendations are based on valid data.

[0099] Furthermore, it generates the most direct action trigger signal for the entire plan, directly quantifying the visibility risk level within the effective coverage area of ​​the current road surveillance video. This serves as the core decision-making input for the plan to automatically initiate different levels of emergency plans, issue early warning information, adjust traffic control measures, or notify relevant personnel to intervene.

[0100] Furthermore, the early analysis and early warning results are transformed into control actions to ensure safety, completing the automated closed loop from monitoring to early warning and final regulation, which serves as the final output link of the entire intelligent monitoring solution.

[0101] In general, invalid value filtering removes invalid and interfering information from the data stream in real time, effectively sampling and reflecting real-world visibility conditions. This ensures the purity and credibility of the data relied upon for all subsequent analysis, decision-making, and presentation. It is one of the most fundamental and important guarantees for reliable system output, accurate decision-making, and efficient operations and maintenance.

[0102] In general, warning matching is the key hub that converts the single data stream of the effective visibility sampling set into instructions that drive the efficient operation of the traffic safety assurance system. Through standardized risk level classification, it converts complex data into clear instructions, ensuring that the system can automatically initiate defense and response measures in low visibility.

[0103] In general, graded control is automatically implemented according to the warning level, and fine-tuned in combination with real-time traffic flow data to maximize process efficiency. At the same time, it ensures that the same warning level triggers consistent and optimal control strategies at different times and on different road sections, eliminates differences and delays in human judgment, ensures standardized and reliable responses, and enables the road network to maintain basic and safe operations in harsh environments with low visibility.

[0104] Compared with the prior art, the present invention has the following beneficial effects: 1. This solution breaks through the limitations of traditional single-dimensional perception, characterizing images into multiple indicators and constructing a four-dimensional degradation feature collaborative quantification system. By synchronously integrating four complementary indicators: edge clarity, texture contrast, color saturation, and recognizable distance, it reduces the misjudgment rate due to local feature degradation or single feature failure, avoids single-frame errors, achieves cross-modal feature cross-validation, and improves the stability of the monitoring solution.

[0105] 2. This solution uses a covariance-driven dynamic regression architecture to analyze nonlinear correlations between features, dynamically adjust the weights of four feature categories, fit the linear relationships between degenerate features, and automatically remove highly linearly correlated features, avoiding prediction bias caused by ignoring feature correlations and maximizing feature accuracy.

[0106] like Figure 2 , which is a functional module diagram of road low visibility monitoring based on video analysis provided by one embodiment of the present invention.

[0107] The video analysis-based road low visibility monitoring system 100 described in the present invention can be installed in an electronic device. Depending on the functionality implemented, the video analysis-based road low visibility monitoring system 100 can include a reference object extraction module 101, a degradation value sequence generation module 102, a steady-state feature vector generation module 103, a steady-state feature vector generation module 104, and a warning level generation module 105. A module, also referred to as a unit, is a series of computer program segments that can be executed by an electronic device processor and perform a fixed function. These modules are stored in the electronic device's memory.

[0108] In this embodiment, the functions of each module / unit are as follows: The reference object extraction module 101 is used to dynamically identify the inherent reference objects of the image features in the target road monitoring video to obtain a continuous reference object area in the target road monitoring video; The degradation value sequence generating module 102 is configured to quantify the degradation features of the reference object region to obtain a degradation value sequence of multiple features of the reference object region; The steady-state feature vector generation module 103 is configured to extract spatiotemporal features from the degradation value sequence to obtain a steady-state feature vector of the reference object region; The dynamic visibility index module 104 is configured to perform a multivariate linear regression calculation on the steady-state feature vector to obtain a dynamic visibility index of the target road monitoring video; The warning level generating module 105 is configured to perform warning grading on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain the warning level of the road visibility.

[0109] In the several embodiments provided by the present invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the module division is merely a logical function division, and other division methods may be used in actual implementation.

[0110] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of the modules may be selected to achieve the purpose of the solution of this embodiment according to actual needs.

[0111] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or hardware plus software functional modules.

[0112] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0113] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring low visibility on roads based on video analysis, characterized in that: The method comprises: S1. Dynamically identifying inherent reference objects of image features in a target road surveillance video to obtain continuous reference object regions in the target road surveillance video; S2. quantifying degradation features of the reference object region to obtain a degradation value sequence of multiple features of the reference object region; S3. Extracting spatiotemporal features from the degradation value sequence to obtain a steady-state feature vector of the reference object region; S4, performing a multivariate linear regression calculation on the steady-state eigenvector to obtain a dynamic visibility index of the target road surveillance video; S5. Performing warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain a warning level of the road visibility.

2. The method for monitoring low visibility on roads based on video analysis according to claim 1, wherein: The method of dynamically identifying the inherent reference objects of the image features in the target road monitoring video to obtain a continuous reference object area in the target road monitoring video includes: Detecting the target road surveillance video with the reference object as the target to obtain an initial candidate frame set of the target road surveillance video; Performing multi-frame trajectory association on the initial candidate frame set to obtain motion trajectory data of the reference object; The motion trajectory data of the reference object is subjected to streaming verification to obtain a continuous reference object area in the target road monitoring video.

3. The method for monitoring low visibility on roads based on video analysis according to claim 1, wherein: When the degradation feature of the reference object region is quantified to obtain a degradation value sequence of multiple features of the reference object region, the method includes: Performing single-frame degradation feature extraction on the reference object region to obtain a single-frame feature vector of the reference object region; Quantizing and iterating the single-frame feature vector to obtain a frame sequence feature set of the reference object region; The frame sequence feature set is temporally integrated to obtain a degradation value sequence of multiple features in the reference object area.

4. The method for monitoring low visibility on roads based on video analysis according to claim 3, wherein: The single-frame feature vector includes: Using the gradient amplitude of the grayscale image in the reference object area as the edge sharpness quantization value of the target reference object area; Performing LBP feature extraction on the target area of ​​the reference object to obtain an LBP coding map of the target reference object area; Performing histogram variance statistics on the LBP coding image to obtain a texture contrast quantization value of the target reference object area; extracting a channel mean value of the reference object area after format conversion, and quantizing the channel mean value into a color saturation quantization value; The degradation rate of the visual features of the current frame and the benchmark features in the reference object area is used as the target recognizable distance estimation value of the reference object area.

5. The method for monitoring low visibility on roads based on video analysis according to claim 1, wherein: When performing spatiotemporal feature extraction on the degradation value sequence to obtain a steady-state feature vector of the reference object region, the method includes: Taking the mean of the degradation value sequence as the mean subsequence of the degradation value sequence; The standard deviation of the features in the degradation value sequence under the same sliding window is used as the characteristic fluctuation quantum sequence of the degradation value sequence; The mean subsequence and the characteristic fluctuation subsequence are arranged in time order to obtain a steady-state characteristic vector of the reference object region.

6. The method for monitoring low visibility on roads based on video analysis according to claim 5, wherein: When the mean subsequence and the characteristic fluctuation subsequence are arranged in time order to obtain a steady-state characteristic vector, the method includes: The mean subsequence and the characteristic fluctuation subsequence are time-aligned and stored to obtain a time series label data set of degradation characteristics of the reference object area; Performing spatiotemporal feature extraction on the time series label dataset to obtain a preliminary feature vector of the degradation feature of the reference object region; The preliminary feature vectors are sequentially fused to obtain a steady-state feature vector of the reference object region.

7. The method for monitoring low visibility on roads based on video analysis according to claim 1, wherein: The steady-state feature vector is subjected to a multivariate linear regression calculation to obtain the dynamic visibility index of the target road monitoring video, including: Performing parameter matrix training on the steady-state eigenvector to obtain a regression coefficient vector of the reference object region; Performing weighted summation on the regression coefficient vector to obtain a visibility fusion base value of the reference object area; An activation conversion is performed on the visibility fusion base value to obtain a dynamic visibility index of the reference object area.

8. The method for monitoring low visibility on roads based on video analysis according to claim 7, wherein: When the dynamic visibility index generation module performs parameter matrix training on the steady-state feature vector to obtain the regression coefficient vector of the reference object area, the module includes: splicing the steady-state feature vectors based on the types of feature labels to obtain an augmented matrix of the reference object region; extracting the covariance of the augmented matrix, and constructing a multidimensional covariance matrix of the reference object region based on the covariance; The inverted covariance matrix is ​​multiplied element by element by the covariance matrix to obtain a regression coefficient vector of the reference object region.

9. The method for monitoring low visibility on roads based on video analysis according to claim 1, wherein: When the warning level generation module performs warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain the warning level of the road visibility, the module includes: Filtering invalid values ​​of the real-time visibility index to obtain a valid visibility sampling set in the target road monitoring video; Performing warning matching on the effective visibility sampling set to obtain a warning level of the target road monitoring video; Perform risk assessment on the warning level to obtain signal control of the target road monitoring video.

10. A road low visibility monitoring system based on video analysis, characterized in that: The system comprises: S1. Dynamically identifying inherent reference objects of image features in a target road surveillance video to obtain continuous reference object regions in the target road surveillance video; S2. quantifying degradation features of the reference object region to obtain a degradation value sequence of multiple features of the reference object region; S3. Extracting spatiotemporal features from the degradation value sequence to obtain a steady-state feature vector of the reference object region; S4, performing a multivariate linear regression calculation on the steady-state eigenvector to obtain a dynamic visibility index of the target road surveillance video; S5. Performing warning classification on the road visibility in the target road monitoring video based on the dynamic visibility index to obtain a warning level of the road visibility.

Images