Dangerous waste treatment monitoring and early warning method based on video analysis

Abstract

The invention belongs to the field of video recognition, and particularly relates to a video analysis-based hazardous waste treatment monitoring and early warning method, which comprises the following steps of: carrying out Gaussian mixture background modeling on a video to obtain a predicted gray value of a target position in a current frame; acquiring scene local stability of the target position in the current frame; determining the dynamic anomaly degree of the target position in the current frame, and marking a candidate abnormal region in the current frame according to the dynamic anomaly degree; obtaining the motion vector of each pixel point in the current frame, and determining the motion consistency of the candidate abnormal region; according to the motion consistency and the average dynamic anomaly degree of the pixel points in the candidate abnormal region, determining the anomaly confidence degree of the candidate abnormal region; and carrying out dangerous waste risk assessment according to the abnormal confidence degree. The technical problem that video analysis of hazardous waste treatment is prone to dynamic interference generated by normal production is solved, and the accuracy and reliability of hazardous waste treatment monitoring and early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of video recognition. More specifically, this invention relates to a method for monitoring and early warning of hazardous waste treatment based on video analysis. Background Technology

[0002] Hazardous waste treatment is related to environmental and personnel safety, so video monitoring systems are usually deployed at the treatment site to detect abnormalities such as leaks and combustion in real time. Existing technologies mainly use video analysis methods based on image features, such as color models, texture features or morphological means to identify anomalies in order to provide monitoring and early warning as soon as possible. These methods are effective in general industrial scenarios, but their performance is limited when faced with the complex working conditions of hazardous waste treatment.

[0003] Existing video analysis methods often rely on fixed thresholds or preset rules to judge the lighting, color, or texture features of images. This approach is easily affected by normal production activities in the field; for example, dynamic changes such as water vapor, dust, or light reflection can be misjudged as risk events. Furthermore, when the characteristics of abnormal events are weak or atypical, these methods may fail to accurately identify them, leading to missed detections. Similarly, when a real leak or fire is in its early stages, its visual characteristics may be weak or atypical, and these methods may also fail to accurately identify it because the signal strength does not reach the preset level, resulting in missed detections.

[0004] The root cause of this problem is that traditional algorithms lack the ability to adapt to dynamic environments. Hazardous waste treatment sites are not stable background environments, but constantly changing dynamic scenarios. Algorithms based on fixed parameters and prior rules cannot distinguish between normal unstable phenomena and real anomalies. As a result, false alarms are frequent when there is a lot of environmental noise, while false alarms are easy to be missed when real risks occur but the signals are not obvious, making it difficult to support safety requirements. Summary of the Invention

[0005] To address the technical problem that video analysis of hazardous waste treatment is easily affected by dynamic interference from normal production, this invention provides a video analysis-based monitoring and early warning method for hazardous waste treatment, comprising: A Gaussian mixture (GJ) background model is performed on videos of key monitoring points at hazardous waste treatment facilities to obtain a GJ model for each location in the video. Taking any pixel location in the video as the target location, the mean parameter of the sub-Gaussian model that best matches the grayscale value of the target location in the previous frame is obtained and used as the predicted grayscale value of the target location in the current frame. The grayscale values ​​of the target location in the current frame and several previous frames are used to construct a grayscale sequence for the target location. Based on the mean and standard deviation of the grayscale sequence, and the mean and standard deviation of the differences between two adjacent grayscale values ​​in the grayscale sequence, the scene of the target location in the current frame is determined. Local stability; based on the difference between the predicted and actual grayscale values ​​of the target location in the current frame, and the local stability of the scene, the dynamic anomaly degree of the target location in the current frame is determined, and candidate anomaly regions are marked in the current frame according to the magnitude of the dynamic anomaly degree; the motion vector of each pixel in the current frame is obtained using optical flow, and the motion consistency of the candidate anomaly regions is determined based on the proportion of pixels in the candidate anomaly regions whose vertical components of the motion vectors are greater than 0; the anomaly confidence of the candidate anomaly regions is determined based on the motion consistency and the average dynamic anomaly degree of the pixels in the candidate anomaly regions; and a hazardous waste risk assessment is performed based on the magnitude of the anomaly confidence.

[0006] This invention improves the accuracy and reliability of video monitoring and early warning in complex and dynamic environments such as hazardous waste treatment. First, it calculates a dynamic anomaly score that can suppress normal production disturbances such as water vapor by using a Gaussian mixture model and scene local stability analysis. Second, it uses motion consistency to assess whether the motion pattern of candidate anomaly areas conforms to the physical law of smoke rising upwards. By combining the anomaly score with the smoke rising model, this invention can effectively distinguish between real risk events and physically meaningless artifacts, overcoming the limitations of traditional fixed threshold algorithms. It can improve the accuracy and reliability of monitoring real hazardous events while reducing the false alarm rate.

[0007] Preferably, the step of obtaining the mean parameter of the sub-Gaussian model in the Gaussian mixture model of the target location that best matches the grayscale value of the target location in the previous frame, and using it as the predicted grayscale value of the target location in the current frame, includes: if the grayscale value of the target location in the previous frame has the smallest difference from the mean of a certain sub-Gaussian model, and is between the mean-centered value of the sub-Gaussian model with the smallest mean difference... A match is considered successful if the target position is within a confidence interval of one-times the standard deviation, and the mean of the matched sub-Gaussian model is taken as the predicted gray value of the target position in the current frame. If the gray value of the target position in the previous frame does not match any of the sub-Gaussian models, the predicted gray value of the nearest temporally matched pixel at the same position is selected as the predicted gray value of the target position in the current frame.

[0008] This invention matches the grayscale value of the previous frame with multiple sub-models in a Gaussian mixture model and selects the mean of the best-matching sub-model as the predicted value for the current frame. The mean represents the most stable and core grayscale value of the background state. Using the mean as the predicted value can effectively filter out random noise that may exist in the previous frame image, thus providing a more stable and accurate background benchmark than directly using the grayscale value of the previous frame. When all sub-Gaussian models do not match, the most recently successfully matched predicted value at that position is selected as the current predicted value. This ensures that even if the pixel does not belong to any known background model, an effective predicted value can still be provided, avoiding the problem of subsequent calculation interruption due to matching failure and ensuring the continuity of monitoring and analysis.

[0009] Preferably, the expression for the local stability of the scene is: ;in, It is the frame number of the current frame. Target location Local scene stability in the current frame, It is a preset reference fluctuation benchmark. and These represent the mean and standard deviation of the grayscale values ​​at the target location within the window ending with the current frame, respectively. and It is the mean and standard deviation of the absolute values ​​of grayscale changes between adjacent frames at the target location within a short time window, with the current frame as the last frame. It is a hyperparameter to prevent the denominator from being 0.

[0010] This invention assesses the local stability of a scene by calculating the mean and standard deviation of the grayscale value of the target location within a time window, as well as the mean and standard deviation of the absolute values ​​of grayscale changes between adjacent frames. In brighter areas, a large grayscale fluctuation may be considered normal noise, while in darker areas, the same fluctuation may be a significant event. This invention can also distinguish between smooth and uniform changes and sudden, irregular disturbances, enabling the subsequent calculation of dynamic anomalies to effectively differentiate between grayscale instability caused by normal production activities and instability caused by real anomalies. This reduces false alarms and improves the accuracy of monitoring and early warning in complex dynamic environments.

[0011] Preferably, the expression for the dynamic anomaly degree is: ;in, It is the frame number of the current frame. Target location In the current frame's dynamic anomaly degree, Target location The grayscale value of the current frame; Target location The predicted grayscale value of the current frame, Target location Local scene stability in the current frame, These are parameters that adjust the impact on the local stability of the scene. It is the absolute value symbol.

[0012] This invention calculates dynamic anomaly degree. Stable regions have higher weights, making the system highly sensitive to weak real anomaly signals. In contrast, unstable regions with normal dynamic interference such as water vapor and dust have lower weights, thereby reducing false alarms of normal fluctuations and distinguishing between real risks and normal environmental interference.

[0013] Preferably, marking candidate abnormal regions in the current frame based on the magnitude of dynamic abnormality includes: marking the target location in the current frame in response to the target location having a dynamic abnormality greater than a preset abnormality threshold; and aggregating spatially adjacent marked points into candidate abnormal regions using a connected component algorithm.

[0014] Preferably, determining the motion consistency of the candidate abnormal region includes: using optical flow to obtain the motion vector of each pixel in the current frame and previous consecutive video frames; calculating the ratio of the number of motion vectors in the candidate abnormal region that satisfy the condition that the vertical component is positive to the total number of motion vectors in the candidate abnormal region, as the motion consistency of the candidate abnormal region.

[0015] Preferably, the expression for the anomaly confidence level is: ;in, It is the frame number of the current frame. Indicates the first Candidate abnormal regions The anomaly confidence level in the current frame, Indicates the first Candidate abnormal regions The number of pixels, Indicates candidate anomaly regions Pixels within, It is a pixel. In the current frame's dynamic anomaly degree, It is a candidate anomaly region Motion consistency in the current frame.

[0016] The anomaly confidence level of this invention is obtained by multiplying the average dynamic anomaly level of candidate anomaly regions by the motion consistency. Only when a candidate anomaly region simultaneously meets the two conditions of being anomaly and conforming to the real smoke model will its anomaly confidence level be high, ensuring that the early warning system only targets real and physically significant risk events, thereby improving the accuracy and reliability of monitoring.

[0017] Preferably, the hazardous waste risk assessment based on the anomaly confidence level includes: presetting multi-level risk thresholds: collecting a large number of anomaly confidence levels under normal operating conditions and classifying the data into thresholds: 95th percentile, 99th percentile, and 99.9th percentile; if the anomaly confidence level of a candidate anomaly region is less than or equal to the 95th percentile, the hazardous waste treatment is risk-free; if the anomaly confidence level of a candidate anomaly region is greater than the 95th percentile and less than the 99th percentile, the hazardous waste treatment is of low risk; if the anomaly confidence level of a candidate anomaly region is greater than or equal to the 99th percentile and less than the 99.9th percentile, the hazardous waste treatment is of medium risk; if the anomaly confidence level of a candidate anomaly region is greater than or equal to the 99.9th percentile, the hazardous waste treatment is of high risk.

[0018] Preferably, the process of performing mixed Gaussian background modeling on the videos of key monitoring points of hazardous waste treatment facilities includes setting the number of sub-Gaussian models to 5.

[0019] Preferably, the step of constructing a grayscale sequence of the target position from the grayscale values ​​of the target position in the current frame and the previous several frames includes: constructing a grayscale sequence of the target position from the grayscale values ​​of the target position in the current frame and the previous 50 frames.

[0020] The beneficial effects of this invention are as follows: This invention constructs a scene local stability index and a dynamic anomaly degree, enabling the sensitivity of anomaly detection to be adaptively adjusted in real time according to changes in environmental factors such as ambient light and water vapor. This solves the problem of high false alarm rate in complex dynamic environments by traditional fixed threshold algorithms, and has strong environmental adaptability and robustness.

[0021] This invention introduces a physical model library and calculates anomaly confidence levels. Based on the confidence levels, it distinguishes whether the anomalies in candidate anomaly regions are real or unreal, providing precise guidance for subsequent emergency response.

[0022] This invention can effectively distinguish between real physical events and artifacts with no physical meaning, such as changes in lighting, accurately distinguish between real risks and normal environmental interference, and reduce the false alarm rate caused by random noise and non-critical events, as well as the missed detection rate of anomalies with inconspicuous features. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the video analysis-based hazardous waste treatment monitoring and early warning method of the present invention; Figure 2 This is a schematic diagram illustrating the flowchart of S4 in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] This invention discloses a video analytics-based method for monitoring and early warning of hazardous waste treatment, with reference to... Figure 1 This includes steps S1 to S5: S1. Perform Gaussian mixture background modeling on the video of key monitoring points of hazardous waste treatment facilities, obtain the mean parameter of the sub-Gaussian model that best matches the gray value of the target location in the previous frame in the Gaussian mixture model of the target location, and use it as the predicted gray value of the target location in the current frame.

[0027] Specifically, visible light cameras are deployed at key monitoring points in hazardous waste treatment facilities, including the feed inlet, the outer wall of the incinerator, the perimeter of the chemical reactor, and the ground of the waste storage area.

[0028] A visible light camera is used to continuously acquire video stream data, and the video stream is decoded into a sequence of image frames indexed by frame number.

[0029] Each frame of the image is preprocessed, including image denoising and contrast enhancement, to obtain clear image data for subsequent analysis. In this embodiment, Gaussian denoising algorithm is used for image denoising and histogram equalization algorithm is used for contrast enhancement. Implementers may also use other denoising and contrast enhancement algorithms, such as bilateral filtering algorithm and gamma transform algorithm.

[0030] Gaussian mixture background modeling is performed on the video, with the number of sub-Gaussian models set to 5, resulting in a Gaussian mixture model for each pixel location in the video. Any pixel location is taken as the target location. If the grayscale value of the target location in the previous frame has the smallest difference from the mean of a certain sub-Gaussian model, and is centered at the mean of the sub-Gaussian model with the smallest mean difference... A match is considered successful if the confidence interval is within two standard deviations. The mean of the matched sub-Gaussian models is taken as the predicted gray value of the target position in the current frame. If the gray value of the target position in the previous frame does not match any of the sub-Gaussian models, the predicted gray value of the nearest successfully matched pixel at the same position is selected as the predicted gray value of the target position in the current frame. This embodiment uses 5 sub-Gaussian models and a confidence interval of two standard deviations for illustration. Implementers can also adjust the number of sub-Gaussian models and the standard deviation multiple of the confidence interval according to actual needs.

[0031] S2. Construct a grayscale sequence of the target position from the grayscale values ​​of the target position in the current frame and several previous frames. Determine the local stability of the target position in the current frame based on the mean and standard deviation of the grayscale sequence and the mean and standard deviation of the difference between two adjacent grayscale values ​​in the grayscale sequence.

[0032] It should be noted that when there are no abnormalities in the treatment of hazardous waste, the gray value of the target location is stably distributed within the corresponding time window. When there is a possibility of abnormalities in the treatment of hazardous waste, the gray value of the target location within the corresponding time window will fluctuate. Therefore, this invention measures the stability of the gray value of the target location within the corresponding time window.

[0033] Specifically, local scene stability is constructed for the target location in the current frame of the video image: ; in, It is the frame number of the current frame. Target location Local scene stability in the current frame, It is a preset reference fluctuation benchmark, an empirical value. To avoid division-by-zero errors in the extreme case of complete stillness, and as a normalization benchmark, the combined fluctuations within the brackets are mapped to the final stability score. and These represent the mean and standard deviation of the grayscale values ​​at the target location within a 50-frame window, with the current frame as the last frame. and It is the mean and standard deviation of the absolute values ​​of grayscale changes between adjacent frames at the target location within a short time window, with the current frame as the last frame. It is a hyperparameter to prevent the denominator from being zero; it is an empirical value. In this embodiment , The short window is 50 frames long. In other embodiments, implementers can adjust the reference fluctuation baseline, hyperparameter size, and short window length according to actual conditions.

[0034] in, This represents the relative change in grayscale values ​​at the target location. For example, for a bright area with an average grayscale value of 200, a fluctuation of 10 grayscale values ​​may just be normal noise; but for a dark area with an average grayscale value of 20, the same fluctuation of 10 grayscale values ​​may be a significant event, making the stability assessment adaptive to areas with different grayscale values. Used to measure the smoothness of change; if the scene changes uniformly and slowly, then the amount of grayscale change of the target position between frames will be very similar, with a standard deviation. If the scene changes unevenly and rapidly, the value of the grayscale change between frames will fluctuate greatly, and its standard deviation will also be small. It will be very large, causing this value to increase. As the target location The scene local stability in the current frame has a value range of (0,1).

[0035] The closer the value is to 1, the more stable the target position is within the time window; The closer the value is to 0, the more unstable the target position is within the time window.

[0036] S3. Based on the difference between the predicted grayscale value and the actual grayscale value of the target position in the current frame, and the local stability of the scene, determine the dynamic anomaly degree of the target position in the current frame. This is used to distinguish between stability, normal instability, and abnormal instability. For example, water vapor generated during hazardous waste treatment is normal instability, while smoke with abnormal color mixed in with the water vapor is abnormal instability.

[0037] Specifically, the dynamic anomaly degree satisfies the expression: ; in, It is the frame number of the current frame. Target location In the current frame's dynamic anomaly degree, Target location The grayscale value of the current frame; Target location The predicted grayscale value of the current frame, Target location Local scene stability in the current frame, These are parameters that adjust the impact on local stability of the scene; they are empirical values. , It is the absolute value symbol. This embodiment... In other embodiments, implementers may also make adjustments according to actual circumstances.

[0038] When a local area of ​​the scene is highly stable Approaching 1, enhancement operator Also tending to ,and Low, output dynamic anomaly degree Approaching 0; when local areas of the scene are unstable Approaching 0, When the stability of a local area in a scene approaches zero, instability is categorized into normal instability and abnormal instability. For example, water vapor generated during hazardous waste disposal is normal instability, while water vapor mixed with abnormally colored smoke is abnormal instability. Normal instability has been reclassified as normal through the learning process of the Gaussian mixture model. This difference would be very small. For abnormal instability, this new change does not conform to the normal water vapor model already identified by the GMM. The difference will be relatively large.

[0039] Furthermore, the dynamic anomaly degree of the output under normal unstable conditions is less than that under abnormal unstable conditions, because even though both belong to unstable states, The values ​​are similar, but The values ​​vary significantly, resulting in abnormal and unstable outputs. Larger than normal unstable output .

[0040] It should be further explained that when a weak real anomaly appears in an otherwise stable background, The change is small, but due to the occurrence of anomalies, its value will decrease slightly compared to the absolutely stable state. At this point, the enhancement operator... It will provide a magnification factor greater than 1, which will amplify the originally weak signal. The signal is amplified.

[0041] S4. Mark the candidate abnormal regions in the current frame according to the dynamic abnormality degree, obtain the motion vector of each pixel in the current frame using the optical flow method, determine the motion consistency of the candidate abnormal regions according to the vertical component of the motion vector in the candidate abnormal regions, and determine the abnormality confidence of the candidate abnormal regions according to the motion consistency and the average dynamic abnormality degree of the pixels in the candidate abnormal regions.

[0042] It should be noted that dynamic anomaly assessment primarily focuses on visual anomalies and cannot distinguish non-physical artifacts. For example, momentary noise on a camera sensor or small squares caused by encoding / decoding errors can lead to high local dynamic anomaly values. However, these noises and small squares are not relevant to hazardous waste handling anomalies. Furthermore, dynamic anomaly assessment lacks a realistic contextual change. For instance, the instantaneous switching of a light in the distance, while a real change in light and shadow, does not constitute an event requiring warning in a hazardous waste scenario. In hazardous waste handling, smoke leakage is an event requiring warning in this invention. Therefore, this invention uses a smoke rise model to assess the authenticity of potential anomaly signals, thereby reducing false alarms caused by random noise and non-critical events.

[0043] The flowchart for step S4 is shown below. Figure 2The process includes steps S401 to S404, specifically as follows: S401. Utilize dynamic anomaly degree to filter out candidate anomaly regions in the current frame of the video.

[0044] Specifically, pixels with a dynamic anomaly degree greater than a preset anomaly threshold are marked, and adjacent marked points are aggregated into independent candidate anomaly regions using a connected component algorithm. The empirical value of the anomaly threshold is 30, and implementers can set the anomaly threshold according to the actual situation.

[0045] S402. Use optical flow to obtain the motion vector of each pixel in the current frame of the video image.

[0046] Specifically, given consecutive video frames, such as the current frame and the previous frame, optical flow is used to calculate the motion direction and velocity of each pixel in the image between consecutive frames, forming a two-dimensional motion vector. This embodiment uses the Gunnar Farnebäck optical flow method. In other embodiments, implementers can choose an algorithm to obtain the motion vector of each pixel in the current frame of the video image, such as the Horn-Schunck optical flow method (HS).

[0047] It should be noted that optical flow can generate a dense motion vector field for the entire image, including all pixels within the candidate anomaly region. A candidate anomaly region outputs a motion field containing the motion vector information of all points within the candidate anomaly region.

[0048] S403. Calculate the ratio of the number of motion vectors with positive vertical components within the candidate anomaly region to the total number of motion vectors in the candidate anomaly region, and use this ratio as the motion consistency of the candidate anomaly region.

[0049] Smoke forms a continuous area with volume, and in an indoor environment without strong winds, its overall movement trend is upward due to buoyancy. In contrast, non-physical artifacts such as random noise may appear at a certain position in a frame and disappear in the next frame, and their movement is random, disordered, and discontinuous. Therefore, this invention obtains the motion consistency of candidate abnormal regions based on smoke characteristics.

[0050] Specifically, the total number of motion vectors contained in a candidate anomaly region is denoted as The total number of motion vectors with positive vertical components is denoted as . Then calculate the candidate anomaly region. Motion consistency in the current frame: ; in, It is the frame number of the current frame. It is a candidate anomaly region The total number of motion vectors with positive inner vertical components. It is a candidate anomaly region The larger the ratio of the number of motion vectors with positive vertical components in a candidate anomaly region to the total number of motion vectors in the candidate anomaly region, the more likely the candidate anomaly region is to be a real smoke rising model; conversely, the smaller the ratio, the more likely it is to be random noise.

[0051] S404. Obtain the anomaly confidence of the candidate anomaly region by utilizing the motion consistency of the candidate anomaly region and the average anomaly degree of the candidate anomaly region.

[0052] Specifically, the expression for the anomaly confidence level is: ; in, It is the frame number of the current frame. Indicates the first Candidate abnormal regions The anomaly confidence level in the current frame, Indicates the first Candidate abnormal regions The number of pixels, Indicates candidate anomaly regions Pixels within, It is a pixel. In the current frame's dynamic anomaly degree, It is a candidate anomaly region Motion consistency in the current frame.

[0053] It should be noted that, It calculates the average dynamic anomaly score of all pixels within the candidate anomaly region. The calculation is the product of the motion consistency of the candidate anomaly region and the average dynamic anomaly score of all pixels within the candidate anomaly region. Only when the candidate anomaly region conforms to the smoke rising model and has a high anomaly score can a high anomaly confidence score be output, indicating that the anomaly in the candidate anomaly region is a real anomaly. Conversely, if the anomaly region does not conform to the real physical model or has a low anomaly score, the output anomaly confidence score will be low, indicating that the anomaly in the candidate anomaly region is not a real anomaly. Real anomalies refer to anomalies such as smoke, while unreal anomalies refer to non-physical artifacts such as random noise.

[0054] It should be noted that although water vapor during hazardous waste treatment may conform to the smoke rising model in the current frame of the video, the water vapor situation has been isolated when S401 uses dynamic anomaly degree to filter candidate anomaly regions in the current frame of the video, and water vapor will not appear in the candidate anomaly regions.

[0055] S5. Conduct a risk assessment of hazardous waste treatment based on the level of anomaly confidence.

[0056] Specifically, multiple risk thresholds are preset: collect a large amount of abnormal confidence scores under normal operating conditions, and classify the data into different thresholds: 95th percentile, 99th percentile, and 99.9th percentile.

[0057] If the anomaly confidence level of the candidate anomaly region is less than or equal to the 95th percentile, the hazardous waste treatment is risk-free; if the anomaly confidence level of the candidate anomaly region is greater than the 95th percentile but less than the 99th percentile, the hazardous waste treatment is of low risk; if the anomaly confidence level of the candidate anomaly region is greater than or equal to the 99th percentile but less than the 99.9th percentile, the hazardous waste treatment is of medium risk; if the anomaly confidence level of the candidate anomaly region is greater than or equal to the 99.9th percentile, the hazardous waste treatment is of high risk.

[0058] This completes the risk assessment for hazardous waste treatment.

Claims

1. A method for monitoring and early warning of hazardous waste treatment based on video analysis, characterized in that, include: Gaussian mixture background modeling is performed on the video of key monitoring points of hazardous waste treatment facilities to obtain the Gaussian mixture model of each location in the video; any pixel location in the video is taken as the target location, and the mean parameter of the sub-Gaussian model that best matches the gray value of the target location in the previous frame is obtained as the predicted gray value of the target location in the current frame. The grayscale values ​​of the target location in the current frame and several previous frames are used to form a grayscale sequence of the target location. Based on the mean and standard deviation of the grayscale sequence, and the mean and standard deviation of the difference between two adjacent grayscale values ​​in the grayscale sequence, the local stability of the target location in the current frame is determined. Based on the difference between the predicted grayscale value and the actual grayscale value of the target location in the current frame, and the local stability of the scene, the dynamic anomaly degree of the target location in the current frame is determined, and candidate anomaly regions are marked in the current frame according to the magnitude of the dynamic anomaly degree. The motion vector of each pixel in the current frame is obtained using optical flow. The motion consistency of the candidate anomaly region is determined based on the proportion of pixels whose vertical component of the motion vector is greater than 0. The anomaly confidence of the candidate anomaly region is determined based on the motion consistency and the average dynamic anomaly degree of the pixels in the candidate anomaly region. Hazardous waste risk assessment is performed based on the anomaly confidence level.

2. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The step of obtaining the mean parameter of the sub-Gaussian model in the Gaussian mixture model of the target location that best matches the grayscale value of the target location in the previous frame, and using it as the predicted grayscale value of the target location in the current frame, includes: If the grayscale value of the target position in the previous frame has the smallest difference from the mean of a certain sub-Gaussian model, and is centered on the mean of the sub-Gaussian model with the smallest mean difference... A match is considered successful if the target position is within a confidence interval of one-time standard deviation. The mean of the matched sub-Gaussian model is taken as the predicted gray value of the target position in the current frame. If the gray value of the target position in the previous frame does not match any of the sub-Gaussian models, the predicted gray value of the nearest successfully matched pixel at the same position is selected as the predicted gray value of the target position in the current frame.

3. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The expression for the local stability of the scenario is: ； in, It is the frame number of the current frame. Target location Local scene stability in the current frame, It is a preset reference fluctuation benchmark. and These represent the mean and standard deviation of the grayscale values ​​at the target location within the window ending with the current frame, respectively. and It is the mean and standard deviation of the absolute values ​​of grayscale changes between adjacent frames at the target location within a short time window, with the current frame as the last frame. It is a hyperparameter to prevent the denominator from being 0.

4. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The expression for the dynamic anomaly degree is: ； in, It is the frame number of the current frame. Target location In the current frame's dynamic anomaly degree, Target location The grayscale value of the current frame; Target location The predicted grayscale value of the current frame, Target location Local scene stability in the current frame, These are parameters that adjust the impact on the local stability of the scene. It is the absolute value symbol.

5. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The step of marking candidate abnormal regions in the current frame according to the magnitude of dynamic abnormality includes: In response to the target location having a dynamic anomaly degree greater than a preset anomaly threshold in the current frame, the target location is marked in the current frame; a connected component algorithm is used to aggregate spatially adjacent marked points into candidate anomaly regions.

6. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, Determining the motion consistency of candidate abnormal regions includes: Optical flow is used on the current frame and previous consecutive video frames to obtain the motion vector of each pixel; the ratio of the number of motion vectors with positive vertical components in the candidate anomaly region to the total number of motion vectors in the candidate anomaly region is calculated as the motion consistency of the candidate anomaly region.

7. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The expression for the anomaly confidence level is: ； in, It is the frame number of the current frame. Indicates the first Candidate abnormal regions The anomaly confidence level in the current frame, Indicates the first Candidate abnormal regions The number of pixels, Indicates candidate anomaly regions Pixels within, It is a pixel. In the current frame's dynamic anomaly degree, It is a candidate anomaly region Motion consistency in the current frame.

8. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The hazardous waste risk assessment based on the anomaly confidence level includes: Preset multi-level risk thresholds: Collect a large amount of abnormal confidence scores from normal operating conditions and classify the data into different thresholds: 95th percentile, 99th percentile and 99.9th percentile; If the anomaly confidence level of the candidate anomaly region is less than or equal to the 95th percentile, the hazardous waste treatment is risk-free; if the anomaly confidence level of the candidate anomaly region is greater than the 95th percentile but less than the 99th percentile, the hazardous waste treatment is of low risk; if the anomaly confidence level of the candidate anomaly region is greater than or equal to the 99th percentile but less than the 99.9th percentile, the hazardous waste treatment is of medium risk; if the anomaly confidence level of the candidate anomaly region is greater than or equal to the 99.9th percentile, the hazardous waste treatment is of high risk.

9. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The method of performing Gaussian mixture background modeling on videos of key monitoring points of hazardous waste treatment facilities includes: Set the number of sub-Gaussian models to 5.

10. The hazardous waste treatment monitoring and early warning method based on video analysis according to claim 1, characterized in that, The step of constructing a grayscale sequence of the target location from the grayscale values ​​of the target location in the current frame and several previous frames includes: The grayscale values ​​of the target location in the current frame and the previous 50 frames are used to construct the grayscale sequence of the target location.

Images