A method for identifying flow observers at chemical plants based on inspection robots
By combining image stability judgment with grayscale difference and optical flow feature fusion on the inspection robot, the reliable identification problem of flow observers in chemical plants was solved, and low-cost and efficient flow status monitoring was achieved.
Patent Information
- Application Number
- CN202611477203.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively identify the fluid state of flow observers in chemical plants, especially under interference from factors such as glass reflection, sudden changes in illumination, residual vibrations, and diverse target motion patterns. Existing robot vision systems suffer from high false alarm and false negative rates, making it impossible to achieve reliable immediate detection and identification upon stopping.
A flow observer identification method based on inspection robots is adopted for chemical plants. By judging the stability of the image, fusing the gray-scale difference and optical flow field characteristics in the circular detection area, and combining adaptive threshold and illumination change protection, reliable identification of the flow observer is achieved.
It reduces the deployment and maintenance costs of flow monitoring at chemical plants, improves the reliability and stability of identification, reduces false alarm and missed detection rates, and provides continuous and objective flow status records.
Smart Images

Figure CN123118600A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial visual inspection technology, and in particular to a method for identifying flow observers in chemical plants based on inspection robots. Background Technology
[0002] Chemical plants, such as oil depots, gas stations, CNG stations, and chemical production areas, contain numerous process pipelines. Many process nodes are equipped with flow meters, which have transparent windows containing small balls or impellers. Operators can visually determine if there are any abnormalities in the fluid flow by observing whether the target within the window is moving. Currently, the inspection of flow meters still relies primarily on manual patrols. This manual method has significant shortcomings: chemical plants are flammable and explosive hazardous locations, severely limiting the frequency and duration of manual inspections; some observation points are located at high altitudes with poor visibility, making observation difficult; manual judgment is influenced by experience and subjective factors, making it difficult to create continuous, objective, and traceable records; and changes in the fluid state between inspections are unpredictable, hindering the timely detection and analysis of abnormal conditions.
[0003] To reduce manual labor, some stations have fixed cameras installed at key observation points for video monitoring. However, fixed camera solutions involve numerous locations, requiring the laying of a large number of power supply and network lines, resulting in high construction and maintenance costs. The fixed field of view of the cameras is easily blocked by pipes and structures, and the observation objects cannot be flexibly changed according to the process flow. For in-service stations with many points and wide coverage, and limited explosion-proof modifications, fixed camera solutions are difficult to achieve comprehensive coverage.
[0004] In recent years, wheeled inspection robots have been gradually promoted and applied in chemical plants, replacing manual labor in tasks such as identifying instrument readings, equipment appearance defects, oil leaks, and flames and smoke. However, for special targets such as flow meters, existing robot vision systems still lack effective automatic identification methods. The main reasons include: First, the flow meter window is a curved transparent glass, which is subject to interference from reflections, shadows, and background transmission, resulting in low contrast of the target within the window. Second, the lighting conditions in the station change significantly with the alternation of day and night, the start and stop of lighting, and vehicle headlight illumination, and sudden changes in screen brightness can easily lead to misjudgments. Third, after the robot stops, there are still residual micro-movements in the chassis and gimbal, which can cause the overall shaking of the screen to be confused with the actual movement of the target within the window. Fourth, the target within the window is small in size and has diverse movement patterns, including both small displacements and continuous rotations.
[0005] Conventional video motion detection methods, such as inter-frame difference, background modeling, and optical flow, have significant drawbacks when directly applied to the above scenarios: inter-frame difference can only detect changes in the frame and cannot distinguish between the overall motion caused by residual robot jitter and the actual motion of the target within the viewport; background modeling tends to absorb moving targets into the background model when they are in continuous motion, resulting in missed detections, and its adaptability to changes in lighting is limited; optical flow is sensitive to reflections and noise, and has a high computational load, making it difficult to run in real time on the embedded computing platform carried by the inspection robot; using a single fixed threshold for decision-making makes it difficult to adapt to the differences in noise levels under different cameras and shooting distances, resulting in high false alarm and false negative rates.
[0006] Therefore, there is an urgent need for a method that can be mounted on a wheeled inspection robot and reliably identify the fluid state of the flow observer in a stop-and-test working mode, in order to overcome the interference caused by factors such as glass reflection, sudden changes in illumination, residual vibration, and the diverse motion patterns of the target. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying flow observers in chemical plants based on inspection robots. This method can reliably identify the fluid state of flow observers in chemical plants under the "stop and measure" working mode, and overcome interference caused by factors such as glass reflection, sudden changes in illumination, residual vibration, and diverse target motion patterns.
[0008] The technical solution adopted in this invention is: A method for identifying flow observers at chemical plants based on inspection robots, comprising the following steps: S1: The wheeled inspection robot travels to the observation point in the chemical plant and controls the gimbal camera set on the wheeled inspection robot to rotate to the preset position, aim at the flow observer installed on the pipeline, and collect the image frame sequence of the flow observer window; S2: Determine the stability of the image frame sequence, and only execute subsequent steps after determining that the image is stable; S3: Determine the circular detection region corresponding to the flow observer window in a stable image frame sequence; S4: Calculate two types of complementary motion features within the circular detection region frame by frame: The first motion feature is the percentage of moving pixels in the circular detection area where the gray-level difference between the current frame and the reference frame exceeds an adaptive threshold. The adaptive threshold is determined adaptively based on the mean and standard deviation of the gray-level difference. The second motion characteristic is the average amplitude of the optical flow field within the circular detection region; S5: Obtain the change in the average brightness of the entire screen in the current frame relative to the reference frame. When the change exceeds the illumination change threshold, discard the motion determination result of the current frame and update the reference frame. S6: Fuse the first motion feature and the second motion feature according to the preset fusion rules to obtain the single-frame motion determination result. Perform statistical voting on the single-frame motion determination result within the sliding window, and combine the hysteresis time constraint to output the fluid state recognition result of the flow observer. The fluid state recognition result includes flow state and no flow state.
[0009] Specifically, in step S1, the corresponding docking pose, gimbal preset position parameters, and observer calibration information are pre-stored for each observation point; after the wheeled inspection robot arrives at the observation point and docks, it controls the gimbal camera to rotate and align with the flow observer according to the gimbal preset position parameters, and records a preset duration T. r Image frame sequence, preset duration T r Not less than the hysteresis time threshold T h .
[0010] Specifically, in step S2, the image stability determination includes: denoted as I_t_th frame full-screen grayscale image. t Let the set of pixels in the entire image be A. Calculate the average grayscale difference between adjacent frames, ΔI = (1 / |A|)·Σ|I t (x,y)-I t-1 (x,y)|, when the average grayscale difference of the entire image in multiple consecutive frames is less than the stable threshold T s If the image is stable, proceed to the next step; otherwise, wait for a preset time and re-acquire the image, or perform global translation and registration compensation on the image before making a judgment.
[0011] Specifically, determining the circular detection area Ω in step S3 includes: reading the pre-stored observer calibration position and mapping it to the current screen to obtain the initial circular area; or performing circular detection on the current screen and correcting the position and radius of the circular detection area based on the detected window circular outline.
[0012] Specifically, in step S4, the first motion feature extraction process is as follows: The images of the current frame and the reference frame within the circular detection area are converted to grayscale and filtered for noise reduction, respectively; the absolute value of the pixel-by-pixel difference between the two grayscale images is calculated to obtain the grayscale difference image D; Based on the mean μ of the gray-level difference image within the circular detection region D With standard deviation σ D Press T D =max(T min , μ D +K·σ D Determine the adaptive threshold T D And the adaptive threshold is not less than the preset lower limit T min, K is a coefficient; According to the adaptive threshold T DBinarize the grayscale difference image D to obtain a motion mask. After morphological denoising of the motion mask, calculate the proportion of moving pixels to the total number of pixels in the circular detection region to obtain the moving pixel ratio ρ. Then, apply exponential smoothing to the moving pixel ratio ρ. t =α·ρ+(1-α)·ρ t-1 α is the smoothing coefficient, and ρ is the smoothed result. t As the first characteristic of motion.
[0013] Specifically, in step S4, the second motion feature extraction process is as follows: The image within the circular detection region Ω is downsampled by a factor of s. The pyramid-layered dense optical flow algorithm is used to calculate the optical flow field (u,v) of the current frame relative to the reference frame after downsampling. The average amplitude of the optical flow field of each pixel within the circular detection region Ω is then calculated. As the second characteristic of motion.
[0014] Specifically, in step S5, the average brightness of the entire screen in frame t is denoted as B. t When the change |B t -B t-1 |Exceeding the light mutation threshold T B When a sudden lighting event occurs, the motion determination result of the current frame is discarded and the current frame is not included in the statistical voting of the sliding window. The reference frame is then reconstructed using the current frame.
[0015] Specifically, in step S6, the preset fusion rule can be configured as any one of logical OR, logical AND-OR weighted confidence, specifically: Logical OR means that if at least one of the first motion feature and the second motion feature exceeds its corresponding determination threshold, the current frame is determined to be a motion frame; The logical AND operation means that when both the first motion feature and the second motion feature exceed their corresponding determination thresholds, the current frame is determined to be a motion frame. The weighted confidence score refers to the motion confidence score obtained by weighted summation after normalizing the first motion feature and the second motion feature respectively. When the motion confidence score exceeds the confidence score threshold, the current frame is determined to be a motion frame.
[0016] Specifically, in step S6, the statistical voting and hysteresis time constraint includes: statistically analyzing the proportion of frames judged as motion frames within the sliding window; when the proportion is not less than a first proportion threshold, determining that the system is in a state with traffic and recording the start time of traffic; during the period of being in a state with traffic, if the duration of consecutive non-motion frames exceeds the hysteresis duration threshold T... h When exiting the traffic state, the system determines that the traffic has ended and records the time when the traffic ends, generating a traffic event that includes the time when the traffic started and the time when the traffic ended.
[0017] Specifically, the fluid state identification result output by the flow observer in step S6 includes at least one of fluid state, flow event, frame-by-frame motion feature recording and labeled video, which is used for inspection log recording, abnormal alarm linkage and remote monitoring display.
[0018] Due to the adoption of the technical solution described above, the present invention has the following advantages: This invention utilizes a wheeled inspection robot as a carrier, enabling flow observer identification with a single-point docking. This eliminates the need for fixed cameras at each observation point, significantly reducing the deployment and maintenance costs of flow monitoring in chemical plants. Through image stability judgment and a re-sampling mechanism, it overcomes motion artifacts caused by residual shaking of the chassis and gimbal after robot docking, achieving near-fixed camera detection reliability for the mobile platform. It extracts and fuses two complementary motion features—grayscale difference motion ratio and average optical flow amplitude—within the circular detection area of the viewing window, making it applicable to both spherical and impeller-type observers. It is sensitive to both small displacement and continuous rotation motion patterns, while suppressing local interference such as glass reflection. The adaptive threshold automatically adjusts according to the image noise level, and the illumination change protection mechanism discards the judgment result and reconstructs a reference frame during brightness changes, significantly improving judgment stability under day / night cycles and lighting start / stop conditions. The combination of sliding window statistical voting and hysteresis constraints avoids false alarms caused by single-frame noise while ensuring rapid response to changes in flow status, outputting flow events with start and end times for easy inspection log recording and alarm linkage. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0020] Figure 2 This is a schematic diagram of the application scenario structure of the present invention.
[0021] Figure 3 This is a block diagram of the dual-channel motion feature extraction and fusion determination method of the present invention.
[0022] Figure 4 This is a schematic diagram of the small ball-type flow observer and its circular detection area Ω of the present invention.
[0023] Figure 5 This is a schematic diagram of the impeller-type flow observer and its circular detection area according to the present invention.
[0024] Figure 6 This is a comparison diagram of the sliding window voting and hysteresis states of the present invention.
[0025] In the diagram: 1-Wheeled inspection robot, 2-Pan-Tilt camera, 3-Pipeline, 4-Flow observer, 5-View window, 6-Observation point. Detailed Implementation
[0026] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments. However, this should not be construed as limiting the scope of protection of the present invention. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention.
[0027] Combined with appendix Figure 1-6 The method for identifying flow observers at chemical plants based on inspection robots, as shown, includes the following steps: S1: Pre-store the corresponding docking pose, gimbal preset position parameters, and flow observer 4 calibration information for each observation point 6. The calibration information includes at least the observer type and the initial position and radius of the window. The wheeled inspection robot 1 travels to the observation point 6 in the chemical plant, brakes and stops, and controls the gimbal camera 2 set on the wheeled inspection robot 1 to rotate and align with the flow observer 4 installed on the pipeline 3 according to the gimbal preset position parameters. The recording frame rate is 25 frames / second, and the recording duration is set to the preset duration T. r Take a sequence of image frames from window 5 of the flow rate observer for 6-10 seconds, with a preset duration T. r Not less than the hysteresis time threshold T h Flow observer 4 can be Figure 4 The small ball type shown in Figure 5 or the impeller type shown in Figure 5; in the small ball type observer, the ball continuously moves and bounces within the viewing window under the propulsion of the fluid; in the impeller type observer, the impeller continuously rotates around the axis under the propulsion of the fluid.
[0028] S2: Image stability determination is used to isolate the residual jitter of the robot's brakes and the mechanical oscillation of the gimbal. It performs image stability determination on the image frame sequence and executes subsequent steps only after the image is determined to be stable. Image stability determination includes: Let I be the grayscale image of the entire image in frame t. t Let the set of pixels in the entire image be A. Calculate the average grayscale difference between adjacent frames, ΔI = (1 / |A|)·Σ|I t (x,y)-I t-1 (x,y)|, when the average grayscale difference of the entire image in multiple consecutive frames is less than the stable threshold T s If the image is stable, proceed to the next step; otherwise, wait for a preset time before re-acquiring, or perform global translation and registration compensation on the image before making a judgment. Since the movement of the target within the viewport is restricted to a circular detection area, the overall image statistics are much more sensitive to overall shaking than to local target movement, so this judgment will not mistakenly affect the actual movement within the viewport.
[0029] S3: Determine the circular detection region Ω corresponding to the flow observer window in a stable image frame sequence; Determining the circular detection region Ω involves: Ω being described by its center coordinates and radius, with the radius set to completely cover the inner circle of the glass window while excluding the metal frame; reading the pre-stored observer calibration position and mapping it to the current screen to obtain the initial circular region; when the pre-stored calibration is unavailable or has deviations, performing circular detection on the current screen, such as Hough circle transform, edge extraction, and circle fitting, and correcting the position and radius of the circular detection region based on the detected circular outline of the window. The detection results can be written back to the calibration library to achieve continuous self-calibration of the observer position.
[0030] S4: Calculate the two complementary motion features within the circular detection region frame by frame, such as Figure 3 As shown: The first motion feature is the percentage of moving pixels in the circular detection region where the gray-level difference between the current frame and the reference frame exceeds an adaptive threshold. The adaptive threshold is determined adaptively based on the mean and standard deviation of the gray-level difference; specifically: The images of the current frame and the reference frame within the circular detection region Ω are converted to grayscale and filtered for noise reduction. The grayscale values are obtained by weighting the R, G, and B components. The absolute value of the pixel-by-pixel difference between the two grayscale images is calculated to obtain the grayscale difference image D. Based on the mean μ of the gray-level difference image within the circular detection region D With standard deviation σ D Press T D =max(T min , μ D +K·σ D Determine the adaptive threshold T D And the adaptive threshold is not less than the preset lower limit T min K is a coefficient; According to the adaptive threshold T D Binarize the grayscale difference image D to obtain a motion mask. After morphological denoising of the motion mask, calculate the proportion of moving pixels to the total number of pixels in the circular detection region to obtain the moving pixel ratio ρ. Then, apply exponential smoothing to the moving pixel ratio ρ. t =α·ρ+(1-α)·ρ t-1 α is the smoothing coefficient, and ρ is the smoothed result. t As the first motion feature, since the threshold automatically floats with the noise level of the image, this method can adapt to noise differences under different cameras, shooting distances and illumination levels, and its robustness is significantly better than the fixed threshold frame difference method.
[0031] The second motion feature is the average amplitude of the optical flow field within the circular detection region; specifically, the image within the circular detection region Ω is downsampled by a factor of s, and the optical flow field (u,v) of the current frame relative to the reference frame after downsampling is calculated using the pyramid layered dense optical flow algorithm. The average amplitude of the optical flow field of each pixel within the circular detection region Ω is then calculated. As a second motion feature, the optical flow channel is more sensitive to smooth motions such as slow translation of the ball and continuous rotation of the impeller. It complements the frame difference channel and can significantly reduce the probability of missed detection under slight motion conditions. Thanks to the limitation of the circular detection area and downsampling processing, the optical flow calculation is controlled at the millisecond level, which meets the real-time computing requirements of the inspection robot embedded platform.
[0032] S5: Obtain the change in the average brightness of the entire screen in the current frame relative to the reference frame. When the change exceeds the illumination change threshold, discard the motion determination result of the current frame and update the reference frame; specifically: Let the average brightness of the entire screen in frame t be B. t When the change |B t -B t-1 |Exceeding the light mutation threshold T B When a sudden change in illumination occurs, the motion determination result of the current frame is discarded and the current frame is not included in the statistical voting of the sliding window. The reference frame is then reconstructed using the current frame. This mechanism, in principle, prevents sudden changes in brightness from being mistaken for motion.
[0033] S6: Fuse the first motion feature and the second motion feature according to the preset fusion rules to obtain the single-frame motion determination result. Perform statistical voting on the single-frame motion determination result within the sliding window, and combine the hysteresis time constraint to output the fluid state recognition result of the flow observer. The fluid state recognition result includes flow state and no flow state.
[0034] Specifically, the preset fusion rule can be configured as any one of logical OR, logical AND-OR weighted confidence, as follows: Logical OR means that if at least one of the first motion feature and the second motion feature exceeds its corresponding determination threshold, the current frame is determined to be a motion frame; The logical AND operation means that when both the first motion feature and the second motion feature exceed their corresponding determination thresholds, the current frame is determined to be a motion frame. The weighted confidence score refers to the motion confidence score obtained by weighted summation after normalizing the first motion feature and the second motion feature respectively. When the motion confidence score exceeds the confidence score threshold, the current frame is determined to be a motion frame.
[0035] Specifically, the statistical voting and hysteresis time constraints include: the proportion of frames judged as moving frames within a statistical sliding window; when the proportion is not less than a first proportion threshold, the system is determined to enter a traffic-carrying state and the start time of traffic is recorded; during the traffic-carrying state, if the duration of consecutive frames judged as non-moving frames exceeds the hysteresis duration threshold T... h When exiting the flow state, the system determines that the flow has ended and records the flow end time, generating a flow event containing both the flow start and end times. The combination of sliding window voting and hysteresis constraints ensures that single-frame noise does not alter the state output, while actual flow changes are responded to within approximately one second. The timing relationship is as follows: Figure 5 As shown.
[0036] Specifically, the output flow observer fluid state identification results include at least one of fluid state, flow events, frame-by-frame motion feature records and labeled videos, which are used for inspection log recording, abnormal alarm linkage and remote monitoring display.
[0037] Taking the small ball-shaped flow observer on the outlet pipeline of the oil depot loading pump as an example, the inspection robot travels to the observation point according to the preset route, and the gimbal camera is aligned with the observer to record an 8-second video. During the recognition process, the image is stable when the pump is not running. The frame difference motion ratio in the circular detection area is about 0.0005 after smoothing, and the average amplitude of optical flow is about 0.03 pixels / frame, both of which are lower than the corresponding thresholds. The sliding window voting ratio is always zero, and the state machine remains in a no-flow state. After the pump starts, the ball bounces back and forth in the window, and the motion ratio and optical flow amplitude increase significantly and exceed the threshold. The motion frame ratio in the sliding window quickly reaches the first ratio threshold, and the state machine flips to a flow state in about 1 second and records the start time. After the pump stops again, the ball tends to be stationary. After the hysteresis time constraint, the state machine exits the flow state and records the end time. There are no false alarms or missed alarms throughout the entire process.
[0038] Taking the impeller-type flow observer on the cooling water pipeline in the chemical plant area as an example, the identification process is the same as that of the ball-type, the only difference being the adaptation of parameters and fusion rules: the flow channel response is more significant when the impeller is rotating continuously, so the amplitude threshold can be lowered (e.g., to 0.20 pixels / frame), and the fusion rule can be configured as logical OR; the periodic brief period of no movement generated when the impeller blades sweep across the gap of the viewing window is naturally smoothed by the sliding window voting, and will not cause repeated state jumps; when the impeller starts and stops intermittently due to the decrease in flow, the state machine outputs the corresponding flow event sequence, which can provide a basis for process personnel to judge flow fluctuations.
[0039] Furthermore, when multiple flow observers exist in the same frame, multiple circular detection areas can be established based on the calibration information. Steps S4 to S6 are executed in parallel for each area, and the recognition results are output separately. When the camera focal length or shooting distance changes, the parameters of the circular detection area can be converted according to the calibration ratio, and the method itself does not need to be modified.
[0040] The parts of this invention not described in detail are prior art.
[0041] The embodiments selected herein for the purpose of disclosing the inventive objectives are currently considered suitable; however, it should be understood that the invention is intended to include all variations and modifications of the embodiments that fall within the scope of this concept and invention.
Claims
1. A method for identifying flow observers at chemical plant stations based on inspection robots, characterized in that, The specific steps are as follows: S1: The wheeled inspection robot travels to the observation point in the chemical plant and controls the gimbal camera set on the wheeled inspection robot to rotate to the preset position, aim at the flow observer installed on the pipeline, and collect the image frame sequence of the flow observer window; S2: Determine the stability of the image frame sequence, and only execute subsequent steps after determining that the image is stable; S3: Determine the circular detection region corresponding to the flow observer window in a stable image frame sequence; S4: Calculate two types of complementary motion features within the circular detection region frame by frame: The first motion feature is the percentage of moving pixels in the circular detection area where the gray-level difference between the current frame and the reference frame exceeds an adaptive threshold. The adaptive threshold is determined adaptively based on the mean and standard deviation of the gray-level difference. The second motion characteristic is the average amplitude of the optical flow field within the circular detection region; S5: Obtain the change in the average brightness of the entire screen in the current frame relative to the reference frame. When the change exceeds the illumination change threshold, discard the motion determination result of the current frame and update the reference frame. S6: Fuse the first motion feature and the second motion feature according to the preset fusion rules to obtain the single-frame motion determination result. Perform statistical voting on the single-frame motion determination result within the sliding window, and combine the hysteresis time constraint to output the fluid state recognition result of the flow observer. The fluid state recognition result includes flow state and no flow state.
2. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S1, the corresponding docking pose, gimbal preset position parameters, and observer calibration information are pre-stored for each observation point. After the wheeled inspection robot arrives at the observation point and docks, it controls the gimbal camera to rotate and align with the flow observer according to the gimbal preset position parameters, and records a preset duration T. r Image frame sequence, preset duration T r Not less than the hysteresis time threshold T h .
3. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S2, the image stability determination includes: denoted as I as the grayscale image of the entire image in frame t. t Let the set of pixels in the entire image be A. Calculate the average grayscale difference between adjacent frames, ΔI = (1 / |A|)·Σ|I t (x,y)-I t-1 (x,y)|, when the average grayscale difference of the entire image in multiple consecutive frames is less than the stable threshold T s If the image is stable, proceed to the next step; otherwise, wait for a preset time and re-acquire the image, or perform global translation and registration compensation on the image before making a judgment.
4. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: The step S3 of determining the circular detection area Ω includes: reading the pre-stored observer calibration position and mapping it to the current screen to obtain the initial circular area; or performing circular detection on the current screen and correcting the position and radius of the circular detection area according to the detected window circular outline.
5. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S4, the first motion feature extraction process is as follows: The images of the current frame and the reference frame within the circular detection area are converted to grayscale and filtered for noise reduction, respectively. Calculate the absolute value of the pixel-by-pixel difference between two grayscale images to obtain the grayscale difference image D; Based on the mean μ of the gray-level difference image within the circular detection region D With standard deviation σ D Press T D =max(T min , μ D +K·σ D Determine the adaptive threshold T D And the adaptive threshold is not less than the preset lower limit T min, K is a coefficient; According to the adaptive threshold T D The motion mask is obtained by binarizing the grayscale difference image D. After morphological denoising of the motion mask, the proportion of moving pixels to the total number of pixels in the circular detection area is counted to obtain the proportion of moving pixels ρ. And perform exponential smoothing on the proportion of moving pixels ρ t =α·ρ+(1-α)·ρ t-1 α is the smoothing coefficient, and ρ is the smoothed value. t As the first characteristic of motion.
6. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S4, the second motion feature extraction process is as follows: The image within the circular detection region Ω is downsampled by a factor of s. The pyramid-layered dense optical flow algorithm is used to calculate the optical flow field (u,v) of the current frame relative to the reference frame after downsampling. The average amplitude of the optical flow field of each pixel within the circular detection region Ω is then calculated. As the second characteristic of motion.
7. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S5, the average brightness of the entire screen in frame t is denoted as B. t When the change |B t -B t-1 |Exceeding the light mutation threshold T B When a sudden lighting event occurs, the motion determination result of the current frame is discarded and the current frame is not included in the statistical voting of the sliding window. The reference frame is then reconstructed using the current frame.
8. The method for identifying flow observers at chemical plants based on inspection robots according to claim 1, characterized in that: In step S6, the preset fusion rule can be configured as any one of logical OR, logical AND-OR weighted confidence, specifically: Logical OR means that if at least one of the first motion feature and the second motion feature exceeds its corresponding determination threshold, the current frame is determined to be a motion frame; The logical AND operation means that when both the first motion feature and the second motion feature exceed their corresponding determination thresholds, the current frame is determined to be a motion frame. The weighted confidence score refers to the motion confidence score obtained by weighted summation after normalizing the first motion feature and the second motion feature respectively. When the motion confidence score exceeds the confidence score threshold, the current frame is determined to be a motion frame.
9. The method for identifying flow observers at chemical plants based on inspection robots according to claim 8, characterized in that: In step S6, the statistical voting and hysteresis time constraint includes: statistically analyzing the proportion of frames judged as motion frames within the sliding window; when the proportion is not less than a first proportion threshold, determining that the system is in a state with traffic and recording the start time of traffic; during the period of being in a state with traffic, if the duration of consecutive non-motion frames exceeds the hysteresis duration threshold T... h When exiting the traffic state, the system determines that the traffic has ended and records the time when the traffic ends, generating a traffic event that includes the time when the traffic started and the time when the traffic ended.
10. The method for identifying flow observers at chemical plants based on inspection robots according to claim 9, characterized in that: The fluid state identification result output by the flow observer in step S6 includes at least one of fluid state, flow event, frame-by-frame motion feature recording and labeled video, which is used for inspection log recording, abnormal alarm linkage and remote monitoring display.
