Activated carbon replacement monitoring method

Abstract

This invention discloses a method for monitoring activated carbon replacement. It employs an image recognition algorithm based on multi-level deep learning to automatically identify activated carbon replacement operations by detecting a continuous sequence of actions: "opening the door → reaching out → retrieving the object → removing it." The accuracy rate is over 95%, and the response time is less than 2 seconds. Compared to traditional manual recording methods, the system can objectively and promptly capture every replacement action, avoiding omissions, errors, and falsified records. Using an event-triggered recording mode, recording only starts when activated carbon replacement is detected. Compared to traditional monitoring systems that record continuously for 24 hours, this saves over 95% of storage space, significantly reducing video storage costs and improving retrieval efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of activated carbon replacement monitoring, and more particularly to a method for monitoring activated carbon replacement. Background Technology

[0002] With increasing environmental awareness, industrial waste gas treatment has become a crucial aspect of enterprise production and operation. Activated carbon adsorption, as a mature and efficient waste gas treatment technology, is widely used in the treatment of organic waste gas in industries such as chemical, spraying, printing, and pharmaceutical.

[0003] In an activated carbon adsorption system, the activated carbon adsorption device refers to a complete equipment unit comprising an adsorption bed, inlet and outlet gas pipelines, valves, and instruments. The activated carbon adsorption bed is the core functional area inside the adsorption device, filled with activated carbon. Activated carbon is the adsorption material filled in the adsorption bed. These three elements form a hierarchical relationship of "device-bed-material": the activated carbon adsorption device contains an activated carbon adsorption bed, which is filled with activated carbon, and the activated carbon removes pollutants from the waste gas through physical adsorption.

[0004] However, activated carbon gradually becomes saturated during use and must be replaced regularly to ensure that the waste gas treatment effect meets environmental emission standards. Currently, enterprises mainly use the following four technical solutions for activated carbon replacement management, but all of them have significant shortcomings:

[0005] (1) Manual record management method

[0006] The current system relies on manual paper-based records of information changes, which are then compiled and archived by environmental administrators. The main drawbacks of this system are: authenticity cannot be guaranteed, leading to the possibility of retroactive or falsified records; process supervision is lacking, making it impossible to confirm the completeness and standardization of the changes; search efficiency is low, with paper records difficult to retrieve; and evidentiary value is weak, as paper records are easily lost or altered.

[0007] (2) Ordinary video surveillance system

[0008] The area is monitored using continuous or timed recording. The main drawbacks of this approach are: high storage costs and the generation of massive amounts of invalid data from continuous recording; low retrieval efficiency, requiring manual playback of each segment to find the target footage; lack of intelligent recognition, unable to automatically distinguish between activated carbon replacement and other activities; and limited remote functionality, lacking advanced features such as intelligent retrieval and event push notifications.

[0009] (3) Sensor-based automatic monitoring system

[0010] Automatic detection is achieved by installing pressure, temperature, and gas concentration sensors. The main drawbacks of this approach are: it can only provide alerts, not supervision; it cannot record the actual replacement process; it cannot record replacement details, thus failing to create a complete traceable record; it has a high false alarm rate; the sensors are susceptible to environmental factors; and the retrofit cost is high, making it difficult to modify existing equipment.

[0011] (4) RFID tag identification system

[0012] RFID tags are attached to activated carbon packaging. The main drawbacks of this approach are: the replacement process cannot be monitored, and it cannot be confirmed whether the activated carbon issued is actually used for replacement; it is easily circumvented, as operators may issue the carbon but not actually replace it; the tags are costly, as each bag of activated carbon needs to be tagged; and there is a lack of visual evidence, as there is no video recording.

[0013] In summary, existing technical solutions cannot simultaneously meet the multiple requirements of intelligent identification, automatic recording, remote monitoring, and evidence preservation. Therefore, there is an urgent need to develop an activated carbon replacement monitoring solution that can intelligently identify activated carbon replacement actions, automatically trigger recording, efficiently store key videos, support remote monitoring and viewing, and facilitate evidence preservation and traceability. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for monitoring activated carbon replacement.

[0015] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0016] A method for monitoring activated carbon replacement includes:

[0017] During the real-time monitoring phase, video footage of the activated carbon storage area is continuously collected and analyzed in real time at a set frame sampling rate.

[0018] In the action recognition stage, when the motion intensity in the region of interest exceeds the threshold, the frame sampling rate is increased, and a multi-level deep learning algorithm is run to identify the action sequence of activated carbon being taken out of the storage box and to obtain the trigger score.

[0019] During the recording phase, the trigger score is used to determine whether the target action is successfully matched. If the match is successful, the recording begins.

[0020] During the continuous recording and judgment phase, the decision to continue or stop recording is made based on the activity on the screen, the status of the activated carbon storage box door, and the presence of personnel.

[0021] During the data storage phase, video files are saved and metadata containing timestamps, device information, and recognition results is generated;

[0022] During the data upload phase, the video recording files and metadata are uploaded to the remote monitoring platform via the network;

[0023] During the remote viewing phase, managers can view the monitoring footage in real time or retrieve and play back historical recordings through the platform.

[0024] Furthermore, the multi-level deep learning algorithm used in the action recognition stage includes:

[0025] Object detection:

[0026] The YOLO algorithm is used to detect human bodies, hands, activated charcoal packaging bags, and activated charcoal storage box doors in video frames, and a set of detection boxes is output. Each detection box include:

[0027] 1) Detection box coordinates: , The coordinates of the detection box center are, and These are the width and height, respectively.

[0028] 2) Detection confidence level: ,in The probability that a target exists in the region. The conditional probability of correctly identifying the target category;

[0029] 3) Category probability vector: ,in , For the first The first detection box Logit output for each category, =5, representing the number of target categories, corresponding to human body, hand, activated carbon packaging bag, storage box door, and background; and satisfying the following conditions: ;

[0030] Motion analysis:

[0031] The target trajectory is calculated using the optical flow method by minimizing the energy function. Solving the optical flow field :

[0032] ;

[0033] in, The images are respectively in direction, Direction and Time Partial derivatives in direction, and They are respectively and The velocity component of the direction of motion, For smoothing weight parameters, and The spatial gradient of the optical flow; the motion matching score is calculated based on the optical flow field. ;

[0034] Action recognition:

[0035] Using a temporal convolutional network (TCN) or a long short-term memory network (LSTM), the output is an action recognition probability score. ;

[0036] Calculate the trigger score : ;

[0037] in, The target detection confidence score is calculated from the detection confidence score. Aggregation yields; For the weighting coefficients, satisfying .

[0038] Furthermore, the target detection confidence score is calculated. The process includes:

[0039] For each type of target These correspond to the human body, hands, activated carbon packaging bags, and activated carbon storage box doors, respectively. The highest confidence level for that category is taken from all detection frames.

[0040] ;

[0041] The overall target detection confidence score is calculated using a weighted average:

[0042] ;

[0043] in, For the first The weight coefficients of the class targets satisfy the following conditions: , For single-class confidence threshold; This is an indicator function; its value is 1 if the condition is met, and 0 otherwise.

[0044] For the recent Frames Time-series smoothing of values:

[0045] ;

[0046] in, To smooth the window length, The time decay factor, This is the raw score for a single frame.

[0047] Furthermore, the motion matching score The following sub-indicators were weighted and combined to obtain the following:

[0048] ;

[0049] in, , , These are the weighting coefficients;

[0050] The trajectory integrity score is obtained through the following process:

[0051] The DeepSORT algorithm is used to continuously track the detected target and establish the hand trajectory. and activated carbon trajectory Define the internal area of ​​the activated carbon storage box. and external areas ,calculate:

[0052] ;

[0053] in, This indicates that the trajectory point falls within the region. The quantity within; =5, which is the threshold for valid trajectory points;

[0054] The motion direction matching score is obtained through the following process:

[0055] Calculate the hand displacement vector and activated carbon displacement vector , and the standard direction vector points into the box and pointing out of the box Calculate cosine similarity:

[0056] ;

[0057] The score for the reasonableness of the movement speed is obtained through the following process:

[0058] Calculate average speed The evaluation is performed using a piecewise function:

[0059] ;

[0060] in, For minimum speed, That is the maximum speed.

[0061] Furthermore, the action recognition probability score It is obtained through the following process:

[0062] Define standard action template sequence ,in For the first A standard state vector; actual observation sequence ; Calculate DTW cumulative distance:

[0063] ;

[0064] in, The distance is Euclidean; the normalized DTW matching score is:

[0065] ;

[0066] Action recognition probability score The result obtained through fusion:

[0067] ;

[0068] in, For the output of the neural network, To integrate weights, This represents the maximum normalized distance.

[0069] Furthermore, a three-level triggering mechanism is adopted for the recording phase:

[0070] Pre-triggered: Detection of motion intensity in the ROI region Furthermore, when a human body is detected, the frame sampling rate is increased and a loop buffer is initiated; The set threshold for exercise intensity;

[0071] Among them, exercise intensity The calculation formula is:

[0072] ; This represents the total number of pixels within the ROI region. and The first Motion velocity components of each pixel;

[0073] Soft trigger: When an open door or a hand is detected entering the enclosure, and... At that time, mark the suspected replacement time. Lock the cache; This is the soft trigger threshold;

[0074] Hard trigger: When it is confirmed that the activated carbon has been removed and Immediately start recording and merge the cached video with the live video; This is the hard trigger threshold.

[0075] Furthermore, during the continuous recording judgment phase, after recording begins, the system continuously monitors changes in the video feed and stops recording when the following stopping conditions are met:

[0076] Motion intensity within the region of interest (ROI) And it continues for longer than the static waiting time. Furthermore, the storage box door becomes closed, and no human target is detected within the ROI area, exceeding the waiting time for personnel to leave. .

[0077] Furthermore, before triggering recording, a multi-dimensional false trigger suppression mechanism is activated, and the recording stage is entered only after the corresponding conditions are met.

[0078] The multi-dimensional false trigger suppression mechanism is as follows:

[0079] Object size filtering: The activated carbon packaging bag detection frame must meet the following requirements. , Aspect Ratio ; The width of the inspection frame for activated carbon packaging bags; The height of the detection frame for activated carbon packaging bags; The minimum effective width of the activated carbon packaging bag; This is the maximum effective width of the activated carbon packaging bag; This refers to the minimum effective height of the activated carbon packaging bag; This is the maximum effective height of the activated carbon packaging bag; The minimum aspect ratio threshold; The maximum aspect ratio threshold;

[0080] Action duration filtering: Total duration of the action sequence ; This is the minimum threshold for the action duration; This represents the maximum threshold for action duration.

[0081] Time window filtering: Only during preset permitted periods Internal triggering; The start time for changing the license period; This is the end time for the changed license period.

[0082] Furthermore, the DeepSORT algorithm is used to continuously track the detected target, including:

[0083] Assign a unique tracking ID to each detected target and record its complete motion trajectory:

[0084] ;

[0085] in, For the first The target center coordinates of the frame;

[0086] Determine the type of action by analyzing changes in trajectory direction:

[0087] If the hand trace is from the outer area of ​​the activated carbon storage box Move to the inner area of ​​the activated carbon storage box This is judged as a "penetration" action;

[0088] If the trajectory of the activated carbon packaging bag is from the internal area of ​​the activated carbon storage box Move to the outer area of ​​the activated carbon storage box. This is determined to be a "take out" action;

[0089] The region determination is achieved by comparing the coordinates of the detection frame center with the preset coordinates of the box region. Comparison implementation, The x-coordinate of the upper left corner of the box area. The vertical coordinate of the upper left corner of the box-shaped area. The x-coordinate of the lower right corner of the box area. This represents the ordinate of the lower right corner of the box-shaped area.

[0090] Furthermore, the data storage stage adopts the H.265 / HEVC encoding standard, and the bitrate is adaptively adjusted according to the complexity of the image. :

[0091] ;

[0092] in, For the set low bitrate, For the set medium bitrate, For setting a high bitrate;

[0093] Simultaneously generate a low-bitrate thumbnail video, reducing the bitrate. for:

[0094] ;

[0095] in, The compression factor is 1.

[0096] Compared with existing technologies, the principles and advantages of this technical solution are as follows:

[0097] 1. Employing a multi-level deep learning-based image recognition algorithm, the system automatically identifies activated carbon replacement operations by detecting the continuous action sequence of "opening the door → reaching out → taking the item → removing it." The recognition accuracy reaches over 95%, with a response time of less than 2 seconds. Compared to traditional manual recording methods, the system can objectively and promptly capture every replacement action, avoiding omissions, errors, and falsified records.

[0098] Simultaneously, when recording is triggered, the system rewinds 10-30 seconds of buffered video to ensure a complete record of the entire process from unpacking to retrieval, without missing any critical operational details. Compared to the continuous or timed recording methods of ordinary video surveillance systems, this automatic triggering mechanism can accurately capture target events and improve recording integrity by 100%.

[0099] 2. It adopts an event-triggered recording mode, which only starts recording when the activated carbon replacement action is detected. Compared with the traditional monitoring system that records continuously for 24 hours, it can save more than 95% of storage space, significantly reducing video storage costs and improving retrieval efficiency.

[0100] Furthermore, using the H.265 encoding standard can save 40-50% of storage space compared to H.264 encoding. Under the premise of ensuring the same image quality, 256GB of local storage can save hundreds of replacement records. Attached Figure Description

[0101] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0102] Figure 1 This is a flowchart illustrating the principle of an activated carbon replacement monitoring method according to the present invention. Detailed Implementation

[0103] The present invention will be further described below with reference to specific embodiments:

[0104] like Figure 1 As shown in this embodiment, an activated carbon replacement monitoring method includes the following steps:

[0105] S1. Real-time monitoring phase: Continuously collect video footage of the activated carbon storage area and perform real-time analysis at a frame sampling rate of 25fps.

[0106] S2, Action Recognition Stage: When the motion intensity in the region of interest exceeds the threshold, the frame sampling rate is increased, and a multi-level deep learning algorithm is run to recognize the action sequence of activated carbon being taken out of the storage box and to obtain the trigger score.

[0107] In this step, the action recognition stage employs a multi-level deep learning algorithm, including:

[0108] Object detection:

[0109] The YOLO algorithm is used to detect human bodies, hands, activated charcoal packaging bags, and activated charcoal storage box doors in video frames, and a set of detection boxes is output. Each detection box include:

[0110] 1) Detection box coordinates: , The coordinates of the detection box center are, and These are the width and height, respectively.

[0111] 2) Detection confidence level: ,in The probability that a target exists in the region. The conditional probability of correctly identifying the target category;

[0112] 3) Category probability vector: ,in , For the first The first detection box Logit output for each category, =5, representing the number of target categories, corresponding to human body, hand, activated carbon packaging bag, storage box door, and background; and satisfying the following conditions: ;

[0113] Motion analysis:

[0114] The target trajectory is calculated using the optical flow method by minimizing the energy function. Solving the optical flow field :

[0115] ;

[0116] in, The images are respectively in direction, Direction and Time Partial derivatives in direction, and They are respectively and The velocity component of the direction of motion, For smoothing weight parameters, and The spatial gradient of the optical flow; the motion matching score is calculated based on the optical flow field. ;

[0117] Action recognition:

[0118] Using a temporal convolutional network (TCN) or a long short-term memory network (LSTM), the output is an action recognition probability score. ;

[0119] Calculate the trigger score : ;

[0120] in, The target detection confidence score is calculated from the detection confidence score. Aggregation yields; For the weighting coefficients, satisfying .

[0121] Calculate the target detection confidence score The process includes:

[0122] For each type of target These correspond to the human body, hands, activated carbon packaging bags, and activated carbon storage box doors, respectively. The highest confidence level for that category is taken from all detection frames.

[0123] ;

[0124] The overall target detection confidence score is calculated using a weighted average:

[0125] ;

[0126] in, For the first The weight coefficients of the class targets satisfy the following conditions: , For single-class confidence threshold; This is an indicator function; its value is 1 if the condition is met, and 0 otherwise.

[0127] For the recent Frames Time-series smoothing of values:

[0128] ;

[0129] in, To smooth the window length, The time decay factor, This is the raw score for a single frame.

[0130] Motion matching score The following sub-indicators were weighted and combined to obtain the following:

[0131] ;

[0132] in, , , These are the weighting coefficients;

[0133] The trajectory integrity score is obtained through the following process:

[0134] The DeepSORT algorithm is used to continuously track the detected target and establish the hand trajectory. and activated carbon trajectory Define the internal area of ​​the activated carbon storage box. and external areas ,calculate:

[0135] ;

[0136] in, This indicates that the trajectory point falls within the region. The quantity within; =5, which is the threshold for valid trajectory points;

[0137] The above describes the use of the DeepSORT algorithm for continuous tracking of the detected target, including:

[0138] Assign a unique tracking ID to each detected target and record its complete motion trajectory:

[0139] ;

[0140] in, For the first The target center coordinates of the frame;

[0141] Determine the type of action by analyzing changes in trajectory direction:

[0142] If the hand trace is from the outer area of ​​the activated carbon storage box Move to the inner area of ​​the activated carbon storage box This is judged as a "penetration" action;

[0143] If the trajectory of the activated carbon packaging bag is from the internal area of ​​the activated carbon storage box Move to the outer area of ​​the activated carbon storage box. This is determined to be a "take out" action;

[0144] The region determination is achieved by comparing the coordinates of the detection frame center with the preset coordinates of the box region. Comparison implementation, The x-coordinate of the upper left corner of the box area. The vertical coordinate of the upper left corner of the box-shaped area. The x-coordinate of the lower right corner of the box area. This represents the ordinate of the lower right corner of the box-shaped area.

[0145] The motion direction matching score is obtained through the following process:

[0146] Calculate the hand displacement vector and activated carbon displacement vector , and the standard direction vector points into the box and pointing out of the box Calculate cosine similarity:

[0147] ;

[0148] The score for the reasonableness of the movement speed is obtained through the following process:

[0149] Calculate average speed The evaluation is performed using a piecewise function:

[0150] ;

[0151] in, For minimum speed, That is the maximum speed.

[0152] Action recognition probability score It is obtained through the following process:

[0153] Define standard action template sequence ,in For the first A standard state vector; actual observation sequence ; Calculate DTW cumulative distance:

[0154] ;

[0155] in, The distance is Euclidean; the normalized DTW matching score is:

[0156] ;

[0157] Action recognition probability score The result obtained through fusion:

[0158] ;

[0159] in, For the output of the neural network, To integrate weights, This represents the maximum normalized distance.

[0160] Before recording is triggered, a multi-dimensional false triggering suppression mechanism is activated, and the recording stage is entered only after the corresponding conditions are met.

[0161] The multi-dimensional false trigger suppression mechanism is as follows:

[0162] Object size filtering: The activated carbon packaging bag detection frame must meet the following requirements. , Aspect Ratio ; The width of the inspection frame for activated carbon packaging bags; The height of the detection frame for activated carbon packaging bags; The minimum effective width of the activated carbon packaging bag; This is the maximum effective width of the activated carbon packaging bag; This refers to the minimum effective height of the activated carbon packaging bag; This is the maximum effective height of the activated carbon packaging bag; The minimum aspect ratio threshold; The maximum aspect ratio threshold;

[0163] Action duration filtering: Total duration of the action sequence ; This is the minimum threshold for the action duration; This represents the maximum threshold for action duration.

[0164] Time window filtering: Only during preset permitted periods Internal triggering; The start time for changing the license period; This is the end time for the changed license period.

[0165] S3. During the trigger recording phase, the trigger score is used to determine whether the target action is successfully matched. If the match is successful, the formal recording is started.

[0166] This step uses a three-level triggering mechanism:

[0167] Pre-triggered: Detection of motion intensity in the ROI region Furthermore, when a human body is detected, the frame sampling rate is increased and a loop buffer is initiated; The set threshold for exercise intensity;

[0168] Among them, exercise intensity The calculation formula is:

[0169] ; This represents the total number of pixels within the ROI region. and The first Motion velocity components of each pixel;

[0170] Soft trigger: When an open door or a hand is detected entering the enclosure, and... At that time, mark the suspected replacement time. Lock the cache; This is the soft trigger threshold;

[0171] Hard trigger: When it is confirmed that the activated carbon has been removed and Immediately start recording and merge the cached video with the live video; This is the hard trigger threshold.

[0172] S4. During the continuous recording judgment stage, decide whether to continue or stop recording based on the activity on the screen, the status of the activated carbon storage box door, and the presence of personnel.

[0173] In this step, after recording starts, the screen continuously monitors for changes, and recording ends when the following stopping conditions are met:

[0174] Motion intensity within the region of interest (ROI) And it continues for longer than the static waiting time. Furthermore, the storage box door becomes closed, and no human target is detected within the ROI area, exceeding the waiting time for personnel to leave. .

[0175] S5. Data saving stage: Save video files and generate metadata containing timestamps, device information, and recognition results;

[0176] In this step, the H.265 / HEVC encoding standard is used, and the bitrate is adaptively adjusted according to the complexity of the image. :

[0177] ;

[0178] in, For the set low bitrate, For the set medium bitrate, For setting a high bitrate;

[0179] Simultaneously generate a low-bitrate thumbnail video, reducing the bitrate. for:

[0180] ;

[0181] in, The compression factor is 1.

[0182] S6. During the data upload phase, the video recording files and metadata are uploaded to the remote monitoring platform via the network.

[0183] S7. During the remote viewing phase, managers can view the monitoring screen or retrieve and play back historical recordings in real time through the platform.

[0184] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for monitoring activated carbon replacement, characterized in that, include: During the real-time monitoring phase, video footage of the activated carbon storage area is continuously collected and analyzed in real time at a set frame sampling rate. In the action recognition stage, when the motion intensity in the region of interest exceeds the threshold, the frame sampling rate is increased, and a multi-level deep learning algorithm is run to identify the action sequence of activated carbon being taken out of the storage box and to obtain the trigger score. During the recording phase, the trigger score is used to determine whether the target action is successfully matched. If the match is successful, the recording begins. During the continuous recording and judgment phase, the decision to continue or stop recording is made based on the activity on the screen, the status of the activated carbon storage box door, and the presence of personnel. During the data storage phase, video files are saved and metadata containing timestamps, device information, and recognition results is generated; During the data upload phase, the video recording files and metadata are uploaded to the remote monitoring platform via the network; During the remote viewing phase, managers can view the monitoring footage in real time or retrieve and play back historical recordings through the platform.

2. The activated carbon replacement monitoring method according to claim 1, characterized in that, The multi-level deep learning algorithm used in the action recognition stage includes: Object detection: The YOLO algorithm is used to detect human bodies, hands, activated charcoal packaging bags, and activated charcoal storage box doors in video frames, and outputs a set of detection boxes. Each detection box include: 1) Detection box coordinates: , The coordinates of the detection box center are, and These are the width and height, respectively. 2) Detection confidence level: ,in The probability that a target exists in the region. The conditional probability of correctly identifying the target category; 3) Category probability vector: ,in , For the first The first detection box Logit output for each category, =5, representing the number of target categories, corresponding to human body, hand, activated carbon packaging bag, storage box door, and background; and satisfying the following conditions: ; Motion analysis: The target trajectory is calculated using the optical flow method by minimizing the energy function. Solving the optical flow field : ； in, The images are respectively in direction, Direction and Time Partial derivatives in direction, and They are respectively and The velocity component of the direction of motion, For smoothing weight parameters, and The spatial gradient of the optical flow; the motion matching score is calculated based on the optical flow field. ; Action recognition: Using a temporal convolutional network (TCN) or a long short-term memory network (LSTM), the output is an action recognition probability score. ; Calculate the trigger score : ; in, The target detection confidence score is derived from the detection confidence score. Aggregation yields; Let be the weighting coefficient, satisfying .

3. The activated carbon replacement monitoring method according to claim 2, characterized in that, Calculate the target detection confidence score The process includes: For each type of target These correspond to the human body, hands, activated carbon packaging bags, and activated carbon storage box doors, respectively. The highest confidence level for that category is taken from all detection frames. ； The overall target detection confidence score is calculated using a weighted average: ； in, For the first The weight coefficients of the class targets satisfy the following conditions: , For single-class confidence threshold; This is an indicator function; its value is 1 if the condition is met, and 0 otherwise. For the recent Frames Time-series smoothing of values: ； in, To smooth the window length, The time decay factor, This is the raw score for a single frame.

4. The activated carbon replacement monitoring method according to claim 2, characterized in that, The motion matching score The following sub-indicators were weighted and combined to obtain the following: ； in, , , These are the weighting coefficients; The trajectory integrity score is obtained through the following process: The DeepSORT algorithm is used to continuously track the detected target and establish the hand trajectory. and activated carbon trajectory Define the internal area of ​​the activated carbon storage box. and external areas ,calculate: ； in, This indicates that the trajectory point falls within the region. The quantity within; =5, which is the threshold for valid trajectory points; The motion direction matching score is obtained through the following process: Calculate the hand displacement vector and activated carbon displacement vector , and the standard direction vector points into the box and pointing out of the box Calculate cosine similarity: ； The score for the reasonableness of the movement speed is obtained through the following process: Calculate average speed A piecewise function is used for evaluation: ； in, For minimum speed, That is the maximum speed.

5. The activated carbon replacement monitoring method according to claim 2, characterized in that, The action recognition probability score It is obtained through the following process: Define standard action template sequence ,in For the first A standard state vector; actual observation sequence ; Calculate DTW cumulative distance: ； in, The distance is Euclidean; the normalized DTW matching score is: ； Action recognition probability score The result obtained through fusion: ； in, For the output of the neural network, To integrate weights, This represents the maximum normalized distance.

6. A method for monitoring activated carbon replacement according to any one of claims 2-5, characterized in that, The recording phase employs a three-level triggering mechanism: Pre-triggered: Detection of motion intensity in the ROI region Furthermore, when a human body is detected, the frame sampling rate is increased and a loop buffer is initiated; The set threshold for exercise intensity; Among them, exercise intensity The calculation formula is: ; This represents the total number of pixels within the ROI region. and The first Motion velocity components of each pixel; Soft trigger: When an open door or a hand is detected entering the enclosure, and... At that time, mark the suspected replacement time. Lock the cache; This is the soft trigger threshold; Hard trigger: When it is confirmed that the activated carbon has been removed and Immediately start recording and merge the cached video with the live video; This is the hard trigger threshold.

7. The activated carbon replacement monitoring method according to claim 6, characterized in that, During the continuous recording judgment phase, after recording starts, the screen changes are continuously monitored, and recording ends when the following stopping conditions are met: Motion intensity within the region of interest (ROI) And it continues for longer than the static waiting time. Furthermore, the storage box door becomes closed, and no human target is detected within the ROI area, exceeding the waiting time for personnel to leave. .

8. The activated carbon replacement monitoring method according to claim 2, characterized in that, Before recording is triggered, a multi-dimensional false triggering suppression mechanism is activated, and the recording stage is entered only after the corresponding conditions are met. The multi-dimensional false trigger suppression mechanism is as follows: Object size filtering: The activated carbon packaging bag detection frame must meet the following requirements. , Aspect Ratio ; The width of the inspection frame for activated carbon packaging bags; The height of the detection frame for activated carbon packaging bags; The minimum effective width of the activated carbon packaging bag; This is the maximum effective width of the activated carbon packaging bag; This refers to the minimum effective height of the activated carbon packaging bag; This is the maximum effective height of the activated carbon packaging bag; The minimum aspect ratio threshold; The maximum aspect ratio threshold; Action duration filtering: Total duration of the action sequence ; This is the minimum threshold for the action duration; This represents the maximum threshold for action duration. Time window filtering: Only during preset permitted periods Internal triggering; The start time for changing the license period; This is the end time for the changed license period.

9. The activated carbon replacement monitoring method according to claim 4, characterized in that, The DeepSORT algorithm is used to continuously track the detected target, including: Assign a unique tracking ID to each detected target and record its complete motion trajectory: ； in, For the first The target center coordinates of the frame; Determine the type of action by analyzing changes in trajectory direction: If the hand trace is from the outer area of ​​the activated carbon storage box Move to the inner area of ​​the activated carbon storage box This is judged as a "penetration" action; If the trajectory of the activated carbon packaging bag is from the internal area of ​​the activated carbon storage box Move to the outer area of ​​the activated carbon storage box. This is determined to be a "take out" action; The region determination is achieved by comparing the coordinates of the detection frame center with the preset coordinates of the box region. Comparison implementation, The x-coordinate of the upper left corner of the box area. The vertical coordinate of the upper left corner of the box-shaped area. The x-coordinate of the lower right corner of the box area. This represents the ordinate of the lower right corner of the box-shaped area.

10. The activated carbon replacement monitoring method according to claim 1, characterized in that, The data storage stage uses the H.265 / HEVC encoding standard, and the bitrate is adaptively adjusted according to the complexity of the image. : ; in, For the set low bitrate, For the set medium bitrate, For setting a high bitrate; Simultaneously generate a low-bitrate thumbnail video, reducing the bitrate. for: ； in, The compression factor is 1.

Images