A video surveillance system and a video surveillance method

Abstract

The present invention discloses a video surveillance system and a video surveillance method, belonging to the technical field of video surveillance, including a video surveillance module, an image processing module, a region association module, an abnormal behavior recognition module, a parameter storage module, a target determination module, a controller, and a warning module. The advantages of the present invention are as follows: it can comprehensively evaluate the characteristics of different types of abnormal behaviors to generate strong warning signals or weak warning signals; and it can make up for the problem of low accuracy of the deep learning algorithm in the initial stage and gradually improve the accuracy of the deep learning algorithm for the recognition of abnormal behaviors.

Description

Technical Field

[0001] The present invention relates to the technical field of video surveillance. Background Art

[0002] At present in China, video surveillance has been basically realized in public places with dense crowds such as shopping malls, railway stations, and subway stations. However, the current video surveillance systems can often only perform the functions of shooting and storing, and cannot capture and detect abnormal behaviors in a timely manner. At present, there have been research reports on applying AI to the scenario of abnormal behavior recognition. However, due to the complexity of human behaviors and the variability of scenarios, relying solely on AI recognition technology has problems such as low accuracy, inability to comprehensively evaluate and give early warnings for specific scenarios and behaviors. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a video surveillance system that can comprehensively evaluate the characteristics of different types of abnormal behaviors to generate strong or weak early warning signals; and can make up for the problem of low accuracy of the deep learning algorithm in the initial stage and gradually improve the accuracy of the deep learning algorithm for abnormal behavior recognition.

[0004] To solve the above technical problem, the technical solution of the present invention is a video surveillance system, including a video surveillance module, an image processing module, a region association module, an abnormal behavior recognition module, a parameter storage module, a target determination module, a controller, and an early warning module;

[0005] The video surveillance module is used to obtain real-time video streams of their respective monitored areas through cameras;

[0006] The image processing module is used to split the real-time video stream into continuous frames to obtain each frame of image and perform image preprocessing;

[0007] The region association module is used to associate different cameras according to preset rules, and the mutually associated cameras form an associated region;

[0008] The abnormal behavior recognition module is used to extract and match features of each frame of image processed by the image processing module by using a trained convolutional neural network model, determine whether there is an abnormal behavior in the image. If there is an abnormal behavior, it will output the type of abnormal behavior, the occurrence probability p, and send the abnormal image to the target determination module;

[0009] The parameter storage module is used to store the risk coefficient k, monitoring period T, critical abnormal time t cv , critical abnormal value U cv and critical probability P cv ;

[0010] The target determination module is used to determine whether the target person can be confirmed based on the features extracted from the abnormal image; if not, it retrieves the images captured by all cameras in the associated area within the monitoring period T according to the time and location information of the abnormal image, matches them with the abnormal image, confirms the target person, and sends the result to the controller;

[0011] The controller is used to calculate the duration t of the abnormal behavior based on all the abnormal images received for the target person, calculate the abnormal value U, where U = p * k * t; according to the relationships among p, t, U, P cv 、t cv and U cv to determine whether to send a strong warning signal or a weak warning signal to the warning module.

[0012] Furthermore, if t ≥ t cv or p ≥ P cv or U ≥ U cv then send a strong warning signal to the warning module; otherwise, send a weak warning signal to the warning module.

[0013] Furthermore, the features extracted from the abnormal image include facial features, body features, clothing features, and gender features.

[0014] Furthermore, when a weak warning signal is generated, the controller is also used to send all the videos containing the target person captured by all cameras in the associated area within the monitoring period T to the cloud management platform, so that the management personnel can judge whether it is necessary to correct the recognition result of the abnormal behavior and feedback the recognition result to the controller.

[0015] Furthermore, it also includes a correction times statistics module; the correction times statistics module is used to count the correction times n corresponding to each type of abnormal behavior. The statistical method of the correction times n is as follows: if the recognition result received by the controller is to correct the abnormal behavior of type A to B, then the correction times for the abnormal behavior of type A is -1, and the correction times for the abnormal behavior of type B is +1; if the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times for the abnormal behavior of type A is -1; the controller judges whether it is necessary to adjust the critical abnormal time tcv, critical abnormal value Ucv, and critical probability Pcv corresponding to each type of abnormal behavior according to the correction times n of each type of abnormal behavior; if adjustment is needed, the correction times n of this type of abnormal behavior is cleared after each adjustment.

[0016] Furthermore, the critical abnormal time t cv 、critical abnormal value U cv and critical probability P cvThe adjustment method is as follows: when the correction times n of a certain type of abnormal behavior is positive and reaches or exceeds the preset value N1, the critical abnormal time t corresponding to this type of abnormal behavior is reduced. cv , the critical abnormal value U cv and the critical probability P cv ; when the correction times n of a certain type of abnormal behavior is negative and reaches or exceeds the preset value N2, the critical abnormal time t corresponding to this type of abnormal behavior is increased. cv , the critical abnormal value U cv and the critical probability P cv .

[0017] Furthermore, when adjusting the values of the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv , the absolute value of the initial adjustment amplitude is set manually. In subsequent adjustments, if the direction of this adjustment is the same as that of the previous adjustment, the absolute value of this adjustment amplitude remains unchanged. If the direction of this adjustment is different from that of the previous adjustment, the absolute value of this adjustment amplitude becomes half of the absolute value of the previous adjustment amplitude.

[0018] When the correction times statistical module has not reached or exceeded the preset N1 and N2 for the correction times n corresponding to each type of abnormal behavior within a certain period of time, the manual correction behavior in the weak warning signal state can be stopped.

[0019] Furthermore, it also includes a storage module; when it is necessary to correct the recognition result of abnormal behavior, the controller sends the corresponding abnormal image and correction result to the storage module for storage and feeds them as new data to the convolutional neural network model for learning.

[0020] The present invention also provides a video surveillance method, including the following steps:

[0021] S1. Obtain images: The image processing module splits the real-time video stream obtained by the video surveillance module into continuous frames to obtain each frame of image, and preprocesses the image.

[0022] S2. Identify abnormal behavior: The abnormal behavior recognition module uses the trained convolutional neural network model to extract features and match each frame of preprocessed image to determine whether there is abnormal behavior in the image. If there is abnormal behavior, it will output the type and occurrence probability p of the abnormal behavior, and send the abnormal image to the target determination module; the features extracted from the abnormal image include facial features, body features, clothing features, and gender features.

[0023] S3. Confirm the target person: The target determination module determines whether the target person can be confirmed based on the features extracted from the abnormal image. If not, it retrieves the images captured by all cameras in the associated area within the monitoring period T according to the time and location information of the abnormal image, matches them with the abnormal image to confirm the target person, and sends the result to the controller.

[0024] S4. Generate a warning signal: The controller calculates the duration t of the abnormal behavior based on all the abnormal images received for the target person, calculates the abnormal value U, where U = p * k * t. According to the relationships among p, t, U, P cv , t cv and U cv , it determines whether to send a strong warning signal or a weak warning signal to the warning module; k, P cv , t cv and U cv are respectively the risk coefficient, critical probability, critical abnormal time, and critical abnormal value corresponding to each type of abnormal behavior. If t ≥ t cv or p ≥ P cv or U ≥ U cv then it sends a strong warning signal to the warning module; otherwise, it sends a weak warning signal to the warning module.

[0025] S5. Warning: The warning module issues a warning based on the received warning signal.

[0026] Furthermore, after step S5, the following steps are also included:

[0027] S6. Result correction: When a weak warning signal is generated, the controller sends all the videos containing the target person captured by all cameras in the associated area within the monitoring period T to the cloud management platform. The management personnel determine whether it is necessary to correct the recognition result of the abnormal behavior and feedback the recognition result to the controller.

[0028] S7. Update the database: When it is necessary to correct the recognition result of the abnormal behavior, the controller sends the corresponding abnormal image and the correction result to the storage module for storage and feeds them as new data to the convolutional neural network model for learning.

[0029] S8. Correction times statistics: The correction times statistics module counts the correction times n corresponding to each type of abnormal behavior. The statistical method for the correction times n is as follows: If the recognition result received by the controller is to correct the abnormal behavior of type A to B, then the correction times for the abnormal behavior of type A are decreased by 1, and the correction times for the abnormal behavior of type B are increased by 1. If the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times for the abnormal behavior of type A are decreased by 1.

[0030] S9. Adjustment Parameters: The controller determines whether to adjust the critical abnormal time $t_{cv}$, the critical abnormal value $U_{cv}$, and the critical probability $P_{cv}$ corresponding to each type of abnormal behavior based on the number of corrections $n$ for each type of abnormal behavior. If adjustment is required, the number of corrections $n$ for that type of abnormal behavior is cleared after each adjustment. When the number of corrections $n$ for a certain type of abnormal behavior is positive and reaches or exceeds the preset value $N1$, the critical abnormal time $t$ corresponding to that type of abnormal behavior is reduced. cv cv cv cv cv cv; When the number of corrections $n$ for a certain type of abnormal behavior is negative and reaches or exceeds the preset value $N2$, the critical abnormal time $t$ corresponding to that type of abnormal behavior is increased. cv cv cv cv cv .

[0031] Further, when adjusting the values of the critical abnormal time $t$ cv cv cv cv cv and the critical probability $P$cv, the absolute value of the initial adjustment amplitude is set manually. In subsequent adjustments, if the direction of this adjustment is the same as the previous adjustment, the absolute value of this adjustment amplitude remains unchanged. If the direction of this adjustment is different from the previous adjustment, the absolute value of this adjustment amplitude becomes half of the absolute value of the previous adjustment amplitude.

[0032] When the correction number statistics module has not reached or exceeded the preset $N1$ and $N2$ for the correction number $n$ corresponding to each abnormal behavior type within a certain period of time, steps S6 - S9 can be cancelled.

[0033] Advantages of the Invention:

[0034] The present invention uses a deep learning algorithm to identify whether there is abnormal behavior in the monitored image, and comprehensively evaluates it in combination with the characteristics of different types of abnormal behavior to generate a strong warning signal or a weak warning signal. For some abnormal behaviors, manual correction can be combined. According to the correction results, the system is continuously optimized. On the one hand, it can make up for the problem of low accuracy of the deep learning algorithm in the initial stage and gradually improve the accuracy of the deep learning algorithm for identifying abnormal behavior. On the other hand, it can also cancel the manual correction behavior after the number of corrections is stable in the later stage, reducing the labor intensity. Brief Description of the Drawings

[0035] Figure 1 It is a structural block diagram of Embodiment 1. Detailed Embodiment

[0036] The following further describes the specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted here that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation on the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. Embodiment

[0037] A video surveillance system, as Figure 1 shown, includes a video surveillance module, an image processing module, a region association module, an abnormal behavior recognition module, a parameter storage module, a target determination module, a controller, and a warning module;

[0038] The video surveillance module is used to obtain the real-time video stream of its respective surveillance area through a camera;

[0039] The image processing module is used to split the real-time video stream into continuous frames to obtain each frame of image and perform image preprocessing; in this embodiment, the image preprocessing includes image denoising and enhancement;

[0040] The region association module is used to associate different cameras according to a preset rule, and the mutually associated cameras form an associated region; the preset rule can be set according to the geographical location;

[0041] The abnormal behavior recognition module is used to extract and match the features of each frame of image processed by the image processing module by using a trained convolutional neural network model, judge whether there is an abnormal behavior in the image, and if there is an abnormal behavior, output the type of the abnormal behavior, the occurrence probability p, and send the abnormal image to the target determination module;

[0042] The parameter storage module is used to store the risk coefficient k, the monitoring period T, the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv ;

[0043] The target determination module is used to judge whether the target person can be confirmed according to the features extracted from the abnormal image; if not, it retrieves the images taken by all cameras in the associated region within the monitoring period T according to the time and position information of the abnormal image, and matches them with the abnormal image to confirm the target person, and sends the result to the controller; just relying on an abnormal picture taken by a certain camera may not be enough to determine the target person. At this time, calling the images of other cameras can confirm the target person;

[0044] The controller is used to calculate the duration t of the abnormal behavior according to all the abnormal images received for the target person, calculate the abnormal value U, U = p * k * t; according to p, t, U and Pcv , t cv and U cv Based on the relationships among t, p, and U, determine whether to send a strong warning signal or a weak warning signal to the warning module. The calculation method for the duration t of the abnormal behavior is to summarize the abnormal images captured by each camera. The time difference between the time information of the first and the last abnormal images is the time of the abnormal behavior captured by that camera. Summing up the times of the abnormal behaviors captured by all cameras in the associated area for the target person can obtain the duration t of the abnormal behavior.

[0045] Furthermore, if t ≥ t cv or p ≥ P cv or U ≥ U cv then send a strong warning signal to the warning module; otherwise, send a weak warning signal to the warning module.

[0046] Furthermore, the features extracted from the abnormal images include facial features, body features, clothing features, and gender features.

[0047] Furthermore, when generating a weak warning signal, the controller is also used to send all the videos containing the target person captured by all cameras in the associated area during the monitoring period T to the cloud management platform, so that the management personnel can judge whether it is necessary to correct the recognition result of the abnormal behavior and feedback the recognition result to the controller.

[0048] Furthermore, it also includes a correction times statistics module. The correction times statistics module is used to count the correction times n corresponding to each type of abnormal behavior. The statistical method for the correction times n is as follows: If the recognition result received by the controller is to correct the abnormal behavior of type A to B, then the correction times for the abnormal behavior of type A are -1, and the correction times for the abnormal behavior of type B are +1; If the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times for the abnormal behavior of type A are -1. The controller judges whether it is necessary to adjust the critical abnormal time tcv, the critical abnormal value Ucv, and the critical probability Pcv corresponding to this type of abnormal behavior according to the correction times n of each type of abnormal behavior. If adjustment is needed, then clear the correction times n of this type of abnormal behavior after each adjustment.

[0049] Furthermore, the adjustment method for the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv is as follows: When the correction times n of a certain type of abnormal behavior are positive and reach or exceed the preset value N1, then reduce the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv; When the number of corrections n for a certain type of abnormal behavior is negative and reaches or exceeds the preset value N2, the critical abnormal time t corresponding to this type of abnormal behavior is increased. cv the critical abnormal value U cv and the critical probability P cv .

[0050] Furthermore, when adjusting the values of the critical abnormal time t cv the critical abnormal value U cv and the critical probability P cv , the absolute value of the initial adjustment amplitude is set manually. In subsequent adjustments, if the direction of this adjustment is the same as the previous adjustment, the absolute value of this adjustment amplitude remains unchanged. If the direction of this adjustment is different from the previous adjustment, the absolute value of this adjustment amplitude becomes half of the absolute value of the previous adjustment amplitude.

[0051] In a certain embodiment, there are a total of five types of abnormal behavior, namely A, B, C, D, and E. The value of N1 is set to 5, and the value of N2 is set to -5. For the critical abnormal time t cv the critical abnormal value U cv and the critical probability P cv of the abnormal behavior of a given type, the absolute values of the adjustment amplitudes are Ht, Hu, and Hp respectively. When the number of corrections for type A abnormal behavior in the correction number statistics module reaches 5, the values of the critical abnormal time t cv the critical abnormal value U cv and the critical probability P cv are reduced, t cv = t cv - Ht, U cv = U cv - Hu, P cv = P cv - Hp; If after a period of time, the number of corrections for type A abnormal behavior in the correction number statistics module reaches 5 again, the values of the critical abnormal time t cv the critical abnormal value U cv and the critical probability P cv are continued to be reduced, t cv = t cv - Ht, U cv = U cv - Hu, P cv = P cv - Hp; If after a period of time, the number of corrections for type A abnormal behavior in the correction number statistics module reaches -5, the values of the critical abnormal time t cv the critical abnormal value U cv and the critical probability P cv are increased, t cv = tcv +Ht / 2, U cv =U cv +Hu / 2, P cv =P cv +Hp / 2.

[0052] When the correction times statistical module has not reached or exceeded the preset N1 and N2 for the correction times n corresponding to each abnormal behavior type within a certain period of time, it indicates that the recognition result of the convolutional neural network model for abnormal behaviors is accurate enough. At this time, the manual correction behavior in the weak warning signal state can be stopped.

[0053] Furthermore, it also includes a storage module; when it is necessary to correct the recognition result of an abnormal behavior, the controller sends the corresponding abnormal image and the correction result to the storage module for storage, and feeds them as new data to the convolutional neural network model for learning. By manually judging the abnormal behaviors presenting weak warning signals and determining whether correction is needed according to the judgment results, the recognition accuracy of the convolutional neural network model can be gradually improved. Embodiment

[0054] The present invention also provides a video monitoring method, including the following steps:

[0055] S1. Obtain images: The image processing module splits the real-time video stream obtained by the video monitoring module into continuous frames to obtain each frame of image, and preprocesses the image;

[0056] S2. Identify abnormal behaviors: The abnormal behavior recognition module uses the trained convolutional neural network model to extract features and match each preprocessed frame of image, and determines whether there are abnormal behaviors in the image. If there are abnormal behaviors, it will output the type and occurrence probability p of the abnormal behavior, and send the abnormal image to the target determination module; the features extracted from the abnormal image include facial features, body features, clothing features, and gender features;

[0057] S3. Confirm the target person: The target determination module determines whether the target person can be confirmed according to the features extracted from the abnormal image; if not, it will retrieve the images captured by all cameras in the associated area within the monitoring period T according to the time and position information of the abnormal image, and match them with the abnormal image to confirm the target person, and send the result to the controller;

[0058] S4. Generate warning signals: The controller calculates the duration t of the abnormal behavior based on all the abnormal images received for the target person, calculates the abnormal value U, U = p * k * t; according to the relationships among p, t, U and P cv , t cv and U cv to determine whether to send a strong warning signal or a weak warning signal to the warning module; k, Pcv , t cv and U cv are the risk coefficient, critical probability, critical abnormal time, and critical abnormal value corresponding to each type of abnormal behavior respectively; if t ≥ t cv or p ≥ P cv or U ≥ U cv then a strong warning signal is sent to the warning module; otherwise, a weak warning signal is sent to the warning module;

[0059] S5. Warning: The warning module gives a warning according to the received warning signal;

[0060] S6. Result correction: When a weak warning signal is generated, the controller sends all videos containing the target person of all cameras in the associated area within the monitoring period T to the cloud management platform, and the management personnel judge whether it is necessary to correct the recognition result of the abnormal behavior and feedback the recognition result to the controller;

[0061] S7. Update database: When it is necessary to correct the recognition result of the abnormal behavior, the controller sends the corresponding abnormal image and correction result to the storage module for storage and feeds them as new data to the convolutional neural network model for learning;

[0062] S8. Correction times statistics: The correction times statistics module counts the correction times n corresponding to each type of abnormal behavior. The statistical method of the correction times n is as follows: If the recognition result received by the controller is to correct the abnormal behavior of type A to B, then the correction times for the abnormal behavior of type A are -1, and the correction times for the abnormal behavior of type B are +1; If the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times for the abnormal behavior of type A are -1;

[0063] S9. Parameter adjustment: The controller judges whether it is necessary to adjust the critical abnormal time tcv, critical abnormal value Ucv, and critical probability Pcv corresponding to this type of abnormal behavior according to the correction times n of each type of abnormal behavior; If adjustment is needed, the correction times n of this type of abnormal behavior are cleared after each adjustment; When the correction times n of a certain type of abnormal behavior are positive and reach or exceed the preset value N1, then reduce the critical abnormal time t cv , critical abnormal value U cv and critical probability P cv ; When the correction times n of a certain type of abnormal behavior are negative and reach or exceed the preset value N2, then increase the critical abnormal time t cv , critical abnormal value U cv and critical probability P cv .

[0064] When the correction times statistical module has not reached or exceeded the preset N1 and N2 for the correction times n corresponding to each abnormal behavior type within a certain period of time, steps S6 - S9 can be cancelled.

[0065] Adjust the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv When adjusting the numerical values, the absolute value of the initial adjustment amplitude is set manually. In subsequent adjustments, if the direction of this adjustment is the same as that of the previous adjustment, the absolute value of this adjustment amplitude remains unchanged. If the direction of this adjustment is different from that of the previous adjustment, the absolute value of this adjustment amplitude becomes half of the absolute value of the previous adjustment amplitude.

[0066] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principles and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments still fall within the protection scope of the present invention.

Claims

1. A video surveillance system, characterized in that: It includes a video monitoring module, an image processing module, a region association module, an abnormal behavior recognition module, a parameter storage module, a target determination module, a controller, and a warning module; The video monitoring module is used to obtain real-time video streams of their respective monitoring areas through cameras; The image processing module is used to split the real-time video stream into continuous frames to obtain each frame of image and perform image preprocessing; The region association module is used to associate different cameras according to preset rules, and the mutually associated cameras form an associated region; The abnormal behavior recognition module is used to extract features and match each frame of image processed by the image processing module by using a trained convolutional neural network model, determine whether there is an abnormal behavior in the image. If there is an abnormal behavior, it will output the type of the abnormal behavior, the occurrence probability p, and send the abnormal image to the target determination module; The parameter storage module is used to store the risk coefficient k, monitoring period T, critical abnormal time t corresponding to each type of abnormal behavior cv , critical abnormal value U cv and critical probability P cv ; The target determination module is used to determine whether the target person can be confirmed according to the features extracted from the abnormal image; if not, it will retrieve the images captured by all cameras in the associated region within the monitoring period T according to the time and position information of the abnormal image, and match them with the abnormal image to confirm the target person, and send the result to the controller; The controller is used to calculate the abnormal behavior duration t based on all the abnormal images received for the target person, calculate the abnormal value U, where U = p * k * t; and determine whether to send a strong warning signal or a weak warning signal to the warning module according to the relationships among p, t, U, and P cv , t cv and U cv ; If t ≥ t cv or p ≥ P cv or U ≥ U cv then send a strong warning signal to the warning module; otherwise, send a weak warning signal to the warning module; When a weak warning signal is generated, the controller is further used to send all the videos containing the target person captured by all cameras in the associated region within the monitoring period T to the cloud management platform, so that the management personnel can judge whether it is necessary to correct the recognition result of the abnormal behavior, and feedback the recognition result to the controller; It also includes a correction times statistics module; The correction times statistics module is used to count the correction times n corresponding to each type of abnormal behavior. The statistical method of the correction times n is as follows: if the recognition result received by the controller is to correct the abnormal behavior of type A to B, then the correction times of the abnormal behavior of type A are decreased by 1, and the correction times of the abnormal behavior of type B are increased by 1; if the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times of the abnormal behavior of type A are decreased by 1; the controller judges whether it is necessary to adjust the critical abnormal time tcv, the critical abnormal value Ucv, and the critical probability Pcv corresponding to this type of abnormal behavior according to the correction times n of each type of abnormal behavior; If adjustment is needed, the correction times n of this type of abnormal behavior are cleared after each adjustment; The critical abnormal time t cv 、the critical abnormal value U cv and the critical probability P cv are adjusted as follows: when the number of corrections n of a certain type of abnormal behavior is positive and reaches or exceeds the preset value N1, the critical abnormal time t cv 、the critical abnormal value U cv and the critical probability P cv corresponding to this type of abnormal behavior are decreased; when the number of corrections n of a certain type of abnormal behavior is negative and reaches or exceeds the preset value N2, the critical abnormal time t cv 、the critical abnormal value U cv and the critical probability P cv corresponding to this type of abnormal behavior are increased; when adjusting the values of the critical abnormal time t cv 、the critical abnormal value U cv and the critical probability P cv , the absolute value of the initial adjustment amplitude is set manually. In subsequent adjustments, if the direction of this adjustment is the same as that of the previous adjustment, the absolute value of the adjustment amplitude for this time remains unchanged; if the direction of this adjustment is different from that of the previous adjustment, the absolute value of the adjustment amplitude for this time becomes half of the absolute value of the previous adjustment amplitude.

2. The video monitoring system according to claim 1, wherein: The features extracted from the abnormal image include facial features, body features, clothing features, and gender features.

3. The video surveillance system according to claim 1, wherein: It also includes a storage module; when it is necessary to correct the recognition result of the abnormal behavior, the controller sends the corresponding abnormal image and the correction result to the storage module for storage, and feeds them as new data to the convolutional neural network model for learning.

4. A monitoring method applicable to the video monitoring system according to any one of claims 1-3, characterized in that: It includes the following steps: S1. Obtain images: The image processing module splits the real-time video stream obtained by the video monitoring module into continuous frames to obtain each frame of image and preprocesses the image; S2. Identify abnormal behavior: The abnormal behavior recognition module uses the trained convolutional neural network model to extract features and match each preprocessed frame of the image to determine whether there is abnormal behavior in the image. If there is abnormal behavior, the type of abnormal behavior and the occurrence probability p will be output, and the abnormal image will be sent to the target determination module. The features extracted from the abnormal image include facial features, body features, clothing features, and gender features. S3. Confirm the target person: The target determination module determines whether the target person can be confirmed based on the features extracted from the abnormal image. If not, it will retrieve the images captured by all cameras in the associated area within the monitoring period T according to the time and location information of the abnormal image, match them with the abnormal image, confirm the target person, and send the result to the controller. S4. Generate a warning signal: The controller calculates the duration t of the abnormal behavior based on all the abnormal images received for the target person, calculates the abnormal value U, where U = p * k * t; according to the relationships among p, t, U, and P cv , t cv and U cv , determines whether to send a strong warning signal or a weak warning signal to the warning module; k, P cv , t cv and U cv are respectively the risk coefficient, critical probability, critical abnormal time, and critical abnormal value corresponding to each type of abnormal behavior; if t ≥ t cv or p ≥ P cv or U ≥ U cv then send a strong warning signal to the warning module; Otherwise, send a weak warning signal to the warning module. S5. Warning: The warning module issues a warning based on the received warning signal.

5. The monitoring method of the video monitoring system according to claim 4, characterized in that: After step S5, the following steps are also included: S6. Result correction: When a weak warning signal is generated, the controller sends all the videos containing the target person captured by all cameras in the associated area within the monitoring period T to the cloud management platform. The management personnel will judge whether it is necessary to correct the recognition result of the abnormal behavior and feedback the recognition result to the controller. S7. Update the database: When it is necessary to correct the recognition result of the abnormal behavior, the controller sends the corresponding abnormal image and the correction result to the storage module for storage and feeds them as new data to the convolutional neural network model for learning. S8. Correction times statistics: The correction times statistics module counts the correction times n corresponding to each type of abnormal behavior. The statistical method for the correction times n is as follows: If the recognition result received by the controller is to correct the abnormal behavior of type A to type B, then the correction times for the abnormal behavior of type A are decreased by 1, and the correction times for the abnormal behavior of type B are increased by 1. If the recognition result received by the controller is to correct the abnormal behavior of type A to none, then the correction times for the abnormal behavior of type A are decreased by 1. S9. Adjust parameters: The controller determines whether it is necessary to adjust the critical abnormal time tcv, the critical abnormal value Ucv, and the critical probability Pcv corresponding to each type of abnormal behavior based on the correction times n of each type of abnormal behavior. If adjustment is needed, clear the correction count n of this type of abnormal behavior after each adjustment; when the correction count n of a certain type of abnormal behavior is positive and reaches or exceeds the preset value N1, reduce the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv ; when the correction count n of a certain type of abnormal behavior is negative and reaches or exceeds the preset value N2, increase the critical abnormal time t cv , the critical abnormal value U cv and the critical probability P cv .

Images