Perimeter intrusion hybrid detection method and perimeter intrusion hybrid detection system

By nesting the motion target analysis algorithm and neural network algorithm, combined with the anti-overreport detection method, the high false alarm and underreport rate problems in the existing boundary intrusion detection are solved, and high frequency, high accuracy and low cost are achieved detection effect.

CN113158800BActive Publication Date: 2024-11-01SHANGHAI YUNSAI ZHILIAN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110298172.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-03-19
Publication Date
2024-11-01
Estimated Expiration
2041-03-19

AI Technical Summary

Technical Problem

The existing periphery intrusion detection methods have high false alarm rates and missed alarm rates, especially in severe weather and lighting conditions, and the neural network technology based on video analysis is costly and slow in analysis.

Method used

Using nested detection methods, combined with motion target analysis algorithm and neural network algorithm, the initial judgment is made through the motion target analysis algorithm, the secondary judgment is made through the neural network algorithm, and the results are optimized through the anti-report detection method to reduce the false alarm rate and missed report Rate.

Benefits of technology

It realizes high frequency, high accuracy and low cost perimeter intrusion detection effects, reduces false alarms and missed alarms, and improves the real-time and accuracy of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113158800B_ABST
    Figure CN113158800B_ABST
Patent Text Reader

Abstract

The present invention relates to a hybrid perimeter intrusion detection method and a hybrid perimeter intrusion detection system. A hybrid perimeter intrusion detection method, characterized in that it comprises step one: a nested detection method: step a, using a motion target analysis algorithm to perform perimeter intrusion detection on a surveillance image of a perimeter, and initially determining whether there is an invasion; step b, if the initial determination is yes, using a neural network algorithm to perform perimeter intrusion detection on the image determined to be yes, and secondarily determining whether there is an invasion. Different from the prior art, in the hybrid perimeter intrusion detection method and hybrid perimeter intrusion detection system provided by the present invention, a motion target analysis algorithm and a neural network algorithm are nested and used in coordination, combining the advantages of the motion target analysis algorithm and the neural network algorithm: on the one hand, the neural network algorithm can more robustly detect specific intrusion targets of interest to the user - people, vehicles, animals, etc.; on the other hand, the motion target analysis algorithm has a fast detection speed and is sensitive to the target.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a perimeter intrusion hybrid detection method and a perimeter intrusion hybrid detection system. Background Art

[0002] Perimeter intrusion detection is a particularly important item in security prevention and control. The current main deployment methods are based on signal detection and video analysis detection.

[0003] Existing perimeter intrusion detection methods mainly include two categories:

[0004] (1) Physical method: Infrared sensing equipment is used to detect perimeter intrusion. This method has the advantages of strong real-time performance and is not affected by light, but it is prone to false alarms due to rain, snow, tree cover, etc. Many airports and railways have deployed this equipment as an auxiliary detection device in the early days, but due to the high false alarm rate and high frequency of alarm messages, it takes a lot of manpower to investigate in actual application.

[0005] (2) Visual method: Use video analysis methods to detect perimeter intrusions, such as algorithms based on motion target detection. This algorithm is fast and has good detection effects under good imaging conditions, but it has the following disadvantages: it cannot determine whether the detected target is a person or other object. When small animals (such as birds) or trees, flowers and plants pass in front of the perimeter camera lens, false alarms may occur. The algorithm is sensitive to conditions such as weather and lighting.

[0006] In addition, neural network technology based on video analysis has the advantages of high detection accuracy and the ability to identify a variety of targets, but it has the disadvantages of slow analysis speed, high latency, and easy underreporting. Moreover, the algorithm requires the use of a large number of high-configuration graphics cards, and the cost is dozens of times the price of ordinary video analysis hardware. Summary of the invention

[0007] One of the purposes of the present invention is to overcome at least one deficiency in the prior art and to provide a perimeter intrusion hybrid detection method and a perimeter intrusion hybrid detection system.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] According to a first aspect of the present invention, a perimeter intrusion hybrid detection method and a perimeter intrusion hybrid detection system are provided. The perimeter intrusion hybrid detection method and the perimeter intrusion hybrid detection system include:

[0010] A hybrid detection method for perimeter intrusion, characterized by comprising:

[0011] Step 1: Nested detection method:

[0012] Step a: Use a moving target analysis algorithm to perform perimeter intrusion detection on the monitoring image of the perimeter to initially determine whether there is an intrusion;

[0013] Step b: If the initial judgment is yes, a neural network algorithm is used to perform boundary intrusion detection on the image judged to be yes, and a second judgment is made as to whether there is intrusion.

[0014] Preferably, step a comprises:

[0015] Get the video stream from the perimeter camera and initialize the video stream image;

[0016] Use the motion target analysis algorithm to detect the boundary intrusion of the boundary image, initially determine whether it is an intrusion, and obtain the accuracy rate P m ;

[0017] The step b comprises:

[0018] The intrusion image judged as yes in step a is transmitted to the neural network algorithm, and the neural network algorithm performs boundary intrusion detection and tests the confidence Pri of the boundary box. dl ;

[0019] The neural network algorithm includes combining and determining the detection results using the following formula:

[0020] Confidence level P merge =(P m +Pri dl )*(1-C)|C∈(0,0.5), where C represents the sensitivity coefficient; Formula (I)

[0021] When the confidence level P merge ≥1, it is judged as yes, there is a target invading the perimeter; when P merge <1, judge as no.

[0022] Preferably, the neural network algorithm comprises:

[0023] Segment the intrusion image determined as yes in step a into a plurality of cells, each of the cells being responsible for detecting a target whose center point falls within the cell;

[0024] Each cell predicts several bounding boxes and the confidence level of the bounding boxes Pri dl , the confidence Where Pr represents the possibility that the bounding box contains the target; Indicates the accuracy of the bounding box, which is represented by the intersection of the predicted box and the actual box; Formula (2)

[0025] The confidence level Pri dlThe value of is between 0 and 1. The larger the value, the higher the credibility of the detected target.

[0026] Preferably, the moving target analysis algorithm includes:

[0027] Initialize a sample set for each pixel point of the bounding image, the sampling values ​​in the sample set are the historical pixel values ​​of the pixel point and the pixel values ​​of its neighborhood, and compare the pixel point sample sets of the previous and next frames to determine whether there is a moving target;

[0028] When the video stream is t=0, the background model formula of the pixel is expressed as:

[0029] BK M 0 =f 0 (x i ,y i )|(x i ,y i )∈N G (x,y), where BK M 0 Represents the mask of the current frame, f 0 (x i ,y i ) represents the current pixel value, N G (x, y) are adjacent pixel values, N G (x,y) is initialized N times and represented as N G (x i ,y i ); Formula (III)

[0030] After the image is initialized, the foreground target segmentation is performed on the image of the subsequent video stream t; when t = k, the mask of the pixel point (x, y) of the current frame is BK M k-1 (x,y), the corresponding pixel value is f 0 (x,y), the formula to determine whether it is the foreground is:

[0031] Where r is the number of initializations, and T is the threshold; Formula (IV)

[0032] When BK M k-1 (x r ,y r ) is less than or equal to T, f k (x,y) is the pixel of the mask;

[0033] When BK M k-1 (x r ,y r ) is greater than T, fk (x,y) is the foreground moving target pixel.

[0034] Preferably, the method further includes step 2: anti-missing detection method:

[0035] Step c: if the second judgment is no, determine whether the third judgment condition is met; if the third judgment condition is met, correct the coefficients in the neural network algorithm, and reuse the neural network algorithm to perform boundary intrusion detection on the detection image judged as no in step b, and determine whether there is intrusion three times.

[0036] Preferably, step c comprises:

[0037] According to the accuracy of the nested detection method, the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are obtained;

[0038] Establish a mixed probability model to prevent underreporting, which is the formula: Correct judgment percentage

[0039] The confidence level determined by the nested detection method is P merge , let the judgment formula be: Thr=P merge ×Pre dl ; Formula (VI)

[0040] When Thr≥0.5, the three judgment conditions are met, the sensitivity coefficient C in the judgment formula of formula (I) is reduced, and the neural network algorithm is used again to detect boundary intrusion for the detection image judged as no in the step b, and whether there is intrusion is judged three times.

[0041] Preferably, if the three judgments are yes, there is underreporting in the detection result of the nested detection method, and the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are adjusted.

[0042] Preferably, the false negative rate is assumed to occur once in M ​​detection samples, expressed as L m , update the TP parameters according to the following formula:

[0043] TP=TP+L m ×C step , where C step For L m The amplification factor, C step=(10,100,1000…), as L m The smaller the value of is, the larger it is. Formula (VII)

[0044] Preferably, the neural network algorithm is reused for perimeter intrusion detection only once per perimeter image frame.

[0045] A perimeter intrusion hybrid detection system, characterized by comprising:

[0046] A moving target analysis algorithm module (10), the moving target analysis algorithm module being configured to: use a moving target analysis algorithm to detect enclosure intrusion in real time on the surveillance image of the enclosure, and initially determine whether there is intrusion;

[0047] A neural network algorithm module (20), the neural network algorithm module being configured to: use a neural network algorithm to perform boundary intrusion detection on the image determined as yes by the moving target analysis algorithm module, and to make a secondary determination as to whether there is intrusion; and

[0048] The anti-missing detection module (30) is configured to: when the neural network algorithm module judges as no, judge whether three judgment conditions are met; if the three judgment conditions are met, correct the coefficients in the neural network algorithm so that the neural network algorithm module can re-use the neural network algorithm to detect boundary intrusion on the detection image judged as no, and judge whether there is intrusion three times.

[0049] Different from the prior art, in the hybrid perimeter intrusion detection method and hybrid perimeter intrusion detection system provided by the present invention, the motion target analysis algorithm and the neural network algorithm are nested and used in coordination, combining the advantages of the motion target analysis algorithm and the neural network algorithm: on the one hand, the neural network algorithm can more robustly detect the specific intrusion targets of interest to the user - people, vehicles, animals, etc.; on the other hand, the motion target analysis algorithm has a fast detection speed and is sensitive to the target. The hybrid perimeter intrusion detection method can achieve the ideal detection effect of high frequency, high accuracy and low cost.

[0050] The motion target analysis algorithm and the neural network algorithm are nested and used in conjunction with each other, and then an anti-missing detection method is used to perform an anti-missing detection analysis on the judgment of the neural network algorithm in the nested detection method. This can achieve an ideal detection effect with high frequency, high accuracy, low missing reports and low cost.

[0051] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a perimeter intrusion hybrid detection method provided by an embodiment of the present invention.

[0053] Figure 2 A determination flow chart of a perimeter intrusion hybrid detection method provided by an embodiment of the present invention.

[0054] Figure 3 A flow chart of anti-missing detection in a perimeter intrusion hybrid detection method provided by an embodiment of the present invention.

[0055] Figure 4 A schematic diagram of a hybrid perimeter intrusion detection system provided by an embodiment of the present invention.

[0056] Among them, 10: motion target analysis algorithm module; 20: neural network algorithm module; 30: anti-missing detection module; 32: anti-missing mixed probability model module. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with the accompanying drawings:

[0058] Embodiment 1:

[0059] According to one aspect of the present invention, a hybrid perimeter intrusion detection method is provided. Figure 1 As shown, the perimeter intrusion hybrid detection method includes the following steps:

[0060] Step 1: Nested detection method. Step 1 includes step a and step b.

[0061] In step a, a motion target analysis algorithm is used to perform perimeter intrusion detection on the monitoring image of the perimeter to initially determine whether there is an intrusion.

[0062] In step b, if the initial judgment is yes, a neural network algorithm is used to perform boundary intrusion detection on the image judged to be yes, and a secondary judgment is made as to whether there is intrusion.

[0063] The motion target analysis algorithm and the neural network algorithm in the hybrid detection method for perimeter intrusion in this embodiment are nested and used in coordination, combining the advantages of the motion target analysis algorithm and the neural network algorithm: on the one hand, the neural network algorithm can more robustly detect specific intrusion targets of interest to the user - people, vehicles, animals, etc.; on the other hand, the motion target analysis algorithm has a fast detection speed and is sensitive to targets. The hybrid detection method for perimeter intrusion can achieve an ideal detection effect with high frequency, high accuracy and low cost.

[0064] In this embodiment, Figure 2 As shown, step a may include the following steps:

[0065] First, obtain the video stream from the perimeter camera and initialize the video stream image.

[0066] Then, the motion target analysis algorithm is used to perform real-time boundary intrusion detection on the initialized boundary image to initially determine whether there is intrusion and obtain the accuracy rate P m The precision rate P m The moving target detection method can be tested using the test samples.

[0067] Step b may include the following steps:

[0068] The intrusion image judged as "yes" in step a is transmitted to the neural network algorithm, which performs boundary intrusion detection on it and obtains the confidence Pri of the bounding box dl .

[0069] The neural network algorithm includes the following formula to combine and judge the detection results:

[0070] Confidence P merge =(P m +Pri dl )*(1-C)|C∈(0,0.5), where C represents the sensitivity coefficient. Formula (I)

[0071] When the confidence level P merge ≥1, it is judged as yes, indicating that a target has invaded the perimeter. merge <1, the judgment is no, and the judgment of no can be further performed to prevent missed alarm detection.

[0072] Specifically, the moving target analysis algorithm may include the following steps:

[0073] Initialize a sample set for each pixel of the bounding image. The sample values ​​in the sample set are the historical pixel values ​​of the pixel and the pixel values ​​of its neighborhood. Comparing the pixel sample sets of the previous and next frames can determine whether there is a moving target. When the moving target analysis algorithm determines that it is, the result is passed to the neural network algorithm.

[0074] When the video stream is t=0, the background model formula of the pixel is expressed as:

[0075] BK M 0 =f 0 (x i ,y i )|(x i ,y i )∈N G (x,y), where BK M 0 Represents the mask of the current frame, f 0 (x i ,y i ) represents the current pixel value, N G(x, y) are adjacent pixel values, N G (x,y) is initialized N times and represented as N G (x i ,y i ). Formula (III)

[0076] After the image is initialized, the foreground object segmentation is performed on the image of the subsequent video stream t. When t = k, the mask of the pixel point (x, y) of the current frame is BK M k-1 (x,y), the corresponding pixel value is f 0 (x,y), the formula to determine whether it is the foreground is:

[0077] Where r is the number of initializations, and T is the threshold. Formula (IV)

[0078] When BK M k-1 When (x,y) is less than or equal to T, f k (x,y) is the pixel of the mask.

[0079] When BK M k-1 When (x,y) is greater than T, f k (x,y) is the foreground moving target pixel.

[0080] Specifically, neural network algorithms may include:

[0081] The intrusion image judged as yes in step a is divided into several cells, and each cell is responsible for detecting the targets whose center points fall within the cell.

[0082] Each cell predicts several bounding boxes and the confidence of the bounding boxes Pri dl .

[0083] Confidence Where Pr represents the possibility that the bounding box contains the target, Indicates the accuracy of the bounding box, which is represented by the intersection of the predicted box and the actual box. Formula (II)

[0084] The confidence level Pri dl The value of is between 0 and 1. The larger the value, the higher the credibility of the detected target.

[0085] In order to prevent the nested detection method from missing out on perimeter intrusion, preferably, the perimeter intrusion hybrid detection method further includes step 2: an anti-missing detection method. The anti-missing detection method includes:

[0086] Step c: If the second judgment is no, determine whether the third judgment condition is met.

[0087] If the three judgment conditions are met, the coefficients in the neural network algorithm are corrected, and the neural network algorithm is used again to perform boundary intrusion detection on the detection image judged as no in step b, and whether there is intrusion is judged three times.

[0088] If the three determination conditions are not met, then the determination of "Step 1: the second determination in the nested detection method is no" can be determined to be true.

[0089] The motion target analysis algorithm and the neural network algorithm are nested and used in conjunction with each other, and then an anti-missing detection method is used to perform an anti-missing detection analysis on the judgment of the neural network algorithm in the nested detection method. This can achieve an ideal detection effect with high frequency, high accuracy, low missing reports and low cost.

[0090] Specifically, step c may include the following steps:

[0091] Firstly, according to the nested detection method, a mixed probability model for preventing under-reporting is established, and a threshold is set in the mixed probability model for preventing under-reporting.

[0092] Then, it is determined whether the actual measured value reaches the threshold of the mixed probability model for preventing missed reports. When the actual measured value reaches the threshold of the mixed probability model for preventing missed reports, the three-time determination condition is met, the coefficient C in formula (I) is reduced, and the neural network algorithm is used again to detect the boundary intrusion for the detection image judged as no in step b, and whether it is intrusion is determined three times.

[0093] Furthermore, if Figure 3 As shown, step c may include the following steps:

[0094] According to the accuracy of the nested detection method, the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are obtained. More preferably, the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) can be obtained through N tests using test samples based on the accuracy of the prior nested detection method.

[0095] Establish a mixed probability model to prevent underreporting, which is the formula: Correct judgment percentage

[0096] The confidence level determined by the nested detection method is P merge , let the judgment formula be: Thr=P merge ×Pre dl .

[0097] Formula (VI)

[0098] The threshold value of the mixed probability model for preventing false negatives can be set to 0.5. When Thr ≥ 0.5, the three-time judgment condition is met, the sensitivity coefficient C in formula (I) is reduced, and the image judged as no in step b is re-used by the neural network algorithm to detect the boundary intrusion, and the intrusion is judged three times. Among them, each frame of the boundary image is re-used by the neural network algorithm to detect the boundary intrusion only once.

[0099] If the three judgments are yes, it can be determined that there are underreporting in the detection results of the nested detection method, and the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are adjusted accordingly to approach the optimal parameter value and achieve an ideal anti-underreporting detection effect.

[0100] More specifically, let the false negative rate be once in M ​​detection samples, expressed as L m The TP parameters can be updated according to the following formula:

[0101] TP=TP+L m ×C step , where C step For L m The amplification factor, C step =(10,100,1000…), as L m The smaller the value of is, the larger it is. Formula (VII)

[0102] Of course, in other embodiments, the TP parameters may also be adjusted according to other calculation methods; the update methods of TN (True Negative), FP (Fales Positive), and FN (False Negative) may be set similarly or in other ways, which will not be elaborated here.

[0103] Embodiment 2:

[0104] According to the same inventive concept, the present invention also provides a perimeter intrusion hybrid detection system, which is used to implement the perimeter intrusion hybrid detection method in the first embodiment. Figure 4 As shown, the perimeter intrusion hybrid detection system includes a motion target analysis algorithm module 10, a neural network algorithm module 20 and an anti-missing detection module 30.

[0105] The moving target analysis algorithm module 10 is used to: use the moving target analysis algorithm to detect the perimeter intrusion in real time on the monitoring image of the perimeter, and initially determine whether there is an intrusion.

[0106] The neural network algorithm module 20 is used to: use the neural network algorithm to perform boundary intrusion detection on the image determined as yes by the motion target analysis algorithm module 10, and make a secondary determination as to whether there is intrusion.

[0107] The anti-missing detection module 30 is used to: when the neural network algorithm module 20 judges as no, judge whether the three-time judgment condition is met; if the three-time judgment condition is met, correct the coefficients in the neural network algorithm so that the neural network algorithm module 20 can use the neural network algorithm to detect the boundary intrusion again for the detection image judged as no, and judge whether there is intrusion three times. The anti-missing detection module 30 may include an anti-missing mixed probability model module 32, which is used to implement the anti-missing detection method in the first embodiment.

[0108] The above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modification, equivalent substitution or improvement within the spirit of the present invention is included in the scope of the claims of the present invention.

Claims

1. A hybrid perimeter intrusion detection method, characterized in that: include Step 1: Nested detection method: Step a: Use a moving target analysis algorithm to perform perimeter intrusion detection on the monitoring image of the perimeter to initially determine whether there is an intrusion; Step b: If the initial judgment is yes, use a neural network algorithm to perform boundary intrusion detection on the image judged to be yes, and make a secondary judgment on whether it is an intrusion; The step a comprises: Get the video stream from the perimeter camera and initialize the video stream image; Use the motion target analysis algorithm to detect the boundary intrusion of the boundary image, initially determine whether it is an intrusion, and obtain the accuracy rate P m ; The moving target analysis algorithm includes: Initialize a sample set for each pixel point of the bounding image, the sampling values ​​in the sample set are the historical pixel values ​​of the pixel point and the pixel values ​​of its neighborhood, and compare the pixel point sample sets of the previous and next frames to determine whether there is a moving target; When the video stream is t=0, the background model of the pixel is expressed as the third formula: BK M 0 =f 0 (x i ,y i )|(x i ,y i )∈N G (x,y) where BK M 0 Represents the mask of the current frame, f 0 (x i ,y i ) represents the current pixel value, N G (x, y) are adjacent pixel values, N G (x,y) is initialized N times and represented as N G (x i ,y i ); After the image is initialized, the foreground target segmentation is performed on the image of the subsequent video stream t; when t = k, the mask of the pixel point (x, y) of the current frame is BK M k-1 (x,y), the corresponding pixel value is f 0 (x, y), the fourth formula to determine whether it is the foreground is: Where r is the number of initializations, and T is the threshold; When BK M k-1 (x r ,y r ) is less than or equal to T, f k (x,y) is the pixel of the mask; When BK M k-1 (x r ,y r ) is greater than T, f k (x,y) is the foreground moving target pixel.

2. The perimeter intrusion hybrid detection method according to claim 1, It is characterized by: The step b comprises: The intrusion image judged as yes in step a is transmitted to the neural network algorithm, and the neural network algorithm performs boundary intrusion detection and tests the confidence Pri of the boundary box. dl ; The neural network algorithm includes combining and determining the detection results using the following first formula: Confidence level P merge =(P m +Pri dl )*(1-C)|C∈(0,0.5), where C represents the sensitivity coefficient; When the confidence level P merge ≥1, it is judged as yes, there is a target invading the perimeter; when P merge <1, judge as no.

3. The perimeter intrusion hybrid detection method according to claim 2, characterized in that: The neural network algorithm includes: Segment the intrusion image determined as yes in step a into a plurality of cells, each of the cells being responsible for detecting a target whose center point falls within the cell; Each cell predicts several bounding boxes and the confidence level of the bounding boxes Pri dl , the second formula of the confidence is: Where Pr represents the possibility that the bounding box contains the target; Indicates the accuracy of the bounding box, which is represented by the intersection of the predicted box and the actual box; The confidence Pri dl The value of is between 0 and 1. The larger the value, the higher the credibility of the detected target.

4. The hybrid perimeter intrusion detection method according to any one of claims 1 to 3, characterized in that: Also includes step 2: anti-missing detection method: Step c: if the second judgment is no, determine whether the third judgment condition is met; if the third judgment condition is met, correct the coefficients in the neural network algorithm, and reuse the neural network algorithm to perform boundary intrusion detection on the detection image judged as no in step b, and determine whether there is intrusion three times.

5. The perimeter intrusion hybrid detection method according to claim 4, characterized in that: The step c comprises: According to the accuracy of the nested detection method, the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are obtained; Establish a mixed probability model to prevent underreporting, which is the fifth formula: Correct judgment percentage The confidence level determined by the nested detection method is P merge , let the sixth formula of threshold determination be: Thr=P merge ×Pre dl ; When the threshold Thr≥0.5, the three judgment conditions are met, the sensitivity coefficient C in the first judgment formula is reduced, and the neural network algorithm is used again to detect boundary intrusion for the detection image judged as no in the step b, and whether there is intrusion is judged three times.

6. The perimeter intrusion hybrid detection method according to claim 5, characterized in that: If the three judgments are yes, then there is underreporting in the detection result of the nested detection method, and the coefficients TP (True Positive), TN (True Negative), FP (Fales Positive), and FN (False Negative) in the nested detection method are adjusted.

7. The perimeter intrusion hybrid detection method according to claim 6, characterized in that: Assume that the false negative rate occurs once in M ​​detection samples, expressed as L m , update the TP parameters according to the following seventh formula: TP=TP+L m ×C step , where C step For L m The amplification factor, C step =(10,100,1000…), as L m The smaller the value, the larger it is.

8. The perimeter intrusion hybrid detection method according to claim 5, characterized in that: The neural network algorithm is reused for perimeter intrusion detection only once per perimeter image frame.

9. A perimeter intrusion hybrid detection system, characterized in that: include: A moving target analysis algorithm module (10), the moving target analysis algorithm module being configured to: use a moving target analysis algorithm to detect enclosure intrusion in real time on the surveillance image of the enclosure, and initially determine whether there is intrusion; The moving target analysis algorithm module is configured to: Get the video stream from the perimeter camera and initialize the video stream image; Use the motion target analysis algorithm to detect the boundary intrusion of the boundary image, initially determine whether it is an intrusion, and obtain the accuracy rate P m ; The moving target analysis algorithm includes: Initialize a sample set for each pixel point of the bounding image, the sampling values ​​in the sample set are the historical pixel values ​​of the pixel point and the pixel values ​​of its neighborhood, and compare the pixel point sample sets of the previous and next frames to determine whether there is a moving target; When the video stream is t=0, the background model of the pixel is expressed as the third formula: BK M 0 =f 0 (x i ,y i )|(x i ,y i )∈N G (x,y) where BK M 0 Represents the mask of the current frame, f 0 (x i ,y i ) represents the current pixel value, N G (x, y) are adjacent pixel values, N G (x,y) is initialized N times and represented as N G (x i ,y i ); After the image is initialized, the foreground target segmentation is performed on the image of the subsequent video stream t; when t = k, the mask of the pixel point (x, y) of the current frame is BK M k-1 (x,y), the corresponding pixel value is f 0 (x, y), the fourth formula to determine whether it is the foreground is: Where r is the number of initializations, and T is the threshold; When BK M k-1 When (x,y) is less than or equal to T, f k (x,y) is the pixel of the mask; When BK M k-1 When (x,y) is greater than T, f k (x, y) is the foreground moving target pixel; A neural network algorithm module (20), the neural network algorithm module being configured to: use a neural network algorithm to perform boundary intrusion detection on the image determined as yes by the moving target analysis algorithm module, and to make a secondary determination as to whether there is intrusion; and The anti-missing detection module (30) is configured to: when the neural network algorithm module judges as no, judge whether three judgment conditions are met; if the three judgment conditions are met, correct the coefficients in the neural network algorithm so that the neural network algorithm module can re-use the neural network algorithm to detect boundary intrusion on the detection image judged as no, and judge whether there is intrusion three times.