Artificial intelligence security monitoring and early warning system suitable for smart park

Abstract

The invention relates to the technical field of monitoring and early warning, in particular to an artificial intelligence security monitoring and early warning system suitable for a smart park. The method comprises the following steps: a video acquisition unit acquires video pictures in a park in real time through a camera mounted at a fixed position of the park; the image processing and analysis unit detects personnel in a video picture based on a regional convolutional neural network model, and identifies abnormal behaviors of the personnel based on a first-class support vector machine model; and the alarm management and transmission unit triggers an alarm and generates alarm information when the image processing and analysis unit detects an abnormal condition, and sends the alarm information to an alarm system of the central station through a wireless network. According to the design of the invention, the weight of the support vector machine model is adjusted by introducing the behavior speed factor, optimization is carried out especially for the behavior mode difference of young people and old people, the ability of the system to understand the behaviors of people of different age groups is enhanced, and the accuracy of abnormal behavior recognition is also improved.

Description

Technical Field

[0001] The present invention relates to the field of monitoring and early warning technology, and in particular to an artificial intelligence security monitoring and early warning system suitable for a smart park. Background Art

[0002] Traditional security monitoring systems mostly rely on manual viewing of video streams, which is not only inefficient, but also prone to negligence and unable to detect and respond to abnormal situations in a timely manner. In addition, traditional systems lack intelligent analysis capabilities and are difficult to automatically identify complex and diverse abnormal behaviors, especially in smart parks with large traffic and frequent activities. With the development of smart parks, the facilities in the parks are becoming more and more complex, and the flow of personnel is also increasing. This puts higher requirements on the security management of the park. An intelligent system that can monitor in real time, automatically identify abnormal behaviors, and respond quickly is needed to improve the efficiency and response speed of security management and ensure the safe operation of the park. There may be groups of different age groups such as young people and the elderly in smart parks at the same time. Their behavior patterns are significantly different. A system that can perform intelligent analysis based on the behavioral characteristics of people of different age groups is needed to accurately identify abnormal behaviors and avoid false alarms or missed reports. Therefore, an artificial intelligence security monitoring and early warning system suitable for smart parks is designed. Summary of the invention

[0003] The purpose of the present invention is to provide an artificial intelligence security monitoring and early warning system suitable for smart parks, so as to solve the limitations of traditional security monitoring systems and the problems of low security management efficiency and response speed proposed in the above background technology.

[0004] To achieve the above objectives, the present invention provides an artificial intelligence security monitoring and early warning system suitable for smart parks, comprising:

[0005] A video acquisition unit, wherein the video acquisition unit acquires video images in the park in real time through a camera installed at a fixed position in the park;

[0006] An image processing and analysis unit, wherein the image processing and analysis unit detects people in the video screen based on a regional convolutional neural network model, identifies abnormal behaviors of people based on a type of support vector machine model, and introduces a behavior speed factor into the type of support vector machine model based on differences in behavior patterns between young people and the elderly;

[0007] The alarm management and transmission unit triggers an alarm and generates alarm information when the image processing and analysis unit detects an abnormal situation, and sends the alarm information to the alarm system of the central station through a wireless network.

[0008] As a further improvement of the technical solution, the video images collected in the park include: personnel behavior and environmental information.

[0009] As a further improvement of the technical solution, the image processing and analysis unit includes a personnel analysis module, a behavior analysis module and a location analysis module;

[0010] Wherein, the personnel analysis module detects personnel in the video screen based on a regional convolutional neural network model;

[0011] The behavior analysis module identifies abnormal behaviors of personnel based on a type of support vector machine model;

[0012] The position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video image to calculate the position of the abnormal person in the actual space.

[0013] As a further improvement of the technical solution, the personnel analysis module detects personnel in the video screen based on the regional convolutional neural network model, including the following steps:

[0014] S1.1, extract key frames from the video and pre-process the original video images;

[0015] S1.2. Select the deep learning network ResNet as the backbone network to extract the features of the input video image;

[0016] S1.3. Based on the backbone network, add the region proposal network RPN to generate candidate regions containing objects, and dynamically adjust the size and proportion of the anchor points used in the region proposal network RPN according to the actual distribution characteristics of the target in the video screen;

[0017] S1.4, from the candidate regions generated by the region proposal network RPN, use the region of interest pooling layer to map each candidate region to a fixed-size feature map;

[0018] S1.5, further process the output of the region of interest pooling layer through the fully connected layer, and generate the category label and bounding box coordinates of each candidate region through the output layer;

[0019] S1.6. Define a loss function to measure the difference between the model prediction value and the true value, and consider the characteristic differences of the candidate regions to further optimize the loss function and perform an adaptive non-maximum suppression strategy on the candidate regions generated by the region proposal network RPN.

[0020] As a further improvement of the technical solution, in S1.6, the loss function is:

[0021] ;

[0022] in, Represents the category label; Represents the probability value predicted by the model; represents the classification loss; Indicates the number of categories;

[0023] Considering the characteristic differences of the candidate regions, the loss function is further optimized:

[0024] ;

[0025] in, represents the classification loss after considering the characteristic differences of the candidate regions; Represents the function that adjusts the loss contribution of different candidate regions.

[0026] As a further improvement of the technical solution, the behavior analysis module identifies abnormal behaviors of personnel based on a type of support vector machine, including the following steps:

[0027] S2.1. Collect video clips of normal and abnormal behaviors, annotate the normal and abnormal behaviors in the videos, extract key features, and use gait analysis technology to identify gait features;

[0028] S2.2, normalize the key features;

[0029] S2.3, using the kernel function to train a class of support vector machine models, and introducing a behavior speed factor into the class of support vector machine models according to the differences in behavior patterns between young people and the elderly;

[0030] S2.4. Use the trained first-class support vector machine model to predict new video images and determine whether they are abnormal behaviors. During the training process, ensure that the model can distinguish the behavior patterns of different age groups.

[0031] As a further improvement of the technical solution, in S2.3, a type of support vector machine model is:

[0032] ;

[0033] in, Determine a new input Decision function for whether it belongs to the same class in the training dataset; represents the total number of support vectors; Indicates The coefficients associated with the support vectors; represents the kernel function; Represents a new input sample; Indicates the first support vectors; Represents the decision function The threshold value of

[0034] In view of the differences in behavior patterns between young people and the elderly, a behavior speed factor is introduced into a class of support vector machine models:

[0035] ;

[0036] in, represents the decision function after the behavior speed factor is introduced; A measure of the speed of behavior; Indicates the speed adjustment factor.

[0037] As a further improvement of the technical solution, the position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video screen to calculate the position of the abnormal person in the actual space, including the following steps:

[0038] S3.1. Calibrate multiple cameras to obtain internal and external parameters of the cameras;

[0039] S3.2, obtaining the two-dimensional coordinates of the abnormal person in the video image;

[0040] S3.3, converting the image coordinates into normalized coordinates;

[0041] S3.4. Use the external parameters of the camera and the data from multiple cameras for fusion. Calculate the three-dimensional coordinates of abnormal personnel in the actual space through triangulation, and optimize the triangulation method based on the distortion of the park camera lens.

[0042] As a further improvement of the technical solution, in S3.4, the triangulation method is:

[0043] ;

[0044] in, Indicates the three-dimensional coordinates of the abnormal person in the world coordinate system; Indicates The projection function of the camera; Indicates The two-dimensional coordinates observed by the camera; Indicates the index of the camera; Indicates the number of cameras involved in the fusion;

[0045] Optimize the triangulation method to address the distortion of the park camera lens:

[0046] ;

[0047] in, Indicates the three-dimensional coordinates after the camera lens distortion of the park; Represents the weight factor that balances the impact of the original triangulation error and the distortion correction term; Represents the lens distortion correction term.

[0048] As a further improvement of the technical solution, the alarm management and transmission unit sends the alarm location information to the alarm system of the central station through the wireless network, including the following steps:

[0049] S4.1. When the system detects abnormal behavior and confirms it as an alarm event, the alarm transmission unit starts working;

[0050] S4.2. Obtaining the location information of abnormal persons from the location analysis module;

[0051] S4.3, packaging the acquired alarm information into a data packet in a standard format;

[0052] S4.4. Connect to the alarm system of the central station via a wireless network and send data packets to the alarm system of the central station.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. In this AI security monitoring and early warning system for smart parks, the weight of a type of support vector machine model is adjusted by introducing a behavior speed factor, and the behavior pattern differences between young people and the elderly are optimized. This not only enhances the system's ability to understand the behavior of people of different age groups, but also improves the accuracy of abnormal behavior recognition, especially in different functional areas such as leisure areas and sports areas. Such a design can help the system better adapt to the complex park environment, reduce false alarms and missed alarms, and provide more reliable security protection.

[0055] 2. This AI security monitoring and early warning system for smart parks uses data from multiple cameras for fusion, and uses optimized triangulation and lens distortion correction technology to accurately calculate the three-dimensional coordinates of abnormal personnel in the actual space. This method not only solves the problem of limited viewing angle of a single camera, but also further improves the accuracy of positioning by considering the influence of lens distortion. This is crucial for quickly responding to security incidents and accurately locating the location of abnormal behavior, greatly improving the security management level and emergency response efficiency of smart parks. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention;

[0057] The meaning of each number in the figure is:

[0058] 1. Video acquisition unit; 2. Image processing and analysis unit; 3. Alarm management and transmission unit. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Example: See Figure 1 As shown, an artificial intelligence security monitoring and early warning system suitable for a smart park is provided, including:

[0061] The video acquisition unit 1 collects video images in the park in real time through a camera installed at a fixed position in the park;

[0062] In this embodiment, the video images collected in the park include: personnel behavior (personnel movements, personnel postures, personnel interactions, etc.) and environmental information (people flow density, object recognition, area functions, etc.).

[0063] The image processing and analysis unit 2 detects people in the video screen based on a regional convolutional neural network model, and identifies abnormal behaviors of people based on a type of support vector machine model;

[0064] In this embodiment, the image processing and analysis unit 2 includes a personnel analysis module, a behavior analysis module and a location analysis module;

[0065] The personnel analysis module detects personnel in the video based on the regional convolutional neural network model (Faster R-CNN), and the regional convolutional neural network model (Faster R-CNN) mainly includes: feature extraction network, regional proposal network, region of interest pooling and classification and regression layer;

[0066] The behavior analysis module identifies abnormal behaviors of personnel based on a type of support vector machine model;

[0067] The position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video image to calculate the position of the abnormal person in the actual space.

[0068] Among them, the regional convolutional neural network model and its improved version extract rich features through multi-layer convolutional neural networks, which can capture the multi-scale and multi-level information of the target, thereby improving the detection accuracy. The regional convolutional neural network model can generate high-quality candidate regions, reduce the interference of background noise, and improve the accuracy of detection; through multi-scale feature maps and multi-scale candidate region generation, the regional convolutional neural network model can effectively detect targets of different sizes and shapes and adapt to the complex and changeable campus environment; the personnel analysis module detects personnel in the video screen based on the regional convolutional neural network model, including the following steps:

[0069] S1.1. Extract key frames from the video and preprocess the original video. The preprocessing includes normalizing the image pixel values ​​to the range of [0, 1] or [-1, 1] to meet the input requirements of the model, scaling the image to the size required by the model input, usually a fixed size (such as 224x224, 600x600, etc.), to increase the diversity of the data and the generalization ability of the model;

[0070] S1.2. Select the deep learning network ResNet as the backbone network to extract the features of the input video image. ResNet can generate high-quality multi-scale feature maps, providing strong support for subsequent detection and classification modules;

[0071] S1.3. On the basis of the backbone network, add the region proposal network RPN to generate candidate regions containing objects (generate multiple anchor points on the feature map. These anchor points are predefined bounding boxes with different scales and aspect ratios. Classify each anchor point to determine whether it contains the target object (binary classification problem). Perform bounding box regression on each anchor point, adjust the position and size of the anchor point to make it closer to the real target bounding box), and dynamically adjust the size and proportion of the anchor points used in the region proposal network RPN according to the actual distribution characteristics of the target in the video screen (including target size distribution, target shape distribution, target density distribution, etc.), so that the anchor points are more in line with the real shape of the target, reduce the number of candidate regions, and improve the detection speed. The region proposal network RPN outputs a series of rectangular regions, each of which has an object score indicating whether the region contains an object;

[0072] S1.4. From the candidate regions generated by the region proposal network RPN, use the region of interest pooling layer to map each candidate region to a fixed-size feature map (map each candidate region to the corresponding region on the feature map, use a pooling operation (such as maximum pooling) to scale each region to a fixed-size feature map, and output a fixed-size feature map to represent the features of each candidate region) for subsequent processing;

[0073] S1.5. The output of the region of interest pooling layer is further processed through the fully connected layer, and the category label and bounding box coordinates (i.e., the coordinates of the person in the video screen) of each candidate region are generated through the output layer;

[0074] S1.6. Define a loss function to measure the difference between the model prediction value and the true value (the prediction includes the category label and the bounding box coordinates), and consider the characteristic differences of the candidate regions. Further optimize the loss function and perform an adaptive non-maximum suppression strategy on the candidate regions generated by the region proposal network RPN to remove overlapping candidate regions and retain the regions most likely to contain the target object.

[0075] Furthermore, the loss function is:

[0076] ;

[0077] in, represents the class label (e.g., 0 for background, 1 for person), Represents the probability value predicted by the model; represents the classification loss; Indicates the number of categories;

[0078] Taking into account the differences in the characteristics of the candidate areas (some candidate areas may be located in the center of the picture and are more easily accurately identified as human areas, while some candidate areas located at the edge or with occlusion may have higher uncertainty), the loss function is further optimized: by considering the specific characteristics of the candidate areas, such as position and occlusion, the model can better learn the common features in different scenarios during the training process, which helps the model to maintain good performance when facing new environments or unseen data; the optimized loss function will give higher weights to the center and unobstructed areas, prompting the model to pay more attention to these areas, thereby improving the overall detection accuracy, which can help the model more accurately distinguish normal and abnormal behaviors, especially for those located at the edge of the picture or with partial occlusion. By appropriately reducing the weights of these areas, the occurrence of false alarms can be reduced and the reliability of the system can be improved;

[0079] ;

[0080] Considering the location information of the candidate area, if the center position of the picture is , candidate region The center coordinates of , the width of the region is , the height is ,definition ; The meaning of this formula is that when the center of the candidate area is closer to the center of the picture, The larger the value (closer to 1), the easier it is for the candidate region to be accurately identified, and the larger its weight in the loss calculation; when the center of the candidate region is farther from the center of the picture, The smaller the value of , the smaller the weight in the loss calculation;

[0081] Considering the occlusion of the candidate area, define an occlusion coefficient , if the candidate region is not occluded, If there is occlusion, ,definition ; There is no occluded candidate area , the weight in the loss calculation is normal; the candidate area with occlusion , the weight in the loss calculation is relatively reduced, so that the model pays more attention to the candidate areas without occlusion, improving the overall detection accuracy;

[0082] Combining the above two situations, we define a comprehensive ;

[0083] Among them, represents the classification loss after considering the characteristic differences of the candidate regions; Represents a function that adjusts the loss contribution of different candidate regions, which represents the candidate region The characteristics of the candidate area include the location (center position, edge position, etc.) and whether there is occlusion.

[0084] Among them, the first-class support vector machine model separates the normal behavior area from the abnormal behavior area by learning the boundary of normal behavior. In the test phase, if a sample falls outside the normal behavior area, it is considered to be abnormal behavior; the first-class support vector machine model can combine multiple features (such as gait features, movement features, posture features, etc.) for comprehensive analysis to improve the accuracy of abnormal behavior recognition; the behavior analysis module identifies abnormal behavior of personnel based on the first-class support vector machine model, including the following steps:

[0085] S2.1. Collect video clips of normal and abnormal behaviors to ensure that all normal and abnormal behaviors are covered, annotate normal and abnormal behaviors in the video, extract key features (including motion features, posture features, time features, etc.), and use gait analysis technology to identify gait features. Gait features include step size, frequency, stability, etc. Gait analysis technology extracts and analyzes people's walking characteristics to identify individuals or determine whether the behavior is abnormal. Gait features (such as step size, frequency, stability, etc.) extracted through gait analysis technology are very effective in identifying abnormal behaviors, especially for the differences in behavior patterns between the elderly and young people;

[0086] S2.2, normalize the key features and normalize the feature values ​​to the range of [0,1] or [-1,1];

[0087] S2.3, use the kernel function (RBF kernel) to train a class of support vector machine models, and introduce the behavior speed factor into the class of support vector machine models according to the differences in behavior patterns between young people and the elderly;

[0088] Furthermore, a type of support vector machine model is:

[0089] ;

[0090] in, Determine a new input A decision function for whether it belongs to the same category as in the training dataset (i.e., whether it is normal behavior); It represents the total number of support vectors, that is, the total number of key samples selected from the training set. The support vectors represent the key features of normal behavior. Indicates The coefficients associated with the support vectors; Represents the kernel function, which is used to calculate the input With sample similarities between; Represent new input samples, such as the behavioral characteristics of a person in a video; Indicates the first support vectors; Represents the decision function The threshold is used to distinguish positive samples (normal behavior) from negative samples (abnormal behavior);

[0091] There are significant differences in the behavior patterns of young people and the elderly. Young people usually move faster, while the elderly move slower. By introducing the behavior speed factor, the model can better capture these differences, thereby improving the ability to distinguish the behaviors of people of different age groups. For example, fast running may be more common among young people, but it may be abnormal behavior among the elderly. There may be multiple functional areas in a smart park, such as office areas, leisure areas, sports areas, etc. The behavior patterns of people in different areas are also different. By introducing the behavior speed factor, the model can better adapt to these different scenarios and improve the detection accuracy in specific areas. For example, in the leisure area, young people may prefer to move quickly, while the elderly prefer to walk slowly.

[0092] In view of the differences in behavior patterns between young people and the elderly, a behavior speed factor is introduced into a class of support vector machine models:

[0093] ;

[0094] First, define a function related to the behavior speed , which can be based on the input sample (representing the characteristic vector of personnel behavior) calculates the relevant measure of behavior speed;

[0095] In a smart park, people may have different behavior patterns in different areas. The behavior patterns can be adjusted according to the length of time they stay in different functional areas (such as office areas, leisure areas, etc.). The definition of It's the staff The length of stay in different areas is the displacement within these regions, defined as This definition takes into account the differences in movement speeds of people in different areas and can reflect the possible behavioral differences between young people and the elderly in different functional areas. For example, young people may move faster and stay shorter in the leisure area, while the elderly may move slower and stay longer in the leisure area;

[0096] Then, according to the different behavior speed characteristics of young people and the elderly, a speed-based weight adjustment factor is introduced , for relatively active (faster) behaviors of young people, may be given a smaller weight; whereas for the relatively slow (slower) behavior of the elderly, A larger weight may be assigned;

[0097] in, represents the decision function after the behavior speed factor is introduced; Represents a measure of the speed of the behavior, which is based on the input sample (i.e., character behavior characteristics) calculated to measure the character's movement speed; represents the speed adjustment factor;

[0098] Considering that the gathering and flow patterns of people are larger than usual, the support vector machine model is further optimized:

[0099] During large-scale events, a large number of people may gather in the park, which is very different from the daily dispersed flow of people. During large-scale events, some behaviors may be considered abnormal at ordinary times, but they are normal during the event. For example, fast movement and gathering during the event may be considered abnormal at ordinary times, but they are common phenomena during the event. By optimizing the model, these behaviors can be distinguished more accurately to avoid misjudging normal activity behaviors as abnormal behaviors, thereby reducing false alarms. During large-scale events, the security risks of the park may increase, so it is necessary to improve the detection sensitivity of the model to detect potential abnormal behaviors in a timely manner.

[0100] ;

[0101] The park is divided into areas according to activity venues, non-activity venues, etc., and the regional indicator function is set ,if The person's position is within the activity area. ,otherwise ;

[0102] in, Represents the decision function for different gathering and flow patterns of people; Area indicator function, used to judge the input sample Whether it is located in the activity area;

[0103] Add the two features of the number of people within a certain range around the character and the average distance between the characters to the decision function:

[0104] Define the number of people within a certain range around a person: If there is a video frame, each detected person has a position coordinate , define a radius , calculate the radius around each person Number of people in scope : ,in It is an indicator function that returns 1 if the condition is true, otherwise it returns 0;

[0105] Similarly, calculate the radius of each person and its surroundings The average distance characteristic of all other people in range : ;if ,but Set to a larger value (such as the diagonal length of the video frame);

[0106] If the original eigenvector Include Now we expand it to: , and will be extended to the feature vector Substitute into the decision function, where the decision function remains unchanged, but the eigenvector and support vector All have new features included.

[0107] S2.4. Use the trained first-class support vector machine model to predict new video images and determine whether they are abnormal behaviors. During the training process, ensure that the model can distinguish the behavior patterns of different age groups.

[0108] The position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video image to calculate the position of the abnormal person in the actual space, including the following steps:

[0109] S3.1. Calibrate multiple cameras to obtain the internal and external parameters of the cameras. The internal parameters include focal length (focal length of the camera in the x-axis and y-axis directions) and principal point (coordinates of the center point of the image). The external parameters include rotation matrix (describing the rotation of the camera relative to the world coordinate system) and translation vector (describing the translation of the camera relative to the world coordinate system).

[0110] S3.2, obtaining the two-dimensional coordinates of the abnormal person in the video image;

[0111] S3.3, converting the image coordinates into normalized coordinates;

[0112] S3.4. Use the external parameters of the camera (rotation matrix R and translation vector t), fuse the data of multiple cameras, calculate the three-dimensional coordinates of the abnormal person in the actual space through triangulation, and optimize the triangulation method according to the distortion of the camera lens in the park;

[0113] Furthermore, the triangulation method uses images taken from different angles by multiple cameras to calculate the position of the target in three-dimensional space through geometric relationships. The fusion of multiple views can provide richer information, reduce the error of a single camera, and improve the accuracy of positioning. Even if the image quality of a camera is poor or blocked, the data of other cameras can still provide supplementary information to ensure the accuracy of positioning; the triangulation method can adapt to camera configurations of different numbers and layouts. Whether it is a fixed camera or a mobile camera, accurate three-dimensional positioning can be achieved by adjusting the algorithm parameters; the triangulation method is:

[0114] ;

[0115] in, Indicates the three-dimensional coordinates of the abnormal person in the world coordinate system; Indicates The projection function of a camera, if is a 3D point in the world coordinate system, The internal parameter matrix of a camera is , the external parameter is the rotation matrix and translation vector , then the projection function for: ; Indicates The two-dimensional coordinates observed by the camera; Indicates the index of the camera; Indicates the number of cameras involved in the fusion;

[0116] Camera lens distortion can cause straight lines in an image to become curves after imaging, thus affecting the accuracy of the position of objects in the image. Lens distortion can lead to the accumulation of measurement errors, especially in the case of multi-camera fusion. The optimized triangulation method can effectively reduce this error and improve the robustness of the system by introducing distortion correction items, which helps the system to operate more stably in complex environments and reduce false positives or negatives caused by lens distortion.

[0117] Optimize the triangulation method to address the lens distortion of the campus camera (lens distortion means that due to the optical properties of the lens, straight lines in the image may become curves after imaging, causing the geometric shape of the image to be deformed):

[0118] ;

[0119] in, Indicates the three-dimensional coordinates after the camera lens distortion of the park; Represents the weight factor that balances the impact of the original triangulation error and the distortion correction term; Represents the lens distortion correction term.

[0120] When the image processing and analysis unit 2 detects an abnormal situation, the alarm management and transmission unit 3 triggers an alarm and generates alarm information, including the alarm type, occurrence location, timestamp, etc., and sends the alarm information to the alarm system of the central station through the wireless network;

[0121] In this embodiment, the alarm management and transmission unit 3 sends the alarm location information to the alarm system of the central station through the wireless network, including the following steps:

[0122] S4.1. When the system detects abnormal behavior and confirms it as an alarm event, the alarm transmission unit starts working;

[0123] S4.2. Obtain the location information of the abnormal person from the location analysis module. The location information includes the three-dimensional coordinates of the person, camera ID, timestamp and other relevant data (such as behavior type, severity, etc.);

[0124] S4.3, packaging the acquired alarm information into a data packet in a standard format;

[0125] S4.4. Connect to the alarm system of the central station via the wireless network (LoRaWAN) and send data packets to the alarm system of the central station.

[0126] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention, and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. An artificial intelligence security monitoring and early warning system suitable for smart parks, characterized in that: include: A video acquisition unit (1), wherein the video acquisition unit (1) acquires video images in the park in real time through a camera installed at a fixed position in the park; An image processing and analysis unit (2), wherein the image processing and analysis unit (2) detects people in the video image based on a regional convolutional neural network model, identifies abnormal behavior of people based on a first-class support vector machine model, and introduces a behavior speed factor into the first-class support vector machine model based on the difference in behavior patterns between young people and the elderly; An alarm management and transmission unit (3) triggers an alarm and generates alarm information when the image processing and analysis unit (2) detects an abnormal situation, and sends a signal of the alarm information to an alarm system of a central station via a wireless network.

2. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 1 is characterized in that: The video images collected in the park include: personnel behavior and environmental information.

3. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 2 is characterized by: The image processing and analysis unit (2) comprises a personnel analysis module, a behavior analysis module and a location analysis module; Wherein, the personnel analysis module detects personnel in the video screen based on a regional convolutional neural network model; The behavior analysis module identifies abnormal behaviors of personnel based on a type of support vector machine model; The position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video image to calculate the position of the abnormal person in the actual space.

4. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 3 is characterized by: The personnel analysis module detects personnel in the video screen based on the regional convolutional neural network model, including the following steps: S1.1, extract key frames from the video and pre-process the original video images; S1.

2. Select the deep learning network ResNet as the backbone network to extract the features of the input video image; S1.

3. Based on the backbone network, a region proposal network RPN is added to generate candidate regions containing objects, and the size and proportion of the anchor points used in the region proposal network RPN are dynamically adjusted according to the actual distribution characteristics of the target in the video screen; S1.4, from the candidate regions generated by the region proposal network RPN, use the region of interest pooling layer to map each candidate region to a fixed-size feature map; S1.5, further process the output of the region of interest pooling layer through the fully connected layer, and generate the category label and bounding box coordinates of each candidate region through the output layer; S1.

6. Define a loss function to measure the difference between the model prediction value and the true value, and consider the characteristic differences of the candidate regions to further optimize the loss function and perform an adaptive non-maximum suppression strategy on the candidate regions generated by the region proposal network RPN.

5. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 4 is characterized in that: In S1.6, the loss function is: ； in, Represents the category label; Represents the probability value predicted by the model; represents the classification loss; Indicates the number of categories; Considering the characteristic differences of the candidate regions, the loss function is further optimized: ； in, represents the classification loss after considering the characteristic differences of the candidate regions; Represents the function that adjusts the loss contribution of different candidate regions.

6. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 5 is characterized by: The behavior analysis module identifies abnormal behaviors of personnel based on a type of support vector machine, including the following steps: S2.

1. Collect video clips of normal and abnormal behaviors, annotate the normal and abnormal behaviors in the videos, extract key features, and use gait analysis technology to identify gait features; S2.2, normalizing the key features and gait features; S2.3, using the kernel function to train a class of support vector machine models, and introducing a behavior speed factor into the class of support vector machine models according to the differences in behavior patterns between young people and the elderly; S2.

4. Use the trained first-class support vector machine model to predict new video images and determine whether they are abnormal behaviors. During the training process, ensure that the model can distinguish the behavior patterns of different age groups.

7. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 6 is characterized by: In S2.3, a type of support vector machine model is: ； in, Determine a new input Decision function for whether it belongs to the same class in the training dataset; represents the total number of support vectors; Indicates The coefficients associated with the support vectors; represents the kernel function; Represents a new input sample; Indicates the first support vectors; Represents the decision function The threshold value of In view of the differences in behavior patterns between young people and the elderly, a behavior speed factor is introduced into a class of support vector machine models: ； in, Represents the decision function after the behavior speed factor is introduced; A measure of the speed of behavior; Indicates the speed adjustment factor.

8. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 7 is characterized in that: The position analysis module combines the position information of the camera and the coordinates of the abnormal person in the video image to calculate the position of the abnormal person in the actual space, including the following steps: S3.

1. Calibrate multiple cameras to obtain internal and external parameters of the cameras; S3.2, obtaining the two-dimensional coordinates of the abnormal person in the video image; S3.3, converting the image coordinates into normalized coordinates; S3.

4. Use the external parameters of the camera and the data from multiple cameras for fusion. Calculate the three-dimensional coordinates of abnormal personnel in the actual space through triangulation, and optimize the triangulation method based on the distortion of the park camera lens.

9. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 8 is characterized in that: In S3.4, the triangulation method is: ； in, Indicates the three-dimensional coordinates of the abnormal person in the world coordinate system; Indicates The projection function of the camera; Indicates The two-dimensional coordinates observed by the camera; Indicates the index of the camera; Indicates the number of cameras involved in the fusion; Optimize the triangulation method to address the distortion of the park camera lens: ； in, Indicates the three-dimensional coordinates after the camera lens distortion of the park; Represents the weight factor that balances the impact of the original triangulation error and the distortion correction term; Represents the lens distortion correction term.

10. The artificial intelligence security monitoring and early warning system suitable for smart parks according to claim 9 is characterized in that: The alarm management and transmission unit (3) sends the alarm location information to the alarm system of the central station via a wireless network, including the following steps: S4.

1. When the system detects abnormal behavior and confirms it as an alarm event, the alarm transmission unit starts working; S4.

2. Obtaining the location information of abnormal persons from the location analysis module; S4.3, packaging the acquired alarm information into a data packet in a standard format; S4.

4. Connect to the alarm system of the central station via a wireless network and send data packets to the alarm system of the central station.

Images