Security decision platform based on artificial intelligence

Abstract

The invention belongs to the technical field of security and protection management, and discloses a security and protection decision platform based on artificial intelligence. Comprising a data acquisition and processing module used for acquiring real-time behavior attitude data of pedestrians in a monitoring area and preprocessing the real-time behavior attitude data to obtain real-time behavior attitude feature data; the data identification module is used for inputting the real-time behavior posture feature data into the trained identification model to obtain a real-time behavior category; by inputting the real-time behavior posture feature data into the trained recognition model, the behavior category of the pedestrian can be recognized, so that a subsequent security decision platform can conveniently output a corresponding security decision according to the behavior category of the pedestrian; the corresponding security decision can be output in time according to whether the real-time behavior category belongs to the security type set or the danger type set and the real-time environment elements, the security in the monitoring area is improved, and some dangerous accidents can be avoided.

Description

Technical Field

[0001] The present invention relates to the field of security management technology, and more specifically, to a security decision-making platform based on artificial intelligence. Background Art

[0002] Security decision-making refers to the determination of the best security solution or action in the field of security protection and monitoring through comprehensive judgment of data analysis, risk assessment, early warning and response measures. The goal of security decision-making is to respond to potential security threats or emergencies in a timely and effective manner to ensure the safety of people, property and information.

[0003] However, most security decisions are often made by relevant personnel manually identifying whether there are dangerous behaviors of pedestrians in the surveillance video. When the surveillance area is large and there are many surveillance videos, it is inconvenient for relevant personnel to monitor the danger in time and make corresponding security decisions, and it is easy to miss some dangerous behaviors of pedestrians. Of course, there are also some security decisions that are made by artificial intelligence to identify and make pedestrians' behavioral postures. These security decisions made by artificial intelligence often do not consider that some safe behaviors may become dangerous behaviors in specific environments, and these dangerous behaviors are easily ignored, which leads to the occurrence of dangerous accidents.

[0004] In view of this, the present invention proposes a security decision-making platform based on artificial intelligence to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: a security decision-making platform based on artificial intelligence, comprising:

[0006] A data collection and processing module is used to collect real-time behavior and posture data of pedestrians in the monitoring area, and pre-process the real-time behavior and posture data to obtain real-time behavior and posture feature data;

[0007] A data recognition module is used to input the real-time behavior posture feature data into the trained recognition model to obtain the real-time behavior category;

[0008] A matching module is used to preset a safety type set and a danger type set. If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario. If the real-time behavior category belongs to the safety type set, the real-time behavior category is matched with a corresponding combined feature path.

[0009] The decision output module is used to output corresponding security decisions through simulated scenarios or combined feature paths, and send the output security decisions to relevant personnel.

[0010] Furthermore, the method of collecting real-time behavior and posture data of pedestrians in the monitoring area includes:

[0011] The monitoring area is collected in real time through monitoring equipment to obtain monitoring video;

[0012] Use the target detection algorithm to identify and extract video clips containing pedestrians from the surveillance video and generate the corresponding video stream;

[0013] Decompose the video stream into v consecutive frames in time sequence, and use the posture estimation algorithm to mark and extract the key points of pedestrians in each frame of the v consecutive frames;

[0014] The key points of each frame of the extracted image are arranged in chronological order to obtain the real-time behavior posture data of the pedestrian.

[0015] Furthermore, the method of preprocessing the real-time behavior posture data to obtain the real-time behavior posture feature data includes:

[0016] For the continuous frame images in the real-time behavior posture data, a fixed reference point is selected as the reference center point, and the relative coordinates of all key points in each frame image relative to the reference point are calculated;

[0017] Use min-max normalization to normalize the relative coordinates of all key points in continuous frame images to the range of [0, 1] to obtain the position information of continuous frame images;

[0018] Based on the position information, the velocity and acceleration of the key points between consecutive frame images are calculated to obtain the velocity feature;

[0019] Calculate the angle changes between key points in consecutive frame images to obtain angle features;

[0020] The speed features and angle features in continuous frame images are aligned in time sequence to obtain real-time behavior posture feature data.

[0021] Furthermore, the training process of the recognition model is as follows:

[0022] Collect several video streams with labeled behavior categories and their corresponding real-time behavior posture feature data, build a behavior sample set, and divide the behavior sample set into a training set, a validation set, and a test set;

[0023] The long short-term memory neural network is used as the basic framework of the recognition model, which includes input layer, LSTM layer, fully connected layer, hidden layer and output layer;

[0024] Among them, the input layer is used to receive real-time behavior posture feature data and convert it into a feature vector P as input. The LSTM layer is used to capture the time dependency and dynamic changes of the real-time behavior posture feature data. The hidden layer is used to further process the feature vector output by the LSTM layer. The fully connected layer is used to map the feature vector output by the hidden layer to the feature space of the classification task and connect the input features with the classification output. The output layer generates the predicted probability of each behavior category through the softmax function and finally outputs the behavior category.

[0025] The loss function of the recognition model is defined as

[0026] Where L is the loss value, N is the number of behavior samples, C is the number of behavior categories, yij is the true label value of the ith behavior sample in the jth behavior category, and pij is the probability that the recognition model predicts that the ith behavior sample belongs to the jth behavior category.

[0027] Furthermore, the process of obtaining the real-time behavior category includes:

[0028] Input the real-time behavior posture feature data into the trained recognition model;

[0029] The output layer of the recognition model outputs the corresponding behavior category, and the output behavior category is the real-time behavior category.

[0030] Furthermore, the preset safety type set and danger type set, if the real-time behavior category belongs to the danger type set, is a way of matching the real-time behavior category with the corresponding simulation scenario, including:

[0031] Preset a behavior category set containing d behavior categories, define a classification standard, and classify the behavior categories in the behavior category set into a safe type set or a dangerous type set according to the classification standard;

[0032] K simulation scenarios are preset, each of which corresponds to a behavior category in the danger type set, wherein each simulation scenario is preset with several evolution paths, each evolution path is preset with a decision, and the several evolution paths are composed of different historical behavior posture feature data corresponding to the same behavior category;

[0033] Compare the real-time behavior categories with the behavior categories in the safety type set and the danger type set respectively;

[0034] If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario.

[0035] Furthermore, if the real-time behavior category belongs to the security type set, the method of matching the real-time behavior category with the corresponding combined feature path includes:

[0036] A combined feature path set is preset, and the combined feature path set includes s combined feature paths;

[0037] Extract the real-time environment elements from the continuous frame images, arrange and combine the extracted real-time environment elements with the real-time behavior categories belonging to the security type set, and obtain the real-time combined feature path:

[0038] Obtaining a combined feature path in the combined feature path set that is identical to the real-time combined feature path and contains the most environmental elements, and then matching the combined feature paths that are identical and contain the most environmental elements with the real-time behavior category;

[0039] The method of extracting real-time environmental elements from continuous frame images includes:

[0040] Perform feature extraction on continuous frame images to obtain the feature vector of each frame image;

[0041] Preset a similarity threshold, and compare the feature vector of each frame image with the feature vector in a pre-built environmental element database for similarity, wherein the environmental element database includes Q environmental elements and feature vectors corresponding to the Q environmental elements;

[0042] The feature vectors in the environmental element database that are greater than or equal to the similarity threshold are screened out, and the corresponding environmental elements are output, that is, all the real-time environmental elements in the continuous frame images are obtained.

[0043] Furthermore, the method for obtaining the combined characteristic path includes:

[0044] Collect R historical video streams corresponding to the behavior categories belonging to the safety type set in chronological order, and classify each historical video stream into low risk or potential danger;

[0045] Extracting behavior categories and environmental elements from each historical video stream that is potentially dangerous, and using the extracted behavior categories and environmental elements as initial combined feature paths;

[0046] Get the probability of the initial combined feature path appearing in R historical video streams, preset the probability threshold, and use the initial combined feature path greater than or equal to the probability threshold as the combined feature path:

[0047] Among them, each historical video stream that is potentially dangerous is preset with a decision, and the decision is associated with the combined feature path corresponding to the historical video stream.

[0048] Furthermore, the method of outputting corresponding security decisions by simulating scenarios or combining feature paths includes:

[0049] When the real-time behavior category belongs to the dangerous type set, the key points of each frame of the behavior posture data are used as nodes of the evolution path in chronological order, and each node is labeled with the sequence number of the evolution path-frame number;

[0050] According to the time sequence, the key point of each frame image corresponding to the real-time behavior posture data is taken as a node, that is, v nodes are obtained;

[0051] A sliding window is constructed with v nodes. For each preset evolution path, the similarities between the v nodes in the sliding window and the corresponding nodes in several evolution paths are calculated one by one to obtain several similarity sets.

[0052] All similarities of each similarity set in a plurality of similarity sets are summed up, and the plurality of similarity sets are sorted in descending order according to the summed up result;

[0053] The evolution path corresponding to the first-ranked similarity set is used as the evolution path of the real-time behavior posture data, and the decision corresponding to this evolution path is output as the security decision;

[0054] When the real-time behavior category belongs to the security type set, the decision corresponding to the combined feature path in the combined feature path set that is the same as the real-time combined feature path is output as a security decision.

[0055] Furthermore, the method of sending the output security decision to relevant personnel includes:

[0056] Obtain the coordinates of the monitoring device to obtain the first coordinates, obtain the coordinates of f related persons to obtain f second coordinates;

[0057] Get the distance between the first coordinate and f second coordinates, and get f distances;

[0058] Sort the f distances in ascending order, send the security decision and the first coordinate to the relevant personnel corresponding to the first distance in the sorting, and the relevant personnel will execute the security decision in the monitoring area corresponding to the first coordinate.

[0059] The technical effects and advantages of the security decision-making platform based on artificial intelligence of the present invention are as follows:

[0060] By inputting real-time behavior posture feature data into the trained recognition model, the behavior category of pedestrians can be identified, so that the subsequent security decision-making platform can output corresponding security decisions based on the pedestrian's behavior category; it can output corresponding security decisions in a timely manner based on whether the real-time behavior category belongs to the safety type set or the danger type set and the real-time environmental elements, thereby improving the safety in the monitoring area and avoiding the occurrence of some dangerous accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a structural schematic diagram of a security decision-making platform based on artificial intelligence of the present invention;

[0062] Figure 2 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0063] 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.

[0064] Example 1

[0065] See also Figure 1 and Figure 2 As shown, the security decision-making platform based on artificial intelligence described in this embodiment includes:

[0066] A data collection and processing module is used to collect real-time behavior and posture data of pedestrians in the monitoring area, and pre-process the real-time behavior and posture data to obtain real-time behavior and posture feature data;

[0067] The data recognition module is used to input the real-time behavior posture feature data into the trained recognition model to obtain the real-time behavior category.

[0068] Methods for collecting real-time behavior and posture data of pedestrians in the monitoring area include:

[0069] The monitoring area is collected in real time through monitoring equipment to obtain monitoring video, with the frame rate set to 30fps and the resolution set to 1080p;

[0070] Use the target detection algorithm to identify and extract video clips containing pedestrians from the surveillance video and generate the corresponding video stream;

[0071] Decompose the video stream into v consecutive frames in chronological order, and apply a posture estimation algorithm (such as OpenPose) to annotate and extract the key points of pedestrians in each of the v consecutive frames. The key points include the head, neck, shoulders, elbows, wrists, hips, knees, ankles, etc. The coordinates of these key points can be expressed in the form of (x, y, z);

[0072] Use target tracking algorithms (such as DeepSORT) to mark pedestrians, and arrange the key points of each frame of the extracted image in chronological order to obtain the real-time behavior posture data of pedestrians.

[0073] The method of preprocessing the real-time behavior posture data to obtain the real-time behavior posture feature data includes:

[0074] For the continuous frame images in the real-time behavior posture data, a fixed reference point is selected as the reference center point, and the relative coordinates of all key points in each frame image relative to the reference point are calculated;

[0075] Use min-max normalization to normalize the relative coordinates of all key points in continuous frame images to the range of [0, 1] to obtain the position information of continuous frame images;

[0076] Based on the position information, the velocity and acceleration of the key points between consecutive frame images are calculated to obtain the velocity feature;

[0077] Calculate the angle changes between key points in consecutive frame images to obtain angle features;

[0078] Align the speed features and angle features in the continuous frame images in time sequence to obtain real-time behavior posture feature data;

[0079] Specifically, high-definition cameras are installed in monitoring areas such as trails, grassy areas, and riverbanks in the park. These cameras can continuously collect pedestrian movements in the area 24 hours a day. The frame rate of the monitoring equipment is set to 30 frames per second (fps) and the resolution is 1080p to ensure sufficient image quality to capture pedestrian details such as gait and posture. The cameras begin to collect pedestrian activities in the monitoring area in real time and generate video data. Using target detection algorithms (such as YOLO or Faster R-CNN) identifies pedestrians from surveillance videos and extracts video clips containing pedestrians; assuming that v consecutive frames of images (for example, v = 10 frames) are processed each time, these images are used to capture the continuous movements of pedestrians, and a posture estimation algorithm (such as OpenPose) is applied to each frame of the image to annotate the key points of the pedestrians, and the key points of the pedestrians' head, neck, shoulders, elbows, wrists, hips, knees, ankles, etc. are annotated, and the coordinates of these key points are represented as (x, y); when the pedestrian raises his arm, the relative position change of the elbow and wrist can be detected, and a target tracking algorithm (such as DeepSORT) is used to assign a unique identifier to each pedestrian so that the system can continuously track the movements of the same pedestrian between different frames. In this way, the key point information of the same pedestrian in different frames of the video stream can be associated in time sequence to form a series of continuous key point data. After the above steps, the behavior posture data of each frame of the pedestrian is obtained; these data include the coordinate information (x, y) of each key point of the pedestrian, and the time sequence change of the key points between different frames.The pedestrian's neck is selected as the reference center point, and the relative coordinates of all other key points in each frame of the image will be calculated relative to this reference point; this can eliminate the impact of changes in camera perspective or overall movement of the pedestrian, and focus on the changes in the pedestrian's posture itself; for the key points in each frame of the image, the relative coordinates relative to the neck are calculated; for example, assuming that the coordinates of the pedestrian's right knee in a certain frame of the image are (200, 300) and the coordinates of the neck are (150, 250), then the relative coordinates of the right knee are (50, 50); the relative coordinates of each key point are normalized to the range of [0, 1]; this can eliminate the differences caused by the height and body shape of different pedestrians, making it easier to compare the postures of different pedestrians; for example, the relative coordinates (50, 50) may be converted to (0.5, 0 .5); Based on the normalized position information, calculate the speed and acceleration of each key point between consecutive frames; for example, if the right knee moves 0.1 unit distance between the 1st frame and the 2nd frame, and the time interval is 0.033 seconds (30fps frame rate), then its speed is 3 units / second. Continue to calculate the changes between multiple frame images, and then the acceleration can be obtained; for different key point pairs (such as shoulders and elbows, knees and ankles), calculate the angle changes of these key points between consecutive frame images. For example, if a pedestrian raises his arm, the angle between the shoulder and elbow will change over time; align the speed features and angle features in all frame images in chronological order, that is, obtain continuous behavioral posture feature data, which can reflect the dynamic behavioral posture of pedestrians when walking in the park.

[0080] The construction and training process of the recognition model is:

[0081] Collect several video streams with labeled behavior categories and their corresponding real-time behavior posture feature data. The video streams record pedestrians fainting, climbing fences, fighting, pushing and slapping, etc. The corresponding behavior categories can be fainting, climbing, fighting, pushing and slapping, etc. Construct a behavior sample set, divide the behavior sample set into a training set, a validation set and a test set, and the ratio of the training set, the validation set and the test set is 70%, 15% and 15% respectively;

[0082] The long short-term memory neural network is used as the basic framework of the recognition model, which includes input layer, LSTM layer, fully connected layer, hidden layer and output layer;

[0083] Among them, the input layer is used to receive real-time behavior posture feature data and convert it into a feature vector P as input. The LSTM layer is used to capture the time dependency and dynamic changes of the real-time behavior posture feature data. The hidden layer is used to further process the feature vector output by the LSTM layer. The fully connected layer is used to map the feature vector output by the hidden layer to the feature space of the classification task and connect the input features with the classification output. The output layer generates the predicted probability of each behavior category through the softmax function and finally outputs the behavior category. For example, if there are 5 behavior categories (fainting, climbing, fighting, pushing, and slapping), the output layer will generate the predicted probabilities of these 5 categories.

[0084] The loss function of the recognition model is defined as

[0085] Where L is the loss value, N is the number of behavior samples, C is the number of behavior categories, yij is the true label value of the i-th behavior sample in the j-th behavior category, and pij is the probability that the recognition model predicts that the i-th behavior sample belongs to the j-th behavior category;

[0086] Input the training set data into the recognition model, and gradually adjust the weights and biases of the recognition model until the loss function converges to a smaller value. During the training process, the model parameters are continuously updated through back propagation and optimization algorithms (such as Adam, etc.);

[0087] After each training cycle, the validation set is used to evaluate the performance of the recognition model, and evaluation indicators such as accuracy, precision, recall and F1 score are calculated. The hyperparameters of the recognition model are adjusted to avoid overfitting. The accuracy reflects the overall classification accuracy of the model, and the F1 score comprehensively considers the precision and recall, which is particularly suitable for cases of class imbalance. After the recognition model training is completed, the test set is used for final evaluation to verify the generalization ability of the recognition model on unseen data.

[0088] The process of obtaining the real-time behavior category includes:

[0089] Input the real-time behavior posture feature data into the trained recognition model;

[0090] The output layer of the recognition model outputs the corresponding behavior category, and the output behavior category is the real-time behavior category.

[0091] In this embodiment, by acquiring the real-time behavior posture data of pedestrians and performing preprocessing, real-time behavior posture feature data can be obtained. By inputting the real-time behavior posture feature data into the trained recognition model, the behavior category of the pedestrian can be identified, thereby facilitating the subsequent security decision-making platform to output corresponding security decisions based on the pedestrian's behavior category.

[0092] Example 2

[0093] See also Figure 1 and Figure 2 As shown, the security decision-making platform based on artificial intelligence described in this embodiment includes:

[0094] A matching module is used to preset a safety type set and a danger type set. If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario. If the real-time behavior category belongs to the safety type set, the real-time behavior category is matched with a corresponding combined feature path.

[0095] The decision output module is used to output the corresponding security decision through simulation scenarios or combined feature paths, and send the output security decision to relevant personnel:

[0096] A safety type set and a danger type set are preset. If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with the corresponding simulation scenario, including:

[0097] Preset a behavior category set containing d behavior categories and define the classification standard. The classification standard is set by relevant personnel according to the actual situation. For example, directly dangerous behavior categories such as fighting, pushing, slapping and fainting are classified into the dangerous type set, and non-directly dangerous behavior categories such as walking and running are classified into the safe type set. According to the classification standard, the behavior categories in the behavior category set are classified into the safe type set or the dangerous type set;

[0098] K simulation scenarios are preset, each of which corresponds to a behavior category in the danger type set, wherein each simulation scenario is preset with several evolution paths, each evolution path is preset with a decision, and the several evolution paths are composed of different historical behavior posture feature data corresponding to the same behavior category;

[0099] Compare the real-time behavior categories with the behavior categories in the safety type set and the danger type set respectively;

[0100] If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario.

[0101] If the real-time behavior category belongs to the security type set, the real-time behavior category matches the corresponding combined feature path, including:

[0102] A combined feature path set is preset, and the combined feature path set includes s combined feature paths;

[0103] Extract the real-time environment elements from the continuous frame images, arrange and combine the extracted real-time environment elements with the real-time behavior categories belonging to the security type set, and obtain the real-time combined feature path:

[0104] Obtaining a combined feature path in the combined feature path set that is identical to the real-time combined feature path and contains the most environmental elements, and then matching the combined feature paths that are identical and contain the most environmental elements with the real-time behavior category;

[0105] The method of extracting real-time environmental elements from continuous frame images includes:

[0106] Perform feature extraction on continuous frame images to obtain the feature vector of each frame image;

[0107] Preset a similarity threshold, which is set by relevant personnel and is generally set to 60% or above. (Use cosine similarity) compare the feature vector of each frame image with the feature vector in a pre-built environmental element database. The environmental element database contains Q environmental elements (thunderstorm weather, roads with a traffic volume threshold greater than the threshold, sharp objects, ditches, deep pits, dangerous plants such as thorns, etc.) and the feature vectors corresponding to the Q environmental elements;

[0108] The feature vectors in the environmental element database that are greater than or equal to the similarity threshold are screened out, and the corresponding environmental elements are output, that is, all the real-time environmental elements in the continuous frame images are obtained.

[0109] The methods for obtaining the combined feature path include:

[0110] Collect R historical video streams corresponding to the behavior categories belonging to the safety type set in chronological order, and classify each historical video stream into low risk or potential danger;

[0111] Extracting behavior categories and environmental elements from each historical video stream that is potentially dangerous, and using the extracted behavior categories and environmental elements as initial combined feature paths;

[0112] Get the probability of the initial combined feature path appearing in R historical video streams, preset the probability threshold, and use the initial combined feature path greater than or equal to the probability threshold as the combined feature path (for example, running under the environmental elements of thunderstorm weather, jumping on uneven stone slabs, etc.):

[0113] Among them, each historical video stream belonging to potential danger is preset with a decision (such as increasing patrol frequency, sending alarms, persuading people to leave, etc.), and the decision is associated with the combined feature path corresponding to the historical video stream;

[0114] Specifically, R historical video streams are collected, where R is greater than 500. The R historical video streams cover different scenes and environmental elements, for example: Scene 1: running in thunderstorm weather, 2: jumping on uneven stone roads, and each historical video stream is marked as "low risk" or "potentially dangerous" according to the behavior category and environmental elements; for example, the historical video stream of running in thunderstorm weather is marked as "potentially dangerous;

[0115] For the historical video streams marked as "potentially dangerous", the following features were extracted: behavior categories: such as running, jumping, etc., environmental elements: such as thunderstorms, cobblestone roads, etc.; the initial combined feature path extracted from the historical video stream of "running in thunderstorm weather" is: behavior category: running, environmental element: thunderstorm;

[0116] The occurrence frequencies of these initial combined feature paths in all R historical video streams are counted. Assuming that a total of 500 historical video streams are analyzed, 65 of them have the combined feature of “running in thunderstorms”. The occurrence probability of these combined features is calculated: occurrence probability = 65 / 500 = 13%. The preset probability threshold is 10%. Therefore, “running in thunderstorms” is taken as the combined feature path. For the historical video stream of “running in thunderstorms”, the preset decision is “send an alarm”, and the system sends an alarm to relevant personnel for reminders.

[0117] The methods of outputting corresponding security decisions by simulating scenarios or combining feature paths include:

[0118] When the real-time behavior category belongs to the dangerous type set, the key points of each frame of the behavior posture data are used as nodes of the evolution path in chronological order, and each node is labeled with the sequence number of the evolution path-frame number;

[0119] According to the time sequence, the key point of each frame image corresponding to the real-time behavior posture data is taken as a node, that is, v nodes are obtained;

[0120] A sliding window is constructed with v nodes. For each preset evolution path, the similarities between the v nodes in the sliding window and the corresponding nodes in several evolution paths are calculated one by one to obtain several similarity sets.

[0121] All similarities of each similarity set in a plurality of similarity sets are summed up, and the plurality of similarity sets are sorted in descending order according to the summed up result;

[0122] The evolution path corresponding to the first-ranked similarity set is used as the evolution path of the real-time behavior posture data, and the decision corresponding to this evolution path is output as the security decision;

[0123] When the real-time behavior category belongs to the security type set, the decision corresponding to the combined feature path in the combined feature path set that is the same as the real-time combined feature path is output as a security decision.

[0124] Specifically, if the real-time behavior category is fighting, the evolution path of the fighting behavior category includes evolution path 1 and evolution path 2; evolution path 1: standing face to face - raising arms - punching - physical contact - pushing each other - falling to the ground and tangling, evolution path 2: walking side by side - suddenly turning around - grabbing the other party's collar - punching - kicking each other - falling to the ground and tangling;

[0125] Assume a real-time behavior posture data sequence, including 10 consecutive frames (v=10), each frame contains the coordinates of 17 key points, and these 10 frames are used as a sliding window; this sliding window is compared with the above two preset evolution paths;

[0126] For each preset evolution path, calculate the similarity between each node (frame) in the sliding window and the corresponding node in the evolution path;

[0127] Use Euclidean distance to calculate similarity:

[0128]

[0129] In the formula, xsd is the similarity, jl is the Euclidean distance;

[0130]

[0131] Where m` is the number of key points, x`i` and y`i` are the i`th key point coordinates of the corresponding node (frame) in the sliding window and evolution path respectively;

[0132] For each evolution path, a similarity set containing 10 similarity values ​​is obtained, the sum of the similarities in each similarity set is calculated, and the two similarity sets are sorted in descending order according to the sum; for example, the sliding window data is simplified to 3 key points: right hand, left hand, head: frame 1; [(2, 3), (4, 3), (3, 5)], frame 2; [(2, 4), (4, 4), (3, 5)], ..., frame 10; [(1, 2) , (5, 2), (3, 4)]; Comparison of evolution path 1: frame 1 similarity; 0.8, frame 2 similarity: 0.75, ..., frame 10 similarity: 0.6, total: 7.2; Comparison of evolution path 2: frame 1 similarity: 0.7, frame 2 similarity: 0.8, ..., frame 10 similarity: 0.65, total: 7.5; Sorting results: evolution path 2 (fighting): 7.5, evolution path 1 (fighting): 7.2;

[0133] Based on this result, it is determined that the current behavior is closest to "evolution path 2", and the decision corresponding to evolution path 2 is selected as the security decision.

[0134] The output security decisions are sent to relevant personnel, including:

[0135] Obtain the coordinates of the monitoring device to obtain the first coordinates, obtain the coordinates of f related persons to obtain f second coordinates;

[0136] Get the distance between the first coordinate and f second coordinates, and get f distances;

[0137] The f distances are sorted in ascending order, and the security decision and the first coordinate are sent to the relevant personnel corresponding to the first distance in the sorting order. The relevant personnel then executes the security decision in the monitoring area corresponding to the first coordinate.

[0138] In this embodiment, by dividing the real-time behavior categories into a danger type set and a safety type set, the real-time behavior categories can be matched with simulation scenarios or combined feature paths respectively; when the real-time behavior category belongs to the danger type set and matches the simulation scenario, the real-time behavior posture data corresponding to the real-time behavior category is compared with several evolution paths contained in the simulation space for similarity, and the decision corresponding to the evolution path with the highest similarity in the similarity set is used as the security decision, so that the security decision belonging to the danger type set can be output in time; when the real-time behavior category belongs to the safety type set and matches the combined feature path, the real-time environmental elements can be combined with the real-time behavior category to obtain a real-time combined feature path. If the combined feature path set has the same combined feature path, the decision corresponding to the same combined feature path is used as the security decision. The corresponding security decision can be output in time according to whether the real-time behavior category belongs to the safety type set or the danger type set and the real-time environmental elements, thereby improving the safety in the monitoring area and avoiding the occurrence of some dangerous accidents.

[0139] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0140] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only one, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0141] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0142] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A security decision-making platform based on artificial intelligence, characterized in that: include: A data collection and processing module is used to collect real-time behavior and posture data of pedestrians in the monitoring area, and pre-process the real-time behavior and posture data to obtain real-time behavior and posture feature data; A data recognition module is used to input the real-time behavior posture feature data into the trained recognition model to obtain the real-time behavior category; A matching module is used to preset a safety type set and a danger type set. If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario. If the real-time behavior category belongs to the safety type set, the real-time behavior category is matched with a corresponding combined feature path. The decision output module is used to output corresponding security decisions through simulated scenarios or combined feature paths, and send the output security decisions to relevant personnel.

2. The artificial intelligence-based security decision-making platform according to claim 1, characterized in that: The method of collecting real-time behavior and posture data of pedestrians in the monitoring area includes: The monitoring area is collected in real time through monitoring equipment to obtain monitoring video; Use the target detection algorithm to identify and extract video clips containing pedestrians from the surveillance video and generate the corresponding video stream; Decompose the video stream into v consecutive frames in time sequence, and use the posture estimation algorithm to mark and extract the key points of pedestrians in each frame of the v consecutive frames; The key points of each frame of the extracted image are arranged in chronological order to obtain the real-time behavior posture data of the pedestrian.

3. The artificial intelligence-based security decision-making platform according to claim 2 is characterized in that: The method of preprocessing the real-time behavior posture data to obtain the real-time behavior posture feature data includes: For the continuous frame images in the real-time behavior posture data, a fixed reference point is selected as the reference center point, and the relative coordinates of all key points in each frame image relative to the reference point are calculated; Use min-max normalization to normalize the relative coordinates of all key points in continuous frame images to the range of [0, 1] to obtain the position information of continuous frame images; Based on the position information, the velocity and acceleration of the key points between consecutive frame images are calculated to obtain the velocity feature; Calculate the angle changes between key points in consecutive frame images to obtain angle features; The speed features and angle features in continuous frame images are aligned in time sequence to obtain real-time behavior posture feature data.

4. The artificial intelligence-based security decision-making platform according to claim 3 is characterized in that: The training process of the recognition model is as follows: Collect several video streams with labeled behavior categories and their corresponding real-time behavior posture feature data, build a behavior sample set, and divide the behavior sample set into a training set, a validation set, and a test set; The long short-term memory neural network is used as the basic framework of the recognition model, which includes input layer, LSTM layer, fully connected layer, hidden layer and output layer; Among them, the input layer is used to receive real-time behavior posture feature data and convert it into a feature vector P as input. The LSTM layer is used to capture the time dependency and dynamic changes of the real-time behavior posture feature data. The hidden layer is used to further process the feature vector output by the LSTM layer. The fully connected layer is used to map the feature vector output by the hidden layer to the feature space of the classification task and connect the input features with the classification output. The output layer generates the predicted probability of each behavior category through the softmax function and finally outputs the behavior category. The loss function of the recognition model is defined as Where L is the loss value, N is the number of behavior samples, C is the number of behavior categories, yij is the true label value of the ith behavior sample in the jth behavior category, and pij is the probability that the recognition model predicts that the ith behavior sample belongs to the jth behavior category.

5. The artificial intelligence-based security decision-making platform according to claim 4 is characterized in that: The process of obtaining the real-time behavior category includes: Input the real-time behavior posture feature data into the trained recognition model; The output layer of the recognition model outputs the corresponding behavior category, and the output behavior category is the real-time behavior category.

6. The artificial intelligence-based security decision-making platform according to claim 5, characterized in that: The preset safety type set and danger type set, if the real-time behavior category belongs to the danger type set, is a way of matching the real-time behavior category with the corresponding simulation scenario, including: Preset a behavior category set containing d behavior categories, define a classification standard, and classify the behavior categories in the behavior category set into a safe type set or a dangerous type set according to the classification standard; K simulation scenarios are preset, each of which corresponds to a behavior category in the danger type set, wherein each simulation scenario is preset with several evolution paths, each evolution path is preset with a decision, and the several evolution paths are composed of different historical behavior posture feature data corresponding to the same behavior category; Compare the real-time behavior categories with the behavior categories in the safety type set and the danger type set respectively; If the real-time behavior category belongs to the danger type set, the real-time behavior category is matched with a corresponding simulation scenario.

7. The artificial intelligence-based security decision-making platform according to claim 6 is characterized in that: If the real-time behavior category belongs to the security type set, the method of matching the real-time behavior category with the corresponding combined feature path includes: A combined feature path set is preset, and the combined feature path set includes s combined feature paths; Extract the real-time environment elements from the continuous frame images, arrange and combine the extracted real-time environment elements with the real-time behavior categories belonging to the security type set, and obtain the real-time combined feature path: Obtaining a combined feature path in the combined feature path set that is identical to the real-time combined feature path and contains the most environmental elements, and then matching the combined feature paths that are identical and contain the most environmental elements with the real-time behavior category; The method of extracting real-time environmental elements from continuous frame images includes: Perform feature extraction on continuous frame images to obtain the feature vector of each frame image; Preset a similarity threshold, and compare the feature vector of each frame image with the feature vector in a pre-built environmental element database for similarity, wherein the environmental element database includes Q environmental elements and feature vectors corresponding to the Q environmental elements; The feature vectors in the environmental element database that are greater than or equal to the similarity threshold are screened out, and the corresponding environmental elements are output, that is, all the real-time environmental elements in the continuous frame images are obtained.

8. The artificial intelligence-based security decision-making platform according to claim 7 is characterized in that: The method for obtaining the combined feature path includes: Collect R historical video streams corresponding to the behavior categories belonging to the safety type set in chronological order, and classify each historical video stream into low risk or potential danger; Extracting behavior categories and environmental elements from each historical video stream that is potentially dangerous, and using the extracted behavior categories and environmental elements as initial combined feature paths; Get the probability of the initial combined feature path appearing in R historical video streams, preset the probability threshold, and use the initial combined feature path greater than or equal to the probability threshold as the combined feature path: Among them, each historical video stream that is potentially dangerous is preset with a decision, and the decision is associated with the combined feature path corresponding to the historical video stream.

9. The artificial intelligence-based security decision-making platform according to claim 8, characterized in that: The method of outputting corresponding security decisions by simulating scenarios or combining feature paths includes: When the real-time behavior category belongs to the dangerous type set, the key points of each frame of the behavior posture data are used as nodes of the evolution path in chronological order, and each node is labeled with the sequence number of the evolution path-frame number; According to the time sequence, the key point of each frame image corresponding to the real-time behavior posture data is taken as a node, that is, v nodes are obtained; A sliding window is constructed with v nodes. For each preset evolution path, the similarities between the v nodes in the sliding window and the corresponding nodes in several evolution paths are calculated one by one to obtain several similarity sets. All similarities of each similarity set in a plurality of similarity sets are summed up, and the plurality of similarity sets are sorted in descending order according to the summed up result; The evolution path corresponding to the first-ranked similarity set is used as the evolution path of the real-time behavior posture data, and the decision corresponding to this evolution path is output as the security decision; When the real-time behavior category belongs to the security type set, the decision corresponding to the combined feature path in the combined feature path set that is the same as the real-time combined feature path is output as a security decision.

10. The artificial intelligence-based security decision-making platform according to claim 1, characterized in that: The method of sending the output security decision to relevant personnel includes: Obtain the coordinates of the monitoring device to obtain the first coordinates, obtain the coordinates of f related persons to obtain f second coordinates; Get the distance between the first coordinate and f second coordinates, and get f distances; The f distances are sorted in ascending order, and the security decision and the first coordinate are sent to the relevant personnel corresponding to the first distance in the sorting order. The relevant personnel then executes the security decision in the monitoring area corresponding to the first coordinate.

Images