A surveillance video intelligent inspection management method and system based on artificial intelligence

Abstract

The present invention relates to the field of intelligent monitoring management technology, and discloses an intelligent patrol management method and system for monitoring videos based on artificial intelligence. The method collects historical video monitoring data, establishes a sample set, and pre-processes the historical video monitoring data in the sample set. At the same time, based on the pre-processed historical video monitoring data, the abnormal behavior characteristics and spatiotemporal characteristics in the historical video monitoring data are extracted by machine learning; at the same time, the reminder frequency of the monitoring camera is set based on the spatiotemporal characteristics extracted from the historical video monitoring data, and by setting a video monitoring event, when the administrator selects a frame of video monitoring data in the corresponding time period, the selected video screen is identified and patrol records are generated based on machine learning and data matching methods. Finally, by setting patrol evaluation management standards and evaluating and optimizing the patrol records generated by the administrator, the accuracy of intelligent patrol of monitoring videos is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring intelligent management, and in particular to an artificial intelligence-based monitoring video intelligent inspection management method and system. Background Art

[0002] Traditional surveillance video intelligent inspection mainly relies on manual work, with low inspection efficiency and low inspection accuracy. With the development of intelligence, surveillance video intelligent inspection has also become a difficult problem for intelligent monitoring, but due to the lack of accuracy in real-time analysis and processing of video data, it cannot meet actual needs in some complex scenarios.

[0003] The existing Chinese patent application CN117575230A, this method builds a remote inspection intelligent perception terminal, collects multi-frequency inspection information and intelligently recognizes and perceives it, configures video analysis parameters and areas, and sorts and executes the recognition results after the video analysis is completed, and completes the remote inspection of the area through manual review and early warning. However, due to the low accuracy of intelligent recognition perception in complex scenes, it has certain limitations. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the prior art, the present invention provides a monitoring video intelligent patrol management method and system based on artificial intelligence, which has the advantages of accuracy, efficiency, and real-time, and solves the problem that the real-time analysis and processing of video data is insufficient in precision and cannot meet actual needs in some complex scenarios.

[0006] (II) Technical solution

[0007] In order to solve the above technical problem that the accuracy of real-time analysis and processing of video data is insufficient and cannot meet the actual needs in some complex scenes, the present invention provides the following technical solutions:

[0008] This embodiment discloses a surveillance video intelligent inspection management method based on artificial intelligence, which specifically includes the following steps:

[0009] S1. Collect historical video surveillance data, establish a sample set based on the collected historical video surveillance data, and preprocess the historical video surveillance data in the sample set to obtain preprocessed historical video surveillance data;

[0010] S2. Based on the pre-processed historical video surveillance data, extract abnormal behavior characteristics and spatiotemporal characteristics in the historical video surveillance data through machine learning;

[0011] S3, setting the reminder frequency of surveillance cameras based on the spatiotemporal features of abnormal behaviors extracted from historical video surveillance data;

[0012] S4. Set a video surveillance event. When the administrator selects a frame of video surveillance data in the corresponding time period, the video screen of the frame of video surveillance data is automatically enlarged, and the selected video screen is identified based on the machine learning method and data matching method to obtain the identified video surveillance data;

[0013] S5. Generate and save inspection records based on the identified video surveillance data;

[0014] S6. Set inspection evaluation management standards and evaluate and optimize the inspection records generated by administrators.

[0015] The present invention collects historical video surveillance data and establishes a sample set, pre-processes the historical video surveillance data in the sample set, and extracts abnormal behavior characteristics and spatiotemporal characteristics in the historical video surveillance data through machine learning based on the pre-processed historical video surveillance data; at the same time, the reminder frequency of the surveillance camera is set based on the spatiotemporal characteristics of the abnormal behavior extracted from the historical video surveillance data, and by setting a video surveillance event, when the administrator selects a frame of video surveillance data in the corresponding time period, the video picture of the frame of video surveillance data is automatically enlarged, and the selected video picture is identified and an inspection record is generated based on a machine learning method and a data matching method, and finally, the inspection evaluation management standard is set and the inspection record generated by the administrator is evaluated and optimized, thereby improving the accuracy of intelligent inspection of surveillance video.

[0016] Preferably, the collecting of historical video surveillance data, establishing a sample set based on the collected historical video surveillance data, and preprocessing the historical video surveillance data in the sample set to obtain the preprocessed historical video surveillance data comprises the following steps:

[0017] S11. Select historical video surveillance data with abnormal behavior in the sample set;

[0018] S12, performing image enhancement on the selected historical video surveillance data to obtain enhanced historical video surveillance data;

[0019] S13: setting the enhanced historical video surveillance data as the pre-processed historical video surveillance data.

[0020] Preferably, the performing image enhancement on the selected historical video surveillance data to obtain the enhanced historical video surveillance data comprises the following steps:

[0021] S121, based on the gray value distribution in the selected historical video surveillance data, calculating the pixel probability distribution of the selected historical video surveillance data, the gray value being the value of the pixel point satisfying the condition R=G=B;

[0022] The pixel probability distribution formula of the selected historical video surveillance data is as follows:

[0023] ;

[0024] in, represents the pixel probability of the selected historical video surveillance data, Indicates the gray value is The number of pixels, is the total number of pixels;

[0025] S122, constructing an image cumulative distribution function according to the calculated pixel probability distribution of the selected historical video surveillance data;

[0026] The formula for constructing the cumulative distribution function of the image is as follows:

[0027] ;

[0028] in, Represents the total number of gray values ​​in the image, Indicates The gray value after the gray value is mapped by the cumulative distribution function;

[0029] S123, performing image enhancement on the selected historical video surveillance data based on the constructed image cumulative distribution function;

[0030] Multiply the gray value mapped by the cumulative distribution function by ( -1);

[0031] The transformation function formula is as follows:

[0032] ;

[0033] in, Indicates about The transformation function is used to make the gray level of the equalized image consistent with the original image before normalization;

[0034] According to the mapping relationship between the image cumulative distribution function and the transformation function, the image after histogram equalization is output with reference to the pixels in the original image.

[0035] The present invention completes the processing of the historical video surveillance data by selecting historical video surveillance data with abnormal behavior in a sample set and performing image enhancement on the selected historical video surveillance data, thereby ensuring the accuracy of the processing of the historical video surveillance data.

[0036] Preferably, the extracting abnormal behavior features and spatiotemporal features in the historical video surveillance data by machine learning based on the pre-processed historical video surveillance data comprises the following steps:

[0037] S21, taking the obtained pre-processed historical video surveillance data as input and inputting it into the three-dimensional convolutional neural network;

[0038] S22. Extract features from the input historical video surveillance data through a three-dimensional convolutional neural network.

[0039] Preferably, the feature extraction of the input historical video surveillance data by using a three-dimensional convolutional neural network comprises the following steps:

[0040] S221, set the input historical video surveillance data size to ;

[0041] in, Indicates the number of channels, Indicates the frame length of historical video surveillance data. Indicates the height of the frame. Indicates the width of the frame;

[0042] Set the convolution kernel size of the 3D convolutional neural network to ;

[0043] in, represents the time depth of the convolution kernel, Indicates the size of the convolution kernel space;

[0044] S222, after receiving the input historical video surveillance data, the convolution layer moves on the input historical video surveillance data through the set convolution kernel according to the set step size, and performs convolution calculation on the corresponding area of ​​each step with the feature value of the area, so as to realize feature extraction of each historical video surveillance data;

[0045] The convolution calculation formula is as follows:

[0046] ;

[0047] in, Represents the input historical video surveillance data, represents the weight of the corresponding convolution kernel, b Indicates the bias value, f represents the output features;

[0048] S223, in the process of continuous convolution of the three-dimensional convolutional neural network, the output of the previous three-dimensional convolutional neural network will be used as the input of the next three-dimensional convolutional neural network, and the data will be processed by the activation function and the residual function in the process of inputting the data into the next three-dimensional convolutional neural network;

[0049] S224, extracting features of historical video surveillance data in a continuously stacked convolution and pooling process;

[0050] S225, passing the extracted features of the historical video surveillance data into the fully connected layer;

[0051] S226, expanding and combining the data through a fully connected layer to obtain a set of feature data, and saving it;

[0052] According to the set convolution kernel, a set of feature data is obtained, the first third of which represents the spatiotemporal features, and the last two thirds represent the abnormal behavior features.

[0053] Preferably, the processing of the data through the activation function and the residual function in the process of inputting the data into the lower three-dimensional convolutional neural network includes the following steps:

[0054] The activation function processing includes the following steps:

[0055] Set the input value to (λ=1,2,……,L), each output value The corresponding input weight is , b is the offset, and the output result obtained after inputting the input value into the three-dimensional convolutional neural network is:

[0056] ;

[0057] in, is the corresponding activation function, is the output result of the activation function;

[0058] The processing of the residual function includes the following steps:

[0059] ;

[0060] in, represents the output of the residual module, represents the residual function.

[0061] The present invention ensures the reliability of data feature extraction by inputting the obtained preprocessed historical video surveillance data into a three-dimensional convolutional neural network and performing abnormal behavior feature extraction and spatiotemporal feature extraction on the input historical video surveillance data through the three-dimensional convolutional neural network.

[0062] Preferably, the method of identifying the selected video images based on a machine learning method and a data matching method to obtain the identified video surveillance data comprises the following steps:

[0063] S41, extracting abnormal behavior features of the selected video images in real time by machine learning;

[0064] S42, matching the abnormal behavior features of the selected video images extracted in real time with the abnormal behavior features in the stored historical video surveillance data by means of data matching.

[0065] Preferably, matching the abnormal behavior features of the real-time extracted selected video images with the abnormal behavior features in the stored historical video surveillance data by means of data matching comprises the following steps:

[0066] The similarity calculation method is as follows:

[0067] ;

[0068] in, For vector and vector The similarity between represents the coordinates of the i-th point, Indicates the total number of pixels;

[0069] A similarity threshold is set. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is greater than or equal to the set similarity threshold, it indicates that the selected video screen has abnormal behavior. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is less than the set similarity threshold, it indicates that the selected video screen does not have abnormal behavior.

[0070] The present invention extracts abnormal behavior features of selected video images in real time through machine learning, and matches the abnormal behavior features of the selected video images extracted in real time with abnormal behavior features in stored historical video surveillance data through data matching, and analyzes the selected video images in real time according to the matching results, thereby improving the real-time performance of the selected video image recognition.

[0071] Preferably, the setting of inspection evaluation management standards and evaluating and optimizing the inspection records generated by the administrator includes the following steps:

[0072] Calculate the recognition accuracy under the current surveillance camera reminder frequency;

[0073] ;

[0074] Among them, TP represents the number of recognition results that are consistent with the actual results; FN represents the number of missed detections; R is the accuracy rate;

[0075] Set the accuracy rate as the inspection evaluation management standard and set the accuracy rate threshold. When the accuracy rate in the inspection record generated by the administrator is less than the set accuracy rate threshold, increase the reminder frequency of the surveillance camera; when the accuracy rate in the inspection record generated by the administrator is greater than or equal to the set accuracy rate threshold, it means that the inspection record is qualified.

[0076] The present invention evaluates and optimizes the inspection records generated by the administrator by calculating the recognition accuracy under the current surveillance camera reminder frequency, setting an accuracy threshold and judging whether the accuracy in the currently generated inspection records meets the set threshold, thereby improving the efficiency of inspection record evaluation.

[0077] This embodiment also discloses an artificial intelligence-based surveillance video intelligent inspection management system, including: a data acquisition module, a data processing module, a feature extraction module, a monitoring adjustment module, an identification and matching module, and an inspection evaluation management module;

[0078] The data acquisition module is used to collect video surveillance data in real time and transmit the collected video surveillance data to the data processing module;

[0079] The data processing module is used to process the received video surveillance data;

[0080] The feature extraction module is used to extract abnormal behavior features and spatiotemporal features from the processed video surveillance data;

[0081] The monitoring adjustment module is used to perform monitoring adjustment according to the extracted spatiotemporal features;

[0082] The identification and matching module is used to perform identification and matching based on the extracted abnormal behavior features;

[0083] The inspection evaluation management module is used to evaluate the accuracy of the identification and matching results.

[0084] (III) Beneficial effects

[0085] Compared with the prior art, the present invention provides a surveillance video intelligent inspection management method and system based on artificial intelligence, which has the following beneficial effects:

[0086] 1. The invention collects historical video surveillance data and establishes a sample set, pre-processes the historical video surveillance data in the sample set, and extracts abnormal behavior characteristics and spatiotemporal characteristics in the historical video surveillance data through machine learning based on the pre-processed historical video surveillance data; at the same time, the reminder frequency of the surveillance camera is set based on the spatiotemporal characteristics of abnormal behavior extracted from the historical video surveillance data, and by setting a video surveillance event, when the administrator selects a frame of video surveillance data in the corresponding time period, the video screen of the frame of video surveillance data is automatically enlarged, and the selected video screen is identified and an inspection record is generated based on the machine learning method and data matching method, and finally, the inspection evaluation management standard is set and the inspection records generated by the administrator are evaluated and optimized, thereby improving the accuracy of intelligent inspection of surveillance video.

[0087] 2. The present invention completes the processing of historical video surveillance data by selecting historical video surveillance data with abnormal behavior in a sample set and performing image enhancement on the selected historical video surveillance data, thereby ensuring the accuracy of the processing of historical video surveillance data.

[0088] 3. The invention ensures the reliability of data feature extraction by inputting the preprocessed historical video surveillance data into a three-dimensional convolutional neural network and performing abnormal behavior feature extraction and spatiotemporal feature extraction on the input historical video surveillance data through the three-dimensional convolutional neural network.

[0089] 4. The invention uses machine learning to extract abnormal behavior features of selected video images in real time, and at the same time matches the abnormal behavior features of the selected video images extracted in real time with the abnormal behavior features in the saved historical video surveillance data through data matching, and analyzes the selected video images in real time based on the matching results, thereby improving the real-time performance of the selected video image recognition.

[0090] 5. The invention evaluates and optimizes the inspection records generated by the administrator by calculating the recognition accuracy under the current surveillance camera reminder frequency, setting an accuracy threshold and judging whether the accuracy in the currently generated inspection records meets the set threshold, thereby improving the efficiency of inspection record evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a schematic diagram of the process structure of the monitoring video intelligent inspection management method of the present invention. DETAILED DESCRIPTION

[0092] 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.

[0093] Example 1: Please refer to Figure 1 This embodiment discloses a surveillance video intelligent inspection management method based on artificial intelligence, which specifically includes the following steps:

[0094] S1. Collect historical video surveillance data, establish a sample set based on the collected historical video surveillance data, and preprocess the historical video surveillance data in the sample set to obtain preprocessed historical video surveillance data;

[0095] Collecting historical video surveillance data, establishing a sample set based on the collected historical video surveillance data, and preprocessing the historical video surveillance data in the sample set to obtain the preprocessed historical video surveillance data includes the following steps:

[0096] S11. Select historical video surveillance data with abnormal behavior in the sample set;

[0097] S12, performing image enhancement on the selected historical video surveillance data to obtain enhanced historical video surveillance data;

[0098] S121, based on the gray value distribution in the selected historical video surveillance data, calculating the pixel probability distribution of the selected historical video surveillance data, the gray value being the value of the pixel point satisfying the condition R=G=B;

[0099] The pixel probability distribution formula of the selected historical video surveillance data is as follows:

[0100] ;

[0101] in, represents the pixel probability of the selected historical video surveillance data, Indicates the gray value is The number of pixels, is the total number of pixels;

[0102] S122, constructing an image cumulative distribution function according to the calculated pixel probability distribution of the selected historical video surveillance data;

[0103] The formula for constructing the image cumulative distribution function is as follows:

[0104] ;

[0105] in, Represents the total number of gray values ​​in the image, Indicates The gray value after the gray value is mapped by the cumulative distribution function;

[0106] S123, performing image enhancement on the selected historical video surveillance data based on the constructed image cumulative distribution function;

[0107] Multiply the gray value mapped by the cumulative distribution function by ( -1);

[0108] The transformation function formula is as follows:

[0109] ;

[0110] in, Indicates about The transformation function is used to make the gray level of the equalized image consistent with the original image before normalization;

[0111] According to the mapping relationship between the image cumulative distribution function and the transformation function, refer to the pixels in the original image and output the image after histogram equalization;

[0112] S13, setting the enhanced historical video surveillance data as the pre-processed historical video surveillance data;

[0113] S2. Based on the pre-processed historical video surveillance data, extract abnormal behavior characteristics and spatiotemporal characteristics in the historical video surveillance data through machine learning;

[0114] Based on the pre-processed historical video surveillance data, extracting abnormal behavior features and spatiotemporal features in the historical video surveillance data by machine learning includes the following steps:

[0115] S21, taking the obtained pre-processed historical video surveillance data as input and inputting it into the three-dimensional convolutional neural network;

[0116] S22, extracting features from the input historical video surveillance data through a three-dimensional convolutional neural network;

[0117] S221, set the input historical video surveillance data size to ;

[0118] in, Indicates the number of channels, Indicates the frame length of historical video surveillance data. Indicates the height of the frame. Indicates the width of the frame;

[0119] Set the convolution kernel size of the 3D convolutional neural network to ;

[0120] in, represents the time depth of the convolution kernel, Indicates the size of the convolution kernel space;

[0121] S222, after receiving the input historical video surveillance data, the convolution layer moves on the input historical video surveillance data through the set convolution kernel according to the set step size, and performs convolution calculation on the corresponding area of ​​each step with the feature value of the area, so as to realize feature extraction of each historical video surveillance data;

[0122] The convolution calculation formula is as follows:

[0123] ;

[0124] in, Represents the input historical video surveillance data, represents the weight of the corresponding convolution kernel, b Indicates the bias value, f represents the output features;

[0125] S223, during the continuous convolution of the three-dimensional convolutional neural network, the output of the previous three-dimensional convolutional neural network is used as the input of the next three-dimensional convolutional neural network, and the data is processed by the activation function and the residual function in the process of inputting the data into the next three-dimensional convolutional neural network;

[0126] The activation function processing includes the following steps:

[0127] Set the input value to (λ=1,2,……,L), each output value The corresponding input weight is , b is the offset, and the output result obtained after inputting the input value into the three-dimensional convolutional neural network is:

[0128] ;

[0129] in, is the corresponding activation function, is the output result of the activation function;

[0130] The processing of the residual function includes the following steps:

[0131] ;

[0132] in, represents the output of the residual module, represents the residual function;

[0133] S224, extracting features of historical video surveillance data in a continuously stacked convolution and pooling process;

[0134] S225, passing the extracted features of the historical video surveillance data into the fully connected layer;

[0135] S226, expanding and combining the data through a fully connected layer to obtain a set of feature data, and saving it;

[0136] According to the set convolution kernel, a set of feature data is obtained, the first third of which represents spatiotemporal features, and the last two thirds represent abnormal behavior features;

[0137] S3, setting the reminder frequency of surveillance cameras based on the spatiotemporal features of abnormal behaviors extracted from historical video surveillance data;

[0138] Setting the surveillance camera reminder frequency based on the spatiotemporal features of abnormal behaviors extracted from historical video surveillance data includes the following steps:

[0139] Based on the frequency of the spatiotemporal characteristics of abnormal behaviors extracted from historical video surveillance data, determine the time period when the abnormal behaviors mainly occur, and set the surveillance camera reminder frequency based on the determined time period when the abnormal behaviors mainly occur;

[0140] S4. Set a video surveillance event. When the administrator selects a frame of video surveillance data in the corresponding time period, the video screen of the frame of video surveillance data is automatically enlarged, and the selected video screen is identified based on the machine learning method and data matching method to obtain the identified video surveillance data;

[0141] The selected video images are identified based on the machine learning method and the data matching method, and the identified video surveillance data is obtained, including the following steps:

[0142] S41, extracting abnormal behavior features of the selected video images in real time by machine learning;

[0143] S42, matching the abnormal behavior features of the selected video screen extracted in real time with the abnormal behavior features in the stored historical video surveillance data by means of data matching;

[0144] The similarity between the abnormal behavior features extracted from the selected video screen in real time and the abnormal behavior features in the saved historical video surveillance data is determined by calculating the cosine distance between the features;

[0145] The similarity calculation method is as follows:

[0146] ;

[0147] in, For vector and vector The similarity between represents the coordinates of the i-th point, Indicates the total number of pixels;

[0148] A similarity threshold is set. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is greater than or equal to the set similarity threshold, it indicates that the selected video screen has abnormal behavior. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is less than the set similarity threshold, it indicates that the selected video screen does not have abnormal behavior.

[0149] S5. Generate and save inspection records based on the identified video surveillance data;

[0150] S6. Set inspection evaluation management standards and evaluate and optimize the inspection records generated by administrators;

[0151] Setting inspection evaluation management standards and evaluating and optimizing the inspection records generated by administrators includes the following steps:

[0152] Calculate the recognition accuracy under the current surveillance camera reminder frequency;

[0153] ;

[0154] Among them, TP represents the number of recognition results that are consistent with the actual results; FN represents the number of missed detections; R is the accuracy rate;

[0155] Set the accuracy rate as the inspection evaluation management standard and set the accuracy rate threshold. When the accuracy rate in the inspection record generated by the administrator is less than the set accuracy rate threshold, increase the reminder frequency of the surveillance camera; when the accuracy rate in the inspection record generated by the administrator is greater than or equal to the set accuracy rate threshold, it means that the inspection record is qualified;

[0156] Embodiment 2: This embodiment also discloses an artificial intelligence-based surveillance video intelligent inspection management system, including: a data acquisition module, a data processing module, a feature extraction module, a monitoring adjustment module, an identification and matching module, and an inspection evaluation management module;

[0157] The data acquisition module is used to collect video surveillance data in real time and transmit the collected video surveillance data to the data processing module;

[0158] The data processing module is used to process the received video surveillance data;

[0159] The feature extraction module is used to extract abnormal behavior features and spatiotemporal features from the processed video surveillance data;

[0160] The monitoring adjustment module is used to perform monitoring adjustment according to the extracted spatiotemporal features;

[0161] The identification and matching module is used to perform identification and matching based on the extracted abnormal behavior features;

[0162] The inspection evaluation management module is used to evaluate the accuracy of the identification and matching results.

[0163] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent inspection management method for surveillance video based on artificial intelligence, characterized in that: The following steps are involved: S1. Collect historical video surveillance data, establish a sample set based on the collected historical video surveillance data, and preprocess the historical video surveillance data in the sample set to obtain preprocessed historical video surveillance data; S11. Select historical video surveillance data with abnormal behavior in the sample set; S12, performing image enhancement on the selected historical video surveillance data to obtain enhanced historical video surveillance data; S121, based on the gray value distribution in the selected historical video surveillance data, calculating the pixel probability distribution of the selected historical video surveillance data, the gray value being the value of the pixel point satisfying the condition R=G=B; The pixel probability distribution formula of the selected historical video surveillance data is as follows: ； in, represents the pixel probability of the selected historical video surveillance data, Indicates the gray value is The number of pixels, is the total number of pixels; S122, constructing an image cumulative distribution function according to the calculated pixel probability distribution of the selected historical video surveillance data; The formula for constructing the cumulative distribution function of the image is as follows: ； in, Represents the total number of gray values ​​in the image. Indicates The gray value after the gray value is mapped by the cumulative distribution function; S123, performing image enhancement on the selected historical video surveillance data based on the constructed image cumulative distribution function; Multiply the gray value mapped by the cumulative distribution function by ( -1); The transformation function formula is as follows: ； in, Indicates about The transformation function is used to make the gray level of the equalized image consistent with the original image before normalization; According to the mapping relationship between the image cumulative distribution function and the transformation function, refer to the pixels in the original image and output the image after histogram equalization; S13, setting the enhanced historical video surveillance data as the pre-processed historical video surveillance data; S2. Based on the pre-processed historical video surveillance data, extract abnormal behavior characteristics and spatiotemporal characteristics in the historical video surveillance data through machine learning; S3, setting the reminder frequency of surveillance cameras based on the spatiotemporal features of abnormal behaviors extracted from historical video surveillance data; S4. Set a video surveillance event. When the administrator selects a frame of video surveillance data in the corresponding time period, the video screen of the frame of video surveillance data is automatically enlarged, and the selected video screen is identified based on the machine learning method and data matching method to obtain the identified video surveillance data; S5. Generate and save inspection records based on the identified video surveillance data; S6. Set inspection evaluation management standards and evaluate and optimize the inspection records generated by administrators.

2. According to the artificial intelligence-based surveillance video intelligent inspection management method of claim 1, it is characterized in that: The method of extracting abnormal behavior features and spatiotemporal features from the historical video surveillance data by machine learning based on the pre-processed historical video surveillance data includes the following steps: S21, taking the obtained pre-processed historical video surveillance data as input and inputting it into the three-dimensional convolutional neural network; S22. Extract features from the input historical video surveillance data through a three-dimensional convolutional neural network.

3. According to claim 2, a surveillance video intelligent inspection management method based on artificial intelligence is characterized in that: The feature extraction of the input historical video surveillance data by the three-dimensional convolutional neural network comprises the following steps: S221, set the input historical video surveillance data size to ; in, Indicates the number of channels, Indicates the frame length of historical video surveillance data. Indicates the height of the frame. Indicates the width of the frame; Set the convolution kernel size of the 3D convolutional neural network to ; in, represents the time depth of the convolution kernel, Indicates the size of the convolution kernel space; S222, after receiving the input historical video surveillance data, the convolution layer moves on the input historical video surveillance data through the set convolution kernel according to the set step size, and performs convolution calculation on the corresponding area of ​​each step with the feature value of the area, so as to realize feature extraction of each historical video surveillance data; The convolution calculation formula is as follows: ； in, Represents the input historical video surveillance data, represents the weight of the corresponding convolution kernel, b Indicates the bias value, f represents the output features; S223, in the process of continuous convolution of the three-dimensional convolutional neural network, the output of the previous three-dimensional convolutional neural network will be used as the input of the next three-dimensional convolutional neural network, and the data will be processed by the activation function and the residual function in the process of inputting the data into the next three-dimensional convolutional neural network; S224, extracting features of historical video surveillance data in a continuously stacked convolution and pooling process; S225, passing the extracted features of the historical video surveillance data into the fully connected layer; S226, expanding and combining the data through a fully connected layer to obtain a set of feature data, and saving it; According to the set convolution kernel, a set of feature data is obtained, the first third of which represents the spatiotemporal features, and the last two thirds represent the abnormal behavior features.

4. The method for intelligent inspection and management of surveillance video based on artificial intelligence according to claim 3 is characterized in that: The process of inputting data into the lower three-dimensional convolutional neural network through the activation function and the residual function includes the following steps: The activation function processing includes the following steps: Set the input value to (λ=1,2,……,L), each output value The corresponding input weight is , b is the offset, and the output result obtained after inputting the input value into the three-dimensional convolutional neural network is: ； in, is the corresponding activation function, is the output result of the activation function; The processing of the residual function includes the following steps: ； in, represents the output of the residual module, represents the residual function.

5. The method for intelligent inspection and management of surveillance video based on artificial intelligence according to claim 1 is characterized in that: The method of identifying the selected video images based on the machine learning method and the data matching method to obtain the identified video surveillance data includes the following steps: S41, extracting abnormal behavior features of the selected video images in real time by machine learning; S42, matching the abnormal behavior features of the selected video images extracted in real time with the abnormal behavior features in the stored historical video surveillance data by means of data matching.

6. The method for intelligent inspection and management of surveillance video based on artificial intelligence according to claim 5 is characterized in that: The method of matching the abnormal behavior features of the real-time extracted selected video screen with the abnormal behavior features in the stored historical video surveillance data by data matching includes the following steps: The similarity between the abnormal behavior features extracted from the selected video screen in real time and the abnormal behavior features in the saved historical video surveillance data is determined by calculating the cosine distance between the features; The similarity calculation method is as follows: ； in, For vector and vector The similarity between represents the coordinates of the i-th point, Indicates the total number of pixels; A similarity threshold is set. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is greater than or equal to the set similarity threshold, it indicates that the selected video screen has abnormal behavior. When the similarity between the abnormal behavior characteristics of the selected video screen and the abnormal behavior characteristics in the saved historical video surveillance data is less than the set similarity threshold, it indicates that the selected video screen does not have abnormal behavior.

7. The method for intelligent inspection and management of surveillance video based on artificial intelligence according to claim 1 is characterized in that: The setting of inspection evaluation management standards and the evaluation and optimization of inspection records generated by administrators include the following steps: Calculate the recognition accuracy under the current surveillance camera reminder frequency; ； Among them, TP represents the number of recognition results that are consistent with the actual results; FN represents the number of missed detections; R is the accuracy rate; Set the accuracy rate as the inspection evaluation management standard and set the accuracy rate threshold. When the accuracy rate in the inspection record generated by the administrator is less than the set accuracy rate threshold, increase the reminder frequency of the surveillance camera; when the accuracy rate in the inspection record generated by the administrator is greater than or equal to the set accuracy rate threshold, it means that the inspection record is qualified.

8. A system for implementing the surveillance video intelligent inspection management method based on artificial intelligence as described in any one of claims 1 to 7, characterized in that: include: Data acquisition module, data processing module, feature extraction module, monitoring and adjustment module, identification and matching module, and inspection and evaluation management module; The data acquisition module is used to collect video surveillance data in real time and transmit the collected video surveillance data to the data processing module; The data processing module is used to process the received video surveillance data; The feature extraction module is used to extract abnormal behavior features and spatiotemporal features from the processed video surveillance data; The monitoring adjustment module is used to perform monitoring adjustment according to the extracted spatiotemporal features; The identification and matching module is used to perform identification and matching based on the extracted abnormal behavior features; The inspection evaluation management module is used to evaluate the accuracy of the identification and matching results.

Images