Monitoring video information intelligent processing system based on cloud computing

Abstract

The invention discloses a monitoring video information intelligent processing system based on cloud computing, which relates to the technical field of monitoring video information processing and comprises a monitoring information processing platform. The monitoring information processing platform is in communication connection with a video data acquisition module, a data transmission module, a cloud data storage module, a data stream processing module, a target analysis module, an anomaly detection module and an intelligent alarm management module. Through training the deep neural network model and learning the features of normal and abnormal behaviors, the method achieves the precise recognition and tracking of a target object in a video, detects a dynamic target in the video, extracts the motion and appearance features of the dynamic target, recognizes the abnormal behaviors through behavior analysis, calculates an abnormal score, and generates alarm information in time. The accuracy of the monitoring system is improved, the situations of false alarm and missing alarm are reduced, and safety monitoring is more efficient and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring video information processing, and in particular to a monitoring video information intelligent processing system based on cloud computing. Background Art

[0002] As a model of providing computing resources and services through the Internet, cloud computing has the main characteristics of virtualization, elastic expansion, pay-as-you-go and automated management. With the rapid development of cloud computing technology, its powerful computing and storage capabilities provide strong support for the intelligent processing of surveillance video information. With the acceleration of urbanization, the requirements for video surveillance systems are becoming higher and higher. With the continuous evolution of security technology, video surveillance technology has developed from analog and digital to networking and intelligent. The introduction of cloud computing technology provides new solutions for the intelligent processing of video surveillance systems.

[0003] In the prior art, when surveillance video information is acquired, it is easily affected by various factors such as lighting changes and occlusions, which leads to insufficient accuracy during analysis and processing, resulting in missed reports and false alarms when the surveillance video identifies the target object. Therefore, how to weaken the impact of environmental factors on surveillance video information and improve the accuracy of surveillance video target object recognition is the problem we need to solve. To this end, a surveillance video information intelligent processing system based on cloud computing is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a cloud computing-based surveillance video information intelligent processing system to solve the problems raised in the above-mentioned background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A cloud computing-based surveillance video information intelligent processing system includes a surveillance information processing platform, wherein the surveillance information processing platform is communicatively connected with a video data acquisition module, a data transmission module, a cloud data storage module, a data stream processing module, a target analysis module, an anomaly detection module, and an intelligent alarm management module; The video data acquisition module is used to collect monitoring video data from multiple cameras and sensors. The cameras are distributed in different monitoring areas and capture video signals in real time to ensure that the system can receive high-quality original video streams; The data transmission module is used to transmit the collected video data to the cloud server through a wireless network. It is necessary to ensure the stable and fast transmission of data for subsequent processing, ensure the timeliness and integrity of the video data, and provide a reliable data source for subsequent intelligent processing; The cloud data storage module is used to utilize the distributed storage capability of the cloud computing platform to perform large-scale and efficient storage of video data transmitted to the cloud server, thereby solving the capacity limitation and cost problems faced by traditional storage methods, and to perform classification, indexing, retrieval and management operations on the stored video data, so as to facilitate users to quickly find the required videos, provide almost unlimited storage space, reduce storage costs and improve data access efficiency; The data stream processing module is used to perform preprocessing operations on the video data to improve image quality and provide clearer images for subsequent analysis. The preprocessing operations include noise reduction, deblurring, contrast enhancement, etc. The target analysis module is used to analyze the video data using deep learning technology, identify and remove the static background in the video data, separate the dynamic target, reduce the false alarms caused by background changes, and adaptively process the environmental factors such as lighting changes and occlusions, significantly improve the accuracy of target object recognition, and reduce the occurrence of missed alarms and false alarms; The anomaly detection module is used to track the target object in the video in real time, detect abnormal behaviors or events in the video through in-depth analysis of the video content, discover potential safety hazards and abnormal situations, respond quickly to abnormal situations, and reduce missed reports; The intelligent alarm management module is used to generate alarm information according to abnormal detection results and preset rules, and manage the alarm process to improve the accuracy and response speed of the alarm.

[0006] A further improvement of the technical solution of the present invention is that in the video data acquisition module, the acquisition process of the monitoring video data is: Deploy cameras and sensors in different monitoring areas to ensure that all areas that need to be monitored are covered. Use cameras to capture video signals in the monitoring area in real time, and use sensors to capture sensor data in the monitoring area. Sensors include infrared sensors and motion sensors to provide more comprehensive environmental information. The captured video signal is preliminarily processed, and automatic exposure and white balance adjustments are performed to improve the image quality. The preliminarily processed video signal is then converted into a digital signal and coded and compressed to effectively reduce the file size while ensuring the video quality for easy storage and transmission; Through the wireless network, the encoded video data and sensor data are encapsulated into network data packets and transmitted to the data transmission module.

[0007] A further improvement of the technical solution of the present invention is that: in the cloud data storage module, the process of storing video data is: The cloud server receives the video data and sensor data from the data transmission module, and performs preliminary processing on the received video data and sensor data, including data format verification and integrity check, to ensure the accuracy and integrity of the data; Perform data slicing on video data and sensor data, divide the video data and sensor data into multiple data blocks, and use redundant storage technology to copy each data block to multiple storage nodes, and perform distributed storage on multiple storage nodes to prevent data loss caused by single point failures, which helps improve data reliability and access speed; Analyze the video data processed by data slicing, classify the video data according to its content, source, and time attributes, and build indexes for the classified video data. Create index information, including keyword index, time index, and metadata index, to facilitate quick retrieval; A user interface is set up to enable users to quickly retrieve required video clips by using indexes and metadata to search in multiple ways such as time range, camera location, and event type. Video clip preview and extraction functions are also provided to enable users to select and extract required video clips as needed.

[0008] A further improvement of the technical solution of the present invention is that: in the data stream processing module, the process of preprocessing the video data is: The data stream processing module receives the video data stream and sensor data from the cloud server, and decodes the encoded and compressed video data to restore it to the original video format; Extract individual video frames from a video stream and apply a smoothing filter to reduce noise and artifacts in the image; If the video data is blurred or motion blurred, a blind deblurring algorithm is used to estimate the blur kernel through an iterative optimization algorithm and restore a clear image; By adjusting the contrast of the video frame, the image is enhanced to make the details in the image more prominent and obvious, so as to optimize the visual effect of the image, and the brightness is automatically adjusted according to the video content and ambient lighting conditions; The preprocessed video frames are repackaged and combined into a video stream, and the preprocessed video data is transmitted to the target analysis module and the anomaly detection module for subsequent analysis and processing.

[0009] A further improvement of the technical solution of the present invention is that in the target analysis module, the process of separating dynamic targets is: Receive the preprocessed video data stream from the data stream processing module, and decode the video data stream to obtain continuous video image frames; Use convolutional neural networks to learn background frames in video data streams, build static background models, and identify static backgrounds; For the continuous video image frames of the subsequent video data stream, they are compared with the static background model, the dynamic target is separated by the background difference method, and the dynamic target is tracked by the target tracking algorithm to continuously identify the same target in the video data stream; The detection result of the target object is adjusted according to the lighting conditions of the image through brightness normalization, the lighting component and reflection component of the scene are estimated, and adaptive lighting processing is performed to improve the lighting conditions of the image. The deep learning model is used to analyze the partial occlusion of the target, identify the occluded area in the video image frame, and restore the occluded target information. The expression for estimating the lighting component and reflection component of the scene and performing adaptive lighting processing is: ; in, Represents the adjusted pixel value, represents the pixel value of the original image, represents the estimated reflection component; The separated targets are identified using a deep learning model to determine the target’s category and attributes, and to extract the target’s feature vector. The analysis and identification results are then output to the anomaly detection module to provide a basis for further analysis and decision-making. The target’s feature vector includes the target’s motion features and appearance features.

[0010] A further improvement of the technical solution of the present invention is that the process of establishing a static background model and identifying the static background is as follows: Select the initial background frame from the video data stream and preprocess the selected background frame, including scaling, normalization, data enhancement, etc., to improve the generalization ability of the model; Construct a CNN (Convolutional Neural Network) architecture for background learning, including multiple convolutional layers, activation layers, pooling layers, and fully connected layers. The convolutional layer extracts image features through filters, and there are multiple filters, each of which is responsible for extracting specific features in the image. The activation layer introduces nonlinearity to enable the model to learn complex features. The pooling layer is used to reduce the spatial dimension of the features and increase the invariance to position changes. The fully connected layer is used to map the features extracted by the convolutional layer and the pooling layer to the final output. Define the loss function to measure the difference between the model prediction and the actual label, input the preprocessed background frame into the CNN, calculate the output through forward propagation, calculate the loss, and update the network weights through backpropagation; Use the trained background model to compare with the new frame and detect dynamic targets by background difference method; The expression of the convolutional layer is: ; in, Indicates Layer in position The output, Indicates Layer in position Input, Indicates The convolution kernel weights of the layer, and Represents the horizontal and vertical offset of the convolution kernel, Indicates The bias term of the layer, Represents activation functions, including ReLU 、Sigmoid , Tanh Any one of; The pooling layer selects the maximum pooling, and its expression is: ; in, Indicates Layer in position The pooled output of Indicates Layer in position Input, and Represents the position offset within the pooling window.

[0011] A further improvement of the technical solution of the present invention is that the process of continuously identifying the same target in the video data stream is: Combined with the training, a static background model is obtained. For each new frame of video image, the background model is updated, and the difference between the current frame and the background model is calculated to obtain a differential image. A dynamic analysis threshold is set, and the correlation between the differential image and the dynamic analysis threshold is analyzed. Pixels above the dynamic analysis threshold are considered as dynamic targets, and a foreground mask is generated to separate the dynamic targets. Perform connected domain analysis on the foreground mask to identify independent objects in the foreground area, detect the target area in the foreground mask, extract the features of the target, and use the feature matching algorithm of Kalman filter to match the target in consecutive frames; The tracking Kalman filter updates the position and features of the target, allowing the same target to be continuously identified in the video sequence; The updating formula of the background model is: ; in, represents the pixel value of the background model, Represents the pixel value of the current frame, Represents the update factor ; The calculation expression of the difference image is: ; in, represents the difference result, Represents the pixel value of the current frame, Represents the pixel value of the background model; The judgment expression for generating the foreground mask is: ; in, represents the pixel value of the foreground mask, Indicates the dynamic analysis threshold.

[0012] A further improvement of the technical solution of the present invention is that in the anomaly detection module, the detection process of abnormal behavior or events in the video is: Using the detection results output by the target analysis module, the target object in the video is determined, and the feature vector of the target is extracted, including the motion features and appearance features of the target. By comparing the distance of the feature vectors, the target features are matched in consecutive frames. Among them, the appearance features include color histogram, texture features and shape descriptors, and the motion features include the speed, acceleration and trajectory of the target. Use machine learning to model normal behavior and build a behavior model. By comparing real-time video content with predefined behavior models, abnormal behavior is identified and the deviation between real-time behavior and the normal model is calculated to generate an abnormality score. The threshold of anomaly score is set based on historical data. When the anomaly score exceeds the threshold, it is judged as abnormal behavior. In multi-target tracking, overlapping detection results are removed to retain the best prediction results. Once abnormal behavior is detected, an alarm message is generated and sent to relevant personnel, providing detailed information of the abnormal behavior, including the time, location, and targets involved, for rapid response and processing.

[0013] A further improvement of the technical solution of the present invention is that the calculation process of the abnormality score is: Collect and annotate video data of normal behavior to extract the appearance and motion features of the target; Select the SVM (Support Vector Machine) machine learning model and use the labeled normal behavior video data to train the model to distinguish normal behavior from abnormal behavior; The features of the targets in the real-time video are extracted to obtain the feature vector of the real-time behavior. The feature vector of the real-time behavior is input into the trained model for analysis, the deviation between the real-time behavior features and the normal model is calculated, and the anomaly score is calculated based on the deviation.

[0014] A further improvement of the technical solution of the present invention is that the expression of the behavior model training is: ; in, represents the decision function, represents the test sample, Represents the kernel function, calculates the sample and test samples The similarity between Indicates The feature vector of the training samples, Indicates The labels of training samples refer to normal behavior or abnormal behavior. Indicates The coefficients of the support vectors, Represents the bias term, which is a constant term in the SVM model; The abnormal behavior recognition is achieved by calculating the deviation between the real-time behavior characteristics and the normal model, combined with cosine similarity judgment, and its expression is: ; in, Represents cosine similarity, which measures the similarity between two vectors, with a value range of -1 to 1. represents the target feature vector extracted from the real-time video frame, The feature vector representing the normal behavior model, and Represents vectors and The The eigenvalues ​​represent Represents the dimension of the feature vector, that is, the total number of eigenvalues ​​in the vector, which is used to describe the total number of features of the target feature; The expression for calculating the abnormality score is: ; in, represents the abnormality score, Represents cosine similarity.

[0015] Due to the adoption of the above technical solution, the present invention has the following technical advances compared with the prior art: The present invention provides a monitoring video information intelligent processing system based on cloud computing, which can accurately identify and track target objects in videos by training deep neural network models, learning the characteristics of normal and abnormal behaviors, and extracting their motion and appearance characteristics, identifying abnormal behaviors through behavior analysis, calculating abnormal scores, and generating alarm information in a timely manner, which not only improves the accuracy of the monitoring system, but also reduces the situation of false positives and false negatives, making security monitoring more efficient and reliable. The present invention provides a monitoring video information intelligent processing system based on cloud computing, which can quickly and accurately detect abnormal behaviors in videos by analyzing video data in real time, and process data streams of multiple cameras in parallel through a cloud computing platform, effectively shortening the response time and improving the detection efficiency. At the same time, by collecting and analyzing a large amount of historical video data, a precise behavior pattern recognition model is trained using a machine learning algorithm, and then the characteristics and laws of normal behaviors are learned and understood, so as to more accurately identify abnormal behaviors, and predict future behaviors of identified behavior patterns, and warn potential risk points in advance, so as to provide users with more comprehensive and in-depth security analysis reports, which are helpful to better understand the security status of the monitoring area. The present invention provides a cloud computing-based surveillance video information intelligent processing system. Based on the cloud computing platform, video data is centrally stored and processed to simplify system deployment and management. The status of all surveillance points can be monitored in real time through the cloud management interface, and problems can be quickly located and solved. At the same time, the resources of the cloud computing platform can be allocated on demand to avoid resource waste and ensure that each surveillance point can obtain sufficient computing power, thereby improving the overall monitoring efficiency and effect. In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0016] Figure 1 It is a module diagram of the present invention; Figure 2 A flow chart for separating dynamic targets of the present invention; Figure 3 A flow chart of continuously identifying the same target in a video data stream of the present invention; Figure 4 The figure is a flow chart of detecting abnormal behaviors or events in a video of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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. Examples, such as Figure 1 As shown, the present invention provides a cloud computing-based monitoring video information intelligent processing system, including a monitoring information processing platform, characterized in that: the monitoring information processing platform is communicatively connected with a video data acquisition module, a data transmission module, a cloud data storage module, a data stream processing module, a target analysis module, an anomaly detection module and an intelligent alarm management module.

[0018] The video data acquisition module is used to collect monitoring video data from multiple cameras and sensors. The cameras are distributed in different monitoring areas and capture video signals in real time to ensure that the system can receive high-quality original video streams.

[0019] The data transmission module is used to transmit the collected video data to the cloud server via a wireless network. It is necessary to ensure the stable and fast transmission of data for subsequent processing, ensure the timeliness and integrity of the video data, and provide a reliable data source for subsequent intelligent processing.

[0020] The cloud data storage module is used to utilize the distributed storage capabilities of the cloud computing platform to perform large-scale and efficient storage of video data transmitted to the cloud server, solving the capacity limitations and cost issues faced by traditional storage methods. It also classifies, indexes, and retrieves the stored video data, making it easier for users to quickly find the videos they need, providing almost unlimited storage space, reducing storage costs, and improving data access efficiency.

[0021] The data stream processing module is used to perform preprocessing operations on video data to improve image quality and provide clearer images for subsequent analysis. The preprocessing operations include noise reduction, deblurring, contrast enhancement, etc.

[0022] The target analysis module is used to analyze video data using deep learning technology, identify and remove static backgrounds in video data, separate dynamic targets, reduce false alarms caused by background changes, and adaptively process environmental factors such as lighting changes and occlusions, significantly improving the accuracy of target object recognition and reducing the occurrence of missed alarms and false alarms.

[0023] The anomaly detection module is used to track the target objects in the video in real time. Through in-depth analysis of the video content, it detects abnormal behaviors or events in the video, discovers potential safety hazards and abnormal situations, responds quickly to abnormal situations, and reduces missed reports.

[0024] The intelligent alarm management module is used to generate alarm information based on abnormal detection results and preset rules, and manage the alarm process to improve the accuracy and response speed of the alarm.

[0025] The cloud computing-based intelligent processing system for surveillance video information of the present invention learns the characteristics of normal and abnormal behaviors by training a deep neural network model, thereby realizing accurate recognition and tracking of target objects in the video, detecting dynamic targets in the video, and extracting their motion and appearance characteristics, identifying abnormal behaviors through behavioral analysis, calculating abnormal scores, and generating alarm information in a timely manner, which not only improves the accuracy of the monitoring system, but also reduces false positives and false negatives, making security monitoring more efficient and reliable.

[0026] In some embodiments, in the video data acquisition module, the acquisition process of monitoring video data is: Cameras and sensors are deployed in different monitoring areas to ensure that all areas that need to be monitored are covered. Cameras are used to capture video signals in the monitoring area in real time, and sensors are used to capture sensor data in the monitoring area. Sensors include infrared sensors and motion sensors to provide more comprehensive environmental information. The captured video signals are preliminarily processed, and automatic exposure and white balance adjustments are performed to improve image quality. The preliminarily processed video signals are then converted into digital signals and encoded and compressed to effectively reduce file size while ensuring video quality for easy storage and transmission. The encoded video data and sensor data are encapsulated into network data packets through wireless networks and transmitted to the data transmission module.

[0027] In some embodiments, in the cloud data storage module, the process of storing video data is: The cloud server receives the video data and sensor data from the data transmission module, and performs preliminary processing on the received video data and sensor data, wherein the preliminary processing includes data format verification and integrity check to ensure the accuracy and integrity of the data; performs data slicing processing on the video data and sensor data, divides the video data and sensor data into multiple data blocks, and adopts redundant storage technology to copy each data block to multiple storage nodes, and performs distributed storage on multiple storage nodes to prevent data loss caused by single point failure, which helps to improve data reliability and access speed; analyzes the video data processed by data slicing, classifies the video data according to the content, source, and time attributes, and builds an index for the classified video data, establishes index information, and the index information includes a keyword index, a time index, and a metadata index to facilitate rapid retrieval; sets a user interface to enable users to use indexes and metadata to search in multiple ways by time range, camera location, and event type, quickly retrieve the required video clips, and provide preview and extraction functions for video clips, so that users can select and extract the required video clips as needed.

[0028] In some embodiments, in the data stream processing module, the process of preprocessing the video data is: The data stream processing module receives the video data stream and sensor data from the cloud server, and decodes the encoded and compressed video data to restore it to the original video format; extracts individual video frames from the video stream and applies a smoothing filter to reduce noise and artifacts in the image, thereby improving the clarity and visual quality of the image; if the video data has blur or motion blur problems, a blind deblurring algorithm is used to estimate the blur kernel through an iterative optimization algorithm and restore a clear image; by adjusting the contrast of the video frame, the image is contrast enhanced to make the details in the image more prominent and obvious, so as to optimize the visual effect of the image and facilitate subsequent analysis, and automatically adjusts the brightness according to the video content and ambient lighting conditions to ensure the visual effect of the image; the preprocessed video frames are repackaged and combined into a video stream, and the preprocessed video data is transmitted to the target analysis module and the anomaly detection module for subsequent analysis and processing.

[0029] In some embodiments, Figure 2 As shown in the figure, in the target analysis module, the process of separating dynamic targets is: The preprocessed video data stream is received from the data stream processing module, and the video data stream is decoded to obtain continuous video image frames; the background frames in the video data stream are learned by using a convolutional neural network, a static background model is established, and the static background is identified; for the continuous video image frames of the subsequent video data stream, they are compared with the static background model, the dynamic target is separated by the background difference method, and the dynamic target is tracked by using the target tracking algorithm, and the same target is continuously identified in the video data stream; the detection result of the target object is adjusted according to the lighting conditions of the image by brightness normalization, the lighting component and the reflection component of the scene are estimated, the lighting is adaptively processed, the lighting conditions of the image are improved, and the partial occlusion of the target is analyzed by using a deep learning model, the occluded area in the video image frame is identified, and the occluded target information is restored; the separated target is identified by using a deep learning model, the category and attribute of the target are determined, and the feature vector of the target is extracted, and the analysis and identification result is output to the anomaly detection module to provide a basis for further analysis and decision-making, wherein the feature vector of the target includes the motion feature and appearance feature of the target.

[0030] Furthermore, in some embodiments, the illumination component and the reflection component of the scene are estimated, and the expression for adaptive illumination processing is: ; in, Represents the adjusted pixel value, represents the pixel value of the original image, Represents the estimated reflection component.

[0031] Furthermore, a static background model is established and the process of identifying the static background is as follows: Select the initial background frame from the video data stream and preprocess the selected background frame, including scaling, normalization, data enhancement, etc., to improve the generalization ability of the model; Construct a CNN architecture for background learning, including multiple convolutional layers, activation layers, pooling layers, and fully connected layers. The convolutional layer extracts image features through filters, and there are multiple filters, each of which is responsible for extracting specific features in the image. The activation layer introduces nonlinearity to enable the model to learn complex features. The pooling layer is used to reduce the spatial dimension of the features and increase the invariance to position changes. The fully connected layer is used to map the features extracted by the convolutional layer and the pooling layer to the final output. Define the loss function to measure the difference between the model prediction and the actual label, input the preprocessed background frame into the CNN, calculate the output through forward propagation, calculate the loss, and update the network weights through backpropagation; Use the trained background model to compare with the new frame and detect dynamic targets by background difference method; The expression of the convolutional layer is: ; in, Indicates Layer in position The output, Indicates Layer in position Input, Indicates The convolution kernel weights of the layer, and Represents the horizontal and vertical offset of the convolution kernel, Indicates The bias term of the layer, Represents activation functions, including ReLU 、Sigmoid , Tanh Any of the above, it should be noted that the convolutional layer extracts useful information by learning the features of the input data and maps these features to a higher level representation; The pooling layer selects the maximum pooling, and its expression is: ; in, Indicates Layer in position The pooled output of Indicates Layer in position Input, and Represents the position offset within the pooling window. It should be noted that the pooling layer increases the invariance to position changes by reducing the spatial dimension of the feature, while reducing the amount of calculation and preventing overfitting.

[0032] In addition, if Figure 3 As shown in Figure 1, the process of continuously identifying the same target in the video data stream is: Combined with the training, a static background model is obtained. For each new frame of video image, the background model is updated, and the difference between the current frame and the background model is calculated to obtain a differential image. A dynamic analysis threshold is set, and the correlation between the differential image and the dynamic analysis threshold is analyzed. Pixels above the dynamic analysis threshold are considered as dynamic targets, and a foreground mask is generated to separate the dynamic targets. Perform connected domain analysis on the foreground mask to identify independent objects in the foreground area, detect the target area in the foreground mask, extract the features of the target, and use the feature matching algorithm of Kalman filter to match the target in consecutive frames; The tracking Kalman filter updates the position and features of the target so that the same target is continuously identified in the video sequence.

[0033] The update formula of the background model is: ; in, represents the pixel value of the background model, Represents the pixel value of the current frame, Represents the update factor ; The calculation expression of the difference image is: ; in, represents the difference result, Represents the pixel value of the current frame, Represents the pixel value of the background model; The judgment expression for generating the foreground mask is: ; in, represents the pixel value of the foreground mask, Indicates the dynamic analysis threshold.

[0034] In some embodiments, Figure 4 As shown in the figure, in the anomaly detection module, the detection process of abnormal behaviors or events in the video is: The detection results output by the target analysis module are used to determine the target object in the video, and the feature vector of the target is extracted, including the motion features and appearance features of the target. The target features are matched in consecutive frames by comparing the distance of the feature vectors, where the appearance features include color histograms, texture features and shape descriptors, and the motion features include the speed, acceleration and trajectory of the target. Machine learning is used to model normal behavior and build a behavior model. By comparing the real-time video content with the predefined behavior model, abnormal behavior is identified, and the deviation between the real-time behavior and the normal model is calculated to generate an abnormal score. The threshold of the abnormal score is set according to historical data. When the abnormal score exceeds the threshold, it is determined to be abnormal behavior. In multi-target tracking, overlapping detection results are removed and the best prediction results are retained. Once abnormal behavior is detected, an alarm message is generated and sent to relevant personnel, providing detailed information of the abnormal behavior, including the time, location, and targets involved, for rapid response and processing.

[0035] Furthermore, the calculation process of the anomaly score is: Collect and annotate video data of normal behavior to extract the appearance and motion features of the target; Select the SVM machine learning model and train the model using labeled normal behavior video data to distinguish normal behavior from abnormal behavior; Extract features of targets in real-time video to obtain feature vectors of real-time behaviors. Input the feature vectors of real-time behaviors into the trained model for analysis. Calculate the deviation between the real-time behavior features and the normal model, and calculate the anomaly score based on the deviation. Furthermore, the expression for behavior model training is: ; in, represents the decision function, represents the test sample, Represents the kernel function, calculates the sample and test samples The similarity between Indicates The feature vector of the training samples, Indicates The labels of training samples refer to normal behavior or abnormal behavior. Indicates The coefficients of the support vectors, Represents the bias term, which is a constant term in the SVM model; Abnormal behavior recognition is achieved by calculating the deviation between real-time behavior features and normal models, combined with cosine similarity judgment. The expression is: ; in, Represents cosine similarity, which measures the similarity between two vectors, with a value range of -1 to 1. represents the target feature vector extracted from the real-time video frame, The feature vector representing the normal behavior model, and Represents vectors and The The eigenvalues ​​represent Represents the dimension of the feature vector, that is, the total number of eigenvalues ​​in the vector, which is used to describe the total number of features of the target feature. It should be noted that The closer the value of is to 1, the more similar the two vectors are, that is, the closer the target behavior is to the normal behavior model; The closer the value of is to -1, the less similar the two vectors are, that is, the greater the difference between the target behavior and the normal behavior model; The expression for calculating the anomaly score is: ; in, represents the abnormality score, Represents cosine similarity.

[0036] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cloud computing-based surveillance video information intelligent processing system, including a surveillance information processing platform, characterized in that: The monitoring information processing platform is communicatively connected with a video data acquisition module, a data transmission module, a cloud data storage module, a data stream processing module, a target analysis module, an anomaly detection module and an intelligent alarm management module; The video data acquisition module is used to collect monitoring video data from multiple cameras and sensors, where the cameras are distributed in different monitoring areas and capture video signals in real time; The data transmission module is used to transmit the collected video data to the cloud server via a wireless network; The cloud data storage module is used to utilize the distributed storage capability of the cloud computing platform to store the video data transmitted to the cloud server, and to perform classification, indexing, retrieval and management operations on the stored video data; The data stream processing module is used to perform preprocessing operations on video data to improve image quality; The target analysis module is used to analyze the video data using deep learning technology, identify and remove the static background in the video data, separate the dynamic target, and perform adaptive processing on environmental factors such as lighting changes and occlusion; The anomaly detection module is used to track the target object in the video in real time, detect abnormal behaviors or events in the video through in-depth analysis of the video content, and discover potential safety hazards and abnormal situations; The intelligent alarm management module is used to generate alarm information according to abnormal detection results and preset rules, and manage the alarm process.

2. The cloud computing-based surveillance video information intelligent processing system according to claim 1 is characterized in that: In the video data acquisition module, the acquisition process of monitoring video data is as follows: Deploy cameras and sensors in different monitoring areas, and use cameras to capture video signals in the monitoring area in real time, and use sensors to capture sensor data in the monitoring area. The sensors include infrared sensors and motion sensors. Perform preliminary processing on the captured video signal, perform automatic exposure adjustment and white balance adjustment, and then convert the preliminary processed video signal into a digital signal for encoding and compression; Through the wireless network, the encoded video data and sensor data are encapsulated into network data packets and transmitted to the data transmission module.

3. The cloud computing-based surveillance video information intelligent processing system according to claim 1, characterized in that: In the cloud data storage module, the process of storing video data is as follows: The cloud server receives the video data and sensor data from the data transmission module, and performs preliminary processing on the received video data and sensor data; Perform data slicing on video data and sensor data, divide the video data and sensor data into multiple data blocks, and use redundant storage technology to copy each data block to multiple storage nodes for distributed storage on multiple storage nodes; Analyze the video data processed by data slicing, classify the video data according to its content, source, and time attributes, and build indexes for the classified video data, and establish index information, including keyword index, time index, and metadata index; A user interface is set up to enable users to quickly retrieve required video clips by using indexes and metadata to search in multiple ways such as time range, camera location, and event type. Video clip preview and extraction functions are also provided to enable users to select and extract required video clips as needed.

4. The cloud computing-based surveillance video information intelligent processing system according to claim 1, characterized in that: In the data stream processing module, the process of preprocessing the video data is as follows: The data stream processing module receives the video data stream and sensor data from the cloud server, and decodes the encoded and compressed video data to restore it to the original video format; Extract individual video frames from a video stream and apply a smoothing filter to reduce noise and artifacts in the image; If the video data is blurred or motion blurred, a blind deblurring algorithm is used to estimate the blur kernel through an iterative optimization algorithm and restore a clear image; By adjusting the contrast of the video frame, the image is enhanced to optimize the visual effect of the image, and the brightness is automatically adjusted according to the video content and ambient lighting conditions; The preprocessed video frames are repackaged and combined into a video stream, and the preprocessed video data is transmitted to the target analysis module and the anomaly detection module for subsequent analysis and processing.

5. The cloud computing-based surveillance video information intelligent processing system according to claim 1, characterized in that: In the target analysis module, the process of separating dynamic targets is as follows: Receive the preprocessed video data stream from the data stream processing module, and decode the video data stream to obtain continuous video image frames; Use convolutional neural networks to learn background frames in video data streams, build static background models, and identify static backgrounds; For the continuous video image frames of the subsequent video data stream, they are compared with the static background model, the dynamic target is separated by the background difference method, and the dynamic target is tracked by the target tracking algorithm to continuously identify the same target in the video data stream; Adjust the detection results of the target object according to the lighting conditions of the image through brightness normalization, estimate the lighting component and reflection component of the scene, perform adaptive lighting processing, improve the lighting conditions of the image, and use a deep learning model to analyze the partial occlusion of the target, identify the occluded area in the video image frame, and restore the occluded target information; The separated targets are identified using a deep learning model to determine the target’s category and attributes, and to extract the target’s feature vector. The analysis and identification results are then output to an anomaly detection module, where the target’s feature vector includes the target’s motion features and appearance features.

6. The cloud computing-based surveillance video information intelligent processing system according to claim 5 is characterized in that: The process of establishing a static background model and identifying the static background is as follows: Selecting an initial background frame from the video data stream and preprocessing the selected background frame; Construct a CNN architecture for background learning, including multiple convolutional layers, activation layers, pooling layers, and fully connected layers. The convolutional layer extracts image features through filters, and there are multiple filters, each of which is responsible for extracting specific features in the image. Define the loss function to measure the difference between the model prediction and the actual label, input the preprocessed background frame into the CNN, calculate the output through forward propagation, calculate the loss, and update the network weights through backpropagation; Use the trained background model to compare with the new frame and detect dynamic targets by background difference method; The expression of the convolutional layer is: ； in, Indicates Layer in position The output, Indicates Layer in position Input, Indicates The convolution kernel weights of the layer, and Represents the horizontal and vertical offset of the convolution kernel, Indicates The bias term of the layer, represents the activation function; The pooling layer selects the maximum pooling, and its expression is: ； in, Indicates Layer in position The pooled output of Indicates Layer in position Input, and Represents the position offset within the pooling window.

7. The cloud computing-based surveillance video information intelligent processing system according to claim 5, characterized in that: The process of continuously identifying the same target in the video data stream is as follows: Combined with the training, a static background model is obtained. For each new frame of video image, the background model is updated, and the difference between the current frame and the background model is calculated to obtain a differential image. A dynamic analysis threshold is set, and the correlation between the differential image and the dynamic analysis threshold is analyzed. Pixels above the dynamic analysis threshold are considered as dynamic targets, and a foreground mask is generated to separate the dynamic targets. Perform connected domain analysis on the foreground mask to identify independent objects in the foreground area, detect the target area in the foreground mask, extract the features of the target, and use the feature matching algorithm of Kalman filter to match the target in consecutive frames; The tracking Kalman filter updates the position and features of the target, allowing the same target to be continuously identified in the video sequence; The updating formula of the background model is: ； in, represents the pixel value of the background model, Represents the pixel value of the current frame, Represents the update factor ; The calculation expression of the difference image is: ； in, represents the difference result, Represents the pixel value of the current frame, Represents the pixel value of the background model; The judgment expression for generating the foreground mask is: ； in, represents the pixel value of the foreground mask, Indicates the dynamic analysis threshold.

8. The cloud computing-based surveillance video information intelligent processing system according to claim 1, characterized in that: In the anomaly detection module, the process of detecting abnormal behaviors or events in the video is as follows: Using the detection results output by the target analysis module, the target object in the video is determined, and the feature vector of the target is extracted, including the motion features and appearance features of the target. By comparing the distance of the feature vectors, the target features are matched in consecutive frames. Among them, the appearance features include color histogram, texture features and shape descriptors, and the motion features include the speed, acceleration and trajectory of the target. Use machine learning to model normal behavior and build a behavior model. By comparing real-time video content with predefined behavior models, abnormal behavior is identified and the deviation between real-time behavior and the normal model is calculated to generate an abnormality score. The threshold of anomaly score is set based on historical data. When the anomaly score exceeds the threshold, it is judged as abnormal behavior. In multi-target tracking, overlapping detection results are removed to retain the best prediction results. Once abnormal behavior is detected, an alarm message is generated and sent to relevant personnel, providing detailed information about the abnormal behavior, including the time, location, and targets involved.

9. The cloud computing-based surveillance video information intelligent processing system according to claim 8, characterized in that: The calculation process of the anomaly score is: Collect and annotate video data of normal behavior to extract the appearance and motion features of the target; Select the SVM machine learning model and train the model using labeled normal behavior video data to distinguish normal behavior from abnormal behavior; The features of the targets in the real-time video are extracted to obtain the feature vector of the real-time behavior. The feature vector of the real-time behavior is input into the trained model for analysis, the deviation between the real-time behavior features and the normal model is calculated, and the anomaly score is calculated based on the deviation.

10. The cloud computing-based surveillance video information intelligent processing system according to claim 9, characterized in that: The expression of the behavior model training is: ； in, represents the decision function, represents the test sample, Represents the kernel function, calculates the sample and test samples The similarity between Indicates The feature vector of the training samples, Indicates The labels of training samples refer to normal behavior or abnormal behavior. Indicates The coefficients of the support vectors, Represents the bias term, which is a constant term in the SVM model; The abnormal behavior recognition is achieved by calculating the deviation between the real-time behavior characteristics and the normal model, combined with cosine similarity judgment, and its expression is: ； in, Represents cosine similarity, which measures the similarity between two vectors, with a value range of -1 to 1. represents the target feature vector extracted from the real-time video frame, The feature vector representing the normal behavior model, and Represents vectors and The The eigenvalues ​​represent Represents the dimension of the feature vector, that is, the total number of eigenvalues ​​in the vector, which is used to describe the total number of features of the target feature; The expression for calculating the abnormality score is: ； in, represents the abnormality score, Represents cosine similarity.

Images