Method and system for abnormal recognition of surveillance images based on artificial intelligence

Abstract

The present invention relates to the field of surveillance image recognition, and more specifically, to a surveillance image anomaly recognition method and system based on artificial intelligence. The method comprises: obtaining a preprocessed behavior image according to a preset sampling interval, and obtaining a label of the behavior image; constructing a multi-dimensional behavior feature, and calculating the amount of information of each dimension and the importance of each dimension to the abnormal behavior according to the behavior feature; sorting the dimensions according to the importance, and filtering the dimensions according to the amount of information to obtain the dimension screening result; establishing a behavior recognition model, inputting the feature value sequence corresponding to the dimension into the behavior recognition model, outputting the probability of abnormal behavior, accumulating the probabilities of the same abnormal behavior output by different behavior recognition models to obtain the final abnormal probability, and completing the abnormal recognition. Through the technical solution of the present invention, the accuracy of the abnormal recognition result of the surveillance image can be improved.

Description

Technical Field

[0001] The present invention relates to the field of surveillance image recognition, and more specifically, to a surveillance image anomaly recognition method and system based on artificial intelligence. Background Art

[0002] Traditional surveillance image analysis methods often rely on manual monitoring and identification of abnormal events, which has problems such as low efficiency and delayed response. Therefore, how to achieve efficient and automated detection and identification of abnormal events in surveillance images has become an important research direction in the field of computer vision and artificial intelligence.

[0003] The existing Chinese patent application document with publication number CN117392615A discloses a method and system for abnormality recognition based on surveillance video, including: establishing a noise attenuation model based on an initial surveillance image set, and denoising the images in the initial surveillance image set; training a preset prediction model through a training image set, and determining abnormal video frame images based on the predicted video frame images and the denoised surveillance image set; extracting a target prediction frame of the abnormal video frame image, determining the abnormal target posture based on the target prediction frame, inputting the abnormal target posture into the trained preset recognition model, and outputting the abnormal recognition result.

[0004] However, the above application documents only consider the behavioral characteristics of a single target when identifying abnormal behavior through key point distances, while ignoring the possible interactive relationships between multiple targets. It is difficult to make accurate judgments based only on the action characteristics of a single target, resulting in inaccurate abnormality recognition results in surveillance images. Summary of the invention

[0005] In order to solve the problem of inaccurate monitoring image anomaly recognition results, the present invention proposes a monitoring image anomaly recognition method and system based on artificial intelligence.

[0006] In a first aspect, the present invention discloses a monitoring image anomaly recognition method based on artificial intelligence, comprising: obtaining a preprocessed behavior image according to a preset sampling interval, and obtaining a label of the behavior image; inputting the behavior image into a key point detection model to obtain a number of key points, calculating the Euclidean distance between any two key points, using all Euclidean distances as multi-dimensional behavior features, and calculating the amount of information of each dimension and the importance of each dimension to the abnormal behavior according to the behavior features; sorting the dimensions according to the importance, and filtering the dimensions according to the amount of information to obtain a dimension screening result; establishing a behavior recognition model for each dimension in the dimension screening result, wherein one dimension corresponds to a behavior recognition model, inputting a feature value sequence corresponding to the dimension into the behavior recognition model, outputting the probability of abnormal behavior, accumulating the probabilities of the same abnormal behavior output by different behavior recognition models to obtain a final abnormal probability, and completing abnormal recognition; wherein the final abnormal probability satisfies the relationship:

[0007] , Indicates abnormal behavior The final abnormal probability, Indicates the dimension in the dimension filtering result Abnormal behavior The importance of Indicates the dimension in the dimension filtering result The corresponding behavior recognition model outputs abnormal behavior The probability of Indicates the total number of dimensions in the dimension filtering results.

[0008] First, through key point detection and Euclidean distance calculation, the spatial features in the behavior image are extracted, and a multi-dimensional behavior feature is constructed to accurately describe the subtle differences in behavior patterns. Then, each dimension is sorted and filtered according to the amount of information and importance to ensure that the model focuses on the most critical features for abnormal behavior identification and avoids the interference of redundant features. On this basis, an independent behavior recognition model is constructed for each filtered dimension, so that each dimension can focus on its most recognizable behavior performance. By weighted fusion of abnormal behavior probabilities output by different models, the accuracy and stability of recognition are further improved. Through accumulation and weighting, the final probability of abnormal behavior can not only reflect the contribution of each dimension in identification, but also more accurately reflect the abnormality of the overall behavior.

[0009] Preferably, the preprocessing includes: removing blurred behavioral images, denoising the remaining behavioral images using Gaussian filtering, and enhancing the contrast of the images using histogram equalization.

[0010] Preferably, the amount of information satisfies the relationship:

[0011] , Representation Dimension The amount of information, Representation Dimension The characteristic value of The number of Representation Dimension The characteristic value of The probability of Represents a logarithmic function.

[0012] Preferably, the importance satisfies the relationship:

[0013] , Representation Dimension Abnormal behavior The importance of Indicates abnormal behavior Case Dimension The characteristic value of The probability of Represents a logarithmic function.

[0014] It is possible to identify which features play a key role in the occurrence of abnormal behavior by quantifying uncertainty, thereby improving the accuracy and efficiency of anomaly detection.

[0015] Preferably, the importance further includes: normalizing the variance of all eigenvalues ​​of the same dimension, and using the result as the importance of the dimension to the abnormal behavior.

[0016] Dimensions with smaller variance usually correspond to features with more concentrated and stable information, while dimensions with larger variance indicate that the distribution of the eigenvalues ​​of this dimension is more dispersed and may contribute less to the identification of abnormal behavior.

[0017] Preferably, obtaining the dimension filtering result includes: for any label, sorting all dimensions according to importance, filtering the dimensions according to a preset information threshold to obtain preliminary filtering results; traversing to obtain the preliminary filtering results of each label, and taking the intersection of all preliminary filtering results as the dimension filtering result.

[0018] Not only does it consider the personalized feature requirements of each label, but it also selects dimensions that are highly important under multiple labels through intersection. This method combines the commonalities and differences between labels to ensure that the selected dimensions can effectively characterize the behavior of each label and avoid the selection of too many irrelevant dimensions due to the bias of a single label.

[0019] Preferably, obtaining the dimension screening result also includes: accumulating all importances of the same dimension to obtain a cumulative value, traversing to obtain the cumulative value of each dimension, sorting the dimensions according to the cumulative value, and filtering the dimensions according to a preset information threshold to obtain the dimension screening result.

[0020] The process of accumulating importance reflects the stability and importance of the dimension in multiple different tasks. Selecting dimensions with higher cumulative values ​​can ensure that the selected features have strong discernment and information content in a variety of situations. In this way, the final selected dimensions can more comprehensively represent the key structure of the data and reduce redundant and noisy features.

[0021] In a second aspect, the present invention discloses an artificial intelligence-based surveillance image anomaly recognition system, comprising: a processor; and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the system executes the above-mentioned artificial intelligence-based surveillance image anomaly recognition method.

[0022] Beneficial effects of the present invention:

[0023] The present invention can efficiently and accurately identify abnormal behaviors in surveillance images by combining the extraction and analysis of multi-dimensional behavioral features. First, by preprocessing the behavioral image, blur and noise are removed, and the image contrast is enhanced, thereby improving the accuracy of subsequent behavioral feature extraction. Then, the multi-dimensional behavioral features are obtained through the key point detection model, and the relationship between different key points is calculated using the Euclidean distance, providing a more accurate basis for the judgment of abnormal behavior. Furthermore, the method performs dimension screening by calculating the amount of information of each dimension and the importance of abnormal behavior, ensuring that the features that have the greatest impact on abnormal behavior can be given priority. On this basis, multiple behavior recognition models are used to independently evaluate each screened dimension, and by weighted accumulation of the abnormal probabilities output by each model, a more stable and reliable final abnormal probability is obtained, thereby improving the accuracy and robustness of abnormal behavior recognition. The present invention can not only adapt to diverse behavioral patterns, but also automatically optimize the selection of feature dimensions and reduce the impact of redundant features through the information amount and importance evaluation mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0025] Figure 1 It is a flow chart of a monitoring image anomaly recognition method based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] 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 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 those skilled in the art without creative work are within the scope of protection of the present invention.

[0027] It should be understood that when the terms "first", "second", etc. are used in the claims, descriptions, and drawings of the present invention, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the description and claims of the present invention indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their collections.

[0028] The present invention provides a monitoring image abnormality recognition method based on artificial intelligence. Figure 1 As shown, the monitoring image anomaly recognition method based on artificial intelligence includes steps S1 to S4, which are described in detail below.

[0029] S1, obtaining the preprocessed behavior image according to the preset sampling interval, and obtaining the label of the behavior image.

[0030] In one embodiment, the video data is cut into static images frame by frame. Exemplarily, a behavior image is obtained every 10 frames. Blurred behavior images are removed, Gaussian filtering is used to denoise the remaining behavior images, and histogram equalization is used to enhance the contrast of the image.

[0031] A person skilled in the art labels each clear behavior image to obtain a label of the behavior image, wherein the label includes no abnormality and the abnormality type when an abnormality exists, and the abnormality type includes but is not limited to: falling, bumping, fighting, etc.

[0032] S2, input the behavior image into the key point detection model to obtain several key points, calculate the Euclidean distance between any two key points, use all Euclidean distances as multi-dimensional behavior features, and calculate the information amount of each dimension and the importance of each dimension to abnormal behavior based on the behavior features.

[0033] In one embodiment, the behavior image obtained in step S1 is input into a key point detection model to obtain key points in the behavior image. Exemplarily, the key point detection model is openpose, Hourglass Network, PoseNet, etc. The key point detection model is a well-known technology for those skilled in the art and will not be described in detail here.

[0034] It should be noted that the current abnormal behavior recognition method mainly judges whether a person has abnormal behavior by the distance between the key points of the human body. This method can accurately identify the abnormal behavior of a single person, but when the abnormal behavior involves the interaction between multiple people, the recognition effect of the traditional method is often poor. The reason is that many complex abnormal behaviors are completed through the coordinated actions of two or more individuals, such as physical collisions, pushing, disputes in groups, etc. These behaviors cannot be judged only by the key point information of a single individual. Therefore, when identifying abnormal behavior of people, we cannot only rely on the distance and posture characteristics between the key points of a single person, but also introduce the relative position, motion trajectory and dynamic relationship of interactive behavior between the key points of different people. By analyzing the spatial relationship and action coordination between multiple people, abnormal behavior in group interaction can be judged more accurately, thereby improving the overall recognition effect.

[0035] Calculate the Euclidean distance between any two key points, use all Euclidean distances as multi-dimensional behavior features, and calculate the amount of information in each dimension and the importance of each dimension to abnormal behavior based on the behavior features.

[0036] The amount of information satisfies the relationship:

[0037] , Representation Dimension The amount of information, Representation Dimension The characteristic value of The number of Representation Dimension The characteristic value of The probability of Represents a logarithmic function.

[0038] The relationship between the amount of information quantifies the distribution of each dimension in a multidimensional data system from the perspective of information theory, and can effectively measure the contribution of each feature to the overall amount of information.

[0039] Importance satisfies the relationship:

[0040] , Representation Dimension Abnormal behavior The importance of Indicates abnormal behavior Case Dimension The characteristic value of The probability of Represents a logarithmic function.

[0041] It represents the opposite number of information entropy. The larger the information entropy, the more unstable it is, and the less important this dimension is to a certain abnormal behavior.

[0042] The importance relationship reflects the opposite number of the information entropy of the probability distribution of the dimension when abnormal behavior occurs. The larger the information entropy, the more uncertain or dispersed the distribution of the eigenvalues ​​of the dimension, which means that the dimension contributes less to abnormal behavior. On the contrary, a lower entropy value indicates that the eigenvalue distribution of the dimension is more concentrated and has a stronger ability to predict abnormal behavior. Therefore, the application of the exponential function amplifies this difference, so that the contribution of those feature dimensions with low entropy and more stable is strengthened. This calculation method can identify which features play a key role in the occurrence of abnormal behavior by quantifying uncertainty, thereby improving the accuracy and efficiency of anomaly detection.

[0043] In one embodiment, the importance further includes: normalizing the inverse of the variance of all eigenvalues ​​of the same dimension, and using the result as the importance of the dimension to the abnormal behavior.

[0044] It should be noted that variance essentially reflects the discreteness or volatility of the eigenvalue. When the eigenvalue of a dimension changes greatly, it means that the dimension provides richer information and can reflect more potential changes. Therefore, dimensions with smaller variance usually correspond to features with more concentrated and stable information, while dimensions with larger variance indicate that the eigenvalue distribution of this dimension is more dispersed and may contribute less to the identification of abnormal behavior. By normalizing the inverse of the variance, feature dimensions with smaller variance and more stable changes can be given higher weights, thereby improving the sensitivity and discrimination of these dimensions in abnormal behavior detection.

[0045] S3, sort the dimensions according to their importance, and filter the dimensions according to the amount of information to obtain dimension filtering results.

[0046] In one embodiment, for any label, all dimensions are sorted according to importance, and the dimensions are filtered according to a preset information threshold to obtain preliminary filtering results; the preliminary filtering results of each label are traversed to obtain, and the intersection of all preliminary filtering results is used as the dimension filtering result.

[0047] First, we filter the dimensions of each label independently. The selected dimensions are based on their performance under each label and the information threshold setting to ensure that the important dimensions under each label are retained. Then, we traverse the filter results of all labels and obtain the intersection of the dimensions of all labels. This not only takes into account the personalized requirements of each label for features, but also filters out the dimensions that are highly important under multiple labels through the intersection.

[0048] This method combines the commonalities and differences between labels to ensure that the selected dimensions can effectively characterize the behavior of each label and avoid the selection of too many irrelevant dimensions due to the bias of a single label.

[0049] In one embodiment, obtaining the dimension filtering result also includes: accumulating all importances of the same dimension to obtain a cumulative value, traversing to obtain the cumulative value of each dimension, sorting the dimensions according to the cumulative value, and filtering the dimensions according to a preset information threshold to obtain the dimension filtering result.

[0050] Integrating the importance of dimensions under different labels into a global indicator reflects the comprehensive contribution of each dimension in multi-task or multi-label scenarios. The process of accumulating importance reflects the stability and importance of the dimension in multiple different tasks. Selecting dimensions with higher cumulative values ​​can ensure that the selected features have strong recognition and information content in various situations. In this way, the final selected dimensions can more comprehensively represent the key structure of the data and reduce redundant and noisy features.

[0051] S4, respectively establish a behavior recognition model for each dimension in the dimension screening result, input the feature value sequence corresponding to the dimension into the behavior recognition model, output the probability of abnormal behavior, accumulate the probabilities of the same abnormal behavior output by different behavior recognition models to obtain the final abnormal probability, and complete the abnormal recognition.

[0052] In one embodiment, a behavior recognition model is established for each dimension in the dimension screening result, wherein one dimension corresponds to one behavior recognition model. Exemplarily, the behavior recognition model includes but is not limited to SVM support vector machine, BP network, etc.

[0053] The feature value sequence corresponding to the dimension is input into the behavior recognition model, and the probability of abnormal behavior is output. The probabilities of the same abnormal behavior output by different behavior recognition models are accumulated to obtain the final abnormal probability, which satisfies the relationship:

[0054] , Indicates abnormal behavior The final abnormal probability, Indicates the dimension in the dimension filtering result Abnormal behavior The importance of Indicates the dimension in the dimension filtering result The corresponding behavior recognition model outputs abnormal behavior The probability of Indicates the total number of dimensions in the dimension filtering results.

[0055] The final abnormal probabilities of different abnormal behaviors are compared, and the abnormal behavior corresponding to the maximum value of the final abnormal probability is taken as the abnormal behavior in the monitoring image to complete the abnormality identification.

[0056] An embodiment of the present invention also discloses an artificial intelligence-based surveillance image anomaly recognition system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, an artificial intelligence-based surveillance image anomaly recognition method according to the present invention is implemented.

[0057] The above system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, and their configuration and functions are known in the art, so they will not be described in detail here.

[0058] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module or both. Any such computer storage medium may be part of a device or accessible or connectable to a device.

[0059] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

[0060] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A monitoring image anomaly recognition method based on artificial intelligence, characterized in that: include: Acquire the preprocessed behavior image according to a preset sampling interval, and obtain the label of the behavior image; Input the behavior image into the key point detection model to obtain several key points, calculate the Euclidean distance between any two key points, use all Euclidean distances as multi-dimensional behavior features, and calculate the information content of each dimension and the importance of each dimension to abnormal behavior based on the behavior features; Sort the dimensions according to their importance, and filter the dimensions according to the amount of information to obtain dimension filtering results; A behavior recognition model is established for each dimension in the dimension screening result, where one dimension corresponds to one behavior recognition model. The feature value sequence corresponding to the dimension is input into the behavior recognition model, and the probability of abnormal behavior is output. The probabilities of the same abnormal behavior output by different behavior recognition models are accumulated to obtain the final abnormal probability, thus completing abnormal recognition. Among them, the final abnormal probability satisfies the relationship: , Indicates abnormal behavior The final abnormal probability, Indicates the dimension in the dimension filtering result Abnormal behavior The importance of Indicates the dimension in the dimension filtering result The corresponding behavior recognition model outputs abnormal behavior The probability of Indicates the total number of dimensions in the dimension filtering results.

2. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The pre-processing comprises: The blurred behavioral images were removed, the remaining behavioral images were denoised using Gaussian filtering, and the image contrast was enhanced using histogram equalization.

3. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The amount of information satisfies the relationship: , Representation Dimension The amount of information, Representation Dimension The characteristic value of The number of Representation Dimension The characteristic value of The probability of Represents a logarithmic function.

4. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The importance satisfies the relationship: , Representation Dimension Abnormal behavior The importance of Indicates abnormal behavior Case Dimension The characteristic value of The probability of Represents a logarithmic function.

5. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The importance also includes: The inverse of the variance of all eigenvalues ​​of the same dimension is normalized and the result is used as the importance of the dimension to the abnormal behavior.

6. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The obtaining of dimension screening results includes: For any label, sort all dimensions according to importance, and filter the dimensions according to the preset information threshold to obtain preliminary screening results; Traverse to obtain the preliminary screening results of each label, and use the intersection of all preliminary screening results as the dimension screening result.

7. The method for identifying abnormalities in surveillance images based on artificial intelligence according to claim 1, characterized in that: The obtaining of dimension screening results also includes: All importances of the same dimension are accumulated to obtain a cumulative value, the cumulative value of each dimension is traversed to obtain the cumulative value, the dimensions are sorted according to the cumulative value, and the dimensions are filtered according to the preset information threshold to obtain the dimension filtering result.

8. The monitoring image anomaly recognition system based on artificial intelligence is characterized by: include: Processor; and A memory storing computer instructions, wherein when the computer instructions are executed by a processor, the system executes the monitoring image anomaly recognition method based on artificial intelligence according to any one of claims 1 to 7.

Images