Video anomaly recognition processing method and system based on artificial intelligence

Abstract

The invention relates to the technical field of video anomaly recognition, in particular to a video anomaly recognition processing method and system based on artificial intelligence, and the method comprises the steps: obtaining a monitoring video in a target region, and compressing the monitoring video to obtain a compressed monitoring video; carrying out key frame extraction on the compressed monitoring video to obtain a key frame sequence corresponding to the compressed monitoring video, and carrying out target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; and positioning an abnormal monitoring video clip according to the abnormal key frame sequence and the compressed monitoring video. By compressing the monitoring video, the storage and transmission burden of the video data can be greatly reduced, the compressed video not only saves the storage space, but also accelerates the speed of the subsequent processing step, and is particularly important in the real-time or large-scale processing scene of the monitoring video.

Description

Technical Field

[0001] The present invention relates to the technical field of video anomaly recognition, and in particular to a video anomaly recognition processing method and system based on artificial intelligence. Background Art

[0002] Traditional methods usually rely on frame-by-frame processing of the entire video frame without selective analysis of key frames or key areas in the video. This approach results in a very large amount of calculation, especially when processing long videos or high-resolution videos, the computing resource consumption is extremely serious; and traditional methods often rely on rule-based detection algorithms or simple statistical methods, which may not be able to handle very complex scenes and are prone to more false positives or omissions. For example, ordinary motion detection algorithms may not be able to distinguish between normal background motion and abnormal events, resulting in a large number of false positives; and traditional methods often process spatiotemporal information in a relatively simple manner and cannot simultaneously consider the association between time and space dimensions. For example, traditional motion detection usually relies on simple optical flow methods to analyze the motion of objects, but does not have the ability to understand the pattern and change of motion from a timing perspective like modern deep learning methods; and traditional methods of anomaly detection often rely on fixed background modeling methods, such as background subtraction or difference map methods, which perform poorly in environments with complex or dynamically changing backgrounds. For example, in a multi-object background, background subtraction methods may misjudge the motion of normal objects as abnormal. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a video anomaly recognition and processing method and system based on artificial intelligence.

[0004] The technical solution adopted to solve the above technical problems is: a video anomaly recognition and processing method based on artificial intelligence, including: Acquire a surveillance video in a target area, and compress the surveillance video to obtain a compressed surveillance video; Extracting key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and performing target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; Locating the abnormal monitoring video segment according to the abnormal key frame sequence and the compressed monitoring video; Performing appearance extraction on the abnormal monitoring video clip according to a pre-trained appearance extraction model to obtain spatial appearance features corresponding to the abnormal monitoring video clip, and performing spatiotemporal extraction on the spatial appearance features according to a spatiotemporal attention module to obtain spatiotemporal features corresponding to the abnormal monitoring video clip; Performing optical flow field extraction on the abnormal monitoring video clip according to the TV-L1 algorithm to obtain the optical flow field corresponding to the abnormal monitoring video clip, and performing motion extraction on the optical flow field according to a pre-trained motion extraction model to obtain the motion features of the abnormal monitoring video clip; The spatiotemporal features and the motion features are fused according to a gated fusion mechanism to obtain fused features, and anomaly detection is performed on the fused features according to a preset normal pattern memory library to obtain an anomaly score of the abnormal monitoring video clip.

[0005] Preferably, compressing the surveillance video to obtain a compressed surveillance video includes: Dividing the surveillance video into a plurality of GOPs, wherein the first frame in the GOP is a key frame and the remaining frames in the GOP are non-key frames; Reconstructing the non-key frame a second time according to the non-key frame's own measurement value a priori to obtain the residual of the non-key frame; Denoising the residual of the non-key frame to obtain output features; The reference frame features and non-key frame features are downsampled step by step to obtain a multi-scale feature space. The alignment features of the previous level are fused according to the temporal spatial inner product attention. The non-key frame features are adaptively refined using the dynamic domain prior knowledge to obtain non-key frame refined features. The non-key frame refinement features are reconstructed to obtain compressed non-key frames, and the compressed non-key frames are combined with key frames to obtain compressed surveillance videos.

[0006] Preferably, the expression of the reconstruction process is as follows: ;in, represents the secondary reconstruction of non-keyframes, represents a reconstruction of a non-keyframe, represents the non-keyframe measurement matrix, represents the reference frame measurement matrix, represents the reference frame measurement, represents non-keyframe measurements, Indicated in Extract static domain residuals guided by context information; The expression of the output feature is as follows: ;in, Indicates The reconstructed features extracted are represents the denoising network, and represents the measurement operator, and represents the transpose of the measurement operator, represents feature transformation, represents the inverse feature transform, and It is obtained by repeatedly expanding the key frame and non-key frame measurements. Indicates Extracted in Extracting static domain residuals guided by context information, Indicates Extract the denoising function implemented by the convolutional network again.

[0007] Preferably, the expression of the non-keyframe refinement feature is as follows: ;in, represents the non-keyframe refinement feature, It is composed of two residual blocks to remove the noise introduced by the process. represents non-keyframe features, It means that it consists of 3 convolutional layers with ReLU. represents the upsampling result of the aligned features, and It means that it is realized by 1×1 convolution. Represents the Hadamard product operation.

[0008] Preferably, the image recognition model adopts a convolutional neural network, the appearance extraction model adopts 3D-ResNet, and the motion extraction model adopts a convolutional neural network.

[0009] Preferably, the expression of the spatiotemporal attention module is as follows: ;in, represents the spatiotemporal attention weight, represents the activation function, , , represents the query, key, and value matrix, represents the key vector dimension, used to scale the dot product, Represents the appearance characteristics of the space.

[0010] Preferably, the expression of the fusion feature is as follows: ;in, represents the fusion feature, represents the Sigmoid function, , represents the learnable parameters, Indicates motion characteristics.

[0011] Preferably, the calculation formula of the abnormal score of the abnormal monitoring video clip is as follows: ;in, represents the anomaly score of the abnormal surveillance video clip, Indicates the total number of video frames in the abnormal monitoring video clip. Indicates the normal mode memory bank, Represents the similarity smoothing factor.

[0012] The technical solution adopted to solve the above technical problems is: a video anomaly recognition and processing system based on artificial intelligence, which is applicable to the video anomaly recognition and processing method based on artificial intelligence, including: A video compression unit, the video compression unit is used to obtain a surveillance video in a target area, and compress the surveillance video to obtain a compressed surveillance video; An image recognition unit, the image recognition unit is used to extract key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and perform target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; A video positioning unit, the video positioning unit is used to locate the abnormal monitoring video segment according to the abnormal key frame sequence and the compressed monitoring video; A spatiotemporal extraction unit, the spatiotemporal extraction unit is used to perform appearance extraction on the abnormal monitoring video clip according to a pre-trained appearance extraction model to obtain a spatial appearance feature corresponding to the abnormal monitoring video clip, and perform spatiotemporal extraction on the spatial appearance feature according to a spatiotemporal attention module to obtain a spatiotemporal feature corresponding to the abnormal monitoring video clip; A motion extraction unit, the motion extraction unit is used to perform optical flow field extraction on the abnormal monitoring video segment according to the TV-L1 algorithm to obtain the optical flow field corresponding to the abnormal monitoring video segment, and perform motion extraction on the optical flow field according to a pre-trained motion extraction model to obtain motion features of the abnormal monitoring video segment; An anomaly scoring unit is used to perform feature fusion on the spatiotemporal features and the motion features according to a gated fusion mechanism to obtain a fusion feature, and perform anomaly detection on the fusion feature according to a preset normal pattern memory library to obtain an anomaly score for the abnormal monitoring video clip.

[0013] The beneficial effects of the present invention are as follows: (1) By compressing surveillance videos, the present invention can greatly reduce the storage and transmission burden of video data. The compressed video not only saves storage space, but also speeds up the subsequent processing steps, which is especially important in real-time or large-scale processing scenarios of surveillance videos. In addition, key frame extraction can effectively reduce the computational complexity of video analysis, because key frames usually contain the main information of the video. By extracting key frames and performing target detection, the method can identify abnormal scenes in the shortest time without processing the entire video stream, thereby reducing the waste of computing resources. (2) By using an appearance extraction model, the present invention can extract key spatial features from abnormal surveillance video clips. These features help the system identify and understand the spatial structure of abnormal events. Combined with the spatiotemporal attention module, it can not only analyze spatial information, but also enhance the model's sensitivity to temporal information. By using the TV-L1 algorithm to extract the optical flow field, the motion pattern of the object in the video can be accurately captured, and the motion features can be further obtained through the motion extraction model, so that the motion changes in abnormal events can be identified, whether it is the sudden movement of the object, the change in speed or other dynamic manifestations of abnormal behavior; (3) The present invention integrates the spatiotemporal features and motion features to fully utilize the complementarity of the two and improve the detection accuracy of the model. The spatiotemporal features mainly describe the changes in the scene, while the motion features emphasize the dynamic behavior of the object. By integrating these two features, abnormal events can be identified more accurately. In addition, the abnormal detection uses the preset normal mode memory library to compare the fused features, which can realize intelligent judgment based on historical data, quickly and accurately locate abnormal behavior, and improve the reliability of abnormal detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic flow chart of the steps of an overall method in an embodiment of the present invention; Figure 2 A schematic diagram of the system architecture of an overall system in an embodiment of the present invention.

[0015] Figure numerals: 1. Video compression unit; 2. Image recognition unit; 3. Video positioning unit; 4. Spatiotemporal extraction unit; 5. Motion extraction unit; 6. Abnormality scoring unit. DETAILED DESCRIPTION

[0016] Embodiment 1, as Figure 1 As shown, the video anomaly recognition and processing method based on artificial intelligence proposed by the present invention includes: S1. Obtaining a surveillance video in a target area, and compressing the surveillance video to obtain a compressed surveillance video; S2, extracting key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and performing target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; S3, locating abnormal monitoring video segments according to abnormal key frame sequences and compressed monitoring videos; S4, performing appearance extraction on the abnormal monitoring video clip according to the pre-trained appearance extraction model to obtain the spatial appearance features corresponding to the abnormal monitoring video clip, and performing spatiotemporal extraction on the spatial appearance features according to the spatiotemporal attention module to obtain the spatiotemporal features corresponding to the abnormal monitoring video clip; S5. Perform optical flow field extraction on the abnormal monitoring video clip according to the TV-L1 algorithm to obtain the optical flow field corresponding to the abnormal monitoring video clip, and perform motion extraction on the optical flow field according to the pre-trained motion extraction model to obtain the motion characteristics of the abnormal monitoring video clip; S6. The spatiotemporal features and the motion features are fused according to the gated fusion mechanism to obtain fused features, and anomaly detection is performed on the fused features according to a preset normal pattern memory library to obtain an anomaly score of the abnormal monitoring video clip.

[0017] In the present invention, surveillance video compression is to compress the original video, usually to reduce the cost of storage and processing. In video compression, redundant data will be removed to make the video file smaller, and the compressed surveillance video retains enough information for subsequent analysis; key frames refer to key frames in the video, which are frames representing changes in video content. These frames can usually capture important information or the beginning and end of events in the video. By selecting key frames, the amount of computation required to process the video can be reduced and the main time information can be retained; key frame extraction is to extract the most representative frames from the video, usually through image content change detection, clustering methods or content changes based on video streams; target detection refers to the process of identifying and locating specific targets from images or video frames. For example, people, vehicles or other objects of interest in video frames can be detected. In video anomaly detection, target detection is usually used to identify unusual events or objects in a scene; appearance features refer to visual features extracted from images that describe target objects, such as color, Texture, shape, edge, etc. Appearance extraction is a key step in describing the surface of an object, which is used to identify and distinguish different objects or scenes; appearance extraction model refers to a deep learning or traditional machine learning model used to extract appearance features from images, usually based on convolutional neural networks (CNN) for training; spatiotemporal attention refers to the fact that in the process of video or continuous frame processing, the spatiotemporal attention mechanism focuses on certain information in time and space in the video. When extracting spatiotemporal features, the model must consider not only the spatial features of the image (such as objects in the image), but also the temporal changes (i.e., motion changes between frames). The spatiotemporal attention module helps to extract important spatiotemporal information by adaptively weighting temporal and spatial features; spatiotemporal features are features that contain both spatial information (static features of the image, such as the appearance of an object) and temporal information (changes or motions of objects in different time frames). In abnormal video detection, spatiotemporal features can help distinguish normal from abnormal behavior; the TV-L1 algorithm is an optimization algorithm for optical flow estimation. Optical flow refers to the movement of pixels between consecutive video frames. The TV-L1 algorithm combines total variation (TV) regularization and L1 norm optimization to better handle noise and boundary problems in motion estimation. In video anomaly detection, the TV-L1 algorithm is used to extract optical flow fields in video clips to analyze the motion patterns of the targets. The normal mode memory is a database that stores normal behavior patterns. When detecting anomalies, the system compares the currently extracted features with the normal mode. If the difference from the normal mode is large, it is judged as an anomaly.

[0018] Embodiment 2, the video anomaly recognition and processing method based on artificial intelligence proposed by the present invention, compared with embodiment 1, this embodiment also includes: compressing the monitoring video to obtain a compressed monitoring video, including: A1. Divide the surveillance video into multiple GOPs, wherein the first frame in the GOP is a key frame and the remaining frames in the GOP are non-key frames; A2, reconstructing the non-key frame based on the prior measurement value of the non-key frame itself to obtain the residual of the non-key frame; A3, denoising the residual of non-key frames to obtain output features; A4. Downsample the reference frame features and non-key frame features step by step to obtain a multi-scale feature space, fuse the alignment features of the previous level according to the temporal spatial inner product attention, and adaptively refine the non-key frame features using the dynamic domain prior knowledge to obtain non-key frame refined features; A5. Reconstruct the refined features of the non-key frames to obtain compressed non-key frames, and combine the compressed non-key frames with the key frames to obtain compressed surveillance videos.

[0019] In this embodiment, the term GOP in video coding refers to a group of image frames, which is usually composed of a key frame (I frame) and several non-key frames (P frame and B frame). The first frame in the GOP is the key frame, and the other frames are encoded by the difference of the previous frame, which are usually called non-key frames; secondary reconstruction refers to the process of reconstructing non-key frames through certain prior information, which is usually to restore the details of non-key frames and reduce the loss caused by compression; dynamic domain refers to the dynamic information area that changes over time in video processing. In time series data, the change of objects or scenes occurs over time, and the dynamic domain emphasizes the dynamic change characteristics at different moments; prior knowledge refers to knowledge based on experience or existing knowledge, which helps the model make better predictions during training and reasoning. In video analysis, dynamic domain prior knowledge may include information such as the movement pattern and behavior pattern of the object, which helps to analyze and understand the video more accurately; feature reconstruction refers to the process of restoring the extracted feature information to the original image or video. In this process, the model attempts to restore the visual effect of the original frame through features. The reconstruction process is usually used for video compression and denoising tasks to minimize the information lost during the compression process.

[0020] In an optional embodiment, the expression of the reconstruction process is as follows: ;in, represents the secondary reconstruction of non-keyframes, represents a reconstruction of a non-keyframe, represents the non-keyframe measurement matrix, represents the reference frame measurement matrix, represents the reference frame measurement, represents non-keyframe measurements, Indicated in Extract static domain residuals guided by context information; The expression of the output feature is as follows: ;in, Indicates The reconstructed features extracted are represents the denoising network, and represents the measurement operator, and represents the transpose of the measurement operator, represents feature transformation, represents the inverse feature transform, and It is obtained by repeatedly expanding the key frame and non-key frame measurements. Indicates Extracted in Extracting static domain residuals guided by context information, Indicates Extract the denoising function implemented by the convolutional network again.

[0021] It should be noted that primary reconstruction usually refers to the preliminary process of recovering non-key frames from compressed data (such as video streams or frame encoding). Unlike secondary reconstruction, primary reconstruction is often based on the original information stored during encoding, with the purpose of recovering a reasonable image or video frame; the non-key frame measurement matrix is ​​a matrix that maps the features of non-key frames to a low-dimensional space. This matrix is ​​used to compress or characterize non-key frames. In the process of compressing video, the measurement matrix is ​​used to extract or measure certain features or information in the frame; the reference frame measurement matrix is ​​similar to the non-key frame measurement matrix, but is used for reference frames. Reference frames usually refer to frames that have been completely reconstructed in video encoding, or frames used to predict non-key frames. The reference frame measurement matrix is ​​used to extract The features in these frames are compared and aligned with non-key frames; the reference frame measurements are values ​​extracted through the measurement matrix of the reference frame, which represent the compressed data or feature information of the reference frame, and are very important in video encoding and reconstruction, and are used to extract information from the reference frame to help restore non-key frames; contextual information refers to assisting the processing process through prior knowledge or background information of the current video frame; the measurement operator is a linear transformation used to map the information in an image or video to a new space or dimension. During the compression process, the measurement operator compresses the data into a low-dimensional representation. During the recovery process, the measurement operator can be used to restore the original information; feature transformation refers to a transformation of the input features (such as the features of a video frame) to adapt them to specific processing or analysis requirements.

[0022] In an optional embodiment, the expression of the non-keyframe refinement feature is as follows: ;in, represents the non-keyframe refinement feature, It is composed of two residual blocks to remove the noise introduced by the process. represents non-keyframe features, It means that it consists of 3 convolutional layers with ReLU. represents the upsampling result of the aligned features, and It means that it is realized by 1×1 convolution. Represents the Hadamard product operation.

[0023] It should be noted that upsampling is an operation used to expand a low-resolution feature map to a higher resolution. Common upsampling methods include deconvolution or interpolation methods. Upsampling operations are often used in image generation or feature recovery processes to restore low-resolution features to a higher resolution state.

[0024] In an optional embodiment, the image recognition model adopts a convolutional neural network, the appearance extraction model adopts 3D-ResNet, and the motion extraction model adopts a convolutional neural network.

[0025] It should be noted that 3D-ResNet is an extension of ResNet (residual network) designed for video and 3D data. Traditional ResNet is used for image recognition tasks. By introducing residual blocks (skip connections), the gradient vanishing problem in deep networks is solved, making the network deeper and easier to train. The motion extraction model refers to a model that extracts motion information from videos or continuous image sequences. The goal of motion extraction is to analyze and capture the motion trajectory or changes of objects in images or videos.

[0026] In an optional embodiment, the expression of the spatiotemporal attention module is as follows: ;in, represents the spatiotemporal attention weight, represents the activation function, , , represents the query, key, and value matrix, represents the key vector dimension, used to scale the dot product, Represents the appearance characteristics of the space.

[0027] In an optional embodiment, the expression of the fusion feature is as follows: ;in, represents the fusion feature, represents the Sigmoid function, , represents the learnable parameters, Indicates motion characteristics.

[0028] In an optional embodiment, the calculation formula of the abnormal score of the abnormal monitoring video clip is as follows: ;in, represents the anomaly score of the abnormal surveillance video clip, Indicates the total number of video frames in the abnormal monitoring video clip. Indicates the normal mode memory bank, Represents the similarity smoothing factor.

[0029] Embodiment three, as Figure 2 As shown, the video anomaly recognition and processing system based on artificial intelligence proposed by the present invention is applicable to the video anomaly recognition and processing method based on artificial intelligence, and includes: The video compression unit 1 is used to obtain the surveillance video in the target area and compress the surveillance video to obtain a compressed surveillance video; Image recognition unit 2, image recognition unit 2 is used to extract key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and perform target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; The video positioning unit 3 is used to locate the abnormal monitoring video segment according to the abnormal key frame sequence and the compressed monitoring video; The spatiotemporal extraction unit 4 is used to perform appearance extraction on the abnormal monitoring video clip according to the pre-trained appearance extraction model to obtain the spatial appearance features corresponding to the abnormal monitoring video clip, and perform spatiotemporal extraction on the spatial appearance features according to the spatiotemporal attention module to obtain the spatiotemporal features corresponding to the abnormal monitoring video clip; The motion extraction unit 5 is used to perform optical flow field extraction on the abnormal monitoring video clip according to the TV-L1 algorithm to obtain the optical flow field corresponding to the abnormal monitoring video clip, and perform motion extraction on the optical flow field according to the pre-trained motion extraction model to obtain the motion characteristics of the abnormal monitoring video clip; The anomaly scoring unit 6 is used to fuse the spatiotemporal features and the motion features according to the gated fusion mechanism to obtain a fused feature, and perform anomaly detection on the fused feature according to a preset normal pattern memory library to obtain an anomaly score of the abnormal monitoring video clip.

[0030] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A video anomaly recognition and processing method based on artificial intelligence, characterized in that: include: Acquire a surveillance video in a target area, and compress the surveillance video to obtain a compressed surveillance video; Extracting key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and performing target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; Locating the abnormal monitoring video segment according to the abnormal key frame sequence and the compressed monitoring video; Performing appearance extraction on the abnormal monitoring video clip according to the pre-trained appearance extraction model to obtain the spatial appearance features corresponding to the abnormal monitoring video clip, and performing spatiotemporal extraction on the spatial appearance features according to the spatiotemporal attention module to obtain the spatiotemporal features corresponding to the abnormal monitoring video clip; Performing optical flow field extraction on the abnormal monitoring video clip according to the TV-L1 algorithm to obtain the optical flow field corresponding to the abnormal monitoring video clip, and performing motion extraction on the optical flow field according to a pre-trained motion extraction model to obtain the motion features of the abnormal monitoring video clip; The spatiotemporal features and the motion features are fused according to a gated fusion mechanism to obtain fused features, and anomaly detection is performed on the fused features according to a preset normal pattern memory library to obtain an anomaly score of the abnormal monitoring video clip.

2. The method for video anomaly recognition and processing based on artificial intelligence according to claim 1 is characterized in that: Compressing the surveillance video to obtain a compressed surveillance video includes: dividing the surveillance video into a plurality of GOPs, wherein the first frame in the GOP is a key frame, and the remaining frames in the GOP are non-key frames; Reconstructing the non-key frame a second time according to the non-key frame's own measurement value a priori to obtain the residual of the non-key frame; Denoising the residual of the non-key frame to obtain output features; The reference frame features and non-key frame features are downsampled step by step to obtain a multi-scale feature space. The alignment features of the previous level are fused according to the temporal spatial inner product attention. The non-key frame features are adaptively refined using the dynamic domain prior knowledge to obtain non-key frame refined features. The non-key frame refinement features are reconstructed to obtain compressed non-key frames, and the compressed non-key frames are combined with key frames to obtain compressed surveillance videos.

3. The method for video anomaly recognition and processing based on artificial intelligence according to claim 2 is characterized in that: The expression of the reconstruction process is as follows: ; in, represents the secondary reconstruction of non-keyframes, represents a reconstruction of a non-keyframe, represents the non-keyframe measurement matrix, represents the reference frame measurement matrix, represents the reference frame measurement, represents non-keyframe measurements, Indicated in The static domain residual is extracted under the guidance of context information; the expression of the output feature is as follows: ;in, Indicates The reconstructed features extracted are represents the denoising network, and represents the measurement operator, and represents the transpose of the measurement operator, represents feature transformation, represents the inverse feature transform, and It is obtained by repeatedly expanding the key frame and non-key frame measurements. Indicates Extracted in Extracting static domain residuals guided by context information, Indicates Extract the denoising function implemented by the convolutional network.

4. The method for video anomaly recognition and processing based on artificial intelligence according to claim 3 is characterized in that: The expression of the non-keyframe refinement feature is as follows: ;in, represents the non-keyframe refinement feature, It is composed of two residual blocks to remove the noise introduced by the process. represents non-keyframe features, It means that it consists of 3 convolutional layers with ReLU. represents the upsampling result of the aligned features, and It means that it is implemented by 1×1 convolution. Represents the Hadamard product operation.

5. The method for video anomaly recognition and processing based on artificial intelligence according to claim 1 is characterized in that: The image recognition model adopts a convolutional neural network, the appearance extraction model adopts 3D-ResNet, and the motion extraction model adopts a convolutional neural network.

6. The method for video anomaly recognition and processing based on artificial intelligence according to claim 1 is characterized in that: The expression of the spatiotemporal attention module is as follows: ;in, represents the spatiotemporal attention weight, represents the activation function, , , represents the query, key, and value matrix, represents the key vector dimension, used to scale the dot product, Represents the appearance characteristics of the space.

7. The method for video anomaly recognition and processing based on artificial intelligence according to claim 1 is characterized in that: The expression of the fusion feature is as follows: ;in, represents the fusion feature, represents the Sigmoid function, , represents the learnable parameters, Indicates motion characteristics.

8. The method for video anomaly recognition and processing based on artificial intelligence according to claim 1 is characterized in that: The calculation formula of the abnormal score of the abnormal monitoring video clip is as follows: ;in, represents the anomaly score of the abnormal surveillance video clip, Indicates the total number of video frames in the abnormal monitoring video clip. Indicates the normal mode memory bank, Represents the similarity smoothing factor.

9. An artificial intelligence-based video anomaly recognition and processing system, which is applicable to the artificial intelligence-based video anomaly recognition and processing method according to any one of claims 1 to 8, characterized in that: include: A video compression unit (1), the video compression unit (1) is used to obtain a surveillance video in a target area, and compress the surveillance video to obtain a compressed surveillance video; An image recognition unit (2), the image recognition unit (2) being used to extract key frames from the compressed surveillance video to obtain a key frame sequence corresponding to the compressed surveillance video, and to perform target detection on each key frame in the key frame sequence according to a pre-trained image recognition model to obtain an abnormal key frame sequence; A video positioning unit (3), the video positioning unit (3) being used to locate the abnormal monitoring video segment according to the abnormal key frame sequence and the compressed monitoring video; A spatiotemporal extraction unit (4), the spatiotemporal extraction unit (4) being used to perform appearance extraction on the abnormal monitoring video segment according to a pre-trained appearance extraction model to obtain a spatial appearance feature corresponding to the abnormal monitoring video segment, and to perform spatiotemporal extraction on the spatial appearance feature according to a spatiotemporal attention module to obtain a spatiotemporal feature corresponding to the abnormal monitoring video segment; A motion extraction unit (5), the motion extraction unit (5) being used to perform optical flow field extraction on the abnormal monitoring video segment according to the TV-L1 algorithm to obtain an optical flow field corresponding to the abnormal monitoring video segment, and to perform motion extraction on the optical flow field according to a pre-trained motion extraction model to obtain motion features of the abnormal monitoring video segment; An anomaly scoring unit (6) is used to perform feature fusion on the spatiotemporal features and the motion features according to a gated fusion mechanism to obtain a fusion feature, and perform anomaly detection on the fusion feature according to a preset normal pattern memory library to obtain an anomaly score for the abnormal monitoring video clip.

Images