Video anomaly detection method and system, medium and equipment

Abstract

The invention relates to the technical field of computer vision, and discloses a video anomaly detection method and system, a medium and equipment, and the method comprises the steps: obtaining a continuous video frame sequence; based on the continuous video frame sequence, detecting an abnormal event through a video anomaly detection model; wherein the video anomaly detection model executes a time shift operation between two continuous feature maps, and replaces a part of channels of each feature map with the same channel part from the previous feature map to extract time features; and after the time shift operation is executed, the spatial features are accumulated through a recursive aggregation method so as to capture multi-scale context information. The sensitivity to time information and space information is improved, and the video anomaly detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a video anomaly detection method, system, medium and device. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Video anomaly detection has great application potential in social security, medical monitoring, disaster warning and other fields.

[0004] However, the current video anomaly detection methods have the following problems: insensitivity to temporal and spatial information, relatively single information scale and low capture rate of local features during feature extraction, resulting in low accuracy of video anomaly detection. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a video anomaly detection method, system, medium and device, which combines time shift with dynamic large kernel convolution to improve the sensitivity to time information and spatial information. Time shift is used to change the distribution of input features in different time periods, thereby improving the sensitivity to time information; dynamic large kernel convolution operation is introduced to accumulate spatial features through recursive aggregation method, effectively capture multi-scale context information, enhance the model's ability to understand local information, improve the capture rate of local features, and achieve a larger receptive field.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a video anomaly detection method, comprising:

[0008] Get a sequence of continuous video frames;

[0009] Based on the continuous video frame sequence, abnormal events are detected through the video anomaly detection model;

[0010] Among them, the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; and after performing the time shift operation, the spatial features are accumulated by recursive aggregation to capture multi-scale contextual information.

[0011] Furthermore, the video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating a continuous video frame sequence into a tensor, linear embedding is performed, and feature extraction is performed using a sliding window to obtain a plurality of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, time shift operations, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks and a layer of dynamic large kernel convolution operations to obtain the encoder output.

[0012] Furthermore, a first skip connection is established in the second layer of dynamic large kernel convolution operations in the two layers of dynamic large kernel convolution operations.

[0013] Furthermore, a second skip connection is established between the six-layer ST-3D block and a dynamic large-kernel convolution layer between the two-layer ST-3D block.

[0014] Furthermore, the video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, sequentially undergoing upsampling operation, six layers of ST-3D blocks, one layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks to obtain the decoder output.

[0015] A second aspect of the present invention provides a video anomaly detection system, comprising:

[0016] A data acquisition module is configured to: acquire a sequence of continuous video frames;

[0017] An anomaly detection module is configured to: detect abnormal events based on a continuous video frame sequence through a video anomaly detection model;

[0018] Among them, the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; and after performing the time shift operation, the spatial features are accumulated by recursive aggregation to capture multi-scale contextual information.

[0019] Furthermore, the video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating a continuous video frame sequence into a tensor, linear embedding is performed, and feature extraction is performed using a sliding window to obtain a plurality of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, time shift operations, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks and a layer of dynamic large kernel convolution operations to obtain the encoder output.

[0020] Furthermore, the video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, sequentially undergoing upsampling operation, six layers of ST-3D blocks, one layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks to obtain the decoder output.

[0021] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor, and the program implementing the steps in a video anomaly detection method as described above when executed by the processor.

[0022] A fourth aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, the steps in a video anomaly detection method as described above are implemented.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention combines the time shift technology with dynamic large kernel convolution, improves the sensitivity to time information and spatial information, and improves the accuracy of video anomaly detection.

[0025] The present invention uses time shifting to change the distribution of input features in different time periods, thereby improving the sensitivity to time information.

[0026] The present invention introduces dynamic large kernel convolution and accumulates spatial features through a recursive aggregation method, effectively capturing multi-scale contextual information, enhancing the model's ability to understand local information, improving the capture rate of local features, and achieving a larger receptive field. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings, which constitute a part of the specification of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute limitations of the present invention.

[0028] Figure 1 This is a flow chart of a video anomaly detection method according to the first embodiment of the present invention;

[0029] Figure 2 It is a schematic diagram of a feature map undergoing time shifting and dynamic large kernel convolution processing before the second downsampling in the encoder of the first embodiment of the present invention;

[0030] Figure 3 This is a diagram of experimental results of Example 1 of the present invention;

[0031] Figure 4 This is a test result diagram of Example 1 of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0034] In the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention is further described below with reference to the accompanying drawings and embodiments.

[0035] Embodiment 1

[0036] The purpose of this embodiment is to provide a video anomaly detection method.

[0037] In view of the insensitivity of previous video anomaly detection methods to temporal and spatial information, as well as the problem of extracting local features during feature extraction, this embodiment provides a video anomaly detection method that combines time shift with DLK (dynamic large kernel convolutional layer) to improve the sensitivity to temporal and spatial information, while improving the capture rate of local features and achieving a larger receptive field.

[0038] The present embodiment provides a method for detecting video anomalies, comprising the following steps:

[0039] Step 1: Select the data set and evaluation indicators.

[0040] Three datasets were used: Avenue (a video dataset shot on the campus avenue of the Chinese University of Hong Kong), Ped2 (a sidewalk video dataset shot by a fixed camera), and Shanghai Tech (a dataset on the campus of ShanghaiTech University).

[0041] Avenue (Video dataset shot on the campus avenue of the Chinese University of Hong Kong): 37 videos were released by students of the Chinese University of Hong Kong in 2013. These videos are further classified into 16 normal videos for training purposes and 21 abnormal videos for testing purposes. The database contains RGB (red, green and blue) single scene data, covering a total of 47 abnormal events. The abnormal events can be divided into three categories: abnormal behavior, wrong direction and abnormal objects.

[0042] Ped2 (Sidewalk Video Dataset Taken by Fixed Cameras): The UCSD Anomaly Detection Dataset collected by the University of California, San Diego (UCSD) contains surveillance video clips taken by fixed cameras located at high vantage points overlooking sidewalks. The density of pedestrians on these sidewalks varies, ranging from sparse to dense. In standard scenarios, the video sequences feature only pedestrians. The Peds2 subset includes 16 training video samples and 12 test video samples.

[0043] ShanghaiTech University Dataset: Released in 2016, it is divided into two parts: Part A (Part_A_final) and Part B (Part_B_final). Part A includes 300 training images and 182 test images, while Part B includes 400 training images and 316 test images.

[0044] In this embodiment, the performance evaluation index of the model adopts PSNR (Peak Signal-to-Noise Ratio). PSNR is calculated by comparing the difference between the original image and the reconstructed image. The higher the PSNR value, the smaller the difference between the reconstructed image and the original image, which means the higher the reconstruction quality.

[0045] Video anomaly detection technology based on deep learning uses various neural networks to identify abnormal behavior patterns or object movements in videos. With the success of the Transformer architecture in natural language processing, the computer vision field also applied the Transformer (VIT) to the image classification task of 16×16 images for the first time, demonstrating strong performance. Subsequently, various networks based on the Transformer architecture emerged. In order to capture more complex details, the Transformer in Transformer (a structure in which a layer of Transformer is nested within the Transformer) model further subdivides each image block (visual sentence) into smaller blocks, enabling the model to grasp local details and global context at the same time. Similarly, the Swin Transformer (a Transformer model based on a sliding window) based on the Transformer framework adopts a hierarchical structure to divide the image into smaller blocks, enabling it to capture local feature information while significantly reducing computational complexity. In addition, a sliding window mechanism has been introduced to improve the receptive field of the model.

[0046] Step 2: Data preprocessing.

[0047] For the three datasets, all images are uniformly cropped to the format of 224 × 224 × 3 as input. Then, the images are normalized by scaling the pixel values ​​from [0, 255] to [0, 1], and the dimensions are rearranged from [height, width, number of channels] to [number of channels, height, width] for subsequent processing.

[0048] Step 3: Hyperparameter setting.

[0049] The input video frame is divided into blocks of size (4, 4, 4) and fed into 3D Swin Transformer (ST-3D) blocks based on the three-dimensional scale containing temporal information, where the local attention window size is (8, 7, 7). The number of layers in the 3D Swin Transformer is [3, 6]. The ratio of the hidden dimension size of the MLP (Multi-layer Perceptron) layer to the input dimension size is 4. The number of attention heads in each 3D Swin Transformer block is [6, 12]. During training, the optimizer used is the AdamX optimization algorithm, and the initial learning rate for the three datasets is set to 0.001.

[0050] Step 4: Model improvement.

[0051] In this embodiment, improvements are made in terms of temporal feature extraction and multi-scale context, and ablation experiments are performed to evaluate its performance in experimental results.

[0052] In this embodiment, if Figure 1 As shown in the figure, the backbone network of the video anomaly detection model is based on the 3D-SwinTransformer model, which completes the reconstruction task of the input image through two downsampling and two upsampling operations. The feature extractor in Swin Transformer is combined with the 3D encoder to learn normal behavior patterns from non-abnormal video frames, and the 3D decoder is used to predict future frames.

[0053] The video anomaly detection model in this embodiment adopts an unsupervised network structure decoder based on Swin Transformer (sliding window-based Transformer model) and 3D (three-dimensional), and uses jump connections to connect feature maps from the encoder and decoder. The network's ability to capture multi-scale contextual information is also enhanced by using large and deep convolution kernels of different sizes and expansion rates to process features; using time shift to change the distribution of input features over different time periods, thereby increasing sensitivity to time information.

[0054] A video anomaly detection method provided in this embodiment integrates time shift processing to process channel information across continuous time dimensions, replaces part of the continuous feature map channels, and uses convolution sizes of different scales to adaptively merge multi-scale local feature maps.

[0055] The shifted window masking of the Swin Transformer is tested in the ST-3D model, and experiments are performed to remove the masking in an attempt to retain more feature information.

[0056] A video anomaly detection method provided in this embodiment mainly involves a 3D encoder and decoder based on Swin Transformer, along channel shift and dynamic large kernel convolution (DLK), such as Figure 1 shown.

[0057] (1) Encoder

[0058] (101) First downsampling.

[0059] The encoder inputs a continuous video image frame sequence T = [t 1 ,t 2 ...t n ] where t is the time-ordered image frame, and the size of each video frame is [224, 224, 3]. These consecutive time frames are integrated into a tensor of shape [T, 224, 224, 3]. After linear embedding, the high-dimensional data is mapped to a low-dimensional space to obtain T / 2 feature maps. The periodic 3D sliding window method is used for feature extraction, and the input is converted from [T, 224, 224, 3] to [T / 2, 96, 56, 56].

[0060] (102) After passing through two ST-3D blocks, channel shift (i.e., time domain translation technology) is applied between two consecutive feature maps, and the first eighth of the channel of each feature map is partially replaced with the same channel from the previous feature map. Among them, the time domain translation technology has a wide range of applications in video analysis and understanding. In order to better capture the time dimension of video data, this embodiment transfers part of the channels in the time dimension to the next frame to capture dynamic changes over time, which helps to improve feature learning in time series data. For each frame, this embodiment extracts feature maps from each residual block using the first 1 / 8 channel features of the video frame and caches them in the memory. For the next frame, this embodiment replaces the first 1 / 8 of the current feature map with the cached feature map. This embodiment uses a combination of 7 / 8 of the current feature map and 1 / 8 of the old feature map to generate a feature map for the next layer of input, and repeats this process.

[0061] (103) Two layers of convolution DLK operations are performed on the feature map. The first layer uses a 5×5×5 kernel; while the second layer uses a 7×7×7 kernel with a dilation rate, increasing the receptive field through spatial dilation and group convolution to enhance the model attention mechanism and establish the first skip connection.

[0062] (104) After the feature map undergoes two layers of dynamic large kernel convolution operations, it is downsampled for the second time through Patch Merging, and the input is transformed from [T / 2, 96, 56, 56] to [T / 2, 192, 28, 28].

[0063] (105) After six layers of ST-3D blocks and another dynamic large kernel convolution layer with an established second skip connection, the encoder operation is completed after two layers of ST-3D blocks and convolution using dynamic large kernels.

[0064] This part mainly involves the 3D encoder and the decoder based on Swin Transformer, along-channel shift and dynamic large kernel convolution modules.

[0065] (2)Decoder.

[0066] Similar to the encoder, the decoder also uses the ST-3D module to process 3D image patches. However, it differs in that it gradually upsamples the feature maps and processes them through convolution and self-attention, reduces the number of feature map channels to obtain higher resolution feature maps, and performs two upsampling operations to halve the dimension.

[0067] For the encoder’s data processing, it starts with a cascade using the second skip connection, followed by an upsampling operation using a convolution kernel of size (1, 2, 2) and a stride that doubles the length and width of the feature map, converting it from [T / 2, 192, 28, 28] to [T / 2, 96, 56, 56]; then, it passes through six layers of ST-3D blocks and DLK to enhance the understanding of the spatiotemporal dynamics in the feature map; then, the output of the dynamic large kernel convolution is cascaded with the first skip connection from the encoder output of the same dimension; followed by upsampling and two layers of ST blocks, changing the tensor dimension from [T / 2, 96, 56, 56] to [3, 224, 224] to generate the final image output.

[0068] To better handle contextual information, this embodiment enhances the existing Swin-Transformer module with a dynamic large kernel convolution (DLK) module, which utilizes multiple large depth convolution kernels to extract multi-scale features. Applying DLK to the feature map after performing the time shift operation allows capturing more refined information and achieving a larger receptive field, which helps understand video behavior. Figure 2As shown, this embodiment uses two convolution kernels of sizes 5×5×5 and 7×7×7, respectively, to enable the features to interact between different spatial descriptors. Before the second downsampling in the encoder, the feature map undergoes time shifting and DLK processing. First, the feature map is moved along the channel dimension to extract temporal features, which is effective for analyzing behavioral patterns over long time spans. Then, entering the DLK model, the spatial features are accumulated through a recursive aggregation method, effectively capturing multi-scale contextual information and enhancing the model's ability to understand local information.

[0069] Step 5: Use the training set to train the improved video anomaly detection model in step 4.

[0070] Step 6: The trained video anomaly detection model was verified on three public video datasets: Shanghai Science and Technology, Avenue, and Ped2, and achieved high accuracy.

[0071] Step 7: Perform video anomaly detection on the trained video anomaly detection model. This includes: obtaining a continuous video frame sequence; based on the continuous video frame sequence, detecting abnormal events through the video anomaly detection model; wherein the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; and after performing the time shift operation, the spatial features are accumulated by a recursive aggregation method to capture multi-scale contextual information.

[0072] The video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating a continuous video frame sequence into a tensor, mapping multiple continuous video frames to an embedding dimension through sliding window processing to obtain a number of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, time shift operations, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks and a layer of dynamic large kernel convolution operations to obtain the encoder output. The second layer of convolution operations in the two layers of dynamic large kernel convolution operations establishes a first jump connection. The layer of dynamic large kernel convolution between the six layers of ST-3D blocks and the two layers of ST-3D blocks establishes a second jump connection.

[0073] Among them, the video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, it sequentially undergoes upsampling operation, six layers of ST-3D blocks, a layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks to obtain the decoder output.

[0074] Table 1. Ablation experiment results

[0075] Time displacement Dynamic large kernel convolution Avenue Dataset Ped2 Dataset ShanghaiTech Dataset × × 83.1％ 95.3％ 71.6％ × √ 85.5％ 96.1％ 72.7％ √ × 84.6％ 96.8％ 72.3％ √ √ 86.5％ 97.1％ 73.7％

[0076] As shown in the second row of Table 1, if time offset is not used to extract time dimension information from the video, and the dynamic large kernel convolution module is not used to understand the contextual information, the performance of the three data sets is the worst. In the third and fourth rows, the dynamic large kernel convolution and time offset are combined separately, thereby improving the model performance. In addition, the role of the mask in the Swin Transformer when using a sliding window is tested, with the aim of removing the mask to retain more image information. However, the experimental results show that the mask did not achieve the expected effect. The anomaly detection accuracy on the three data sets of Avenue, Ped2 and Shanghai Tech is 86.5%, 97.7% and 73.7% respectively. It is proved that the model of this embodiment can improve the network reconstruction accuracy and improve the model performance.

[0077] Since the ShanghaiTech dataset contains more video frame data and various complex scene information, it is more challenging for video anomaly detection.

[0078] like Figure 3 As shown in the figure, the reconstruction error heatmaps of Avenue dataset, ped2 dataset and Shanghai Science and Technology dataset are displayed. The method of this embodiment is used to reconstruct abnormal frames and display the re-addition error heatmap. The first row consists of the original image, the second row contains the reconstructed image, and the third row shows the re-addition error heatmap between the two. Among them, running, driving a motorcycle and riding a bicycle are all considered abnormal behavior events.

[0079] like Figure 4 As shown in Figure 1, anomaly detection is performed on the first video dataset of Shanghai Science and Technology using a pre-trained model. The first test video has a total of 265 frames. In normal video frames, the road is relatively empty and there are few pedestrians. When pedestrians appear, the peak signal-to-noise ratio value begins to decrease. Figure 4 In the above figure, the peak signal-to-noise ratio has been normalized. In fact, during the test, a higher peak signal-to-noise ratio value indicates normality, while a decrease indicates abnormality. Around frame 100, pedestrians begin to appear and the value increases. From frame 150, a pedestrian on a bicycle appears, and around frame 230, the cyclist disappears and the value returns to normal.

[0080] like Figure 4 As shown, after testing using the network of this embodiment, for a continuous video sequence with anomalies, the anomaly score will increase when an abnormal frame appears.

[0081] Embodiment 2

[0082] The purpose of this second embodiment is to provide a video anomaly detection system.

[0083] A data acquisition module is configured to: acquire a sequence of continuous video frames;

[0084] An anomaly detection module is configured to: detect abnormal events based on a continuous video frame sequence through a video anomaly detection model;

[0085] Among them, the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; and after performing the time shift operation, the spatial features are accumulated by recursive aggregation to capture multi-scale contextual information.

[0086] Among them, the video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating the continuous video frame sequence into a tensor, linear embedding is performed, and feature extraction is performed using a sliding window to obtain a number of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, time shift operations, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks and a layer of dynamic large kernel convolution operations to obtain the encoder output.

[0087] Among them, the video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, sequentially undergoing upsampling operation, six layers of ST-3D blocks, one layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks to obtain the decoder output.

[0088] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0089] Embodiment 3

[0090] This embodiment provides a computer-readable storage medium on which a computer program is stored. The program is executed by a processor. When the program is executed by the processor, the steps in the video anomaly detection method described in the first embodiment are implemented.

[0091] Embodiment 4

[0092] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the steps in the video anomaly detection method described in the first embodiment are implemented.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0094] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A video anomaly detection method, characterized in that: include: Get a sequence of continuous video frames; Based on the continuous video frame sequence, abnormal events are detected through the video anomaly detection model; Among them, the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; After performing the time shift operation, the spatial features are accumulated through recursive aggregation to capture multi-scale contextual information.

2. A video anomaly detection method as claimed in claim 1, characterized in that: The video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating a continuous video frame sequence into a tensor, linear embedding is performed, and feature extraction is performed using a sliding window to obtain a plurality of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, a time shift operation, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks, and a layer of dynamic large kernel convolution operations to obtain the encoder output.

3. A video anomaly detection method as claimed in claim 1, characterized in that: The second layer of dynamic large kernel convolution operation in the two-layer convolution DLK operation establishes a first skip connection.

4. A video anomaly detection method as claimed in claim 1, characterized in that: A second skip connection is established between the six-layer ST-3D block and the one-layer dynamic large kernel convolution between the two-layer ST-3D block.

5. A video anomaly detection method as claimed in claim 1, characterized in that: The video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, sequentially undergoing upsampling operation, six layers of ST-3D blocks, one layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks, to obtain the decoder output.

6. A video anomaly detection system, characterized in that: include: A data acquisition module is configured to: acquire a sequence of continuous video frames; An anomaly detection module is configured to: detect abnormal events based on a continuous video frame sequence through a video anomaly detection model; Among them, the video anomaly detection model performs a time shift operation between two consecutive feature maps, replacing part of the channels of each feature map with the same channel part from the previous feature map to extract temporal features; After performing the time shift operation, the spatial features are accumulated through recursive aggregation to capture multi-scale contextual information.

7. A video anomaly detection system as claimed in claim 6, characterized in that: The video anomaly detection model includes an encoder, and the operation of the encoder includes: after integrating a continuous video frame sequence into a tensor, linear embedding is performed, and feature extraction is performed using a sliding window to obtain a plurality of feature maps, and the feature maps are sequentially passed through two layers of ST-3D blocks, a time shift operation, two layers of dynamic large kernel convolution operations, downsampling, six layers of ST-3D blocks, a layer of dynamic large kernel convolution layer, two layers of ST-3D blocks, and a layer of dynamic large kernel convolution operations to obtain the encoder output.

8. A video anomaly detection system as claimed in claim 6, characterized in that: The video anomaly detection model includes a decoder, and the operation of the decoder includes: after cascading the encoder output with the second jump connection, sequentially undergoing upsampling operation, six layers of ST-3D blocks, one layer of dynamic large kernel convolution operation, cascading with the first jump connection, upsampling and two layers of ST-3D blocks, to obtain the decoder output.

9. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor, characterized in that: When the program is executed by a processor, the steps in a video anomaly detection method as described in any one of claims 1 to 5 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the video anomaly detection method as described in any one of claims 1-5 are implemented.

Images