Video anomaly detection method using progressive Gaussian noise negative learning

Abstract

The invention discloses a video anomaly detection method using progressive Gaussian noise negative learning, which belongs to the field of image data processing, and comprises the following steps: constructing a data set D; constructing an unsupervised collaborative learning network which comprises a feature extractor, a generator, a discriminator, a pseudo tag generator and a noise scheduler; regarding the pseudo reconstruction target as the final output of a generator, and training the generator and a discriminator in the unsupervised collaborative learning network by using a negative learning method to obtain an anomaly detection model; a to-be-recognized fragment is obtained and sent to the anomaly detection model, a corresponding feature vector is generated through the feature extractor, and a corresponding anomaly probability is output through the discriminator. According to the method, the diffusion probability model is introduced into the collaborative learning method, the cross supervision between the generator and the discriminator is established by using the low occurrence frequency of the anomaly, so that the learning ability of potential data distribution is improved, and on the basis of introducing the diffusion probability model, a negative learning method is designed to improve the sensitivity of anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, and in particular to a video anomaly detection method using progressive Gaussian noise negative learning. BACKGROUND

[0002] With the exponential growth of monitoring video data and the wide demand for video anomaly detection in practical applications, it is almost impossible to label all types of anomalies. In most cases, we cannot clearly define the normal behavior of human activities in life. For example, it is normal to ride a bike on the road, but it is prohibited to ride a bike on the sidewalk. In addition, whether it is semi-supervised video anomaly detection or weakly supervised video anomaly detection, obtaining coarse-grained labels at the video level still requires checking the entire video to determine whether the video is normal, so the labeling cost is still high. Therefore, unsupervised video anomaly detection has important research value.

[0003] The goal of unsupervised video anomaly detection is to directly discover anomalies from completely unlabeled videos in an unsupervised manner, without normal and abnormal videos to construct a training set, and without the need to collect samples, avoiding the heavy labeling work. However, the lack of supervision signals makes unsupervised video anomaly detection more challenging than semi-supervised and weakly supervised video anomaly detection. In addition, this assumption also expands the application field of video anomaly detection, meaning that the detection system can continuously train itself without human intervention.

[0004] Although unsupervised video anomaly detection has made significant progress in recent years, it still faces some challenges. A large number of unsupervised video anomaly detection methods are based on reconstruction, which judges abnormal events by the size of reconstruction error, especially the introduction of diffusion probability model, which significantly improves the performance of unsupervised video anomaly detection. However, the unsupervised video anomaly detection based on reconstruction method often only trains the generation model in the training process, and usually only uses mean square error (MSE) loss to guide the generation model, which makes the generation model still have some deficiencies in distinguishing positive and negative samples.

[0005] Glossary: GCL method (Generative Cooperative Learning for Unsupervised Video Anomaly Detection) is a generative cooperative learning method for unsupervised video anomaly detection, which is used for video anomaly detection. This method does not require labeled training data, and through the cooperative learning of the generator and the discriminator, it effectively identifies abnormal behavior in videos.

[0006] ResNeXt3D (ResNeXt 3D) is a neural network that extends 2D convolution to 3D (spatiotemporal) convolution based on ResNeXt, designed for video understanding, action recognition, etc. It is one of the mainstream backbone networks for current video understanding, action recognition and various 3D vision tasks. Its input is a batch of video clips, and after feature extraction of each video clip, the corresponding feature vector is output.

[0007] Denoising Diffusion Probabilistic Models (DDPM) is a class of generative models, including a forward process and a reverse process. In the forward process, Gaussian noise is gradually added to the original data until it becomes almost pure noise. In the reverse process, a neural network is trained to gradually "denoise" from the noise and ultimately restore clear data (such as images). SUMMARY

[0008] The purpose of the present application is to provide a video anomaly detection method using progressive Gaussian noise negative learning, which can use the low occurrence frequency of anomalies to establish cross-supervision between the generator and the discriminator, and use gradually increasing Gaussian noise to label anomaly pseudo-labels, further improving the convergence and recognition accuracy of the algorithm.

[0009] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a video anomaly detection method using progressive Gaussian noise negative learning, comprising the following steps: S1, constructing a data set D; Obtain I videos for video anomaly detection, divide each video into J segments, and each segment contains p frames of pictures, wherein the jth segment of the ith video is C i,j , 1≤i≤I, 1≤j≤J; S2, constructing an unsupervised co-learning network, including a feature extractor, a generator, a discriminator, a pseudo-label generator and a noise scheduler; The feature extractor is used to randomly input a batch of segments to generate a feature vector for each segment. If the qth segment in the batch is C i,j , then the feature vector is marked as , 1≤q≤b, b is the batch size; The generator is a diffusion probability model, which is used to input , generate corresponding reconstruction vector ; The discriminator is used to input , output the anomaly probability of ; ​The pseudo-tag generator is used to generate The first pseudo tag Second pseudo-label If the reconstruction loss is based on the generator, then... Abnormal, then =1, otherwise 0, if the judgment is based on the binary cross-entropy loss of the discriminator. Abnormal, then =1, otherwise 0; The noise scheduler is used to generate according to the following formula. False reconstruction goals ; , , In the formula, t is the number of iterations, and x t Let be the noisy data from the t-th iteration, and є be standard Gaussian noise. Let be the cumulative attenuation coefficient at step t; S3, treating the pseudo-reconstructed target as the final output of the generator, uses the negative learning method to train the generator and discriminator in the unsupervised collaborative learning network until the iteration ends, thus obtaining the anomaly detection model; S4: The segment to be identified is obtained and fed into the anomaly detection model. The feature extractor generates the corresponding feature vector, and the discriminator outputs the corresponding anomaly probability.

[0010] Preferably, the feature extractor is a pre-trained R3D network.

[0011] Preferably, the pseudo-tag generator generates the tag according to the following formula. The first pseudo tag : , , In the formula, In the generator and Reconstruction loss, The first threshold, , Each batch The mean and standard deviation are given by k, where k is the first constant.

[0012] Preferably, the pseudo-tag generator generates the tag according to the following formula. The second pseudo-tag : , , In the formula, for binary cross-entropy loss of the discriminator, is a second threshold value, , are respectively the mean and standard deviation of a batch of , is a second constant.

[0013] As preferred, in the noise scheduler, is obtained according to the following formula: , , where a s is the attenuation coefficient of the s-th step, 1≤s≤t, a t is the attenuation coefficient of the t-th step.

[0014] As preferred, the negative learning method is; When training the discriminator, the loss function L D of the discriminator is calculated according to the following formula: D and the network parameters of the discriminator are adjusted to minimize L G ; , When training the generator, the loss function L G of the generator is calculated according to the following formula: G and the network parameters of the generator are adjusted to minimize L G ; .

[0015] In the present application, the generator and the discriminator are trained in an adversarial manner.

[0016] Regarding the generator: unlike the GCL method which uses an autoencoder (AE) as the generator, the present application adopts a diffusion probability model, which is stronger in capturing the overall distribution of the training data, and in the present application, the reconstructed features output by the generator are not directly used for adversarial training, but are further added with noise by the noise scheduler to generate pseudo-reconstruction targets, and the pseudo-reconstruction targets are used as the final output of the generator for adversarial training. The pseudo-reconstruction targets are generated based on the noise scheduler.

[0017] Regarding the pseudo-label generator: the purpose is to generate a feature vector a first pseudo-label and a second pseudo-label , the first pseudo-label judges whether it is abnormal based on the reconstruction loss of the generator, the second pseudo-label judges whether it is abnormal based on the binary cross-entropy loss of the discriminator.

[0018] Regarding the noise scheduler: it adds noise to the reconstructed features output by the generator according toSecond pseudo label of Specific value of Generate pseudo reconstruction target When =1, add gradually increasing Gaussian noise to the reconstruction vector x to form x t As a pseudo reconstruction target When =0, it means that the discriminator considers to be normal, and the reconstruction vector of the generator is directly used as the pseudo reconstruction target As a pseudo reconstruction target When =1, it means that the discriminator considers to be abnormal, and Gaussian noise needs to be added to to expand the distance between the positive and negative samples after reconstruction. In addition, the added Gaussian noise is not constant, but gradually increases with the increase of the number of iterations.

[0019] Regarding negative learning: in the existing GCL method, the generator and the discriminator jointly learn, and when training the generator, the idea of negative learning is used. Negative learning uses the prediction score of the discriminator to generate a pseudo label, and uses the pseudo label to optimize the generator through a negative learning method to increase the distinction between normal input and abnormal input reconstruction, and encourages the generator to perform poor reconstruction on samples with abnormal pseudo labels, and to perform regular reconstruction with minimal error on samples with normal pseudo labels. Since the diffusion probability model can better perceive the distribution of the training data, after introducing k-diffusion into the collaborative learning method, the negative learning in the original collaborative learning method is no longer applicable to the new model. This is because, if a full 1 vector or Gaussian noise is used to mark the abnormal pseudo label generated by the discriminator network D, it may lead to the cumulative propagation of false pseudo labels, which will introduce a lot of bias and make the algorithm performance decline. Therefore, in order to enable the present application to still perform negative learning, the present application designs a noise scheduler to generate a pseudo reconstruction target based on and gradually increasing Gaussian noise, instead of the original output of the generator.

[0020] The reason for proposing this method is that video anomalies are usually not "all or nothing" mutations, but gradually develop from slight deviations to significant anomalies (for example, the moving speed of an object gradually increases, and the occlusion gradually increases). By adding gradually increasing Gaussian noise to the anomaly vector, this gradualness can be simulated, so that the generator gradually learns different intensity anomaly patterns during the training process, rather than forcing the reconstruction of a fixed pattern (all 1 vector) at once. This is more in line with the evolution process of anomalies in real scenarios. Moreover, anomaly events are usually located in the low probability region of the normal data distribution. Gradually increasing Gaussian noise will gradually push the anomaly samples away from the normal data manifold, forcing the generator to more sensitively capture this deviation during the denoising process, improving the sensitivity of anomaly detection.

[0021] Compared with the prior art, the present application has the advantages that: The present application introduces a diffusion probability model into the co-learning method, uses the low occurrence frequency of anomalies to establish cross-supervision between the generator and the discriminator, thereby improving the learning ability of the data latent distribution. Compared with the simple fully connected network in the original co-learning method, the diffusion probability model can better capture the latent features of the data.

[0022] After introducing the diffusion probability model, retaining the original negative learning method will introduce a large amount of bias, which will reduce the performance of the algorithm and make the entire model difficult to converge. The present application designs a new negative learning method based on and gradually increasing Gaussian noise to generate a pseudo-reconstruction target , instead of the original output of the generator , so as to simulate the gradualness of anomalies such as "gradual development from slight deviation to significant anomaly" in video anomalies, so that the generator gradually learns different intensity anomaly patterns during the training process, rather than forcing the reconstruction of a fixed pattern at once. This is more in line with the evolution process of anomalies in real scenarios. Moreover, because anomaly events are usually located in the low probability region of the normal data distribution, this gradually increasing Gaussian noise will gradually push the anomaly samples away from the normal data manifold, forcing the generator to more sensitively capture this deviation during the denoising process, improving the sensitivity of anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is an unsupervised co-learning network structure diagram; Figure 2 is a schematic diagram of a pseudo-reconstruction target; Figure 3 is an AUC comparison diagram of different kinds of anomaly events on the Stech dataset; Figure 4 is a ROC curve comparison diagram of four strategies; Figure 5 is a visual detection result diagram of a video using the method of the present application; Figure 6 This is a visualization of the detection results of video 2 using the method of this invention; Figure 7 This is a visualization of the detection results of the video using the method of this invention. Figure 8 This is a visualization of the detection results of the video using the method of this invention; Figure 9 This is a visualization of the detection results of the method of the present invention, based on the video. Figure 10 This is a visualization of the detection results of the method of the present invention, which is based on the Six Classics in video. Figure 11 This is a visualization of the detection results of the method of the present invention, based on video seven. Figure 12 This is a visualization of the detection results of the method of this invention in video. Detailed Implementation

[0024] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0025] Example 1: See Figures 1-2 A video anomaly detection method using progressive Gaussian noise negative learning includes the following steps; S1, construct dataset D; Obtain I videos for anomaly detection, and divide each video into J segments, each segment containing p frames, where the j-th segment of the i-th video is C. i,j , 1≤i≤I, 1≤j≤J; S2, construct an unsupervised collaborative learning network, including a feature extractor, generator, discriminator, pseudo-label generator and noise scheduler; The feature extractor is used to randomly input a batch of segments and generate a feature vector for each segment. If the q-th segment in the batch is C... i,j Then its feature vector is labeled as 1≤q≤b, where b is the batch size; The generator is a diffusion probability model, used as input. ,generate corresponding reconstructed vector ; The discriminator is used for input. Output abnormal probability ; The pseudo-tag generator is used to generate The first pseudo tag Second pseudo-label If the reconstruction loss is based on the generator, then... Abnormal, then =1, otherwise 0, if the judgment is based on the binary cross-entropy loss of the discriminator. Abnormal, then =1, otherwise 0; The noise scheduler is used to generate according to the following formula. False reconstruction goals ; , , In the formula, t is the number of iterations, and x t Let be the noisy data from the t-th iteration, and є be standard Gaussian noise. Let be the cumulative attenuation coefficient at step t; S3, treating the pseudo-reconstructed target as the final output of the generator, uses the negative learning method to train the generator and discriminator in the unsupervised collaborative learning network until the iteration ends, thus obtaining the anomaly detection model; S4: The segment to be identified is obtained and fed into the anomaly detection model. The feature extractor generates the corresponding feature vector, and the discriminator outputs the corresponding anomaly probability.

[0026] In this embodiment, the feature extractor is a pre-trained R3D network.

[0027] The pseudo-tag generator generates according to the following formula. The first pseudo tag : , , In the formula, In the generator and Reconstruction loss, The first threshold, , Each batch The mean and standard deviation are given by k, where k is the first constant.

[0028] The pseudo-tag generator generates according to the following formula. The second pseudo-tag : , ; In the formula, for The binary cross-entropy loss of the discriminator, The second threshold, , Each batch The mean and standard deviation, It is the second constant.

[0029] In the noise scheduler, According to the following formula: , In the formula, a s is the attenuation coefficient of the s-th step, 1≤s≤t, a t is the attenuation coefficient of the t-th step, a t =1-β t .

[0030] The negative learning method is: The loss function L D of the discriminator and the loss function L G of the generator are constructed; , , When training the discriminator, L D is calculated, and the network parameters of the discriminator are adjusted to minimize L D ; When training the generator, L G is calculated, and the network parameters of the generator are adjusted to minimize L G .

[0031] Embodiment 2: Referring to Figures 1-2 , this embodiment selects several semi-supervised, weakly supervised and unsupervised video anomaly detection methods in the prior art and compares them with the method of the present application. The AUC values of different methods on the STech dataset are counted to generate Table 1. The several existing methods in Table 1 are: MemAE (2019): Gong D, Liu L, Le V, et al. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2019: 1705-1714. FFPVAD (2018): Liu W, Luo W, Lian D, et al. Future frame prediction for anomaly detection - a new baseline [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 6536-6545. Chen et al. (2022): CHEN C, HU Y. Video Anomaly Detection Combining Memory-Augmented [J]. Journal of Computer Engineering & Applications, 2022, 58(15). HF2-VAD (2021): Liu Z, Nie Y, Long C, et al. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 13588-13597. Georgescu et al. (2021): Georgescu M I, Ionescu R T, Khan F S, et al. A background-agnostic framework with adversarial training for abnormal event detection in video [J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(9): 4505-4523. STG-NF (2023): Hirschorn O, Avidan S. Normalizing flows for human pose anomaly detection [C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision. 2023: 13545-13554. Noise Cleaner (2022): Li N, Zhong J X, Shu X, et al. Weakly-supervised anomaly detection in video surveillance via graph convolutional label noise cleaning [J]. Neurocomputing, 2022, 481: 154-167. AR-Net (2020): Wan B, Fang Y, Xia X, et al. Weakly supervised video anomaly detection via center-guided discriminative learning [C] / / 2020 IEEE international conference on multimedia and expo (ICME). IEEE, 2020: 1-6. Li et al. (2023): Li Wenzhong, Wu Kewei, Sun Yongxuan, et al. Weakly supervised video anomaly detection based on contrastive memory network [J]. Application Research of Computers / Jisuanji Yingyong Yanjiu, 2023, 40(10). RTFM (2021): Tian Y, Pang G, Chen Y, et al. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 4975-4986. Tur et al. (2023): Tur A O, Dall’Asen N, Beyan C, et al. Exploring diffusion models for unsupervised video anomaly detection [C] / / 2023 IEEE international conference on image processing (ICIP). IEEE, 2023: 2540-2544. Pan et al. (2024): Pan Z, Li Z, Xue C, et al. Unsupervised video anomaly detection based on time-appearance energy diffusion Transformer [J]. Microelectronics & Computer, 2025, 42(02): 68-76. DOI:10.19304 / J.ISSN1000-7180.2024.0005. GCL (2022): Generative Cooperative Learning, generative cooperative learning; STG-NFAllData (2023): Hirschorn O, Avidan S. Normalizing flows for human pose anomaly detection [C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision. 2023: 13545-13554. Nie et al. (2024): Nie Y, Huang H, Long C, et al. Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection [C] / / European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 449-467. AUC: Area Under Curve, area under the ROC curve.

[0032] ROC curve: receiver operating characteristic curve

[0033] Table 1: Comparison of AUC of different methods on the STech dataset , From Table 1, the improved GCL method obtains an AUC value of 88.47% on the STech dataset with relatively simple scene transformation and fewer abnormal types, which is 9.54% higher than the method proposed by Zaheer et al.

[0034] Example 3: see Figure 3 In order to illustrate the sensitivity of the new method to various abnormal events, this embodiment uses two methods to analyze different types of abnormal events in the Stech dataset in detail, and calculates the AUC% of the present method and the original GCL method for different types of abnormal events. The AUC is the area under the ROC curve and the coordinate axis. From Figure 3 It can be seen that the performance of the present application in the fine-grained performance of multiple types of abnormal events is improved compared with the improved GCL method.

[0035] Example 4: see Figure 4 This embodiment evaluates the influence of four negative learning strategies on the performance of the improved model on the STech dataset, which are strategy 1: no negative learning, strategy 2: all one vector replacement, strategy 3: random Gaussian noise replacement, and strategy 4: the gradual Gaussian noise replacement proposed by the present application. The ROC curve comparison chart of the four strategies is as shown in Figure 4 , and the influence table of negative learning on algorithm performance is as shown in Table 2, and Figure 4 In the table, the diagonal dashed line is the diagonal line of the ROC curve chart.

[0036] Table 2: Influence table of negative learning on algorithm performance , From Table 2, it can be seen that the strategy 4 of the present application achieves an AUC value of 88.47%, which is 7.98% higher than the strategy 1 without negative learning. It is particularly worth noting that the traditional negative learning method shows a significant performance degradation phenomenon. The all one vector replacement of strategy 2 reduces the AUC by 2.96% percentage points, and the random Gaussian noise replacement of strategy 3 leads to a more serious performance decline of 7.97% percentage points, Figure 4 which also confirms this trend.

[0037] This phenomenon can be attributed to the interference effect of the negative learning process on the learning of the latent feature distribution. In the model initialization stage, due to the insufficient establishment of the discrimination ability of the discriminator and the generator, the introduction of strong noise disturbance too early will destroy the representation of the latent distribution of the normal frame. Specifically, the all-one vector replacement causes anisotropic distortion of the latent space by introducing a structured bias, and the homogeneous disturbance of the random Gaussian noise destroys the continuity pattern of the temporal features.

[0038] In contrast, strategy 4 proposed in the present application adopts an adaptive noise injection mechanism, which ensures that the model applies weak disturbance at the beginning of training, and gradually increases the noise intensity as the discriminator confidence improves.

[0039] Experiments show that the traditional negative learning method will produce a model degradation effect in the improved model, while the progressive noise injection mechanism proposed in the present application significantly improves the discriminability of abnormal features by dynamically adjusting the disturbance intensity, while ensuring the stability of model convergence. Its Roc curve is significantly better than the other three negative learning methods, as shown in Figure 4 , which is 7.98 percentage points higher than the best performance of 80.49% in the other three negative learning methods.

[0040] Embodiment 5, see Figures 5-12 In order to more comprehensively evaluate the performance of the method of the present application, four videos are selected from the STech and UCFC data sets in this embodiment, respectively labeled as video one to video eight, and the results are shown by using the method of the present application for visual detection. Figures 5-8 The figure in Figures 9-12 The figure data set in Figures 5-12 In

[0041] Figure 5 shows the normal event of a pedestrian crossing the road normally, Figure 6 and Figure 8 shows the abnormal event of a van driving on the sidewalk; Figure 7 shows the abnormal event of a scooter driving on the sidewalk; Figure 9 shows the normal event of working in the office, Figure 10 shows the abnormal event of fighting, Figure 11 and Figure 12 shows the abnormal event of theft. It can be seen that the model can accurately find the abnormality even in a relatively dark environment.

[0042] In Figure 6 and Figure 8It can be seen that the model predicts the first segment of the video as abnormal, which may be because the first segment of the video lacks context information, causing the model to fail to correctly judge. And in Figure 6 The problem of abnormality detection delay occurs in the middle, which may be because the model lacks supervision signals, and the car has not completely entered the monitoring picture at the beginning of the anomaly, so the model cannot normally judge the anomaly. As the part of the car entering the monitoring picture increases, the abnormal score of the model also slowly increases.

[0043] From the visualization results, it can be seen that the prediction score of the model is basically consistent with the actual label: the model predicts a higher abnormal score for the abnormal segment, and a lower score for the normal segment. From Figures 5-12 It can be seen that no matter what type of anomaly, the abnormal score will increase at the beginning of the real anomaly and decrease after the end of the real anomaly. Although there may be some false positives and false negatives, overall, the method proposed in this paper has good performance in video anomaly detection and can accurately identify the abnormal area in the video.

[0044] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for video anomaly detection using progressive Gaussian noise negative learning, characterized in that, The method comprises the following steps: S1, constructing a data set D; obtaining I videos for video anomaly detection, dividing each video into J segments, each segment containing p frame pictures, wherein the jth segment of the ith video is C i,j , 1≤i≤I, 1≤j≤J; S2, constructing an unsupervised co-learning network comprising a feature extractor, a generator, a discriminator, a pseudo-label generator and a noise scheduler; The feature extractor is configured to randomly input a batch of segments, generate a feature vector of each segment, and if the qth segment in the batch is C i,j , mark the feature vector of the qth segment as , 1≤q≤b, and b is the size of the batch. The generator is a diffusion probability model for input , generating a corresponding reconstructed vector ; The discriminator is configured to input , and output an anomaly probability ; The pseudo label generator is configured to generate a first pseudo label and a second pseudo label , where if it is determined that the anomaly is based on a reconstruction loss of the generator , then = 1, otherwise 0, and if it is determined that the anomaly is based on a binary cross-entropy loss of the discriminator , then = 1, otherwise 0. The noise scheduler is configured to generate a pseudo-reconstruction target according to the following equation ， ， In the formula, t is the iteration number, x t is the noise-added data of the tth iteration, and e is a standard Gaussian noise, is the cumulative attenuation coefficient of the tth step. S3, taking a pseudo-reconstruction target as a final output of the generator, training the generator and the discriminator in the unsupervised co-learning network by using a negative learning method until an iteration is ended, and obtaining an anomaly detection model; S4, obtaining a to-be-identified segment, inputting the to-be-identified segment into the anomaly detection model, generating a corresponding feature vector by the feature extractor, and outputting a corresponding anomaly probability by the discriminator. 2.The video anomaly detection method using progressive Gaussian noise negative learning according to claim 1, wherein, The feature extractor is a pre-trained R3D network. 3.The video anomaly detection method using progressive Gaussian noise negative learning according to claim 1, wherein, The pseudo-label generator generates a first pseudo-label of according to the following equation : ， ， In the formula, is the reconstruction loss of the generator with , is the first threshold value, , are the average value and the standard deviation of a batch of , respectively, and k is the first constant.

4. The video anomaly detection method using progressive Gaussian noise negative learning according to claim 1, wherein, The pseudo-label generator generates a second pseudo-label of :​ ， ， wherein is the binary cross-entropy loss of the discriminator, is a second threshold value, , are the mean and standard deviation, respectively, of a batch of data, is a second constant.

5. The video anomaly detection method using progressive Gaussian noise negative learning according to claim 1, wherein, In the noise scheduler, According to the following formula: ， ， where a s is the attenuation coefficient of the s-th step, 1≤s≤t, a t is the attenuation coefficient of the t-th step.

6. The video anomaly detection method using progressive Gaussian noise negative learning according to claim 1, wherein, The negative learning method is When training the discriminator, a loss function L of the discriminator is calculated according to the following formula D and L is minimized D adjusting network parameters of the discriminator; ， When training the generator, a loss function L for the generator is computed according to the following equation G and L is minimized G adjusting network parameters of the generator; 。

Images