Weakly supervised video anomaly detection method and system based on potential field damping dynamics

Abstract

This invention proposes a weakly supervised video anomaly detection method and system based on potential field damping dynamics. The method includes: inputting video into an anomaly detection model to obtain an original video feature sequence; calculating the wave potential energy using the original video feature sequence and further obtaining a global inertial surrogate vector; calculating the interaction potential energy between the global inertial surrogate vector and the original video feature vector; generating a nonlinear adaptive damping force using the interaction potential energy; applying the nonlinear adaptive damping force to the original video feature vector to obtain a purified dynamic feature vector; constructing a loss function using the purified dynamic feature vector and training the model to obtain an optimized model; inputting the video into the optimized model to obtain the final frame-level anomaly detection result. This invention reconstructs anomaly detection from a classification problem into a signal decoupling and energy dissipation problem in a physical system, theoretically explaining the physical essence of the bias removal mechanism.

Description

Technical Field

[0001] This invention relates to the field of computer vision and video analysis technology, and in particular to a weakly supervised video anomaly detection method and system based on potential energy field damping dynamics. Background Technology

[0002] Weakly supervised video anomaly detection (WS-VAD) aims to locate frame-level anomalous events using only video-level labels. The mainstream multiple instance learning (MIL) paradigm suffers from an inherent "stationary trap" in its optimization dynamics: the background environment in surveillance videos (such as walls and sky) exhibits extremely high stability and low entropy over time, forming a "strong inertia" component in the feature space. In contrast, anomalous behavior is often a fleeting "transient perturbation."

[0003] Because deep neural networks tend to prioritize fitting the component with the highest signal-to-noise ratio in a signal, the optimizer naturally takes a shortcut: predicting anomalies by identifying specific contextual biases rather than capturing complex action patterns. This phenomenon causes the model to essentially degenerate into a "scene recognizer," prone to generating a large number of false positives or false negatives in complex scenes. Existing pseudo-label self-training or feature clustering methods are often based on the simple assumption of Euclidean distance, ignoring the continuous variation of environmental features on the manifold and easily introducing noise. Summary of the Invention

[0004] In view of the above, the main objective of this invention is to propose a weakly supervised video anomaly detection method and system based on potential energy field damping dynamics, so as to solve the above-mentioned technical problems.

[0005] This invention proposes a weakly supervised video anomaly detection method based on potential energy field damping dynamics, the method comprising the following steps:

[0006] Step 1: Input the high-resolution image dataset into the reconstruction model, and process the high-resolution image dataset through the image processing module to obtain high-resolution texture images and low-resolution depth images;

[0007] Step 2: Input the high-resolution texture image and the low-resolution depth image into the encoder module to extract shallow features, so as to obtain shallow features of the texture image and shallow features of the depth image.

[0008] Step 3: Use the encoder module to extract deep features from the shallow features of the texture image and the shallow features of the depth image to obtain deep guiding features and deep depth features;

[0009] Step 4: Use the encoder module to fuse the deep guidance features and deep depth features to obtain high-resolution guidance features and low-resolution depth features.

[0010] Step 5: Input the high-resolution guiding features and low-resolution depth features into the decoder module for processing to obtain the super-resolution reconstructed image.

[0011] This invention also proposes a weakly supervised video anomaly detection system based on potential energy field damping dynamics, the system comprising:

[0012] The feature extraction module is used for:

[0013] Feature extraction is performed on the video to obtain the original video feature sequence;

[0014] The inertial attractor module is used for:

[0015] The wave potential energy is calculated and obtained using the original video feature vector in the original video feature sequence; the wave potential energy is converted into the probability of belonging to the background inertia at each time step using the Boltzmann distribution, and then the global inertial proxy vector is obtained.

[0016] Damped dynamics module, used for:

[0017] The interaction potential energy is calculated by the global inertial proxy vector and the original video feature vector; a nonlinear adaptive damping force is generated by the interaction potential energy; the nonlinear adaptive damping force is applied to the original video feature vector to obtain the purified dynamic feature vector.

[0018] The training module is used for:

[0019] A loss function is constructed using the purified dynamic feature vectors, and the video anomaly detection model is trained using the loss function to obtain the optimized model;

[0020] The anomaly scoring module is used for:

[0021] The video is input into the optimized model to obtain the updated purified dynamic feature vector; based on the updated purified dynamic feature vector, the final frame-level anomaly detection result is obtained.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention proposes a physical-inspired paradigm shift that reconstructs anomaly detection from a simple classification problem into a signal decoupling and energy dissipation problem in a physical system, and explains the physical essence of the debiasing mechanism from a theoretical perspective.

[0024] 2. This invention uses second-order statistics (Hessian potential energy) instead of the traditional first-order average to define the background, which can more accurately locate the steady-state inertial center in the video and conforms to the thermodynamic laws of Boltzmann distribution.

[0025] 3. The Lagrange damping dynamics mechanism designed in this invention dynamically "dissipates" background features through nonlinear damping force. Compared with traditional feature subtraction or attention mechanisms, it can more effectively handle complex nonlinear background drift such as illumination changes. Attached Figure Description

[0026] Figure 1 This is a flowchart of the weakly supervised video anomaly detection method based on potential field damping dynamics proposed in this invention;

[0027] Figure 2 This is a schematic diagram of the inertial attractor and damping dynamics module of the weakly supervised video anomaly detection method based on potential energy field damping dynamics proposed in this invention.

[0028] Figure 3 This is a diagram illustrating the gradient reversal and discriminator game principle in the adversarial inertial stripping framework of the weakly supervised video anomaly detection method based on potential field damping dynamics proposed in this invention.

[0029] Figure 4 This is a schematic diagram of the framework of the weakly supervised video anomaly detection system based on potential field damping dynamics proposed in this invention. Detailed Implementation

[0030] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0031] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0032] Please see Figure 1 This invention proposes a weakly supervised video anomaly detection method based on potential energy field damping dynamics, which includes the following steps:

[0033] Step 1: Input the video into the video anomaly detection model and extract features from the video to obtain the original video feature sequence;

[0034] In step 1, the video is input into the video anomaly detection model, and features are extracted from the video to obtain the original video feature sequence. The corresponding process has the following relationship:

[0035] ;

[0036] in, Represents the original video feature sequence. Both represent the original video feature vectors. Indicates a time index; This indicates the total number of moments, i.e., the total number of video segments; Indicates feature dimension as A real vector space (e.g., 2048 or 512).

[0037] Furthermore, in this step, the original video stream is uniformly divided into... The original video feature sequence is obtained by extracting features from non-overlapping segments using a pre-trained deep neural network (such as I3D or CLIP).

[0038] Step 2: Calculate and obtain the wave potential energy using the original video feature vector in the original video feature sequence; use the Boltzmann distribution to convert the wave potential energy into the probability of belonging to the background inertia at each time moment, and further obtain the global inertia proxy vector;

[0039] Please participate Figure 2 In step 2, the wave potential energy is calculated and obtained using the original video feature vector in the original video feature sequence; the wave potential energy is then converted into the probability of belonging to the background inertia at each time step using the Boltzmann distribution, and the global inertial proxy vector is further obtained. The specific steps are as follows:

[0040] The temporal centroid vector is calculated using the original video feature vector;

[0041] Based on the time-series centroid vector, the second-order term of the Taylor expansion is used to measure the stability of the feature at each time step relative to the centroid, so as to obtain the wave potential energy.

[0042] By converting the wave potential energy into a probability distribution using the Boltzmann distribution, the probability of belonging to the background inertia at each moment can be obtained.

[0043] Based on the probability of belonging to the background inertia at each time moment, the original video feature vector is weighted and aggregated to obtain the global inertial proxy vector;

[0044] In the process of calculating the temporal centroid vector using the original video feature vectors, the following relationship exists:

[0045] ;

[0046] in, Represents the temporal centroid vector;

[0047] In the process of obtaining the fluctuation potential energy by using the second-order term of the Taylor expansion to measure the stability of the feature at each time step relative to the centroid based on the time-series centroid vector, the following relationship exists:

[0048] ;

[0049] in, express Fluctuating potential energy at any given moment Indicates a dimension index. The original video feature vector is represented at the th... Scalar values ​​of dimension The temporal centroid vector is represented in the th order. Dimensional variable component values; Represents a small constant that keeps the value stable (e.g.) );

[0050] In the process of converting wave potential energy into a probability distribution through the Boltzmann distribution to obtain the probability of belonging to background inertia at each moment, the following relationship exists:

[0051] ;

[0052] in, express The probability that a given moment belongs to the background inertia. Indicates the temperature coefficient; This represents the time index of the summation traversal in the denominator normalization calculation, indicating the traversal from the 1st time segment to the 2nd time segment. A moment in time; express Fluctuating potential energy at any given moment;

[0053] In the process of weighted aggregation of the original video feature vectors based on the probability of belonging to background inertia at each time moment to obtain the global inertial proxy vector, the following relationship exists:

[0054] ;

[0055] in, This represents the global inertial proxy vector.

[0056] Furthermore, in this step, the wave potential energy The physical meaning is that, The smaller the value, the smaller the curvature of the frame feature on the time axis, indicating that it is at the bottom of the "steady-state trap" and is very likely a static background.

[0057] Specifically, the core idea of ​​this step is to model the video feature manifold as a thermodynamic system, then use the second-order statistical moments of the feature sequence to construct the potential energy surface of temporal fluctuations, and calculate the probability that the features at each time step belong to the background inertia through the Boltzmann distribution, and then aggregate them to obtain the global inertial proxy vector.

[0058] Step 3: Calculate the interaction potential energy using the global inertial proxy vector and the original video feature vector; generate a nonlinear adaptive damping force using the interaction potential energy; apply the nonlinear adaptive damping force to the original video feature vector to obtain the purified dynamic feature vector;

[0059] In step 3, the interaction potential energy is calculated using the global inertial proxy vector and the original video feature vector; a nonlinear adaptive damping force is generated using the interaction potential energy; and the nonlinear adaptive damping force is applied to the original video feature vector to obtain the purified dynamic features. The specific steps are as follows:

[0060] The interaction potential energy is calculated using the global inertial proxy vector and the original video feature vector;

[0061] Based on the interaction potential energy, the energy dissipation rate is obtained through a mapping network;

[0062] By applying a damping force to the original video feature vector using the energy dissipation rate, a purified dynamic feature vector is obtained.

[0063] In the process of calculating the interaction potential energy using the global inertial proxy vector and the original video features, the following relationship exists:

[0064] ;

[0065] in, express The interaction potential energy at time t, and ; Represents a learnable coupled tensor, and ; Indicates the transpose operation; This indicates processing using the hyperbolic tangent function;

[0066] In the process of obtaining the energy dissipation rate based on the interaction potential energy through a mapping network, the following relationship exists:

[0067] ;

[0068] in, express The energy dissipation rate at time t, i.e., the damping coefficient and ; This indicates that the process is handled using the sigmoid activation function; This indicates that the feature dimension and the dimension of the real vector space are processed by a mapping network consisting of two fully connected layers. Consistent;

[0069] In the process of applying a damping force to the original video feature vector using the energy dissipation rate to obtain the purified dynamic feature vector, the following relationship exists:

[0070] ;

[0071] in, This represents the purified dynamic feature vector. This represents element-wise multiplication. This represents the perturbation term.

[0072] Furthermore, in this step, The perturbation terms are generated by a lightweight neural network (e.g., a multilayer perceptron MLP containing one or two fully connected layers or a one-dimensional convolutional layer).

[0073] This lightweight network uses the original video feature vector as... The input and output are perturbation vectors with the same dimension as the features. This design makes the perturbation terms state-dependent, which can be dynamically adjusted according to the characteristics of the current features, thereby enhancing the nonlinear fitting ability of the model to compensate for the shortcomings of linear damping.

[0074] Furthermore, in this step, the purified dynamic feature vector is obtained. The equations show that the stronger the coupling between the feature and the background, the higher the damping coefficient. The larger the value, the more it forces the consumption of the inertial component;

[0075] The physical essence of this equation is that when the feature is strongly coupled with the background ( Large), damping coefficient Increasing the damping force forces the inertial component to be "consumed"; conversely, for anomalous features, the damping is minimal, thus preserving the original signal.

[0076] Step 4: Construct a loss function using the purified dynamic feature vectors, and train the video anomaly detection model using the loss function to obtain the optimized model;

[0077] Please see Figure 3 In step 4, a loss function is constructed using the purified dynamic feature vectors. The video anomaly detection model is then trained using this loss function to obtain an optimized model. The specific process for constructing the loss function used for model training is as follows:

[0078] The purified dynamic feature vectors are aggregated to obtain purified video-level aggregated features;

[0079] Construct a method that minimizes cross-entropy loss using purified video-level aggregated features;

[0080] Construct an adversarial loss by minimizing the cross-entropy loss;

[0081] The ranking loss is constructed by using the highest abnormal score instance corresponding to the abnormal video and the highest score instance corresponding to the normal video.

[0082] A regularized loss is constructed using purified dynamic features;

[0083] The purified dynamic feature sequence was obtained based on the purified dynamic features;

[0084] The purified dynamic feature sequence, ranking loss, regularization loss, and adversarial loss are used together to form the loss for model training;

[0085] In the process of constructing a system that minimizes the cross-entropy loss using purified video-level aggregated features, the following relationship exists:

[0086] ;

[0087] in, This represents minimizing the cross-entropy loss function. Indicates the total number of potential environmental categories; Indicates the first The pseudo-labels corresponding to the class environment serve as the "real labels" of the steady-state trap discriminator, constituting the supervision signal for training; This indicates that the discriminator predicts that it belongs to the first... The probability of a class environment. This represents the purified video-level aggregated features;

[0088] In constructing the system to minimize the cross-entropy loss, the specific construction of the training data involves: the system using K-Means or spectral clustering algorithms to process the global inertial proxy vectors generated from historical iterations. Cluster analysis is performed to divide the feature space into Each cluster is defined, and the cluster index is used as the environment label. ;

[0089] In this process, the purified video-level aggregated features The input training data (i.e., the training set) constitutes the discriminator, while the automatically generated... Then, as the corresponding target label, this self-supervised mechanism enables the model to perform adversarial training without additional manual environmental annotation;

[0090] In the process of constructing adversarial loss by minimizing cross-entropy loss, the following relationship exists:

[0091] ;

[0092] in, Represents the adversarial loss function; The adversarial weight coefficient is a preset hyperparameter used to control the gradient scaling ratio during gradient reversal. In this invention, the weight can be set to a fixed value (e.g., 1.0) based on experimental experience, or set to a value that changes dynamically with the number of training iterations (e.g., gradually increasing from 0 to 1) to balance the relationship between feature confusion and feature preservation.

[0093] In the process of constructing adversarial loss, in order to decouple features from the environment, the feature extractor and damping module need to "trick" the discriminator. A gradient inversion layer (GRL) is introduced in backpropagation. During forward propagation, the GRL keeps the input unchanged, and during backpropagation, it inverts and scales the gradient.

[0094] In constructing the ranking loss using the highest-scoring instance of anomaly videos and the highest-scoring instance of normal videos, the following relationship exists:

[0095] ;

[0096] in, Represents the sorting loss function. This indicates that the data is processed using a classifier scoring function. This represents the instance with the highest anomaly score corresponding to the abnormal video. This represents the highest score instance corresponding to a normal video. Indicates classifier pair The predicted score, Indicates classifier pair The predicted score, This indicates the operation of retrieving the maximum value;

[0097] In the process of constructing the regularized loss using the purified dynamic features, the following relationship exists:

[0098] ;

[0099] in, Represents the regularization loss function. Indicates time The purified dynamic feature vector, Indicates classifier pair The predicted score, Indicates classifier pair The predicted score, Represents the weighting coefficients of the balancing sparsity constraint terms;

[0100] In the process of using the purified dynamic feature sequence, ranking loss, regularization loss, and adversarial loss to jointly construct the loss for model training, the following relationship exists:

[0101] ;

[0102] in, This represents the total loss function used for model training. and All represent the equilibrium hyperparameters. This represents the purified dynamic feature sequence. This indicates a video-level realistic tag.

[0103] Furthermore, this step uses an adversarial learning mechanism to force the purified features to be indistinguishable in terms of environmental attributes, thereby achieving complete debiasing.

[0104] This step constructs an adversarial stripping framework (including a feature encoder and a steady-state trap discriminator). By using a gradient inversion mechanism, an adversarial game is established between the feature encoder and the steady-state trap discriminator, forcing the purified features to be indistinguishable in terms of environmental attributes, thereby completely stripping away the steady-state inertia of the environment.

[0105] Step 5: Input the video into the optimized model to obtain the updated purified dynamic feature vector; based on the updated purified dynamic feature vector, obtain the final frame-level anomaly detection result.

[0106] This step integrates the updated purified dynamic feature vectors into a purified dynamic feature sequence, which is then input into the trained MIL classifier (containing fully connected layers and sigmoid), and outputs the anomaly probability at each time step. ;

[0107] The anomaly probability sequence is smoothed over time using Gaussian smoothing to output the final frame-level anomaly detection results.

[0108] Please see Figure 4 This invention also provides a weakly supervised video anomaly detection system based on potential field damping dynamics, the system comprising:

[0109] The feature extraction module is used for:

[0110] Feature extraction is performed on the video to obtain the original video feature sequence;

[0111] The inertial attractor module is used for:

[0112] The wave potential energy is calculated and obtained using the original video feature vector in the original video feature sequence; the wave potential energy is converted into the probability of belonging to the background inertia at each time step using the Boltzmann distribution, and then the global inertial proxy vector is obtained.

[0113] Damped dynamics module, used for:

[0114] The interaction potential energy is calculated by the global inertial proxy vector and the original video feature vector; a nonlinear adaptive damping force is generated by the interaction potential energy; the nonlinear adaptive damping force is applied to the original video feature vector to obtain the purified dynamic feature vector.

[0115] The training module is used for:

[0116] A loss function is constructed using the purified dynamic feature vectors, and the video anomaly detection model is trained using the loss function to obtain the optimized model;

[0117] The anomaly scoring module is used for:

[0118] The video is input into the optimized model to obtain the updated purified dynamic feature vector; based on the updated purified dynamic feature vector, the final frame-level anomaly detection result is obtained.

[0119] Furthermore, to verify the effectiveness and advancement of this invention, extensive quantitative and qualitative experiments were conducted on two video anomaly detection benchmark datasets with different bias distributions. The datasets include UCF-Crime and ShanghaiTech. UCF-Crime contains a large number of real surveillance videos with severe environmental biases (such as dark nights and complex textured backgrounds), which is a challenging benchmark for testing the "bias removal capability" of this invention. ShanghaiTech was used to verify the feature preservation capability of the model in standard scenarios.

[0120] Under weak supervision, this invention uses the industry-standard frame-level AUC (Area Under the ROCCurve) as the main evaluation metric and compares it with 10 recent state-of-the-art (SOTA) methods, covering a variety of approaches based on I3D, ViT, and CLIP features.

[0121] Table 1 shows the performance comparison results (AUC / %) of the present invention and the current state-of-the-art technology on the UCF-Crime and ShanghaiTech datasets:

[0122] Table 1. Performance comparison of the present invention with the most advanced technologies (AUC / %)

[0123]

[0124] Based on the analysis of the experimental data in Table 1, the following conclusions can be drawn:

[0125] (1) On the highly challenging high-biased dataset UCF-Crime, this invention achieved an AUC of 87.93%, setting a new record for the best performance in the field. Compared with methods that focus on optimizing the internal mechanism of MIL (such as UR-DMU), this invention achieved a significant advantage (+0.96%). This proves that simply patching the aggregation logic of MIL is insufficient to resist severe context bias, while the "physical dynamics debiasing" strategy proposed in this invention is more effective.

[0126] (2) Compared with the HSC method, which attempts to solve scene interference through hierarchical comparison, this invention demonstrates superior robustness. This shows that using the temporal stationarity defined by the "Hessian potential field" as the adversarial target can more accurately remove environmental noise than simple semantic comparison;

[0127] (3) The performance of the present invention is comparable to multimodal methods such as VadCLIP and TPWGN that rely on complex text prompts (PromptEngineering). For example, compared with TPWGN (87.50%), the present invention achieves higher accuracy without the need to design complex text descriptions. This proves that optimizing from the underlying physical logic (energy dissipation of feature manifolds) is an efficient way to improve feature adaptability.

[0128] Furthermore, to verify the necessity and contribution of each core module (Hessian inertial attractor, Lagrange damping dynamics, and antagonistic inertial stripping) in this invention, a decomposition analysis was performed on the UCF-Crime dataset. Table 2 shows the ablation experimental results of the core modules:

[0129] Table 2 Ablation Experiment Results

[0130]

[0131] Table 2 clearly illustrates the independent contribution of each innovative module of this invention:

[0132] (1) The core role of adversarial inertial stripping (AIS): Compared with Experiment 1 and Experiment 3, after removing the adversarial branch, the model degenerates to the standard MIL baseline and the AUC drops to 87.37%. It is believed that the model falls back into "shortcut learning" at this time and tends to use background texture for prediction. The introduction of AIS brings significant performance improvement, proving that forcing features to be indistinguishable in terms of environmental attributes is the key to improving generalization ability.

[0133] (2) Synergistic effect of Lagrange damping and Hessian attractor: Experiment 2 shows that if only physical modules are used without adversarial constraints, the performance fluctuates. This is because physical damping requires the guidance of adversarial gradients to calibrate the "damping coefficient". Only by combining the two (Experiment 3) can the best balance of "suppressing background noise and preserving action semantics" be achieved.

[0134] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0135] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0136] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A weakly supervised video anomaly detection method based on potential energy field damping dynamics, characterized in that, The method includes the following steps: Step 1: Input the video into the video anomaly detection model and extract features from the video to obtain the original video feature sequence; Step 2: Calculate and obtain the wave potential energy using the original video feature vector from the original video feature sequence; convert the wave potential energy into the probability of belonging to the background inertia at each time step using the Boltzmann distribution, and further obtain the global inertial proxy vector. The specific steps are as follows: The temporal centroid vector is calculated using the original video feature vector; Based on the time-series centroid vector, the second-order term of the Taylor expansion is used to measure the stability of the feature at each time step relative to the centroid, so as to obtain the wave potential energy. By converting the wave potential energy into a probability distribution using the Boltzmann distribution, the probability of belonging to the background inertia at each moment can be obtained. Based on the probability of belonging to the background inertia at each time moment, the original video feature vector is weighted and aggregated to obtain the global inertial proxy vector; Step 3: Calculate the interaction potential energy using the global inertial proxy vector and the original video feature vector; generate a nonlinear adaptive damping force using the interaction potential energy; apply the nonlinear adaptive damping force to the original video feature vector to obtain the purified dynamic feature vector. The specific steps are as follows: The interaction potential energy is calculated using the global inertial proxy vector and the original video feature vector; Based on the interaction potential energy, the energy dissipation rate is obtained through a mapping network; By applying a damping force to the original video feature vector using the energy dissipation rate, a purified dynamic feature vector is obtained. Step 4: Construct a loss function using the purified dynamic feature vectors, and train the video anomaly detection model using the loss function to obtain the optimized model. The specific process of constructing the loss function used for model training is as follows: The purified dynamic feature vectors are aggregated to obtain purified video-level aggregated features; Construct a method that minimizes cross-entropy loss using purified video-level aggregated features; Construct an adversarial loss by minimizing the cross-entropy loss; The ranking loss is constructed by using the highest abnormal score instance corresponding to the abnormal video and the highest score instance corresponding to the normal video. A regularized loss is constructed using purified dynamic features; The purified dynamic feature sequence was obtained based on the purified dynamic features; The purified dynamic feature sequence, ranking loss, regularization loss, and adversarial loss are used together to form the loss for model training; Step 5: Input the video into the optimized model to obtain the updated purified dynamic feature vector; based on the updated purified dynamic feature vector, obtain the final frame-level anomaly detection result.

2. The weakly supervised video anomaly detection method based on potential field damping dynamics according to claim 1, characterized in that, In step 1, the video is input into the video anomaly detection model, and features are extracted from the video to obtain the original video feature sequence. The corresponding process has the following relationship: ； in, Represents the original video feature sequence. Both represent the original video feature vectors. Indicates the time index. Represents the total number of moments. Indicates feature dimension as The real vector space.

3. The weakly supervised video anomaly detection method based on potential field damping dynamics according to claim 1, characterized in that, In the process of calculating the temporal centroid vector using the original video feature vectors, the following relationship exists: ； in, Represents the temporal centroid vector; In the process of obtaining the fluctuation potential energy by using the second-order term of the Taylor expansion to measure the stability of the feature at each time step relative to the centroid based on the time-series centroid vector, the following relationship exists: ； in, express Fluctuating potential energy at any given moment Indicates a dimension index. The original video feature vector is represented at the th... Scalar values ​​of dimension The temporal centroid vector is represented in the th order. Dimensional variable component values, This represents a small constant that maintains numerical stability; In the process of converting wave potential energy into a probability distribution through the Boltzmann distribution to obtain the probability of belonging to background inertia at each moment, the following relationship exists: ； in, express The probability that a given moment belongs to the background inertia. Indicates the temperature coefficient. This represents the index of the time traversal during the summation process in the denominator normalization calculation. express Fluctuating potential energy at any given moment; In the process of weighted aggregation of the original video feature vectors based on the probability of belonging to background inertia at each time moment to obtain the global inertial proxy vector, the following relationship exists: ； in, This represents the global inertial proxy vector.

4. The weakly supervised video anomaly detection method based on potential field damping dynamics according to claim 1, characterized in that, In the process of calculating the interaction potential energy using the global inertial proxy vector and the original video features, the following relationship exists: ； in, express The interaction potential energy at time t, Represents a learnable coupled tensor. This indicates the transpose operation. This indicates processing using the hyperbolic tangent function; In the process of obtaining the energy dissipation rate based on the interaction potential energy through a mapping network, the following relationship exists: ； in, express Energy dissipation rate at time t, This indicates that the process is handled using the sigmoid activation function. This indicates processing via a mapping network consisting of two fully connected layers; In the process of applying a damping force to the original video feature vector using the energy dissipation rate to obtain the purified dynamic feature vector, the following relationship exists: ； in, This represents the purified dynamic feature vector. This represents element-wise multiplication. This represents the perturbation term.

5. The weakly supervised video anomaly detection method based on potential field damping dynamics according to claim 1, characterized in that, In the process of constructing a system that minimizes the cross-entropy loss using purified video-level aggregated features, the following relationship exists: ； in, This represents minimizing the cross-entropy loss function. Indicates the total number of potential environmental categories. Indicates the first Pseudo-tags corresponding to class environments This indicates that the discriminator predicts that it belongs to the first... The probability of a class environment. This represents the purified video-level aggregated features; In the process of constructing adversarial loss by minimizing cross-entropy loss, the following relationship exists: ； in, Represents the adversarial loss function. Indicates the adversarial weight coefficient; In constructing the ranking loss using the highest-scoring instance of anomaly videos and the highest-scoring instance of normal videos, the following relationship exists: ； in, Represents the sorting loss function. This indicates that the data is processed using a classifier scoring function. This represents the instance with the highest anomaly score corresponding to the abnormal video. This represents the highest score instance corresponding to a normal video. Indicates classifier pair The predicted score, Indicates classifier pair The predicted score, This indicates the operation of retrieving the maximum value; In the process of constructing the regularized loss using the purified dynamic features, the following relationship exists: ； in, Represents the regularization loss function. Indicates time The purified dynamic feature vector, Indicates classifier pair The predicted score, Indicates classifier pair The predicted score, Represents the weighting coefficients of the balancing sparsity constraint terms; In the process of using the purified dynamic feature sequence, ranking loss, regularization loss, and adversarial loss to jointly construct the loss for model training, the following relationship exists: ； in, This represents the total loss function used for model training. and All represent the equilibrium hyperparameters. This represents the purified dynamic feature sequence. This indicates a video-level realistic tag.

6. A weakly supervised video anomaly detection system based on potential energy field damping dynamics, characterized in that, The system employs any one of the weakly supervised video anomaly detection methods based on potential energy field damping dynamics as described in claims 1 to 5, and the system comprises: The feature extraction module is used for: Feature extraction is performed on the video to obtain the original video feature sequence; The inertial attractor module is used for: The wave potential energy is calculated and obtained using the original video feature vector in the original video feature sequence; the wave potential energy is converted into the probability of belonging to the background inertia at each time step using the Boltzmann distribution, and then the global inertial proxy vector is obtained. Damped dynamics module, used for: The interaction potential energy is calculated by the global inertial proxy vector and the original video feature vector; a nonlinear adaptive damping force is generated by the interaction potential energy; the nonlinear adaptive damping force is applied to the original video feature vector to obtain the purified dynamic feature vector. The training module is used for: A loss function is constructed using the purified dynamic feature vectors, and the video anomaly detection model is trained using the loss function to obtain the optimized model; The anomaly scoring module is used for: The video is input into the optimized model to obtain the updated purified dynamic feature vector; based on the updated purified dynamic feature vector, the final frame-level anomaly detection result is obtained.

Images