Weakly supervised video anomaly detection method and system based on non-equilibrium characteristic flow phase transition topology

Abstract

This invention proposes a weakly supervised video anomaly detection method and system based on non-equilibrium feature flow phase transition topology, belonging to the fields of computer vision and artificial intelligence. The method extracts video segment-level features and compresses them into a low-dimensional latent space; constructs a feature manifold transition matrix to estimate the local drift field; extracts geometric and topological evidence such as action asymmetry; and generates a corrected signal based on the evidence, which is then superimposed on the basic anomaly logic value output by the backbone network to obtain a corrected score. This invention improves the interpretability and robustness of anomaly detection by explicitly quantifying abrupt changes in feature flow dynamics, effectively capturing rare anomalies and suppressing noise-induced false alarms.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, and specifically relates to a weakly supervised video anomaly detection method and system based on non-equilibrium feature flow phase transition topology. Background Technology

[0002] Video anomaly detection is one of the core tasks in the field of intelligent video surveillance. It aims to automatically identify abnormal events in videos that deviate from normal behavioral patterns, and has wide applications in scenarios such as public safety, traffic management, and industrial quality inspection. With the development of deep learning technology, data-driven video anomaly detection methods have become mainstream. Their basic paradigm involves training a model using a large amount of normal video data to identify segments that deviate from normal distribution patterns as anomalies.

[0003] However, existing methods still face the following key technical challenges in practical applications: First, the models lack interpretability. Most mainstream video anomaly detection methods currently employ reconstruction errors based on autoencoders or ranking models based on multi-instance learning. While these methods can output anomaly scores, their decision-making processes typically exhibit an "end-to-end" black-box mapping relationship. When the model generates false positives or false negatives, it is difficult to trace the basis of its judgment. In scenarios such as security checks and judicial evidence collection, where high transparency in decision-making is required, this significantly limits the system's reliability and deployability.

[0004] Secondly, it lacks robustness in detecting rare anomalies. Real-world anomalies are often sparse and diverse, with many not significantly different from normal behavior in appearance (e.g., a pedestrian slowly falling, a disguised intrusion). Methods based on appearance reconstruction or prediction errors are prone to missing such events because the model may have overfitted to normal samples in the training data that appear similar but have different movement patterns.

[0005] Third, there is the problem of over-response to isolated noise. Surveillance videos inevitably contain transient noise such as sudden changes in lighting, camera shake, and encoding artifacts. Existing methods often incorrectly identify these isolated noises as anomalous events. Although some methods introduce temporal smoothing post-processing, this post-processing is disconnected from the model's own scoring mechanism and cannot fundamentally establish the ability to distinguish between "persistent anomalies" and "transient noise."

[0006] Fourth, fine-grained localization under weak supervision is difficult. Under weak supervision with only video-level annotations, existing methods typically rely on attention mechanisms or aggregation strategies based on multiple instances to locate anomalous segments. However, the loss functions of these methods mainly focus on video-level category differentiation, and their ability to detect anomalies at the segment level is limited. There is still considerable room for improvement in the accuracy and boundary integrity of the models in locating anomalous events. Summary of the Invention

[0007] 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 non-equilibrium characteristic flow phase transition topology, so as to solve the above-mentioned technical problems.

[0008] This invention proposes a weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology, the method comprising the following steps: The original video stream is acquired, and segment-level features are extracted to obtain segment-level visual feature sequences; Each segment feature in the segment-level visual feature sequence is dimensionally compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence; In the low-dimensional latent space, the transition matrix of the feature manifold is constructed based on the feature sequence of the latent space, and the local drift field information on the feature manifold is estimated based on the transition matrix; Based on local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; A correction signal is generated based on geometric topological evidence. The correction signal is then superimposed onto the basic anomaly logic value of the segment level generated by the backbone network based on the segment-level visual feature sequence to obtain the corrected segment-level anomaly logic value. Anomaly scores are generated for the video based on the corrected segment-level anomaly logic values.

[0009] This invention also proposes a weakly supervised video anomaly detection system based on non-equilibrium feature flow phase transition topology, wherein the system applies the weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology as described above, and the system includes: The feature extraction and dimensionality reduction module is used for: The original video stream is acquired, and segment-level features are extracted to obtain segment-level visual feature sequences; Each segment feature in the segment-level visual feature sequence is dimensionally compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence; The manifold construction and drift estimation module is used for: In the low-dimensional latent space, the transition matrix of the feature manifold is constructed based on the feature sequence of the latent space, and the local drift field information on the feature manifold is estimated based on the transition matrix; The geometric topology evidence extraction module is used for: Based on local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; The correction and rating generation module is used for: A corrected signal is generated based on geometric topological evidence. The corrected signal is then superimposed on the basic anomaly logic value of the segment level generated by the backbone network based on the segment-level visual feature sequence to obtain the corrected segment-level anomaly logic value.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Existing deep learning-based video anomaly detection methods mostly employ "black box" models, directly mapping from raw features to anomaly scores, making their decision-making logic difficult to trace. This invention explicitly constructs the transition matrix of the feature manifold, estimates the local drift field, and extracts geometric and topological evidence with clear physical meaning, such as motion asymmetry and trajectory curvature. This transforms the question of "why the model determines the segment is anomaly" into a dynamic analysis of "whether the feature flow exhibits irreversible transitions or drastic directional changes at that moment." This makes the basis for anomaly judgment traceable, verifiable, and understandable, solving the interpretability problem of traditional deep learning models.

[0011] 2. Existing methods typically rely on reconstruction or prediction errors, which are insufficient for detecting rare events that appear similar to normal behavior but exhibit abnormal movement patterns (such as slowly occurring anomalies or disguised violent behavior). This invention introduces the "action asymmetry" index, which quantifies the irreversibility of the temporal evolution of feature streams by calculating and comparing the positive and negative action quantities separately. Anomalies are essentially disruptions of normal temporal symmetry; therefore, even if anomalies are not prominent in appearance, their dynamic irreversibility can still be effectively captured by this invention, thereby significantly improving the recall rate for rare anomalies.

[0012] 3. In real-world surveillance scenarios, isolated noises such as lens flicker and sudden changes in illumination often produce transient abnormal responses in the feature space, leading to numerous false alarms using traditional methods. This invention addresses this by differentiating between "real manifold distortion" and "random noise disturbances" through local drift variance estimation. Furthermore, it introduces a multi-scale temporal smoothing response mechanism and event-gated signals, forcing the model to focus on persistent abnormal events rather than instantaneous spikes. This "persistent modeling" strategy significantly suppresses false alarms caused by isolated noise, making the system more reliable in practical deployments.

[0013] 4. Unlike existing technologies that directly splice or simply add multimodal features, this invention employs a "bounded correction" mechanism. Through learnable strength control scalars and numerical pruning operations, a correction signal generated from geometric topological evidence is superimposed onto the basic anomaly logic value of the backbone network. This design offers the following unique advantages: In the early stages of training, the influence of the correction branch tends to zero, and the model degenerates into a mature backbone network, ensuring training stability; as training progresses, the correction signal is gradually introduced, achieving a progressive transition from "data-driven" to "mechanism-driven"; the boundedness of the correction signal ensures that even if an evidence branch makes an incorrect estimate, it will not lead to a collapse in the overall score. This "degenerate, controllable, and collaborative" correction mechanism solves the technical problem that inferior branches in traditional multi-branch fusion can drag down overall performance.

[0014] 5. Existing weakly supervised methods typically only perform video-level or coarse-grained anomaly localization. This invention constructs an event-metric-based transmission separation loss. By explicitly calculating the Wasserstein distance between anomalous and normal videos in the event metric space and separating the two distributions, the model can more accurately learn event-level discriminative features under weak supervision with only video-level labels. Simultaneously, the adaptive multi-loss weighting strategy dynamically adjusts the weights of each sub-loss function according to its convergence speed, avoiding the problem of a single loss function dominating training and causing the overall model to get stuck in local optima. This improves training stability and generalization ability in complex multi-scenario environments.

[0015] 6. This invention maps the original high-dimensional features to a low-dimensional latent space using a dimensionality compression operator before performing dynamic deconstruction, significantly reducing the computational complexity of transition matrix construction and drift field estimation. Dimensionality compression employs non-parametric operators such as channel grouping averaging or one-dimensional adaptive average pooling, avoiding the overfitting risk and training overhead caused by introducing additional learnable parameters. This allows the invention to maintain high detection accuracy while possessing good real-time performance, making it suitable for video anomaly detection deployment at edge devices or in resource-constrained environments.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of the weakly supervised video anomaly detection method for non-equilibrium characteristic flow phase transition topology proposed in this invention; Figure 2 This is a diagram illustrating the overall architecture of the weakly supervised video anomaly detection method for non-equilibrium characteristic flow phase transition topology proposed in this invention. Figure 3 This is a flowchart illustrating the calculation of phase-change topology transmission loss in this invention. Figure 4 This is a schematic diagram of the structure of the weakly supervised video anomaly detection system based on the non-equilibrium characteristic flow phase transition topology proposed in this invention. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] Please see Figure 1 and Figure 2 This embodiment provides a weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology. The method includes the following steps: Step 1: Obtain the original video stream and perform segment-level feature extraction to obtain a segment-level visual feature sequence; First, this invention divides the input raw video stream into T consecutive, non-overlapping segments. A pre-trained visual feature extractor (such as an I3D or C3D encoder) is then used to extract segment-level visual features, resulting in the original feature sequence. Where d is the original feature dimension (e.g., 512) and T is the total number of segments in the original video stream.

[0021] Step 2: Compress the dimensionality of each segment feature in the segment-level visual feature sequence and map it to a low-dimensional latent space to obtain the latent space feature sequence; To reduce the overhead of subsequent complex dynamic calculations, this step introduces a dimension compression operator. Map features to a low-dimensional space , Indicates the dimension of a low-dimensional quantity space: (1) Latent space mapping: First, the original features are projected through a linear layer and the ReLU activation function to obtain the latent representation. The corresponding process has the following relationship: ; in, Represents the ReLU activation function. This represents the bias term of the linear mapping layer. This represents the weight matrix of the linear mapping layer. Implicit representation, This represents the t-th original feature.

[0022] (2) Masking: To eliminate the distribution offset between different videos, the latent representation is masked to obtain the standardized latent representation. The corresponding process has the following relationship: ; in, and Let represent the mean and variance of the valid segments within the video, respectively. This represents the smoothing factor used to ensure numerical stability. This is the standardized implicit representation.

[0023] A dimension compression operator is introduced to reduce the dimensionality of the standardized latent representation, yielding latent space features. The computation process of the dimension compression operator follows the following relationship: ; in, This represents the low-dimensional latent space features corresponding to time step t, and the feature flow of the corresponding video sequence in the latent space is composed of T low-dimensional features; in this embodiment, Set to 32, when the original dimension can be When divisible, Channel-group averaging is used; otherwise, one-dimensional adaptive average pooling is used. This step significantly reduces the computational complexity of subsequent feature manifold construction while preserving core semantic information.

[0024] Step 3: In the low-dimensional latent space, construct the transition matrix of the feature manifold based on the latent space feature sequence, and estimate the local drift field information on the feature manifold based on the transition matrix; This step treats the latent space feature sequence as a discrete feature flow in the latent space and uses the construction of a manifold topology to characterize the evolution of features over time.

[0025] (1) Self-adjusting scale diffusion kernel calculation: Calculate the Euclidean distance between the latent space features of any two valid segments. The corresponding process has the following relationship: ; in, Let Euclidean distance represent the latent space features between any two valid segments. Let represent the latent space features of the i-th and j-th valid segments, respectively; Based on the obtained Euclidean distance, a self-adjusting scale diffusion kernel is constructed, and the corresponding process follows the following relationship: ; in, The similarity strength represents the mutual diffusion intensity (i.e., similarity intensity) of feature points corresponding to the i-th segment and the j-th segment on the feature manifold, where the feature points refer to the vector representations of the video segments mapped to the latent space; Let represent the local scale factors of the i-th and j-th effective segments, respectively, and let be defined as the mean distances from the current segment features to neighboring points; These represent the effective segment masks for the i-th and j-th effective segments, respectively, used to exclude interference from padding segments; Let represent an exponential function with base e; the calculation of the local scale factor of the i-th effective segment follows the following relationship: .

[0026] (2) Construction of the transition matrix: Row normalization is performed on the diffusion kernel to obtain the transition matrix on the characteristic manifold. The corresponding process has the following relationship: ; in, It represents the conditional probability of transitioning from state i corresponding to the i-th valid segment to state j corresponding to the j-th valid segment in the feature space; This represents the similarity strength between the i-th valid segment and the q-th valid segment; q represents the segment index.

[0027] (3) Estimation of local drift mean and diffusion variance: The local drift mean and diffusion variance at each time step are estimated based on the transition matrix. The corresponding process has the following relationship: ; ; in, This represents the mean local drift at each time step, used to characterize the average evolution trend of the feature flow in the local region; This represents the local diffusion variance at each time step, used to characterize the degree of local divergence of the manifold. The mean local drift and diffusion variance at each time step are used as information about the local drift field.

[0028] Step 4: Based on the local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; Abnormal events are often accompanied by the disruption of the characteristic flow equilibrium state, manifested as irreversibility of motion or drastic abrupt changes in direction.

[0029] (1) Calculation of action asymmetry: Based on the local drift mean and diffusion variance at each time step, the positive and negative action of the characteristic flow at time step t are calculated, and the corresponding process has the following relationship: ; ; in, Let represent the positive and negative motion quantities of the feature flow at time step t, respectively. Represents the difference operator. This represents the component of the local drift field in the r-th dimension at time t. This represents the component of the local diffusion variance in the r-th dimension at time t. This represents the component of the characteristic displacement in the r-th dimension between time t and time t+1. This represents the component of the local drift field in the r-th dimension at time t+1. This represents the component of the local diffusion variance in the r-th dimension at time t+1.

[0030] Based on the positive and negative action quantities, evidence of action quantity asymmetry is obtained, and the corresponding process has the following relationship: ; in, This indicates evidence of asymmetric action volume at time step t. The larger the value, the greater the difference in the forward and reverse evolution costs of the feature flow, that is, the higher the degree of local irreversibility, which indicates the occurrence of anomalies; This indicates the operation of taking the absolute value.

[0031] (2) Calculation of trajectory curvature evidence: by calculating the continuous displacement vector and The cosine of the angle between the two points measures the curvature of the characteristic trajectory, providing evidence of trajectory curvature. The corresponding process follows the following relationship: ; in, Evidence of the trajectory curvature at time step t. The larger the value, the more drastic the sudden change in the direction of motion; Ultimately, through normalization fusion and Evidence of a fragment-level phase transition is obtained, and the corresponding process follows the following relationship: ; in, This indicates evidence of fragment-level phase transitions. This indicates a normalization operation.

[0032] To distinguish between isolated noise (such as lens flicker) and persistent anomalous events (such as fighting), this step introduces a persistence modeling mechanism.

[0033] (1) Construction of phase transition potential energy: Combining the characteristics of the characteristic dynamic abrupt change and the local manifold density, the phase transition potential energy is constructed, and the corresponding process has the following relationship: ; in, Represents the potential energy of phase transition. Represents the natural logarithm function. This indicates logarithmic local density evidence. The local manifold density is represented by the following formula, which is used in the calculation of the local manifold density: ; This phase transition potential is designed to highlight segments with high phase transition intensity that are located in regions of low manifold density (i.e., statistically rare regions).

[0034] (2) Multi-scale temporal smoothing response: across multiple time window scales The local peak-valley structure is calculated to obtain the soft persistent response factor, and the corresponding process has the following relationship: ; in, This represents the soft persistent response factor at scale k. Let represent the Sigmoid activation function, and k represent the size of the local time window. Let represent the maximum and minimum phase transition potential energies within the mouth, respectively. The calculation process for the maximum and minimum phase transition potential energies is based on the following relationship: ; in, These represent the maximum value operation and the minimum value operation, respectively.

[0035] The first term in the formula for calculating local peak-valley structure measures the prominence of a segment relative to the local baseline, while the second term measures the severity of overall fluctuations within the window. Finally, the persistence evidence is obtained by averaging the soft persistent response factors across multiple scales, and the corresponding process follows the relationship: ; in, Indicates persistent evidence; After obtaining persistent evidence, a soft event gating mechanism is first constructed using an element-wise multiplicative fusion mechanism. The corresponding process has the following relationship: ; in, This represents soft event gating; by nonlinearly combining time-domain persistence characteristics with phase transition energy intensity, it achieves preliminary suppression of non-persistent burst noise, aiming to lock in potential anomalous segments with stable dynamic structures. Subsequently, a time-domain sliding window filtering strategy is used to smooth the soft event gating signal to generate the final event gating signal. The corresponding process has the following relationship: ; in, Indicates an event gating signal. This represents the average pooling operation, which eliminates transient spikes in the response sequence and significantly enhances the temporal continuity of anomaly detection results by aggregating the gating scores within the local temporal neighborhood.

[0036] Step 5: Generate a correction signal based on geometric topological evidence, and superimpose the correction signal onto the basic anomaly logic value of the fragment level generated by the backbone network based on the fragment level visual feature sequence to obtain the corrected fragment level anomaly logic value. This step designs an auxiliary feature correction branch, which transforms the above geometric topological evidence into a correction signal for the main trunk anomaly logit value.

[0037] (1) Fusion and Convolution Modeling: Construct a 10-dimensional evidence vector consisting of fragment-level phase transition evidence, action asymmetry evidence, trajectory curvature evidence, logarithmic local density evidence, phase transition potential energy, persistence evidence, event gating signals and their interaction terms, and feed it into a lightweight one-dimensional temporal convolutional network to obtain fragment-level bypass evidence logic values. The corresponding process has the following relationship: ; in, Indicates the logical value of the fragment-level bypass evidence. This represents a lightweight one-dimensional temporal convolutional network. Represents a 10-dimensional evidence vector. ; Indicates the interaction between different pieces of evidence; This indicates a vector construction or concatenation operation. The 10-dimensional evidence vector is masked to zero at invalid fragment locations.

[0038] (2) Bounded Correction Logic: The auxiliary feature correction branch transforms the above geometric topological evidence into a correction signal for the main abnormal logic value, resulting in the corrected fragment-level abnormal logic value. The corresponding process has the following relationship: ; in, This represents the corrected fragment-level exception logic value. This represents a learnable bypass strength control scalar. This represents the hyperbolic tangent activation function, used to restrict it to [-1, 1]. This represents a numerical pruning operation used to limit the logical values ​​of evidence within a stable range. This mechanism ensures that the model degenerates into a backbone network during the initial training phase, and then topological evidence is gradually introduced for fine-tuning and correction, thereby enhancing the stability of the detection.

[0039] This represents the basic anomaly logic value output by the main detection path; its prediction process is detailed in the appendix. Figure 2 As shown, it specifically includes: ① Latent feature mapping: The extracted fragment feature sequence X is fed into a linear mapping layer, and the corresponding latent space feature representation is obtained through a non-linear activation function; ② Regularization: Perform a random dropout operation on the latent space feature representation to enhance the model's generalization ability and prevent overfitting under weakly supervised training; ③ Logical value prediction: The processed latent features are fed into the backbone classifier, which performs linear combination calculations through fully connected network layers to output the basic logistic value of the original probability that the representation fragment belongs to an abnormal state. .

[0040] Step 6: Generate anomaly scores for the video based on the corrected segment-level anomaly logic values.

[0041] Specifically, it includes: Based on the corrected segment-level anomaly logic values, anomaly scores are generated for the video. The corresponding process follows the following relationship: ; in, Indicates the segment-level anomaly score; Finally, all segment-level anomaly scores are aggregated to form the video's anomaly score.

[0042] Please see Figure 3 To learn event-level features more effectively under weak supervision, this invention also involves training optimization steps based on weak supervision, as follows: (1) Construction of event measure: Candidate event region construction: Based on segment-level anomaly scores and segment-level phase transition evidence, the video is divided into... A continuous candidate anomaly region To construct candidate event representations for the video, candidate event regions are obtained; the process of constructing candidate event regions includes: 1) Active segment detection: Using event gating signals that incorporate phase transition intensity for threshold discrimination, the set of active segments that deviate from the normal manifold is identified; 2) Temporal continuous aggregation: Clustering active segments with continuous time steps into several candidate anomalous regions, where each region represents a potential dynamic anomalous event; 3) Filler region construction: If no explicit active region is detected in the current video, a short-term candidate region is constructed around the segment with the highest segment-level anomaly score in the video as a filler region to ensure that each video sample contains at least one event atom that can be used for weakly supervised aggregation.

[0043] Regional feature statistics: For each candidate anomaly region Calculate its average abnormal score Average phase transition intensity and duration ratio ; Calculate the region persistence weight Utilizing the persistence of a region The importance weight of the event atom is calculated using the following formula: ; in, Indicates the region persistence weight. This represents the total number of candidate event regions in video k, where m represents the index of the candidate event region. This indicates the durability of a region, used to characterize the temporal stability of that region.

[0044] Constructing event atoms Combining the above statistics yields event atoms that characterize the properties of candidate events, as shown in the following formula: ; Here, [.] represents the eigenvector construction operation.

[0045] (2) Transmission separation loss: Optimal transmission theory is used to measure the distributional differences between videos: Generate video event metrics : Represent the video k as by Discrete measure composed of event atoms: ; in, This represents the Dirac measure.

[0046] Construct the cost matrix : Calculate event atoms of abnormal videos Event atoms in normal video Migration costs between ; Entropy-regularized Sinkhorn iteration: Based on the cost matrix, the Sinkhorn algorithm is used to iterate and generate the optimal transmission plan. ; Calculate transmission distance The Wasserstein distance between abnormal and normal videos is calculated based on the transmission plan, using the following formula: ; in, This indicates the optimal transmission plan between abnormal and normal videos. Calculate transmission separation loss: ; in, The transmission separation loss is represented by P and N, which represent the sets of anomalous and normal video, respectively. Indicates the adaptive interval. This indicates the operation of taking positive values; This represents the number of elements in the set; the loss explicitly separates the two types of videos in the event measure space.

[0047] (3) Adaptive multi-loss weighting: video-level classification loss Normal item suppression loss Anomaly activation loss Separation loss from transmission Through adaptive weights Weighted optimization is performed on the model to achieve model training. The calculation of the adaptive weights follows the following formula: ; in, This indicates that the gradient operation is stopped. to These represent video-level classification losses respectively. Transmission separation loss Normal item suppression loss and anomaly activation loss ; Let m be the sub-loss function. Let r represent the r-th sub-loss function.

[0048] Constructing the joint optimization total loss L: The final joint optimization objective is obtained by weighting and summing the sub-losses according to the calculated adaptive weights, as shown in the following formula: ; The total loss function balances the convergence speed of each task, ensuring that the model can robustly learn anomalous responses with regional and dynamic structural characteristics in a weakly supervised environment.

[0049] Please refer to Figure 4 This embodiment also provides a weakly supervised video anomaly detection system based on non-equilibrium feature flow phase transition topology, the system comprising: The feature extraction and dimensionality reduction module is used for: The original video stream is acquired, and segment-level features are extracted to obtain segment-level visual feature sequences; Each segment feature in the segment-level visual feature sequence is dimensionally compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence; The manifold construction and drift estimation module is used for: In the low-dimensional latent space, the transition matrix of the feature manifold is constructed based on the feature sequence of the latent space, and the local drift field information on the feature manifold is estimated based on the transition matrix; The geometric topology evidence extraction module is used for: Based on local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; The correction and rating generation module is used for: A corrected signal is generated based on geometric topological evidence. The corrected signal is then superimposed on the basic anomaly logic value of the segment level generated by the backbone network based on the segment-level visual feature sequence to obtain the corrected segment-level anomaly logic value.

[0050] Experimental verification; To comprehensively verify the effectiveness, robustness, and industrial application potential of the Non-equilibrium Flow Topology Network (NFT-Net) proposed in this invention, this embodiment conducts in-depth evaluation on several highly representative public datasets and carries out targeted analysis on the core innovations of this invention.

[0051] 1. Experimental Standards and Evaluation System (1) Dataset selection logic: UCF-Crime (Complex Dynamic Environment): This dataset covers extreme environments such as nighttime, low light, and severe camera shake. Existing technologies are prone to false alarms on this dataset due to "spurious salience" caused by sudden changes in lighting or camera autofocus. It is the core benchmark for verifying the robustness of the "dynamic phase transition modeling" of this invention.

[0052] ShanghaiTech (Refined Behavior Recognition): Used to verify the model's ability to capture subtle anomalies (such as cycling, fighting), and to test the filtering accuracy of the "multi-scale temporal smooth response modeling" of this invention on normal background fluctuations.

[0053] (2) Evaluation indicators: This embodiment uses the area under the frame-level AUC (Area Under the ROC Curve) as the primary metric. In addition, a false alarm rate (False Alarm Rate) is introduced for comparison to reflect the performance difference of the model in distinguishing between "normal natural fluctuations" and "genuine abnormal damage."

[0054] 2. Performance Comparison Analysis: Empirical Evidence of Technological Generational Leadership The NFT-Net proposed in this invention is compared with mainstream algorithms in the current academic community. The results are shown in Table 1: Table 1: Performance Comparison with the Most Advanced Technologies (AUC / %)

[0055] Analysis of technical advantages: (1) The leap from “significance” to “physical phase transition”; Traditional RTFM and MGFN rely on enhancing the saliency of fragments, which are easily affected by environmental noise (such as light and shadow flicker) in complex backgrounds. The present invention, NFT-Net, identifies anomalies from a physical dynamics perspective by extracting the "action asymmetry" and "trajectory curvature" of the feature flow, achieving a significant leap in accuracy on the UCF-Crime dataset.

[0056] (2) Beyond multimodal representation depth; Compared to VadCLIP, which uses textual information, this invention achieves superior performance in a purely visual modality. This demonstrates that explicit modeling of the topological dynamics of feature flows is more accurate in capturing the essential patterns of anomalous events than introducing static textual priors.

[0057] 3. Ablation and technology attribution analysis of core innovation points; The technical contributions of the four core innovative designs of this invention were verified through ablation experiments (see Table 2): Table 2: Percentage Contribution of Innovation Modules to Performance Gains (UCF-Crime)

[0058] Summary of key innovations: (1) Robustness of nonequilibrium dynamics evidence: Evidence of asymmetric action volume The introduction of this feature enables the model to distinguish between "rapid but reversible normal movement" and "irreversible aberrations." In experiments, for rapid walking by normal individuals, the evidence logic value output by NFT-Net remained low, effectively reducing false positives.

[0059] (2) Structured filtering for event persistence: Persistent evidence performs exceptionally well in handling spurious anomalies caused by "lens flicker." Through multi-scale local peak-valley difference analysis, the model can quickly identify the lack of "persistent structure" on the time axis in this response, thereby automatically suppressing its score and achieving a second leap in positioning accuracy.

[0060] (3) Robust optimization in event measure space: The transmission separation loss (step 6) abandons the traditional, fragile Top-k aggregation. By modeling the video as an "event metric" and using Sinkhorn iteration for global matching, the model can more fairly focus on stable outlier regions rather than individual noise points, making the weakly supervised training process more efficient and stable.

[0061] 4. Lightweight deployment and industrial value validation: This invention particularly emphasizes lightweight design of the bypass branch while maintaining high precision (see Table 3): Table 2: Percentage Contribution of Innovation Modules to Performance Gains (UCF-Crime)

[0062] Summary of innovative advantages: This invention achieves a technological breakthrough of "simplifying complexity." The auxiliary feature correction branch requires only minimal computational overhead to perform "controlled correction" on the backbone network. This design allows high-precision video anomaly detection algorithms to reside directly within the embedded chip (SOC) of smart cameras, eliminating the need for expensive backend servers and significantly reducing the deployment cost of smart security systems.

[0063] 5. Physical interpretability analysis: Visual experiments revealed that the boundary residual metrics output by NFT-Net have clear physical meaning. During normal video evolution, the feature flow is in equilibrium, and the phase transition potential energy curve is flat; however, at moments such as "vehicle collisions," evidence of motion asymmetry emerges. Evidence of trajectory curvature The surge in collaboration creates significant "topological breakpoints." This output method, based on system dynamic instability, provides security personnel with clear and reliable evidence for anomaly detection, solving the long-standing "black box" problem of deep learning models.

[0064] 6. Final evaluation of the experiment: Experimental data fully demonstrate that this invention, through non-equilibrium feature flow phase transition topology modeling, successfully overcomes the three core challenges of "noise sensitivity, inaccurate classification, and deployment difficulty" in weakly supervised video anomaly detection. This invention not only sets a new state-of-the-art record in accuracy metrics but also provides an innovative solution unmatched by existing technologies in terms of model lightweighting and physical interpretability, possessing extremely high commercial closed-loop value.

[0065] The core innovations and beneficial effects of this invention: Compared with existing technologies, the weakly supervised video anomaly detection method and system based on non-equilibrium characteristic flow phase transition topology proposed in this invention has significant novelty and inventiveness. Its core innovations and beneficial effects are summarized as follows: 1. A fundamental shift in modeling perspective: from "saliency discrimination" to "non-equilibrium dynamic topology analysis"; Limitations of existing technologies: Traditional weakly supervised video anomaly detection (WVAD) mainly follows the logic of "anomaly = visual saliency," relying solely on segment classifiers to score apparent features. This makes the model highly susceptible to generating a large number of false positives when faced with "natural fluctuations" such as sudden changes in lighting, violent background movement, or rapid shot transitions, due to the enhanced visual saliency.

[0066] This invention is innovative in that it re-examines anomaly detection from the perspective of latent space feature flow for the first time, modeling feature sequences as non-equilibrium discrete flows in latent space. By introducing evidence of "action quantity asymmetry" and "manifold curvature," it explicitly characterizes the phase transition intensity of the feature flow from two dimensions: the irreversibility of physical evolution and the abruptness of geometric trajectories.

[0067] Technical Effects: This innovation enables the model to distinguish between "rapid but normal motion" (reversible path, smooth direction) and "destructive abnormal events" (irreversible path, abrupt change in direction) from a dynamic perspective. Experiments show that this mechanism effectively reduces the false alarm rate in complex monitoring scenarios, making anomaly localization more physically reliable.

[0068] 2. Refined balancing of temporal structure modeling: Introducing a multi-scale temporal smooth response gating mechanism; Limitations of existing technologies: Existing methods often rely on single-point responses and lack effective constraints on the duration of abnormal events, resulting in frequent isolated noise pulses in the detection results and difficulty in characterizing the boundary stability of abnormal events.

[0069] This invention is innovative in that it proposes a method for modeling soft-persistent event evidence. By constructing a "phase transition potential field" and combining it with multi-scale local peak-valley structure analysis, a scale-adaptive soft-persistent response is defined. This mechanism not only measures the salience of the current fragment but also evaluates its structural salience within its local temporal neighborhood.

[0070] Technical Effects: This innovation acts as a "temporal structure filter," automatically suppressing transient noise interference that is not persistent and identifying true anomalous regions with stable dynamic structures. This significantly improves the temporal consistency of detection scores, making the anomalous curves output by the model smoother and more consistent with the true temporal sequence of events.

[0071] 3. Controlled stability of the fusion architecture: Design a lightweight auxiliary feature correction branch fusion mechanism; Existing technical shortcomings: Directly modifying the features of the backbone network or introducing complex branch fusion often leads to extreme instability in the early stages of model training, which can easily cause semantic separation between the normal and abnormal feature spaces.

[0072] The innovation of this invention lies in its design of an "auxiliary feature correction branch fusion" architecture. Topological evidence does not directly interfere with the feature extraction process; instead, it generates correction logic values ​​through a lightweight convolutional bypass, performing bounded corrections only on the backbone output at the decision layer. Simultaneously, zero-initialization constraints and a learnable bypass strength control scalar are introduced to achieve progressively controlled learning of the topological information incorporation process.

[0073] Technical Effects: This design, while maintaining the basic discriminative ability of the main detection path, introduces higher-order topological dynamics to supplement information. It ensures both rapid model convergence and the ability to progressively learn how to use topological evidence to correct classification bias, achieving a perfect balance between basic discriminative power and topological awareness.

[0074] 4. Robustness optimization of weakly supervised training paradigm: Propose phase-change topology transmission loss; Limitations of existing technologies: Traditional weakly supervised frameworks often employ a Top-k aggregation strategy, which is highly sensitive to the choice of k value and is easily misled by isolated high-scoring noise fragments within a bag, making it impossible to understand the regional distribution of anomalous events from a global perspective.

[0075] This invention is innovative in that it proposes a phase-change topology transmission loss, representing video as an "event metric" composed of candidate event regions. It utilizes entropy-regularized Sinkhorn iterations to calculate the Wasserstein transmission distance between anomalous and normal video metrics, and incorporates an adaptive weighting mechanism to significantly widen the statistical distance between them.

[0076] Technical Results: This innovation represents a leap from "single-point fragment comparison" to "structured event comparison." The transmission loss forces the model to focus on stable event regions rather than scattered noise, thus achieving more robust suppression of normal events and activation of abnormal events under weak supervision.

[0077] 5. Ultimate value for industrial deployment: a deep integration of high precision and lightweight design; Advantages of this invention: Thanks to the efficient topological feature calculation logic, the evidence branch and loss optimization module of this invention has extremely low parameter overhead (the core evidence branch has only about 0.5M parameters).

[0078] Technical Results: While maintaining state-of-the-art (SOTA) detection accuracy, the model file size is reduced by two orders of magnitude (approximately 2MB) compared to similar multimodal heavy-duty networks. This characteristic gives the algorithm strong potential for edge deployment, enabling it to reside in computing-constrained smart security cameras and achieve ultra-real-time anomaly monitoring with extremely low latency, making it highly valuable for commercial application and promotion.

[0079] 6. Intuitive physical interpretability: Provides geometric evidence for anomaly attribution; Advantages of this invention: The anomaly determination output by this invention is no longer a "black box" probability value, but is accompanied by geometric evidence with clear physical meaning, such as asymmetric action volume, curvature abrupt change, and density transition.

[0080] Technical benefits: By visualizing the phase transition topology curve, security personnel can intuitively observe the complete evolution of the system state from "equilibrium evolution" to "topological instability." This transparent determination mechanism provides scientific and physical support for post-event tracing of abnormal events and evidence extraction.

[0081] 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 non-equilibrium characteristic flow phase transition topology, characterized in that, The method includes the following steps: The original video stream is acquired, and segment-level features are extracted to obtain segment-level visual feature sequences. Each segment feature in the segment-level visual feature sequence is dimensionally compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence; In the low-dimensional latent space, the transition matrix of the feature manifold is constructed based on the feature sequence of the latent space, and the local drift field information on the feature manifold is estimated based on the transition matrix; Based on local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; A correction signal is generated based on geometric topological evidence. The correction signal is then superimposed onto the basic anomaly logic value of the segment level generated by the backbone network based on the segment-level visual feature sequence to obtain the corrected segment-level anomaly logic value. Anomaly scores are generated for the video based on the corrected segment-level anomaly logic values.

2. The weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology according to claim 1, characterized in that, The dimensionality of each segment feature in the fragment-level visual feature sequence is compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence. The specific steps include the following: The latent representation is obtained by projecting each segment feature of the segment-level visual feature sequence through a linear layer and a ReLU activation function; The hidden representation is masked and normalized to eliminate the distribution offset between different videos; The standardized features are mapped to a low-dimensional latent space using channel grouping averaging or one-dimensional adaptive average pooling operators to obtain the latent space feature sequence.

3. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 2, characterized in that, The transition matrix of the feature manifold is constructed based on the latent space feature sequence, and the local drift field information on the feature manifold is estimated based on the transition matrix. The specific steps include the following: Calculate the Euclidean distance between any two feature segments in a latent space feature sequence; Based on Euclidean distance, and combined with the local scale factor of each segment feature and the effective segment mask, a self-adjusting scale diffusion kernel is constructed. The local scale factor is defined as the mean distance from the current segment feature to the neighboring points. Row normalization of the diffusion kernel yields the transition matrix on the characteristic manifold; Based on the transition matrix, the mean local drift and the variance of diffusion at each time segment are estimated as information about the local drift field.

4. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 2, characterized in that, Geometric topological evidence includes evidence of trajectory curvature, evidence of action asymmetry, and evidence of fragment-level phase transitions. The step of extracting geometric topological evidence characterizing the degree of abrupt changes in characteristic hydrodynamics based on local drift field information further includes: Calculate the feature displacement based on the latent space characteristics of adjacent segments; Based on the drift mean and diffusion variance in the local drift field information, the positive and negative action quantities of the feature flow at the time step are calculated respectively, and evidence of action quantity asymmetry is generated based on the ratio of the two. Calculate the cosine of the angle between continuous feature displacement vectors to generate trajectory curvature evidence that measures the degree of curvature of the feature trajectory; By normalizing and fusing evidence of action asymmetry and trajectory curvature, fragment-level phase transition evidence is obtained.

5. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 2, characterized in that, Geometric topological evidence also includes logarithmic local density evidence, phase transition potential energy, persistence evidence, and event-gated signals. The step of extracting geometric topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics based on local drift field information further includes: Based on the local manifold density, construct logarithmic local density evidence; By combining evidence of fragment-level phase transitions with evidence of logarithmic local density, a phase transition potential energy is constructed. At multiple time window scales, a soft, persistent response factor is calculated based on phase transition potential energy to measure the degree of local protrusion and the intensity of fluctuations. The response factors at multiple scales are then fused to obtain evidence of persistence. A soft event-gated signal is constructed based on persistent evidence and phase transition potential energy, and average pooling is used to smooth the abnormal response to obtain the event-gated signal.

6. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 5, characterized in that, A correction signal is generated based on geometric topological evidence. This correction signal is then superimposed onto the basic anomaly logic value at the fragment level, which is generated by the backbone network based on fragment-level visual feature sequences, to obtain the corrected fragment-level anomaly logic value. The specific steps include the following: Construct a multidimensional evidence vector that includes fragment-level phase transition evidence, action quantity asymmetry evidence, trajectory curvature evidence, logarithmic local density evidence, phase transition potential energy, persistence evidence, event-gated signals, and their interaction terms; By inputting multidimensional evidence vectors into a lightweight one-dimensional temporal convolutional network, fragment-level bypass evidence logic values ​​are obtained. Obtain the backbone anomaly logic value generated by the backbone network based on fragment-level visual feature sequences; By using a learnable strength control scalar, the fragment-level bypass evidence logic value is superimposed onto the main trunk anomaly logic value in a bounded manner to obtain the corrected fragment-level anomaly logic value.

7. The weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology according to claim 6, characterized in that, The method also includes a weakly supervised training optimization step, the specific steps of which are as follows: Each video is divided into multiple consecutive candidate anomaly regions; For each candidate anomalous region, the average anomalous score, average phase transition intensity, and duration ratio are calculated, and the region persistence weight is calculated based on the region persistence. The average anomaly score, average phase transition intensity, duration ratio, and regional persistence weight are combined to construct event atoms that characterize the properties of candidate events. Each video is represented as a discrete event measure consisting of multiple event atoms and their corresponding region persistence weights; Construct the migration cost matrix between event atoms of abnormal videos and event atoms of normal videos, and use the entropy regularization Sinkhorn iterative algorithm to generate the optimal transmission plan. Calculate the Wasserstein distance between the abnormal video measure and the normal video measure based on the optimal transmission plan and the migration cost matrix; Based on the Wasserstein distance, a transmission separation loss function is constructed to increase the distance between abnormal and normal videos in the event measure space; The video-level classification loss, transmission separation loss, normal item suppression loss, and abnormal item activation loss are weighted together, and the weights of each loss item are dynamically adjusted through an adaptive weight allocation strategy to obtain the joint optimized total loss. Weakly supervised training optimization is achieved by minimizing the total loss of joint optimization.

8. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 7, characterized in that, The adaptive weight allocation strategy is specifically as follows: For each sub-loss function in the joint optimization total loss, its weight is determined by the ratio of the reciprocal of the stopping gradient value to the sum of the reciprocals of the stopping gradient values ​​of all sub-loss functions.

9. The weakly supervised video anomaly detection method based on non-equilibrium characteristic flow phase transition topology according to claim 8, characterized in that, The process of acquiring the raw video stream and extracting segment-level visual feature sequences involves the following steps: The input raw video stream is divided into multiple consecutive and non-overlapping segments; Using a pre-trained I3D or C3D visual feature extractor, visual features of each segment are extracted to form a segment-level visual feature sequence.

10. A weakly supervised video anomaly detection system based on a non-equilibrium characteristic flow phase transition topology, characterized in that, The system employs the weakly supervised video anomaly detection method based on non-equilibrium feature flow phase transition topology as described in any one of claims 1 to 9, and the system comprises: The feature extraction and dimensionality reduction module is used for: The original video stream is acquired, and segment-level features are extracted to obtain segment-level visual feature sequences. Each segment feature in the segment-level visual feature sequence is dimensionally compressed and mapped to a low-dimensional latent space to obtain the latent space feature sequence; The manifold construction and drift estimation module is used for: In the low-dimensional latent space, the transition matrix of the feature manifold is constructed based on the feature sequence of the latent space, and the local drift field information on the feature manifold is estimated based on the transition matrix; The geometric topology evidence extraction module is used for: Based on local drift field information, extract geometric and topological evidence characterizing the degree of abrupt changes in characteristic hemodynamics; The correction and rating generation module is used for: A corrected signal is generated based on geometric topological evidence. The corrected signal is then superimposed on the basic anomaly logic value of the segment level generated by the backbone network based on the segment-level visual feature sequence to obtain the corrected segment-level anomaly logic value.

Images