Method and model for identifying abnormal behavior under shielding

Abstract

The invention relates to a method and a model for identifying abnormal behaviors under occlusion. The method comprises the following steps: S110, acquiring an abnormal behavior segment containing an occlusion scene of a user; s120, based on the abnormal behavior segment containing the shielding scene, dynamic characteristic parameters are calculated, and the dynamic characteristic parameters comprise one or more of texture changes, shape changes or motion modes; s130, behavior recognition is carried out based on the dynamic characteristic parameters; and S140, the identification result comprises one of a normal behavior, an abnormal behavior or a specific abnormal behavior type. According to the method, the abnormal action of the shielded limb is indirectly deduced by analyzing the dynamic characteristics of the shielding object, and the problem of failure of a traditional method in a shielding scene is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior analysis, and in particular to a method and model for identifying abnormal behavior under occlusion. Background Art

[0002] Epilepsy is a sudden, repetitive, involuntary abnormal bodily movement. Epilepsy attacks are often accompanied by violent body movements. By identifying these behavioral manifestations, it is possible to intuitively determine whether a patient has epileptic seizure symptoms. The behavioral manifestations of epileptic seizures are often highly regular and characteristic. By analyzing the patient's movements and posture changes in video images, it is possible to identify abnormal behaviors in epileptic patients.

[0003] Traditional methods for monitoring epileptic behavior rely heavily on manual observation and video recording, with experts identifying characteristic behaviors during epileptic seizures. While epileptic behavior monitoring technology based on human posture estimation has made some progress in recent years, existing technologies still have some significant shortcomings in practical applications.

[0004] In most real-world scenarios, epileptic patients lie in bed, covered by quilts, sheets, or other obstructions. Traditional behavioral recognition technologies often fail to capture the patient's complete movements or state, resulting in reduced recognition accuracy. For example, certain body movements, facial expressions, or specific behaviors (such as convulsions or strenuous exercise) may be partially or completely obscured by obstructions. This loss of visual information makes it difficult for recognition algorithms to accurately determine whether a patient is experiencing an epileptic seizure. Summary of the Invention

[0005] In order to solve the above-mentioned defects, the present invention proposes a method and model for identifying abnormal behavior under occlusion.

[0006] The technical solution adopted by the present invention is a method for identifying abnormal behavior under occlusion, the method comprising:

[0007] S110, obtaining a user's abnormal behavior segment including an occlusion scene;

[0008] S120. Calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, where the dynamic feature parameters include one or more of texture change, shape change, or motion pattern;

[0009] S130, performing behavior recognition based on the dynamic feature parameters;

[0010] S140. The recognition result includes: normal behavior, abnormal behavior, or a specific abnormal behavior type.

[0011] Preferably,

[0012] The texture change includes LBP texture change, and the LBP texture change is used to quantify the change of surface texture;

[0013] The shape change includes a convex hull shape change, and the convex hull shape change is used to identify a sudden change in a boundary;

[0014] The motion pattern includes an optical flow motion vector, and the optical flow motion vector is used to represent the overall motion trend.

[0015] Preferably, the step S120 of calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene specifically includes:

[0016] S121, using an image segmentation algorithm to process the video frame of the abnormal behavior segment containing the occlusion scene, and segmenting the user and the occlusion area;

[0017] S122: Calculate dynamic feature parameters based on the segmented user and occluder areas.

[0018] Preferably, the S130 specifically includes:

[0019] The dynamic feature parameters are input into an abnormal behavior recognition model under occlusion to perform behavior recognition.

[0020] Preferably,

[0021] The image segmentation algorithm of S121 includes one of SAM2, Mask R-CNN or U-Net; and / or

[0022] The abnormal behavior recognition model under occlusion in S130 is a deep learning model trained based on the Transformer architecture.

[0023] Preferably, the training method of the abnormal behavior recognition model under occlusion in S130 includes:

[0024] Acquire an abnormal behavior segment containing an occlusion scene, and label the abnormal behavior segment with abnormal behavior / abnormal behavior type;

[0025] Calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, wherein the dynamic feature parameters include one or more of texture change, shape change, or motion pattern;

[0026] Constructing a time series feature set of the dynamic feature parameters;

[0027] A deep learning model is designed and trained using the temporal feature set so that the deep learning model can identify the mapping relationship between dynamic feature parameters and abnormal behaviors / abnormal behavior types, thereby obtaining the abnormal behavior recognition model under occlusion.

[0028] The present invention also provides a model for identifying abnormal behavior under occlusion, which is trained by the following method:

[0029] S210: Acquire an abnormal behavior segment containing an occlusion scene, and label the abnormal behavior segment with abnormal behavior / abnormal behavior type;

[0030] S220: Calculate dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, where the dynamic feature parameters include one or more of texture change, shape change, or motion pattern;

[0031] S230, constructing a time series feature set of the dynamic feature parameters;

[0032] S240. Design a deep learning model, and train the deep learning model using the temporal feature set so that the deep learning model can identify the mapping relationship between dynamic feature parameters and abnormal behaviors / abnormal behavior types, thereby obtaining the abnormal behavior recognition model under occlusion.

[0033] Preferably,

[0034] The texture change includes LBP texture change, and the LBP texture change is used to quantify the change of surface texture;

[0035] The shape change includes a convex hull shape change, and the convex hull shape change is used to identify a sudden change in a boundary;

[0036] The motion pattern includes an optical flow motion vector, and the optical flow motion vector is used to represent the overall motion trend.

[0037] Preferably, the dynamic feature parameters of S220 are obtained through a deep learning feature extractor.

[0038] Preferably, after S240, the method further includes: S250, optimizing the abnormal behavior recognition model under occlusion by using transfer learning.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The existing technology relies on directly capturing key points of the human body to judge behavior. When the patient is blocked by quilts, sheets, etc., the limb movement information is seriously lost, resulting in detection failure. The present invention indirectly infers the abnormal movement of the blocked limb by analyzing the dynamic characteristics of the blocking object (such as texture changes, optical flow motion, convex hull morphology changes, etc.). For example, when the patient convulses, it will cause high-frequency jitter of the quilt surface texture and sudden changes in the boundary convex hull area. Even if the limb is not visible, accurate detection can still be achieved through the physical deformation characteristics of the blocking object, solving the failure problem of traditional methods in occlusion scenarios.

[0041] 2. The existing technology only relies on the single-dimensional features of human posture and is easily affected by lighting, viewing angle, etc. The present invention integrates multi-dimensional parameters such as optical flow motion vectors of occluded areas (to capture the overall motion trend), LBP texture changes (to quantify subtle surface jitters), and convex hull shape analysis (to identify boundary mutations), and realizes the fusion analysis of temporal features through the Transformer model. This multimodal feature fusion mechanism significantly enhances the detection robustness in complex scenes (such as low light and partial occlusion) and reduces the false alarm rate and missed detection rate.

[0042] 3. To address the existing reliance on limited annotated data, this paper constructs a specialized temporal feature set for occlusion scenarios, encompassing a variety of occlusion types (quilts and sheets of varying thickness), abnormal behavior patterns (localized convulsions, generalized rigidity, etc.), and complex lighting conditions. This model also expands sample diversity through data augmentation (dynamic noise injection and simulated occlusion synthesis). Furthermore, the model supports transfer learning, enabling rapid adaptation to diverse hospital environments (e.g., differences in bed layouts), significantly improving cross-scenario generalization capabilities. Traditional solutions require re-annotating key point data in new scenarios, resulting in extremely high deployment costs.

[0043] 4. Existing pose estimation algorithms require high-precision keypoint calculations frame by frame, resulting in high computational complexity and difficulty processing long videos in real time. This paper uses the SAM2 image segmentation algorithm to rapidly locate the patient and occluded areas, narrowing the analysis scope to a specific ROI (Region of Interest) and reducing redundant computations. Furthermore, preprocessing such as grayscale conversion and Gaussian filtering reduces data dimensionality, and combining this with a lightweight Transformer model enables efficient time series modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention is described in detail below with reference to the embodiments and accompanying drawings, in which:

[0045] Figure 1 It is a flow chart of a method for identifying abnormal behavior under occlusion;

[0046] Figure 2 This is a flow chart of the training method for the abnormal behavior recognition model under occlusion

[0047] Figure 3 It is a schematic diagram of video data processing. DETAILED DESCRIPTION

[0048] To make the objectives, technical solutions, and advantages of the present invention more apparent, embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar components or components having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0049] The purpose of this invention is to provide an intelligent recognition method and model that can accurately identify abnormal patient behavior even in the presence of occlusion. By combining image segmentation, optical flow, and deep learning techniques, this method addresses the inability of traditional visual recognition techniques to accurately capture patient limb movements under occlusion, thereby improving the accuracy and reliability of abnormal behavior recognition.

[0050] In one embodiment, a method for identifying abnormal behavior under occlusion is provided. Figure 1 , the method comprising:

[0051] S110: Obtain a user's abnormal behavior segment containing an occlusion scene.

[0052] Specifically, clips of abnormal behavior involving occlusion can be provided directly by the user, or video data can be collected using a high-definition camera to ensure video quality suitable for subsequent analysis. Video clips showing abnormal patient behavior (such as convulsions and strenuous movements), particularly those with body occlusion, are selected as the dataset for the subsequent recognition model.

[0053] Furthermore, the selected video clips can be subjected to image preprocessing operations such as denoising, brightness adjustment, and contrast enhancement to improve the accuracy of subsequent analysis.

[0054] S120: Calculate dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene. The dynamic feature parameters include texture changes or shape changes. These parameters reflect changes caused by human behavior in the occlusion scene from different perspectives. When a person is occluded, the texture of the covering object is relatively more representative of the person's posture, while the shape change can better reflect the change in the person's posture.

[0055] Specifically, texture changes can include LBP (local binary pattern) texture changes, which are used to quantify changes in surface texture (such as subtle jitter). During an epileptic seizure, the patient's limb twitching can cause rapid changes in the surface texture of the occluded object. This change can be captured by calculating the difference in LBP features between adjacent frames and quantifying the degree of texture change.

[0056] Shape changes can include changes in the convex hull shape, which is used to identify sudden changes in boundaries. Abnormal behaviors (such as twitching or limb extension) may cause obvious changes in the convex hull of the quilt area, such as irregular sudden changes in boundaries or sudden increases in area.

[0057] Furthermore, dynamic feature parameters can also include motion patterns. These motion patterns can include optical flow motion vectors, which are used to characterize overall motion trends. Optical flow methods can be used to analyze pixel displacements between adjacent frames to generate a motion vector field that describes the direction and speed of movement of the patient and its occluders.

[0058] Furthermore, the dynamic feature parameters may also include other similar parameters.

[0059] S130: Perform behavior recognition based on the dynamic feature parameters.

[0060] Specifically, a threshold value of the dynamic feature parameter may be set, and behavior recognition may be performed by comparing the dynamic feature parameter calculated in S120 with the corresponding threshold value.

[0061] The dynamic feature parameters can also be input into an abnormal behavior recognition model under occlusion for behavior recognition, such as the abnormal behavior recognition model under occlusion disclosed in the following embodiments. The use of advanced deep learning models improves the accuracy of posture estimation and abnormal behavior detection under non-ideal video conditions.

[0062] S140. The recognition result includes: normal behavior, abnormal behavior, or a specific abnormal behavior type.

[0063] When the shape, texture or optical flow of the identified area has no significant changes, the recognition result is normal behavior; when the shape, texture or optical flow of the identified area has significant changes, the recognition result is abnormal behavior or a specific abnormal behavior type (such as convulsions, violent movements, etc.), and an alarm can be further issued.

[0064] Existing technologies rely on directly capturing key points of the human body to make behavioral judgments. When the patient is obscured by quilts, sheets, etc., the limb movement information is severely lost, resulting in detection failure. In scenarios where the patient's limbs are obscured (such as covered by quilts or sheets), this embodiment bypasses the limitation of the lack of direct visual information and uses the dynamic changes of the surface of the obstruction to indirectly infer the patient's abnormal behavior. By analyzing the dynamic characteristics of the obstruction (such as texture changes, optical flow motion, convex hull morphology changes, etc.), the abnormal movements of the obscured limbs are indirectly inferred. For example, when the patient convulses, it will cause high-frequency jitter of the quilt surface texture and sudden changes in the boundary convex hull area. Even if the limbs are not visible, accurate detection can still be achieved through the physical deformation characteristics of the obstruction, solving the failure problem of traditional methods in occlusion scenarios.

[0065] In one embodiment, calculating the dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene in S120 specifically includes:

[0066] S121, using an image segmentation algorithm to process the video frame of the abnormal behavior segment containing the occlusion scene, and segmenting the user and occlusion areas, see Figure 3 .

[0067] Specifically, this embodiment takes into account that there may be multiple objects and people in the video scene, which may have a significant impact on the accuracy of the model's prediction. Therefore, an image segmentation algorithm is used to segment out the area where the patient is located (including the patient's occlusions, such as quilts, etc.). Furthermore, the image segmentation algorithm may include SAM2, Mask R-CNN or U-Net, etc. Among them, SAM2 has a strong zero-sample generalization capability and can segment unseen visual content without additional training for specific objects. By introducing the "memory module" and "occlusion head" model, SAM2 can predict the subsequent behavior of the occluded object and maintain high-precision segmentation even in dynamic scenes. For example, when the target is briefly occluded, the model can still continue to track based on historical frame information, significantly reducing the number of tracking losses.

[0068] S122: Calculate dynamic feature parameters based on the segmented user and occluder areas.

[0069] This embodiment provides a method for identifying abnormal behavior under occlusion based on image segmentation. By performing a series of parameter calculations on the segmented areas, such as the shape change, optical flow change, texture change, and change frequency of the area, image information such as the texture change, contour, and shape change of the occlusion can be extracted, thereby predicting the behavior of people under the occlusion, which helps to solve the problem of being unable to recognize human posture under occlusion.

[0070] In one embodiment, the abnormal behavior recognition model in S130 under occlusion is a deep learning model trained based on the Transformer architecture. The Transformer architecture incorporates a self-attention mechanism that processes all positions in the input sequence in parallel, significantly improving training and inference efficiency. Furthermore, the Transformer architecture utilizes information from all positions in the sequence when calculating the representation of each position, leveraging global information modeling to handle complex contextual relationships and excel at processing long videos.

[0071] Existing technologies rely solely on single-dimensional features of human posture and are susceptible to interference from factors such as lighting and viewing angles. This embodiment integrates multi-dimensional parameters such as optical flow motion vectors (to capture overall motion trends), LBP texture changes (to quantify subtle surface jitter), and convex hull shape analysis (to identify boundary mutations) in occluded areas, and implements fusion analysis of temporal features through the Transformer model. This multimodal feature fusion mechanism significantly enhances detection robustness in complex scenarios (such as low light and partial occlusion) and reduces false alarm and missed detection rates.

[0072] In other embodiments, a deep learning model trained with other classification models may also be used to obtain the abnormal behavior recognition model under occlusion in S130.

[0073] In one embodiment, the present invention also discloses a model for identifying abnormal behavior under occlusion. Figure 2 The recognition model can be applied to step S130 in the above embodiment, and the model is trained by the following method:

[0074] S210: Acquire an abnormal behavior segment containing an occlusion scene, and label the abnormal behavior segment with abnormal behavior / abnormal behavior type.

[0075] Specifically, abnormal behavior clips involving occlusions can be directly retrieved from a video library. Alternatively, high-definition cameras can be used to capture real-time patient video data, filtering out clips of patients exhibiting abnormal behavior, particularly those with body occlusions. These clips are manually annotated, including the type of abnormal behavior, such as convulsions, strenuous movements, and generalized rigidity. The annotations can also include the type of obstructing object, such as a blanket or bed sheet. The annotations can also include lighting conditions. The selected video clips undergo image preprocessing, including denoising, brightness adjustment, and contrast enhancement, to improve the accuracy of subsequent analysis.

[0076] S220: Calculate dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, where the dynamic feature parameters include one or more of texture change, shape change, or motion pattern.

[0077] Specifically, an image segmentation algorithm such as SAM2 can be used to segment the video frame first, extracting the area where the patient is located, including the patient and its obstructions (such as quilts, sheets, etc.), and further extracting the contour and shape information of the obstructions to provide basic data for subsequent parameter calculations.

[0078] By analyzing the image information in the video frames, key parameters reflecting the patient's behavioral characteristics are extracted. These parameters, including the patient's and their occluders' motion patterns, texture changes, and shape changes, are used for subsequent deep learning model training and abnormal behavior detection. Furthermore, the segmented patient and occluder regions can be grayscaled, converting the color image to a grayscale image to reduce computational complexity. Denoising processing (such as Gaussian filtering) is then performed on the grayscale image to eliminate noise interference.

[0079] Furthermore, during an epileptic seizure, the twitching of the patient's limbs can cause rapid changes in the surface texture of the occluder. This change can be captured by calculating the difference in LBP features between adjacent frames and quantifying the degree of texture change. That is, texture changes can include LBP texture changes, which are used to quantify changes in surface texture. LBP (Local Binary Pattern) or improved algorithms (such as LBP++) can be used to calculate the texture differences of the occluder surface between adjacent frames and quantify subtle surface jitters (such as high-frequency texture changes caused by quilt twitching).

[0080] Shape changes can include changes in the convex hull shape, which is used to identify sudden changes in boundaries. A convex hull algorithm (such as the Graham scan method) is used to calculate the convex hull area, perimeter, and vertex count changes in the occluded area to capture sudden changes in boundaries (such as a localized bulge in a quilt caused by limb twitching).

[0081] The motion pattern can include optical flow motion vectors, which are used to characterize the overall motion trend. Optical flow methods, such as Lucas-Kanade, can be used to analyze the displacement of pixels between adjacent frames to generate a motion vector field that describes the direction and speed of movement of the patient and its occluders.

[0082] S230: Construct a time series feature set of the dynamic feature parameters.

[0083] Specifically, the calculated dynamic feature parameters are combined with the temporal correlation of the video frame sequence to construct a temporal feature set. This not only considers the features of a single frame, but also focuses on the changing trends and dependencies of features over time. For example, the variation patterns of optical flow vectors across multiple consecutive frames, as well as the increase and decrease of the convex hull area over time, are analyzed. This time series information is then integrated to form a temporal feature set, providing the model with richer dynamic information.

[0084] S240. Design a deep learning model, and train the deep learning model using the temporal feature set so that the deep learning model can identify the mapping relationship between dynamic feature parameters and abnormal behaviors / abnormal behavior types, thereby obtaining the abnormal behavior recognition model under occlusion.

[0085] Specifically, the constructed time series feature set is used as input to train the deep learning model. During training, the model continuously learns the mapping relationship between dynamic feature parameters and abnormal behaviors / types, adjusting the model parameters to minimize the discrepancy between the predicted results and the annotated labels. After training on a large amount of data, the model can accurately identify the corresponding abnormal behaviors or types based on the input dynamic feature parameters, ultimately achieving a model for identifying abnormal behaviors under occlusion.

[0086] By mapping these dynamic feature parameters to the patient's behavior, we can infer whether the patient has abnormal body movements or behaviors. For example, we can map texture changes to the patient's behavior patterns. High-frequency texture changes may correspond to the patient's twitching behavior, while low-frequency texture changes may indicate the patient is in a state of rest.

[0087] To address the problem that existing technologies rely on limited annotated data, this embodiment constructs a professional occlusion scene temporal feature set, which includes various types of occlusion objects, such as quilts and sheets of different thicknesses; abnormal behavior patterns, such as local convulsions and whole-body rigidity, and complex lighting conditions. It can also expand sample diversity through data enhancement, such as dynamic noise injection and simulated occlusion synthesis.

[0088] In one embodiment, the Transformer architecture is used to train the above-mentioned deep learning model to predict whether a patient has abnormal behavior. The Transformer model can process sequence data and is suitable for analyzing the temporal relationship between video frames. The above-mentioned extracted parameters, such as optical flow changes, texture changes, shape changes, etc., and the labels of the video are used as inputs to the model. Through training with a large amount of annotated data, the model can accurately identify abnormal behavior under occlusion conditions. Furthermore, cross-validation, data enhancement and other technologies can also be used to optimize model performance and improve the generalization ability of the model in different scenarios.

[0089] In one embodiment, the dynamic feature parameters of S220 can also be obtained through a deep learning feature extractor. In addition to optical flow and LBP texture analysis, deep learning feature extractors, such as pre-trained ResNet, can be used to extract spatiotemporal features of occluded areas. Frequency domain analysis, such as Fourier transform, can also be combined to capture periodic twitch frequency, enhancing behavioral characterization capabilities.

[0090] In one embodiment, after S240, the following further includes: S250, optimizing the abnormal behavior recognition model under occlusion using transfer learning. The model supports transfer learning and can quickly adapt to different environments (such as different bed layouts), significantly improving cross-scenario generalization capabilities. There is no need to re-label key point data in new scenarios, reducing deployment costs.

[0091] The abnormal behavior recognition method and model under occlusion described in the above embodiment can be used to identify epileptic seizures in confirmed patients, as well as sleep disorders in people with sleep disorders and infant care. For example, they can identify people turning over under a blanket in a sleeping scene, or violent kicking and beating in sleep behavior disorders.

[0092] The present invention also provides a system for identifying abnormal behavior under occlusion, comprising:

[0093] A video acquisition module configured to acquire user video data in real time;

[0094] An image segmentation module configured to segment user and occluder regions;

[0095] A feature calculation module configured to extract optical flow, LBP texture and convex hull shape parameters;

[0096] a behavior recognition module configured to analyze parameters through a deep learning model and output detection results;

[0097] An alert module, configured to issue alerts when abnormal behavior is detected.

[0098] Specifically, the hardware configuration can be a server equipped with 4 NVIDIA A100s and 1 camera with night vision function.

[0099] In this specification, the use of terms such as "Embodiment 1," "this embodiment," and "in one embodiment" indicates that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in the invention or at least one embodiment or example of the invention. In this specification, schematic representations 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 appropriately combined in any one or more embodiments or examples.

[0100] In the description of this specification, the terms "connect," "install," "fix," "dispose," and "have" are to be understood in a broad sense. For example, "connect" can mean a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.

[0101] In the description of this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.

[0102] The above description of the embodiments is to facilitate ordinary technicians in this technical field to understand and apply the technology of this case. People familiar with the technology in this field can obviously make various modifications to these examples easily and apply the general principles described here to other embodiments without having to go through creative work. Therefore, this case is not limited to the above embodiments. Modifications to the following situations should all be within the scope of protection of this case: ① A new technical solution implemented based on the technical solution of the present invention and combined with existing common knowledge, the technical effect produced by the new technical solution does not exceed the technical effect of the present invention; ② The equivalent replacement of some features of the technical solution of the present invention with the known technology, the technical effect produced is the same as the technical effect of the present invention; ③ The technical solution of the present invention is expandable, and the substantive content of the expanded technical solution does not exceed the technical solution of the present invention; ④ The equivalent transformation made by the content of the description and drawings of the present invention is directly or indirectly applied to other related technical fields.

Claims

1. A method for identifying abnormal behavior under occlusion, characterized in that: The method comprises: S110, obtaining a user's abnormal behavior segment including an occlusion scene; S120. Calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, where the dynamic feature parameters include one or more of texture change, shape change, or motion pattern; S130, performing behavior recognition based on the dynamic feature parameters; S140. The recognition result includes: normal behavior, abnormal behavior, or a specific abnormal behavior type.

2. The identification method according to claim 1, characterized in that The texture change includes LBP texture change, and the LBP texture change is used to quantify the change of surface texture; The shape change includes a convex hull shape change, and the convex hull shape change is used to identify a sudden change in a boundary; The motion pattern includes an optical flow motion vector, and the optical flow motion vector is used to represent the overall motion trend.

3. The identification method according to claim 2, characterized in that The step S120 of calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene specifically includes: S121, using an image segmentation algorithm to process the video frame of the abnormal behavior segment containing the occlusion scene, and segmenting the user and the occlusion area; S122: Calculate dynamic feature parameters based on the segmented user and occluder areas.

4. The identification method according to claim 3, characterized in that The S130 specifically includes: The dynamic feature parameters are input into an abnormal behavior recognition model under occlusion to perform behavior recognition.

5. The identification method according to claim 4, characterized in that: The image segmentation algorithm of S121 includes one of SAM2, Mask R-CNN or U-Net; and / or The abnormal behavior recognition model under occlusion in S130 is a deep learning model trained based on the Transformer architecture.

6. The identification method according to claim 4, characterized in that: The training method of the abnormal behavior recognition model under occlusion in S130 includes: Acquire an abnormal behavior segment containing an occlusion scene, and label the abnormal behavior segment with abnormal behavior / abnormal behavior type; Calculating dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, wherein the dynamic feature parameters include one or more of texture change, shape change, or motion pattern; Constructing a time series feature set of the dynamic feature parameters; A deep learning model is designed and trained using the temporal feature set so that the deep learning model can identify the mapping relationship between dynamic feature parameters and abnormal behaviors / abnormal behavior types, thereby obtaining the abnormal behavior recognition model under occlusion.

7. A model for identifying abnormal behavior under occlusion, characterized by: The model is trained by the following method: S210: Acquire an abnormal behavior segment containing an occlusion scene, and label the abnormal behavior segment with abnormal behavior / abnormal behavior type; S220: Calculate dynamic feature parameters based on the abnormal behavior segment containing the occlusion scene, where the dynamic feature parameters include one or more of texture change, shape change, or motion pattern; S230, constructing a time series feature set of the dynamic feature parameters; S240. Design a deep learning model, and train the deep learning model using the temporal feature set so that the deep learning model can identify the mapping relationship between dynamic feature parameters and abnormal behaviors / abnormal behavior types, thereby obtaining the abnormal behavior recognition model under occlusion.

8. The recognition model according to claim 7, characterized in that The texture change includes LBP texture change, and the LBP texture change is used to quantify the change of surface texture; The shape change includes a convex hull shape change, and the convex hull shape change is used to identify a sudden change in a boundary; The motion pattern includes an optical flow motion vector, and the optical flow motion vector is used to represent the overall motion trend.

9. The recognition model according to claim 7 or 8, characterized in that The dynamic feature parameters of S220 are obtained through a deep learning feature extractor.

10. The recognition model according to claim 7 or 8, characterized in that After S240, the method further includes: S250, optimizing the abnormal behavior recognition model under occlusion by using transfer learning.

Images