A visual-based monitoring video anomaly identification method

Abstract

The application belongs to the technical field of anomaly identification, and particularly relates to a monitoring video anomaly identification method based on vision. The method collects hospital logistics area monitoring video frames, and after spatiotemporal joint interference perception preprocessing, an improved instance segmentation algorithm is introduced, which introduces a frequency perception channel attention, a deformable convolution and a boundary perception mask refining, to accurately locate the device display interface ROI; a lightweight industrial OCR is used to combine multi-frame time sequence verification to extract operation parameters and construct a time sequence; an empirical mode decomposition is used to separate trend and mutation characteristics to generate a comprehensive anomaly score; a new parameter extraction failure reason analysis and shielding personnel identity verification mechanism is added. The application can realize all-weather non-contact monitoring, simultaneously identify long-term gradual failure and instantaneous anomaly, greatly reduce the false detection and missed detection rate, and effectively improve the hospital logistics equipment operation and maintenance efficiency and fault early warning capability.

Description

Technical Field

[0001] This invention belongs to the field of anomaly recognition technology, and particularly relates to a vision-based method for anomaly recognition in surveillance videos. Background Technology

[0002] The HVAC, water, and electricity systems in hospital logistics areas are crucial for ensuring the continuous and stable operation of hospital services. Traditional logistics equipment maintenance relies primarily on regular manual inspections, on-site reading of instrument parameters, and manual recording. This approach suffers from inefficiency, delayed response times, and the risk of human error leading to missed inspections and misjudgments, making it difficult to achieve real-time monitoring and early warning of equipment anomalies. While hospital logistics departments have widely deployed video surveillance systems, these are mostly used for post-event tracing and lack automated visual analysis of equipment display parameters and intelligent anomaly identification capabilities. The core application value of video data remains largely untapped. Summary of the Invention

[0003] In view of the technical problems existing in the background art, the present invention proposes a vision-based method for anomaly identification in surveillance videos.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] S1. Collect real-time monitoring video images from monitoring equipment in various areas of the hospital's logistics department and obtain the original video frame sequence;

[0006] S2. Perform frame-by-frame preprocessing on the original video frame sequence, and for each processed video image, detect and locate all regions of interest (ROI) images in the video frame using an improved instance segmentation algorithm.

[0007] S3. Extract the operating parameter data from the device display interface one by one for the ROI region, arrange the operating parameter data extracted from multiple consecutive frames in time sequence, and construct a device operating data sequence of consecutive time frames.

[0008] S4. Extract features from the continuously running data sequence, analyze the trend features and mutation features, and generate anomaly scores based on the trend features and mutation features;

[0009] S5. Based on the anomaly score, identify anomalies in the equipment operating parameters and obtain the anomaly identification results.

[0010] Preferably, the specific implementation of frame-by-frame preprocessing of the original video frame sequence in step S2 is as follows:

[0011] S21. Construct a spatiotemporal joint interference perception model for the input raw video frames. Where t is the image frame number, The pixel coordinates are used as input, and then multi-scale decomposition is performed to obtain the base layer image. and detail layer images The decomposition formula is: ;

[0012] S22. Based on the temporal consistency constraint of consecutive K frames, perform static interference detection and repair, and calculate the temporal variance of each pixel in the base layer image and detail layer image respectively. The formula is: ,in, Let H be the average pixel value of K consecutive frames, where H = 1 and 2 represent the base layer image and detail layer image, respectively. Regions whose variance is less than a preset temporal variance threshold are identified as static interference regions. These regions are repaired using Poisson fusion based on the normal pixel distribution in the neighborhood. The repaired image is then the static interference-free base layer image. and static disturbance detail layer image ;

[0013] S23, will As a component of light And build based on Adaptive gain function of local standard deviation The formula is: ,in, To remove static interference detail layer Local standard deviation within the window These are the global maximum and minimum values, respectively. These are the maximum and minimum values ​​of the adaptive gain function, respectively.

[0014] S24. Perform global illumination normalization processing to obtain the preprocessed image: ,in, for and The resulting image.

[0015] Preferably, in step S2, the specific implementation of detecting and locating all Regions of Interest (ROIs) in each processed video frame using an improved instance segmentation algorithm is as follows:

[0016] S25. Transfer the preprocessed video frame images Input the backbone feature extraction network to extract a set of multi-scale feature maps; construct a standard feature pyramid through a top-down feature fusion method; apply frequency-aware channel attention to each pyramid feature layer, and generate channel attention weights based on the high-frequency energy ratio of the feature maps to obtain the feature pyramid after display area enhancement.

[0017] S26. At each layer of the enhanced feature pyramid, scale-adaptive anchor boxes are used to generate initial candidate regions; deformable convolution sampling is performed on the features of each candidate position to achieve geometric feature alignment for display panels of different shapes and tilt angles; an additional display panel category confidence branch is introduced in the region proposal stage to perform comprehensive scoring and sorting of candidate boxes and non-maximum suppression to obtain a set of selected candidate boxes.

[0018] S27. Extract dual-scale fusion features for each selected candidate box, and generate multi-granularity mask output through a fully convolutional mask prediction head; first predict the coarse-grained instance mask, then construct the boundary-aware refinement branch, extract boundary-sensitive features and generate the mask boundary response map; perform multiple rounds of iterative refinement on the coarse-grained mask based on the boundary response to obtain the final instance mask with accurate edges;

[0019] S28. Perform geometric integrity verification of the display area on all detected instances in sequence, remove false detection areas, and obtain the final set of regions of interest (ROIs).

[0020] Preferably, step S3 extracts the operating parameter data from the device display interface of the ROI region one by one, and arranges the operating parameter data extracted from multiple consecutive frames in chronological order to construct a sequence of device operating data for consecutive time frames. The specific implementation of this step is as follows:

[0021] S31. Build a lightweight industrial digital OCR recognition model with a built-in character template library; in the single-frame spatial dimension, calculate the similarity between the information recognized by the OCR recognition model and the standard character template, and select the character with the highest similarity as the recognition result;

[0022] S32. Verify the recognition results. If the character recognition results are the same for 5 consecutive frames, they are determined to be valid recognition results.

[0023] S33. Arrange the data of the effective identification results linearly in chronological order to construct a time series data sequence of time-continuous operating parameters.

[0024] As a preferred method, the process of extracting equipment operating parameters is monitored in real time to determine whether the duration of continuous extraction failures exceeds a preset threshold. If the duration of continuous extraction failures exceeds the preset threshold, key video frames within the extraction failure period are located and the reasons for parameter extraction failures are analyzed. If the extraction failure is determined to be caused by personnel obstruction, the behavior analysis of the obstructing personnel is initiated to determine whether it is abnormal. If it is determined to be non-personal obstruction, an abnormal warning is issued directly.

[0025] As a preferred approach, analyze the reasons for parameter extraction failure; if the failure is determined to be due to personnel occlusion, then initiate occlusion behavior analysis to determine whether it is an anomaly. The specific implementation is as follows:

[0026] For key video frames within a continuous extraction failure period, a human detection model is used to detect whether there are human targets in the frames; the bounding boxes of the detected human targets are calculated and determined to be extraction failures caused by human occlusion.

[0027] Extract identity features from obscured personnel, including uniform color, name tag area detection features, and facial features. Match the extracted features with a pre-set authorized personnel feature database to determine whether the person is an authorized personnel.

[0028] If the user is determined to be an authorized person, the operation is considered normal maintenance and the operation log is recorded; if the user is determined to be an unauthorized person, an abnormal event record is generated and an abnormal warning is triggered.

[0029] Preferably, step S4 involves feature extraction from the continuously running data sequence, analyzing trend characteristics and mutation characteristics, and generating anomaly scores based on the trend characteristics and mutation characteristics. The specific implementation of this step is as follows:

[0030] S41. Equipment operation data sequence Dynamic normalization is performed using a sliding time window of length W to obtain a normalized data sequence. ;

[0031] S42. Perform Empirical Mode Decomposition (EMD) on the normalized data sequence to obtain n Intrinsic Mode Functions (IMFs) and 1 Residual Component. ,satisfy: ,in, Let j be the value of the j-th intrinsic mode function at time t; calculate the energy of each intrinsic mode function; adaptively determine the trend abrupt change grouping boundary based on the cumulative energy ratio: ,in, The preset energy percentage threshold is, where, Let j be the energy value of the j-th intrinsic mode;

[0032] S43, Residual component and previous The superposition of low-frequency IMF components is used as a trend feature sequence. This characterizes the long-term trend of equipment operating parameters; the remaining... The superposition of several high-frequency IMF components serves as a mutation characteristic sequence. Characterizing the instantaneous fluctuations and abrupt changes of parameters: , ;

[0033] S44. Calculate the slope of the trend feature sequence within the current window and compare it with the slope of the historical normal baseline to generate a trend anomaly score; use the sliding window standard deviation ratio algorithm to detect the mutation feature sequence, calculate the ratio of the standard deviation of the current window to the standard deviation of the previous window, and determine that a mutation has occurred when the ratio is greater than a set threshold, and generate a mutation anomaly score.

[0034] S45. The mutation anomaly score and the trend anomaly score are weighted and summed according to the preset weight coefficients to generate the final anomaly score.

[0035] Compared with existing technologies, the advantages and positive effects of this invention are as follows: First, it adopts spatiotemporal joint interference perception preprocessing, effectively eliminating environmental interference such as complex lighting and static dust reflection in hospitals through multi-scale decomposition and temporal consistency repair; second, it improves the instance segmentation algorithm by introducing frequency-aware channel attention, deformable convolution, and boundary-aware mask refinement to accurately adapt to display panels of devices with different shapes and tilt angles; third, it combines lightweight template matching OCR with multi-frame temporal verification to achieve low-latency and high-accuracy extraction of operating parameters at the edge; and fourth, it uses trend-abrupt dual-dimensional anomaly assessment based on empirical mode decomposition to simultaneously identify long-term gradual faults and instantaneous anomalies, which, combined with the occlusion personnel identity verification mechanism, significantly reduces the false detection and missed detection rates. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating a vision-based method for anomaly recognition in surveillance videos. Detailed Implementation

[0038] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0039] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0040] This embodiment provides a vision-based method for anomaly identification in surveillance videos. The specific implementation process is as follows: Figure 1As shown, this method accurately locates the ROI region of the device display interface through an improved instance segmentation algorithm, combines lightweight industrial digital OCR technology to automate the extraction of operating parameters, and separates the operating data sequence into trend features and abrupt change features based on empirical mode decomposition, thereby achieving multi-dimensional anomaly assessment of the equipment's operating status. This method can monitor the operating status of key hospital logistics equipment 24 / 7 without contact, effectively improving operation and maintenance efficiency and reducing the risk of equipment failure.

[0041] Real-time monitoring video images from surveillance equipment in various areas of the hospital's logistics system are acquired to obtain raw video frame sequences. Specifically, network-connected high-definition bullet cameras are deployed in all key equipment rooms of the hospital's logistics system (including HVAC rooms, pump rooms, power distribution rooms, medical gas stations, etc.). The cameras use H.265 encoding format to capture real-time video streams with a resolution of 1920×1080, and transmit them to the corresponding edge computing nodes via the hospital's gigabit LAN using the RTSP protocol. The video acquisition module of the edge computing node continuously listens to the designated RTSP stream address, establishes a TCP connection, receives video data packets, and performs decoding. The decoding module extracts keyframes from the continuous video stream at fixed time intervals to generate the raw video frame sequence. To address potential frame loss and out-of-order issues during network transmission...

[0042] The original video frame sequence is preprocessed frame by frame, and for each processed video image, all regions of interest (ROIs) in the video frame are detected and located using an improved instance segmentation algorithm.

[0043] Specifically, a spatiotemporal joint interference perception model is constructed for the input raw video frames. Where t is the image frame number, The pixel coordinates are used as input, and then multi-scale decomposition is performed to obtain the base layer image. and detail layer images The decomposition formula is: Based on the temporal consistency constraint of consecutive K frames, static interference is detected and repaired by calculating the temporal variance of each pixel in the base layer image and detail layer image respectively. The formula is: ,in, Let H be the average pixel value of K consecutive frames, where H = 1 and 2 represent the base layer image and detail layer image, respectively. Regions whose variance is less than a preset temporal variance threshold are identified as static interference regions. These regions are repaired using Poisson fusion based on the normal pixel distribution in the neighborhood. The repaired image is then the static interference-free base layer image. and static disturbance detail layer image ;Will As a component of light And build based on Adaptive gain function of local standard deviation The formula is: ,in, To remove static interference detail layer Local standard deviation within the window These are the global maximum and minimum values, respectively. These represent the maximum and minimum values ​​of the adaptive gain function, respectively; global illumination normalization is then performed to obtain the preprocessed image: ,in, for and The resulting image.

[0044] Furthermore, for each processed video frame, an improved instance segmentation algorithm is used to detect and locate all Regions of Interest (ROIs) within the video frame. The specific implementation of this is as follows: [The text abruptly shifts to a different topic] ...the obtained preprocessed frame image... The improved backbone feature extraction network is input and passed through L-level downsampling convolutions to obtain a multi-scale feature map set. Where L=5; for adjacent layer feature maps and Perform top-down feature fusion to construct a standard feature pyramid. The fusion formula is: ,in, For channel-aligned convolution, To perform bilinear upsampling, a display region frequency response enhancement module is introduced, which enhances the frequency response of the display region in each pyramid feature layer. Apply frequency-aware channel attention to the frequency-aware channel. Perform a two-dimensional discrete cosine transform along the channel dimension to extract frequency domain feature representations. ,in Let be the spatial frequency coordinates, and 'c' be the channel index. The high-frequency energy ratio is defined in the frequency domain. This is used to measure the proportion of high-frequency components in the feature map of this layer, and the formula is: Where H is a preset high-frequency region mask, and the frequency coordinates satisfy... For regions exceeding a preset threshold, F represents the full frequency domain. To prevent extremely small constants with a denominator of 0, frequency channel attention weights are constructed based on the high-frequency energy ratio. The calculation formula is: ,in, This is the frequency domain feature map of the c-th channel. These are frequency-domain global average pooling and global max pooling, respectively. For learnable weights, For learnable bias, The Sigmoid activation function is used. Frequency channel attention weights are applied to the spatial domain pyramid feature, resulting in a feature pyramid with enhanced display area. The calculation method is as follows: .

[0045] In the feature enhancement pyramid On each layer, candidate region proposals are generated based on dense anchor frames; considering the large differences in display panel sizes of equipment in hospital logistics scenarios, a set of anchor frame sizes for each layer is defined. This assigns different pyramid levels the responsibility for detection across different scale ranges, as shown in the formula: ,in, Based on the area of ​​the anchor frame, is the intra-layer scale factor. Features at each candidate location are sampled using deformable convolution to adapt to the irregular shape and tilt angle of the device display panel; a standard is set. The regular sampling grid of the convolution kernel is: The output features of deformable convolution for: ,in, For convolution weights, The learning offset for the nth sampling point is obtained by regressing from the current feature map using an additional convolutional layer: An additional confidence branch is introduced during the region proposal stage, for each candidate box. The confidence score is obtained by regression analysis of its corresponding features: ,in, These are fixed-dimensional candidate box features extracted through deformable convolution alignment followed by the RoI Align operation. For panel classification branch parameters, sort them according to the weighted comprehensive ranking of region proposal score and panel confidence score, select the first preset number of candidate boxes, and apply non-maximum suppression with an IoU threshold of 0.7 to remove redundant boxes, and output a set of selected candidate boxes.

[0046] For each selected candidate box The RoI Align operation extracts dual-scale features from two adjacent levels of the feature pyramid, concatenates them, and then generates multi-granularity mask outputs via a fully convolutional mask prediction head; the coarse-grained mask branch... Resolution prediction coarse-grained instance mask: The number of output channels equals the number of target categories, and the foreground probability is predicted independently for each category; a boundary-aware refinement branch is constructed to upsample the coarse-grained mask features to... Then, it is fused with pyramid features of the same resolution through element-wise multiplication to obtain enhanced boundary-sensitive features. The formula is: ,in, Candidate boxes The corresponding optimal pyramid level, This is an element-wise multiplication. The Laplacian operator is applied to the boundary-sensitive features to extract the mask boundary response map (BI), as shown in the formula: ,in, For the discrete Laplace operator, The boundary enhancement gain coefficient is used. The coarse-grained mask is refined in two rounds. In each round, the current mask and the boundary response map are concatenated and input into a lightweight refinement convolutional block, outputting an updated mask. The formula is as follows: , ,in, This is the boundary response map recalculated based on the current mask in the T-th iteration. Initialize the mask residual correction amount. After two rounds of iteration, a refined mask is obtained, which is then binarized to output the final instance mask.

[0047] Finally, a geometric integrity verification of the display area is performed. Based on the prior characteristic that the display panels of hospital logistics equipment are mostly regular rectangles, the convex hull area ratio and minimum bounding rectangle fit of each instance mask are calculated. Instances with both indicators exceeding the preset threshold are retained, while false detection targets with irregular shapes such as wall stains, equipment shells, and cables are removed. Next, the validity of the display content is verified. The edge density within the mask area is calculated, and the gradient significance threshold is used to determine whether the area contains valid display content such as numbers, scale lines, and indicator lights. Blank panels without content and non-display areas with overly cluttered textures are excluded.

[0048] Then, the operating parameter data from the device display interface is extracted one by one for each ROI region. The operating parameter data extracted from multiple consecutive frames are arranged in chronological order to construct a continuous time frame sequence of device operating data. This is achieved by building a lightweight industrial digital OCR recognition model with a built-in character template library. In the single-frame spatial dimension, the information recognized by the OCR recognition model is compared with the standard character template, and the character with the highest similarity is selected as the recognition result. The recognition result is verified; if the character recognition results are the same for 5 consecutive frames, it is determined to be a valid recognition result. The data of the valid recognition results are linearly arranged in chronological order to construct a continuous time-series sequence of operating parameters. Specifically, non-contact automatic parameter extraction is performed for the output valid ROI region. Through a lightweight template matching OCR architecture combined with multi-frame chronological verification, the accuracy of recognition is guaranteed while meeting the operational requirements of low latency and low computational consumption at the edge. A lightweight digital OCR recognition model adapted to industrial instrument scenarios is constructed. The core of the model adopts a template matching algorithm based on normalized cross-correlation (NCC), which does not require complex deep learning inference, and the single-frame recognition time is less than 5ms. A dedicated industrial character template library is pre-built. In the single-frame spatial dimension, the cropped ROI image undergoes preprocessing: it is converted to an 8-bit grayscale image, and binarized segmentation is performed using the OTSU adaptive thresholding method. Continuous display content is then segmented into individual character regions using horizontal and vertical projection methods, removing noise regions smaller than 10 pixels. Each segmented character region is normalized to 16×16 pixels, and its normalized cross-correlation coefficient is calculated against all standard character templates in the template library. The coefficient ranges from -1 to 1, with higher values ​​indicating higher matching accuracy. The template character corresponding to the maximum coefficient is selected as the recognition result for that position; if the maximum coefficient is lower than the preset threshold of 0.85, the character recognition fails.

[0049] To eliminate misidentification caused by instantaneous lighting fluctuations, minor jitter, or local occlusion in single-frame images, a consistency check mechanism for five consecutive frames is introduced. The system maintains a cyclic queue of recognition results of length 5 for each ROI region. Recognition results for each frame are enqueued chronologically, automatically overwriting the oldest historical results. When the character recognition results at five consecutive positions in the queue are completely identical, it is considered a valid recognition result. If there are differences but a certain result appears ≥3 times, the high-frequency result is taken as the valid recognition result. If all results appear less than 3 times, the parameter extraction at that moment is marked as invalid. Finally, the valid recognition results undergo data type conversion, converting string-formatted parameter values ​​to double-precision floating-point numbers, and associating them with metadata such as a system timestamp accurate to the second, a unique device ID, and parameter type encoding to generate structured data entries. All entries are linearly arranged in ascending order of timestamp and stored in the local InfluxDB time-series database on the edge nodes, forming a time-continuous sequence of device operating parameters. If three consecutive frames are invalid, linear interpolation is used to fill in missing data points, ensuring the integrity and continuity of the sequence.

[0050] In addition, real-time monitoring of the equipment operation parameter extraction process requires adding a check to determine whether the duration of consecutive extraction failures exceeds a preset threshold. If the duration of consecutive extraction failures exceeds the preset threshold, the key video frames within the extraction failure period are located, and the reasons for parameter extraction failures are analyzed. If the extraction failure is determined to be caused by personnel obstruction, the behavior analysis of the obstructing personnel is initiated to determine whether it is abnormal. If it is determined to be non-personal obstruction, an abnormal warning is issued directly.

[0051] Further, the reasons for parameter extraction failure are analyzed. If the failure is determined to be due to personnel occlusion, the behavior analysis of the occluded personnel is initiated to determine whether it is an anomaly. Specifically, for key video frames within a continuous extraction failure period, a human detection model is used to detect whether there are human targets in the frames. The bounding boxes of the detected human targets are calculated to determine that the extraction failure is due to personnel occlusion. Identity discrimination features are extracted from the occluded personnel, including uniform color, name tag area detection features, and facial features. The extracted features are matched with a preset authorized personnel feature library to determine whether the person is an authorized person. If the person is determined to be an authorized person, it is considered a normal maintenance operation, and the operation log is recorded. If the person is determined to be an unauthorized person, an abnormal event record is generated and an abnormal warning is triggered.

[0052] Feature extraction is performed on continuous operation data sequences to analyze trend and abrupt change characteristics, and anomaly scores are generated based on these characteristics. Specifically, this involves analyzing equipment operation data sequences. Dynamic normalization is performed using a sliding time window of length W to obtain a normalized data sequence. Empirical Mode Decomposition (EMD) is performed on the normalized data sequence to obtain n Intrinsic Mode Functions (IMFs) and one residual component. ,satisfy: ,in, Let j be the value of the j-th intrinsic mode function at time t; calculate the energy of each intrinsic mode function: Where w is the index of the time sampling point within the current sliding window, ranging from 1 to W; the trend abrupt change group boundary is adaptively determined based on the cumulative energy ratio: ,in, The preset energy percentage threshold; the residual component and the previous The superposition of low-frequency IMF components is used as a trend feature sequence. This characterizes the long-term trend of equipment operating parameters; the remaining... The superposition of several high-frequency IMF components serves as a mutation characteristic sequence. Characterizing the instantaneous fluctuations and abrupt changes of parameters: , ; Calculate the slope of the trend feature sequence within the current window. and the slope of the historical normal baseline Compare and generate trend anomaly scores The calculation method is as follows: ,in, The standard deviation of the historical normal baseline slope. The duration of the abnormal trend. The time constant is used; this formula takes into account both the degree of trend deviation and the duration, with longer durations resulting in higher anomaly scores.

[0053] The sliding window standard deviation ratio algorithm is used to detect mutation feature sequences, and the ratio of the standard deviation of the current window to the standard deviation of the previous window is calculated. The calculation method is as follows: ,in, The standard deviation of the current window mutation feature sequence. The standard deviation of the mutation feature sequence in the previous window is used. The sliding window moves forward frame by frame with a step size of 1, and adjacent windows overlap by W-1 data points. When the ratio is greater than a set threshold, a mutation is identified, and a mutation anomaly score is generated. ,in, This is the threshold for mutation detection. This represents the historical standard deviation of the mutation feature sequence. Finally, the mutation anomaly score and the trend anomaly score are weighted and summed according to preset weight coefficients to generate the final anomaly score. The weight of the priority mutation anomaly score is 0.6, and the weight of the trend anomaly score is 0.4.

[0054] Finally, based on the anomaly score, the system compares the results with the preset judgment rules to identify anomalies in the equipment operating parameters and obtains the anomaly identification results.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A vision-based method for anomaly identification in surveillance videos, characterized in that, Includes the following steps: S1. Collect real-time monitoring video images from monitoring equipment in various areas of the hospital's logistics department and obtain the original video frame sequence; S2. Perform frame-by-frame preprocessing on the original video frame sequence, and for each processed video image, detect and locate all regions of interest (ROI) images in the video frame using an improved instance segmentation algorithm. S3. Extract the operating parameter data from the device display interface one by one for the ROI region, arrange the operating parameter data extracted from multiple consecutive frames in time sequence, and construct a device operating data sequence of consecutive time frames. S4. Extract features from the continuously running data sequence, analyze the trend features and mutation features, and generate anomaly scores based on the trend features and mutation features; S5. Based on the anomaly score, identify anomalies in the equipment operating parameters and obtain the anomaly identification results.

2. The vision-based anomaly identification method for surveillance videos according to claim 1, characterized in that, The specific implementation of frame-by-frame preprocessing of the original video frame sequence in step S2 is as follows: S21. Construct a spatiotemporal joint interference perception model for the input raw video frames. Where t is the image frame number, The pixel coordinates are used as input, and then multi-scale decomposition is performed to obtain the base layer image. and detail layer images The decomposition formula is: ; S22. Based on the temporal consistency constraint of consecutive K frames, perform static interference detection and repair, and calculate the temporal variance of each pixel in the base layer image and detail layer image respectively. The formula is: ,in, Let H be the mean of the pixels corresponding to K consecutive frames, where H = 1 and 2 represent the base layer image and detail layer image, respectively. Regions whose variance is less than a preset temporal variance threshold are identified as static interference regions. These regions are repaired using Poisson fusion based on the normal pixel distribution in the neighborhood. The repaired image is then the static interference-free base layer image. and static disturbance detail layer image ; S23, will As a component of light And build based on Adaptive gain function of local standard deviation The formula is: ,in, To remove static interference detail layer Local standard deviation within the window These are the global maximum and minimum values, respectively. These are the maximum and minimum values ​​of the adaptive gain function, respectively. S24. Perform global illumination normalization processing to obtain the preprocessed image: ,in, for and The resulting image.

3. The vision-based anomaly identification method for surveillance videos according to claim 1, characterized in that, In step S2, the specific implementation of detecting and locating all Regions of Interest (ROIs) in each processed video frame using the improved instance segmentation algorithm is as follows: S25. Transfer the preprocessed video frame images Input the backbone feature extraction network to extract a set of multi-scale feature maps; construct a standard feature pyramid through a top-down feature fusion method; apply frequency-aware channel attention to each pyramid feature layer, generate channel attention weights based on the high-frequency energy ratio of the feature maps, and obtain the feature pyramid after display area enhancement. S26. At each level of the enhanced feature pyramid, scale-adaptive anchor boxes are used to generate initial candidate regions; deformable convolution sampling is performed on the features at each candidate location to achieve geometric feature alignment for display panels of different shapes and tilt angles. In the region proposal stage, an additional display panel category confidence branch is introduced to perform comprehensive scoring and ranking of candidate boxes and non-maximum suppression to obtain a set of selected candidate boxes. S27. Extract dual-scale fusion features for each selected candidate box, and generate multi-granularity mask output through a fully convolutional mask prediction head; first predict the coarse-grained instance mask, then construct the boundary-aware refinement branch, extract boundary-sensitive features and generate the mask boundary response map; perform multiple rounds of iterative refinement on the coarse-grained mask based on the boundary response to obtain the final instance mask with accurate edges; S28. Perform geometric integrity verification of the display area on all detected instances in sequence, remove false detection areas, and obtain the final set of regions of interest (ROIs).

4. The vision-based anomaly identification method for surveillance videos according to claim 1, characterized in that, The specific implementation of step S3, which extracts the operating parameter data from the device display interface of the ROI region one by one, and arranges the operating parameter data extracted from multiple consecutive frames in chronological order to construct a sequence of device operating data in consecutive time frames, is as follows: S31. Build a lightweight industrial digital OCR recognition model with a built-in character template library; in the single-frame spatial dimension, calculate the similarity between the information recognized by the OCR recognition model and the standard character template, and select the character with the highest similarity as the recognition result; S32. Verify the recognition results. If the character recognition results are the same for 5 consecutive frames, they are determined to be valid recognition results. S33. Arrange the data of the effective identification results linearly in chronological order to construct a time series data sequence of time-continuous operating parameters.

5. The vision-based anomaly identification method for surveillance videos according to claim 1, characterized in that, The process of extracting equipment operating parameters is monitored in real time to determine whether the duration of continuous extraction failures exceeds a preset threshold. If the duration of consecutive extraction failures exceeds a preset threshold, the key video frames within the extraction failure period are located, and the reasons for the parameter extraction failures are analyzed. If the extraction failure is determined to be caused by personnel occlusion, the behavior analysis of the occluding personnel is initiated to determine whether it is abnormal. If the obstruction is determined to be non-personnel obstruction, an abnormal warning will be issued directly.

6. The vision-based anomaly identification method for surveillance videos according to claim 5, characterized in that, Analyze the reasons for parameter extraction failure; if the failure is determined to be due to personnel occlusion, initiate occlusion personnel behavior analysis to determine if it is an anomaly. The specific implementation is as follows: For key video frames within a continuous extraction failure period, a human detection model is used to detect whether there are human targets in the frames; the bounding boxes of the detected human targets are calculated and determined to be extraction failures caused by human occlusion. Extract identity features from obscured personnel, including uniform color, name tag area detection features, and facial features. Match the extracted features with a pre-set authorized personnel feature database to determine whether the person is an authorized personnel. If the user is determined to be an authorized person, the operation is considered normal maintenance and the operation log is recorded; if the user is determined to be an unauthorized person, an abnormal event record is generated and an abnormal warning is triggered.

7. The vision-based anomaly identification method for surveillance videos according to claim 1, characterized in that, Step S4 involves feature extraction from the continuously running data sequence, analyzing trend and mutation characteristics, and generating anomaly scores based on these trend and mutation characteristics. The specific implementation of this step is as follows: S41. Equipment operation data sequence Dynamic normalization is performed using a sliding time window of length W to obtain a normalized data sequence. ; S42. Perform Empirical Mode Decomposition (EMD) on the normalized data sequence to obtain n Intrinsic Mode Functions (IMFs) and 1 Residual Component. ,satisfy: ,in, Let j be the value of the j-th intrinsic mode function at time t; calculate the energy of each intrinsic mode function; adaptively determine the trend abrupt change grouping boundary based on the cumulative energy ratio: ,in, For the preset energy percentage threshold, where, Let j be the energy value of the j-th intrinsic mode; S43, Residual Components and Previous The superposition of low-frequency IMF components is used as a trend feature sequence. This characterizes the long-term trend of equipment operating parameters; the remaining... The superposition of several high-frequency IMF components serves as a mutation characteristic sequence. Characterizing the instantaneous fluctuations and abrupt changes of parameters: , ; S44. Calculate the slope of the trend feature sequence within the current window and compare it with the slope of the historical normal baseline to generate a trend anomaly score; use the sliding window standard deviation ratio algorithm to detect the mutation feature sequence, calculate the ratio of the standard deviation of the current window to the standard deviation of the previous window, and determine that a mutation has occurred when the ratio is greater than a set threshold, and generate a mutation anomaly score. S45. The mutation anomaly score and the trend anomaly score are weighted and summed according to the preset weight coefficients to generate the final anomaly score.

Images