Method and device for filtering human bodies from posters in elevators

Abstract

This application proposes a method and device for filtering people on posters in elevators, including the following steps: obtaining elevator video, using a pre-trained human detection model to detect the elevator video to obtain multiple human detection frames; using human detection frames that simultaneously meet the conditions of average texture complexity less than a first threshold, contour direction consistency less than a second threshold, and highlight area ratio greater than a third threshold as suspected poster-state detection frames; if the coordinate change value of the center point of the suspected poster-state detection frame in consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold, then the suspected poster-state detection frame is used as a confirmed poster-state detection frame; and filtering the people in the confirmed poster-state detection frame. This solution filters poster-state people by analyzing the texture features, contour features, and reflection features of each detection frame.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to a method and device for filtering human bodies from posters in elevators. Background Art

[0002] In current elevator systems, trapped person detection has gradually become one of the important functions of smart elevators. The trapped person detection function relies on the camera to analyze the real-time images inside the elevator car to identify trapped people. However, posters containing human images or static advertisements displayed on electronic screens are often posted inside elevators. Such static human images are easily misidentified as real human bodies by the human detection model of traditional monitoring systems, causing the system to believe that there are trapped people and trigger false alarms. This not only reduces the accuracy of trapped person detection, but also affects the reliability of elevator operation and user experience. This is especially true in complex lighting conditions, such as light reflections, changes in brightness and darkness, or in crowded scenes. The misjudgment problem will be more prominent.

[0003] Existing target detection algorithms lack a specific feature analysis mechanism for elevator scenarios. Therefore, when a poster portrait is similar to a real human body in appearance, the target detection algorithm cannot eliminate the interference of the poster portrait, resulting in a high false alarm rate and difficulty adapting to special environments such as lighting changes in the elevator and the fixed angle of the camera.

[0004] To address the above issues, there is an urgent need for a technical solution that can accurately identify human bodies in elevator posters and filter out their interference, so as to improve the robustness and accuracy of the trapped person detection system in complex scenarios and avoid resource waste and user inconvenience caused by misjudgment of human portraits in static posters. Summary of the Invention

[0005] The embodiments of the present application provide a method and device for filtering people on posters in an elevator, which filters people on posters by analyzing the texture features, contour features, and reflection features of each detection frame.

[0006] In a first aspect, an embodiment of the present application provides a method for filtering people from posters in an elevator, the method comprising:

[0007] Obtain elevator video and use the pre-trained human detection model to detect multiple human detection frames.

[0008] Calculate the average texture complexity, contour direction consistency, and highlight area ratio of each human detection frame. Detection frames that meet the following conditions simultaneously: average texture complexity less than a first threshold, contour direction consistency less than a second threshold, and highlight area ratio greater than a third threshold are considered suspected poster-like detection frames.

[0009] If the coordinate change value of the center point of the suspected poster state detection frame in the consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold, then the suspected poster state detection frame is used as the confirmed poster state detection frame;

[0010] Filter the human body within the detection frame of the confirmed poster state.

[0011] In a second aspect, an embodiment of the present application provides a poster-based human body filtering device in an elevator, comprising:

[0012] The acquisition module is used to obtain elevator videos and use the pre-trained human detection model to detect the elevator videos to obtain multiple human detection frames;

[0013] A calculation module is used to calculate the average texture complexity, contour directional consistency, and highlight area ratio of each human detection frame, and selects human detection frames that meet the following conditions: average texture complexity less than a first threshold, contour directional consistency less than a second threshold, and highlight area ratio greater than a third threshold as suspected poster detection frames;

[0014] The judgment module determines whether the suspected poster state detection frame is a confirmed poster state detection frame if the change value of the detection frame center point coordinate of the suspected poster state detection frame in the consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold.

[0015] The filtering module is used to filter the human body in the detection frame of the confirmation poster state.

[0016] In a third aspect, an embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for filtering people on posters in an elevator.

[0017] The main contributions and innovations of the present invention are as follows:

[0018] The embodiment of the present application uses a pre-trained human detection model combined with multi-dimensional feature analysis such as average texture complexity, contour direction consistency, and saturated pixel ratio to accurately identify human detection frames suspected to be posters. It then verifies dynamic parameters such as the change in the coordinate value of the center point of the detection frame confidence and the confidence fluctuation range within a set time to effectively distinguish between real people and posters, thereby improving filtering accuracy. This solution uses dual transmission channels to transmit elevator video, and only transmits data when key frames are captured, lighting conditions change, or objects move, optimizing data transmission efficiency and reducing system load. Furthermore, during the detection process of the human detection model, the confidence threshold is adaptively adjusted based on the image brightness, and the non-maximum suppression overlap threshold is set to 0.5 to improve detection robustness and adapt to different lighting environments. This solution calculates feature parameters through methods such as circular local binary patterns and the Sobel operator to ensure the scientific and reliable feature extraction, and sets the confirmed poster detection frame area as an invalid area. Combined with the intersection-over-union judgment, the filtering efficiency is further improved. The entire solution performs outstandingly in accuracy, efficiency, adaptability, and reliability, and can effectively solve the problem of false detection of posters in elevators.

[0019] The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0021] Figure 1 This is a flow chart of a method for filtering human bodies from posters in an elevator according to an embodiment of the present application;

[0022] Figure 2 is a structural diagram of a human body detection model according to an embodiment of the present application;

[0023] Figure 3 is a schematic diagram of identifying a poster detection frame according to an embodiment of the present application;

[0024] Figure 4 This is a structural block diagram of a poster-based human body filtering device in an elevator according to an embodiment of the present application;

[0025] Figure 5 Schematic diagram of the hardware structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The implementations described in the following exemplary embodiments are not intended to represent all implementations consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of this specification, as detailed in the appended claims.

[0027] It should be noted that in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the method may include more or fewer steps than those described in this specification. In addition, a single step described in this specification may be broken down into multiple steps for description in other embodiments, and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0028] Example 1

[0029] The embodiment of the present application provides a method for filtering people on posters in elevators, which filters people on posters by analyzing the texture features, contour features, and reflection features of each detection frame. Specifically, Figure 1 , the method comprising:

[0030] Obtain elevator video and use the pre-trained human detection model to detect multiple human detection frames.

[0031] Calculate the average texture complexity, contour direction consistency, and highlight area ratio of each human detection frame. Detection frames that meet the following conditions simultaneously: average texture complexity less than a first threshold, contour direction consistency less than a second threshold, and highlight area ratio greater than a third threshold are considered suspected poster-like detection frames.

[0032] If the coordinate change value of the center point of the suspected poster state detection frame in the consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold, then the suspected poster state detection frame is used as the confirmed poster state detection frame;

[0033] Filter the human body within the detection frame of the confirmed poster state.

[0034] In some specific embodiments, the present solution fixes a camera in the front top corner area inside the elevator car. The camera is installed above the front side of the car and forms a top-down angle with the ground, preferably set to 30°±5°, so as to reduce blind spots in shooting and effectively distinguish between foreground and background areas.

[0035] Specifically, the camera used in this solution has an image output specification of 720×1080 resolution, and the video encoding model adopts the H.264 standard. In order to adapt to the dynamic changes in lighting conditions in the elevator (such as turning lights on and off, external light entering, etc.), the system enables the camera's automatic exposure control mechanism and dynamically adjusts the exposure time. The range is set between 1 / 30 second and 1 / 1000 second to maintain the image brightness at an appropriate level and avoid overexposure or underexposure that affects detection accuracy.

[0036] Specifically, the camera enables "corridor mode", which rotates the image direction to vertical shooting, so that the picture better fits the narrow and long space structure of the elevator car, and improves the vertical image utilization; at the same time, the system also turns on "dynamic encoding mode", allowing the camera to dynamically adjust the compression parameters according to the complexity of the moving area in the image, further optimizing video bandwidth occupancy and reducing redundant data transmission.

[0037] In some specific embodiments, before using the human detection model to process the elevator video, an integrity check is performed on key frames of the elevator video, wherein the key frames are frames containing significant content changes or used for analysis.

[0038] Specifically, a CRC32 check code is attached to the key frame data packet of the elevator video. The CRC32 check code is generated based on the original byte content of the elevator video image frame. Before the elevator video is sent to the human body detection model, the CRC32 check code is checked for consistency. If the check passes, the elevator video is detected using the human body detection model. If the check fails, the elevator video is requested to be retransmitted or the key frame is discarded to prevent erroneous data from participating in subsequent analysis and ensure the accuracy of human body detection.

[0039] In some specific embodiments, a hardware acceleration module is integrated into the camera or server. The hardware acceleration module has functions such as image pre-decoding, cache management and format conversion. The elevator video is pre-processed using the hardware acceleration module and then input into the human detection model for detection.

[0040] Specifically, the hardware acceleration module can directly complete the preliminary processing of image data without the involvement of the CPU, thereby greatly reducing the burden on the main processor and improving the response speed and stability of the overall system.

[0041] In some specific embodiments, when key frame capture of the current elevator video, changes in lighting conditions, and object motion are detected, image stream data of the elevator video is input into the human detection model via a first transmission channel, and metadata and control signaling of the image stream data are input into the human detection model via a second transmission channel.

[0042] Specifically, the metadata includes the timestamp, frame type and illumination value of the corresponding image stream data, and the control signaling includes model request and feedback signal. This solution performs dual-channel transmission through the first transmission channel and the second transmission channel, so as to perform fine control and feedback on the hardware acceleration module integrated in the camera or server without affecting the continuity of the image stream, thereby effectively improving parallel processing capabilities and scalability.

[0043] In other words, this solution does not continuously occupy bandwidth to transmit elevator video to the human detection model. Instead, it only transmits elevator video when key frames are captured, lighting conditions change, or objects move. This effectively reduces resource consumption and improves the response efficiency of the camera's human detection model.

[0044] Specifically, this solution uses the hardware acceleration module on the camera or server to perform data pre-decoding processing to judge key frame capture, changes in lighting conditions or object movement. The changes in multiple frames before and after, pixel brightness, and pixel complexity can be used to judge key frame capture, changes in lighting conditions or object movement. The specific judgment method is the existing technology and this solution will not be described in detail here.

[0045] In some specific embodiments, the human body detection model is used to identify human target areas appearing in video images and output their location information. In this solution, the constructed human body detection model is trained by constructing a training data set including elevator images. The training data set includes image samples collected in various environments such as different floors, different light, and different poster contents. The total number of samples is approximately 200,000, and the image size of each training sample is adjusted to 640×640 resolution to unify the input dimension.

[0046] In some specific embodiments, the human detection model decomposes the elevator video into images and then performs target detection on each frame of the image. The end condition of the human detection model during the training process is that the training accuracy meets the actual use requirements. If it does not meet the requirements, the training is repeated after sample enhancement and parameter optimization. If the accuracy requirements are still not met, the training batch is increased and the training is repeated until the end condition is met. The structure of the human detection model is as follows: Figure 2 As shown, the human detection model sequentially connects the following modules:

[0047] Input module: Input a 3-channel RGB image. A GSConv module is used in the input module for preliminary feature extraction. The number of output channels is 32, the convolution kernel size is 3, and the stride is 2. The input module achieves the first downsampling of the image spatial size and effectively preserves edge and contrast information.

[0048] Basic downsampling module 1: In basic downsampling module 1, the features initially extracted in the input module are fed into the second GSGSConv module, thereby expanding the number of channels to 64. Then, convolution with a stride of 2 is used to further perform spatial compression and extract mid- and low-level semantic features.

[0049] Lightweight residual block: Introduces the C2f module and adopts a branch fusion structure to improve the spatial expression capability of features, enhancing the human detection model's ability to model the contours and posture changes of people in elevator images.

[0050] Basic downsampling module 2: GSConv is used to further downsample the feature map to a 1 / 8 ratio, and the number of channels is expanded to 128 to meet the computational requirements of mid-level semantic expression.

[0051] Double-layer fusion module: Two C2f modules are stacked continuously in the double-layer fusion module to perform cross-layer semantic fusion of features with 128 channels, improving the stability and resolution of the human detection model in complex scenes with overlapping people.

[0052] Lightweight downsampling unit: The SCdown module is used to replace the conventional convolution module, with an output channel number of 256. The SCdown module is a separable convolution with a stride of 2. The SCdown module can further enhance the computational efficiency of the model and reduce memory usage.

[0053] Mid-layer semantic fusion module: Uses two consecutive C2f modules for feature fusion and residual enhancement, thereby improving the human detection model's ability to recognize human targets with large size differences and variable postures.

[0054] Deep feature extraction and downsampling module: The SCDown structure is used twice to reduce the spatial dimension to 1 / 32 and expand the number of channels to 512, thereby enhancing the human detection model's ability to detect small targets in dense areas.

[0055] Fusion attention module: The channel interaction module (CIB) is integrated with the C2f module to guide the human detection model to focus on the key human structures in the occluded area based on the joint enhancement of the channel dimension and the spatial dimension.

[0056] Spatial Pyramid Pooling Module: This module uses a multi-scale pooling strategy to aggregate information within different receptive fields, effectively improving the model's generalized detection capabilities for multi-scale targets, and is particularly adaptable to elevator lighting changes and environmental occlusion.

[0057] Attention Enhancement Module (PSA): Introduces the attention mechanism to reconstruct the channel-dimensional weights of deep semantic features, enhancing the detection response to stationary or edge people in the elevator.

[0058] Upsampling and feature fusion module 1: The deep feature map is upsampled and concatenated with the 256-channel feature map output from the lightweight downsampling unit, and cross-layer alignment is performed at the semantic and structural levels.

[0059] Intermediate fusion module: The C2f module performs convolution integration and channel compression on the 768-dimensional feature map output by the upsampling and feature fusion module 1, and outputs a 256-channel feature map for connecting the upper and lower feature information flows.

[0060] Upsampling and feature fusion module 2: After sampling the 256-channel feature map output by the intermediate fusion module, it is spliced ​​with the 128-dimensional low-level features output by the basic downsampling module 2, thereby realizing unified modeling of multi-layer information.

[0061] Low-level fusion module: The C2f structure is used to fuse and compress the 384-dimensional feature map output by the upsampling and feature fusion module 2, and the final output is a 128-channel feature map, thereby improving the model's ability to recognize small targets and edge figures.

[0062] Feature regression path module 1: Downsample the 128-channel feature map output by the low-level fusion module through GSConv, so as to facilitate cross-scale collaboration in subsequent steps.

[0063] Multi-branch fusion module: The output of the feature regression path module 1 is spliced ​​with the output of the intermediate fusion module to obtain a 384-channel mesoscale feature map, which provides support for subsequent mid-level target modeling.

[0064] Feature re-fusion module: The 384-channel mid-scale feature map output by the multi-branch fusion module is compressed into a 256-channel feature map through C2f for high-level re-fusion.

[0065] Feature regression path module 2: The 256-channel feature map output by the feature re-fusion module is downsampled through the SCDown module, and the downsampling result is spliced ​​with the feature map output by the attention enhancement module at the channel level.

[0066] Deep fusion module: Use the C2fCIB module to fuse the 768-dimensional feature map output by the feature regression path module 2 with the channel and spatial attention mechanism to obtain a 512-dimensional feature map with high semantic representation.

[0067] Detection Head Module: The detection head module uses the outputs of the low-level fusion module, the feature re-fusion module, and the deep fusion module as input to perform object classification and bounding box regression. By performing object classification and bounding box regression on feature maps of different dimensions, the detection head module can identify a variety of human bodies and their associated objects, such as children, elderly people, strollers, pets, and wheelchairs.

[0068] In some embodiments, the human detection model detects the elevator video to obtain multiple first detection target frames and the confidence of each first detection target frame, eliminates the first detection target frames whose confidence is less than the dynamic confidence threshold to obtain multiple second detection target frames, and performs non-maximum suppression on the multiple second detection target frames to obtain multiple human detection frames.

[0069] Furthermore, the image of each first detection target frame is converted into a grayscale image, and the grayscale average of all pixel grayscale values ​​is calculated. When the grayscale average is greater than or equal to the first grayscale threshold, the dynamic confidence threshold is 0.8; when the grayscale average is less than or equal to the second grayscale threshold, the dynamic confidence threshold is 0.6; when the grayscale average is between the first grayscale threshold and the second grayscale threshold, the dynamic confidence fluctuates proportionally between 0.6 and 0.8.

[0070] Specifically, the color first detection target frame image is converted into a grayscale image using the formula Y = 0.299R + 0.587G + 0.114B, and the grayscale average is calculated. In this scheme, the first grayscale threshold is 180 and the second grayscale threshold is 80. Therefore, when the grayscale average is between the first and second grayscale thresholds, that is, between 80 and 180, the dynamic confidence increases by 0.01 for every increase of 5 in the grayscale average.

[0071] That is to say, when the overall image brightness is bright, this solution will eliminate target detection frames with a confidence level less than 0.8. When the overall image brightness is dark, this solution will eliminate target detection frames with a confidence level less than 0.6, thereby ensuring that reasonable detection recall and accuracy can be maintained in low-light or strong-light environments.

[0072] In some embodiments, in the step of performing non-maximum suppression on multiple second detection target frames, the non-maximum suppression overlap threshold is set to 0.5.

[0073] Specifically, for complex scenes such as dense crowds and overlapping occlusions that often occur in elevators, this solution sets the non-maximum suppression overlap threshold of the human detection model detection frame to 0.5, so that more detection frames of close but different people in the elevator can be retained, avoiding missed detections of people in crowded environments and improving overall detection performance.

[0074] In some embodiments, a circular local binary pattern is used to calculate the average texture complexity of each human detection frame, a Sobel operator is used to extract the image edge of each human detection frame and the gradient direction distribution is analyzed to obtain the contour direction consistency, and the proportion of saturated pixels in each human detection frame is calculated as the proportion of the highlight area.

[0075] Specifically, the circular local binary pattern has the ability to express textures of different scales, so it is more suitable for scenarios where light fluctuations and camera viewing angles change in elevators. If the average texture complexity of the human body detection model is less than the first threshold, it means that the corresponding human body detection frame is a low-texture area with a poster-like characteristic tendency. The first threshold in this solution is 0.25.

[0076] Specifically, the Sobel operator is used to extract the image edge of the human body detection frame and analyze the gradient direction distribution. The gradient direction distribution is the directional consistency of the human body contour. When the contour direction consistency in the human body detection frame is less than the second threshold, it means that the human body contour structure is regular and has a poster feature tendency. The second threshold in this scheme is 60%.

[0077] Specifically, the proportion of saturated pixels within each human body detection frame is calculated as the proportion of the highlight area. If the proportion of the highlight area is greater than the third threshold, it indicates that there is a flat printed material with light reflection, which has a poster characteristic tendency.

[0078] Therefore, when a human detection frame has an average texture complexity less than the first threshold, a contour direction consistency less than the second threshold, and a highlight area ratio greater than the third threshold, it means that the human detection frame is likely to be the human body on the poster.

[0079] Furthermore, after obtaining the suspected poster state detection frame, the suspected poster state detection frame is continuously detected within a set time, so as to further confirm the suspected poster state detection frame as a confirmed poster state detection frame or a real person state detection frame.

[0080] Therefore, this solution uses the suspected poster state detection frame whose coordinate change value of the center point of the detection frame confidence within the set time is less than the change threshold and the confidence fluctuation range is less than the fluctuation threshold as the confirmed poster state detection frame. The schematic diagram of identifying the poster detection frame is as follows Figure 3 shown.

[0081] Specifically, the area size change and pixel point change of the poster state detection frame between the current frame and the next frame are calculated. If the area size change exceeds the set threshold and the number of pixel point changes exceeds the set threshold, the poster state detection frame is re-identified as a human body detection frame and the human body detection frame is re-evaluated.

[0082] Specifically, the poster-state detection frame is continuously monitored for changes in area size and pixel points between the current frame and the next frame within a set time. In this solution, the set time is 1 second or 20 frames. When the area size change between the two frames within the set time exceeds a set threshold (5%) or the pixel change exceeds a set threshold (500), the poster-state detection frame is re-identified as a human body detection frame.

[0083] Specifically, when the area size change of the poster state detection frame of the current and the subsequent frames exceeds the set threshold and the number of pixel changes exceeds the set threshold, it means that the poster state detection frame has changed significantly, that is, the person may have moved.

[0084] In some other embodiments, for each area within the poster state detection frame, the motion vector field (Optical Flow Field) is extracted between multiple frames; the vector direction, amplitude, density, etc. are statistically and clustered analyzed to determine whether there is a real motion trend in the area; if orderly and continuous changes in the vector motion direction are found, it is determined that there may be dynamic events in the area such as human passing, camera vibration, or external light interference; at the same time, a direction consistency verification mechanism is introduced. When most optical flow directions are random or disordered, the system can eliminate false alarms caused by periodic reflections or image flickering.

[0085] In some specific embodiments, the time for collecting continuous time frames in this solution is 10 seconds, and the change threshold is adaptively calculated based on the size of the detection frame. The fluctuation threshold in this solution is 5%.

[0086] Specifically, based on the image resolution, the standard detection frame is set to an area ratio of 1% to 5% of the image resolution. When the ratio is greater than 5%, the detection is more stringent; when it is less than 1%, the detection movement is more relaxed, and the detection of distant models is prone to fluctuation. The fluctuation threshold is ±5%. For example, if the area change is set to 5%, then the corresponding area change range is between 4.75% and 5.25% when the frame ratio is between 1% and 5%. If the pixel count changes by 500, then the corresponding pixel count change range is between 475 and 525 when the frame ratio is between 1% and 5%. The lowest value is used when the frame ratio is less than 1%, and the highest value is used when it is greater than 5%.

[0087] That is to say, if the suspected poster state detection frame or the confirmed poster state detection frame simultaneously satisfies the following conditions: the change value of the coordinate of the center point of the detection frame in continuous video frames is greater than the change threshold, the confidence fluctuation range of the detection frame confidence is greater than the fluctuation threshold, and the average texture complexity is greater than or equal to the first threshold, then the current suspected poster state detection frame or the confirmed poster state detection frame is judged as a real human posture detection frame, and is regarded as a real person in the elevator.

[0088] In some specific embodiments, the area where each confirmation poster state detection frame is located in the elevator video is defined as an invalid area. In the step of detecting the elevator video with a human body detection model to obtain multiple human body detection frames, the intersection-and-union ratio of each human body detection frame and each invalid area is calculated, and the human body detection frame whose intersection-and-union ratio with the invalid area is greater than a fourth threshold is directly used as the confirmation poster state detection frame.

[0089] Specifically, the invalid area is confirmed by obtaining the upper left corner coordinates and the lower right corner coordinates of each confirmation poster state detection frame. The fourth threshold in this scheme is 0.4, that is, when the intersection and union ratio of the human body detection frame and the invalid area is greater than 0.4, the human body detection frame is directly used as the confirmation poster state detection frame.

[0090] Furthermore, an effective time and an expiration time are set for each invalid area. If the current time exceeds the expiration time, the corresponding invalid area will automatically become invalid.

[0091] Specifically, the difference between the effective time and the end time in this scheme is 30 seconds. That is to say, the effective time of each invalid detection frame in this scheme is only 30 seconds. After 30 seconds, the invalid area will automatically become invalid, thereby preventing a false detection from permanently filtering the area.

[0092] In some other embodiments, a human activity path heat map is generated based on the movement trajectory of each real human posture detection frame in the corresponding elevator video, and the high-heat area in the human activity path heat map is used as a behavior hot zone through a threshold judgment method, and the confirmation poster posture detection frame in the behavior hot zone is re-judged as a human detection frame.

[0093] Specifically, this solution obtains a human activity path map by recording the center point coordinates and timestamps of all real human detection frames into the trajectory database, and then performs heat zone modeling on the human activity path map through the heat zone modeling method to obtain a time-series-based human activity path heat map.

[0094] Specifically, since humans tend to be concentrated near doors and buttons when performing activities in elevators, this solution uses a threshold judgment method to use high-heat areas in the human activity path heat map as behavioral hot zones, while retaining the confirmation poster state detection box in the behavioral hot zone to ensure that the real human body will be judged as a poster body when it is still for a short time.

[0095] In some other embodiments, the brightness change of each pixel in the elevator video in continuous time frames is determined in real time, and a static background area in the elevator video is obtained by modeling based on the brightness change of each pixel using a mixed Gaussian distribution method.

[0096] In other words, this solution uses a mixed Gaussian distribution to use the long-term stable pixel value distribution as the static background area, and as new time frames are input, the background area is automatically updated, thereby effectively filtering out interference such as occasional brightness changes and image compression noise, and ensuring efficient modeling of fixed structures in elevators, such as billboards and button panels.

[0097] In some specific embodiments, this solution encapsulates structured information such as image samples, detection labels, filtering judgment results, and timestamps uploaded by different elevator terminals and saves them in a database. The encapsulation format includes fields such as terminal ID, image path, detection box information, filtering status, model version and confidence, etc., and supports fast retrieval and indexing based on dimensions such as site ID, time range, model version, and filtering type in the database; data writing adopts a chain log method to ensure sequential integrity, and combines hash signatures to verify the consistency of data during transmission and storage; supports hierarchical management of hot and cold data, archives and compresses historical data, and periodically cleans it to ensure the coexistence of system performance and data tracking capabilities.

[0098] In some specific embodiments, each camera and server performs lightweight model fine-tuning training locally, and only uploads weight differences or gradient information to the cloud; the cloud aggregates the updated parameters of multiple terminals and uses weighted averaging to update the global main model; it simultaneously supports the "personalized model fusion" mechanism, allowing some terminals to retain their exclusive fine-tuning branches to achieve optimal recognition accuracy in heterogeneous environments; the federal optimization cycle can be configured to be triggered by hours, days, or sample size, and has the ability to resume uploads and check model consistency; the overall architecture complies with the federated learning privacy protection specifications, and no original images or personal privacy information are uploaded during any training process.

[0099] In some specific examples, this solution designs a misjudgment sample mining and structure optimization module to conduct centralized analysis on the filtered misjudgments, false detections and missed detections in the uploaded samples: the system classifies and groups the misjudgment samples through a clustering algorithm based on feature vectors or visual similarity; combined with image metadata, it can output an "analysis report on high-incidence areas of misjudgment", such as frequent interference from poster backgrounds on a certain elevator floor, and the impact of reflections at specific angles; further combined with error heat maps and misjudgment distribution statistics, the system can automatically generate model structure optimization suggestions (for example, it is recommended to add a reflection feature branch and enhance the edge module); at the same time, it supports the model incremental update mechanism, and the system automatically compares the differences between the current model and the terminal model, and only sends down the necessary differential weight files.

[0100] In some specific embodiments, this solution provides a cloud-based graphical operations and maintenance platform for visually displaying system operating status and data trends. The platform utilizes modular visual components to display recognition accuracy, filtering error rate, false alarm rate, and number of misjudged samples at different sites. It also supports timeline display of model performance trends, site anomaly alarm records, and uploaded sample growth curves. Operations and maintenance personnel can use a graphical interface to select specific sites, time periods, or specific model versions for diagnostic tracing. The system supports notifications for abnormal events (such as an abnormal decrease in recognition accuracy or a model update failure) and automatically pushes maintenance recommendations. It also integrates operational permissions such as remote model updates, blacklist synchronization, and regional filtering policy issuance to facilitate remote centralized management. Differential files are stored using compressed encoding.

[0101] Example 2

[0102] Based on the same concept, refer to Figure 4 , this application also proposes a poster human body filtering device in an elevator, comprising:

[0103] The acquisition module is used to obtain elevator videos and use the pre-trained human detection model to detect the elevator videos to obtain multiple human detection frames;

[0104] A calculation module is used to calculate the average texture complexity, contour directional consistency, and highlight area ratio of each human detection frame, and selects human detection frames that meet the following conditions: average texture complexity less than a first threshold, contour directional consistency less than a second threshold, and highlight area ratio greater than a third threshold as suspected poster detection frames;

[0105] The judgment module determines whether the suspected poster state detection frame is a confirmed poster state detection frame if the change value of the detection frame center point coordinate of the suspected poster state detection frame in the consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold.

[0106] The filtering module is used to filter the human body in the detection frame of the confirmation poster state.

[0107] Example 3

[0108] This embodiment also provides an electronic device, referring to Figure 5 , includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program, and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0109] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0110] Memory 404 may include a large-capacity memory 404 for data or instructions. By way of example, and not limitation, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to the data processing device. In certain embodiments, memory 404 is non-volatile memory. In certain embodiments, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. In appropriate circumstances, the RAM may be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM may be a fast page mode dynamic random access memory 404 (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0111] The memory 404 may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor 402 .

[0112] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any one of the methods for filtering people from posters in an elevator in the above embodiments.

[0113] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408 , wherein the transmission device 406 is connected to the processor 402 , and the input / output device 408 is connected to the processor 402 .

[0114] Transmission device 406 can be used to receive or transmit data via a network. Specific examples of such networks may include wired or wireless networks provided by the electronic device's communications provider. In one embodiment, the transmission device includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 406 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0115] The input and output device 408 is used to input or output information. In this embodiment, the input information may be an elevator video, a human body detection frame, etc., and the output information may be a poster state detection frame, etc.

[0116] Optionally, in this embodiment, the processor 402 may be configured to execute the following steps through a computer program:

[0117] Obtain elevator video and use the pre-trained human detection model to detect multiple human detection frames.

[0118] Calculate the average texture complexity, contour direction consistency, and highlight area ratio of each human detection frame. Detection frames that meet the following conditions simultaneously: average texture complexity less than a first threshold, contour direction consistency less than a second threshold, and highlight area ratio greater than a third threshold are considered suspected poster-like detection frames.

[0119] If the coordinate change value of the center point of the suspected poster state detection frame in the consecutive time frames is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold, then the suspected poster state detection frame is used as the confirmed poster state detection frame;

[0120] Filter the human body within the detection frame of the confirmed poster state.

[0121] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be repeated here.

[0122] In general, various embodiments may be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention may be implemented in hardware, while other aspects may be implemented in firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that, as non-limiting examples, the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0123] The embodiments of the present invention may be implemented by computer software that is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets and / or macros may be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer executable components that are configured to perform an embodiment when the program is run. One or more computer executable components may be at least one software code or a portion thereof. In addition, it should be noted at this point that, for example, Figure 5 Any block of the logic flow in the program may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on physical media such as memory chips or memory blocks implemented within the processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. Physical media are non-transitory media.

[0124] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] The above embodiments merely illustrate several embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for filtering human bodies from posters in an elevator, characterized in that: The following steps are involved: Obtain elevator video and use the pre-trained human detection model to detect multiple human detection frames. Calculate the average texture complexity, contour direction consistency, and highlight area ratio of each human detection frame. The circular local binary pattern is used to calculate the average texture complexity of each human detection frame. The Sobel operator is used to extract the image edge of each human detection frame and analyze the gradient direction distribution to obtain the contour direction consistency. The ratio of saturated pixels in each human detection frame is calculated as the highlight area ratio. Human detection frames that simultaneously meet the conditions of average texture complexity less than a first threshold, contour direction consistency less than a second threshold, and highlight area ratio greater than a third threshold are regarded as suspected poster-state detection frames. If the coordinate change value of the center point of the suspected poster-state detection frame in the continuous time frame is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold, then the suspected poster-state detection frame is used as the confirmed poster-state detection frame, and the area size change and pixel point change of the suspected poster-state detection frame between the current frame and the next frame are calculated. If the area size change exceeds the set threshold and the number of pixel point changes exceeds the set threshold, the suspected poster-state detection frame is re-identified as a human detection frame; Filter the human body within the detection frame of the confirmed poster state.

2. The method for filtering human bodies in posters in an elevator according to claim 1, characterized in that: When key frame capture, changes in lighting conditions, and object motion are detected in the current elevator video, the image stream data of the elevator video is input into the human detection model via the first transmission channel, and metadata and control signaling of the image stream data are input into the human detection model via the second transmission channel.

3. The method for filtering human bodies in posters in an elevator according to claim 1, characterized in that: The human detection model detects the elevator video to obtain multiple first detection target frames and the confidence of each first detection target frame, eliminates the first detection target frames whose confidence is less than the dynamic confidence threshold to obtain multiple second detection target frames, and performs non-maximum suppression on the multiple second detection target frames to obtain multiple human detection frames.

4. The method for filtering people in posters in elevators according to claim 3, characterized in that: The image within each first detection target frame is converted into a grayscale image, and the grayscale average of all pixel grayscale values ​​is calculated. When the grayscale average is greater than or equal to the first grayscale threshold, the dynamic confidence threshold is 0.8; when the grayscale average is less than or equal to the second grayscale threshold, the dynamic confidence threshold is 0.6; when the grayscale average is between the first grayscale threshold and the second grayscale threshold, the dynamic confidence fluctuates between 0.6 and 0.8 in proportion.

5. The method for filtering people in posters in an elevator according to claim 1, characterized in that: If the suspected poster state detection frame or the confirmed poster state detection frame simultaneously satisfies the following conditions: the change value of the coordinate of the center point of the detection frame in continuous video frames is greater than the change threshold, the confidence fluctuation range of the detection frame confidence is greater than the fluctuation threshold, and the average texture complexity is greater than or equal to the first threshold, then the current suspected poster state detection frame or the confirmed poster state detection frame is determined as a real human posture detection frame.

6. The method for filtering human bodies from posters in an elevator according to claim 1, characterized in that: Define the area where each confirmation poster state detection frame is located in the elevator video as an invalid area. In the step of using the human body detection model to detect the elevator video to obtain multiple human body detection frames, calculate the intersection and union ratio of each human body detection frame with each invalid area, and directly use the human body detection frame whose intersection and union ratio with the invalid area is greater than the fourth threshold as the confirmation poster state detection frame.

7. The method for filtering human bodies from posters in an elevator according to claim 1, characterized in that: Based on the movement trajectory of each real human posture detection frame in the corresponding elevator video, a human activity path heat map is generated. The high-heat area in the human activity path heat map is used as the behavior hot zone through the threshold judgment method, and the confirmation poster posture detection frame in the behavior hot zone is re-judged as the human detection frame.

8. A poster-like human body filtering device in an elevator, characterized in that: include: The acquisition module is used to obtain elevator videos and use the pre-trained human detection model to detect the elevator videos to obtain multiple human detection frames; A calculation module is used to calculate the average texture complexity, contour directional consistency, and highlight area ratio of each human detection frame. The circular local binary pattern is used to calculate the average texture complexity of each human detection frame. The Sobel operator is used to extract the image edge of each human detection frame and analyze the gradient direction distribution to obtain contour directional consistency. The ratio of saturated pixels in each human detection frame is calculated as the highlight area ratio. Human detection frames that simultaneously meet the conditions of average texture complexity less than a first threshold, contour directional consistency less than a second threshold, and highlight area ratio greater than a third threshold are regarded as suspected poster-state detection frames. The judgment module uses the suspected poster state detection frame as the confirmed poster state detection frame if the coordinate change value of the detection frame center point of the suspected poster state detection frame in the continuous time frame is less than the change threshold, and the confidence fluctuation range of the detection frame confidence is less than the fluctuation threshold. The suspected poster state detection frame is used as the confirmed poster state detection frame, and the area size change and pixel point change of the suspected poster state detection frame between the current frame and the next frame are calculated. If the area size change exceeds the set threshold and the number of pixel point changes exceeds the set threshold, the suspected poster state detection frame is re-identified as a human body detection frame. The filtering module is used to filter the human body in the detection frame of the confirmation poster state.

9. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to execute the method for filtering people from posters in an elevator according to any one of claims 1 to 7.

Images