Elevator car abnormal behavior recognition linkage alarm system and method
By combining spatial and temporal feature modeling with a deep neural network recognition module, the problem of recognizing complex abnormal behaviors of elevator cars in elevator monitoring systems has been solved, realizing intelligent and efficient elevator safety management, reducing the false judgment rate, and providing comprehensive safety protection.
Patent Information
- Application Number
- CN202511238048.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-01
- Publication Date
- 2025-12-02
AI Technical Summary
Existing elevator monitoring systems are unable to effectively identify various complex abnormal behaviors of elevator cars and have poor adaptability to environmental changes, resulting in a high misjudgment rate and failing to meet the safety management needs of large-scale elevator networks.
A deep neural network recognition module is used in conjunction with spatial feature extraction and temporal feature modeling. Image data is collected through a high-definition camera, and abnormal behavior is identified using a CBAM attention module and a bidirectional LSTM network. Based on the severity of the behavior, an alarm is triggered.
It enables accurate identification and efficient processing of abnormal elevator car behavior in complex scenarios, reduces the misjudgment rate, improves the level of intelligent elevator safety management, reduces the burden of manual monitoring, and provides comprehensive safety protection.
Smart Images
Figure CN121044444A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of elevator safety monitoring technology, and more specifically, to an elevator car abnormal behavior identification and linkage alarm system and method. Background Technology
[0002] In modern urban buildings, elevators, as a core tool for vertical transportation, are of paramount importance in terms of operational safety and passenger management efficiency. Elevator cars, being relatively enclosed spaces, are characterized by dense human activity and unpredictable behavior, potentially leading to various abnormal situations such as accidental falls, deliberate damage to car facilities, unauthorized prying of doors, and prolonged lingering. Failure to promptly identify and address these abnormal behaviors can not only cause elevator malfunctions and shutdowns but also result in personal injury accidents or property damage. Traditional elevator monitoring systems largely rely on manual monitoring, with security personnel observing surveillance footage in real time to determine the presence of anomalies. This approach is limited by labor costs and attention spans, resulting in inherent drawbacks such as delayed response times, high missed detection rates, and untimely handling, making it difficult to meet the safety management needs of large-scale elevator networks.
[0003] With the development of computer vision technology, some elevator monitoring systems have begun to introduce automated recognition algorithms, attempting to replace manual anomaly detection with machines. Early algorithms were mostly based on traditional computer vision methods, such as detecting moving targets using background subtraction, identifying specific behaviors using Haar feature-based classifiers, or using simple threshold judgments, such as determining overload if the number of people exceeds the limit. However, these methods are poorly adaptable to environmental changes, and are prone to misjudgment or missed detection in scenarios with complex lighting conditions, such as nighttime, backlighting, obstruction by objects inside the car, and diverse behavioral patterns. Furthermore, most existing systems are designed for single types of abnormal behavior, lacking the ability to comprehensively identify multiple complex behaviors, and failing to effectively combine the time-series characteristics of behavior. This makes it difficult to distinguish between behaviors with similar static characteristics but fundamentally different characteristics, such as normal leaning against the door versus malicious prying open the door, limiting the effectiveness of practical applications and failing to provide reliable technical support for elevator safety management.
[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention
[0005] To address the problems in related technologies, this invention proposes an elevator car abnormal behavior recognition and linkage alarm system and method to overcome the aforementioned technical problems existing in the prior art.
[0006] The technical solution of this invention is implemented as follows:
[0007] One aspect of the present invention:
[0008] An elevator car abnormal behavior recognition and linkage alarm system includes:
[0009] The image acquisition module is used to acquire video image data inside the elevator car in real time;
[0010] The preprocessing module is electrically connected to the image acquisition module and is used to preprocess the acquired video images, including enhancing contrast through adaptive histogram equalization, Gaussian filtering for noise reduction, standardizing the image size to 640×640 pixels, and extracting 16 frames in time series to form a sample window.
[0011] The deep neural network recognition module is connected to the preprocessing module and is used to receive sample window data. After processing by the spatial feature extraction module and the temporal feature modeling module, it outputs the abnormal behavior type and confidence level. The spatial feature extraction module includes an improved PANet structure with a CBAM attention module, and the temporal feature modeling module adopts a bidirectional LSTM network.
[0012] The linkage alarm module is connected to the deep neural network recognition module and is used to determine the alarm level based on the type of abnormal behavior and the confidence level, and trigger the corresponding response, including in-car audio and visual prompts, elevator operation status control, and pushing alarm information to multiple terminals.
[0013] The remote monitoring center is communicatively connected to the linkage alarm module, used to receive and display alarm information, store video clips and alarm records, and support remote configuration of system parameters.
[0014] Furthermore, the image acquisition module employs at least two high-definition cameras with a resolution of no less than 1080P and a frame rate of no less than 30fps, installed diagonally on the top of the car, in conjunction with a wide-angle lens to achieve complete coverage of the car interior without blind spots.
[0015] Furthermore, the sample window of the preprocessing module is extracted by overlapping adjacent windows by 8 frames to calibrate temporal continuity.
[0016] Furthermore, the spatial feature extraction module includes: a Backbone network, a Neck network, and a Head output layer, wherein;
[0017] The Backbone network adopts the CSPDarknet53 structure and outputs feature maps of three scales: 80×80×256, 40×40×512, and 20×20×1024.
[0018] The Neck network adds a CBAM attention module after the PANet feature fusion node. The CBAM attention module includes a channel attention submodule and a spatial attention submodule.
[0019] The Head output layer compresses the fused feature map into a 7×7×256-dimensional spatial feature sequence through three convolutional layers.
[0020] Furthermore, the channel attention submodule generates channel weights through global pooling and MLP, and the spatial attention submodule generates spatial weights through channel pooling and 7×7 convolution.
[0021] Furthermore, the temporal feature modeling module includes: an input layer, a bidirectional LSTM layer, and an output layer, wherein;
[0022] The input layer is used to flatten the 16-frame spatial feature sequence into a temporal vector of dimension;
[0023] The bidirectional LSTM layer contains a two-layer stacked structure. The first layer has 64 hidden units and the second layer has 32 hidden units. Each layer contains forward and backward propagation paths.
[0024] The output layer outputs the probability distribution of six abnormal behaviors through a dropout layer and a fully connected layer. The abnormal behaviors include people falling, violent damage, prying open doors, prolonged stay, overloading, and leaving abnormal items behind.
[0025] Furthermore, the alarm levels of the linkage alarm module are divided into three levels, wherein;
[0026] A Level 1 alarm corresponds to a minor abnormality and triggers a voice prompt inside the elevator car.
[0027] A level 2 alarm corresponds to a moderate anomaly, triggering an audible and visual alarm inside the elevator car and sending information to the property monitoring terminal.
[0028] A Level 3 alarm corresponds to a serious anomaly, triggering an audible and visual alarm inside the elevator car, an emergency stop at the nearest floor with the doors opening, and simultaneously sending information to the property management, maintenance center, and public security platform.
[0029] Another aspect of the present invention:
[0030] A method for identifying and triggering alarms for abnormal behavior in elevator cars, used in the aforementioned elevator car abnormal behavior identification and alarm system, includes the following steps:
[0031] Step S1: In advance, the video stream inside the car is acquired in real time through the image acquisition module;
[0032] Step S2: The preprocessing module enhances and normalizes the video frames to form image samples, which are used as input data.
[0033] Step S3: The deep neural network recognition module extracts the spatial features of each frame through the spatial feature extraction module, and then captures the temporal correlation between frames through the temporal feature modeling module, and outputs the abnormal behavior type and confidence level.
[0034] Step S4: The linkage alarm module triggers an alarm response of the corresponding level based on the identification result;
[0035] Step S5: The remote monitoring center receives, displays, and stores alarm information.
[0036] The beneficial effects of this invention are:
[0037] 1. This invention effectively overcomes the limitations of traditional monitoring technologies by combining a spatial feature extraction module and a temporal feature modeling module to construct a deep neural network recognition module with both spatial feature extraction and temporal series modeling capabilities. The spatial feature extraction module introduces a CBAM attention module at the feature fusion node, enabling precise focusing on key areas and feature channels within the car related to abnormal behavior. This significantly enhances the ability to perceive subtle anomalies in complex scenarios. For example, even under conditions of changing lighting or object obstruction, it can clearly capture key information such as changes in posture when a person falls or the hand details of a door-opening action, avoiding the oversensitivity of traditional algorithms to environmental interference. The temporal feature modeling module, through temporal modeling of continuous frame sequences, deeply mines the dynamic evolution features of behavior, accurately distinguishing between statically similar but fundamentally different behaviors, such as clearly differentiating normal leaning from malicious door-opening, and short-term stay from long-term lingering. This fundamentally solves the problem of difficulty in determining the nature of behavior through single-frame image recognition.
[0038] Meanwhile, this invention forms a complete closed loop from anomaly identification to emergency response, significantly improving the intelligence level of elevator safety management. The system can automatically match corresponding alarm response strategies based on the severity of abnormal behavior. Through coordinated actions such as in-car audio-visual prompts, elevator operation status control, and multi-terminal information push, it achieves precise and efficient anomaly handling. Furthermore, the parameter configuration and data storage functions of the remote monitoring center not only facilitate managers in optimizing system performance according to actual scenarios but also provide reliable evidence for post-event traceability and responsibility determination. This reduces the workload of manual monitoring and, through proactive early warning and rapid response, minimizes the safety risks that abnormal behavior may cause, providing comprehensive technical support for elevator passenger safety and operation and maintenance management. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 This is a schematic diagram of an elevator car abnormal behavior recognition and linkage alarm system according to an embodiment of the present invention;
[0041] Figure 2 This is a schematic diagram of a deep neural network recognition module in an elevator car abnormal behavior recognition and linkage alarm system according to an embodiment of the present invention.
[0042] Figure 3 This is a schematic diagram of a spatial feature extraction module of an elevator car abnormal behavior recognition and linkage alarm system according to an embodiment of the present invention;
[0043] Figure 4 This is a flowchart illustrating an elevator car abnormal behavior identification and linkage alarm method according to an embodiment of the present invention. Detailed Implementation
[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.
[0045] According to an embodiment of the present invention, an elevator car abnormal behavior recognition and linkage alarm system is provided.
[0046] like Figure 1 As shown, the elevator car abnormal behavior recognition and linkage alarm system according to an embodiment of the present invention includes an image acquisition module 1, a preprocessing module 2, a deep neural network recognition module 3, a linkage alarm module 4, and a remote monitoring center 5, wherein;
[0047] Image acquisition module 1 is used to acquire video image data inside the elevator car in real time. It uses at least one high-definition camera with wide-angle shooting and low-light compensation functions to ensure clear video images under different lighting conditions.
[0048] Specifically, in implementation, two high-definition cameras with a resolution of 1080P and a frame rate of 30fps can be used, installed diagonally on the top of the car, and used with a wide-angle lens to achieve no blind spots in the car; the cameras have infrared supplementary light function to ensure normal operation at night or in low light conditions.
[0049] The preprocessing module 2 is electrically connected to the image acquisition module 1 and is used to preprocess the acquired video image data, including enhancing image contrast through adaptive histogram equalization, removing noise through Gaussian filtering, standardizing the image size to 640×640 pixels, and extracting 16 frames in time series to form a sample window as input data for the deep neural network recognition module 3.
[0050] Specifically, the embedded processor NVIDIA Jetson Nano can be used for real-time processing. The video stream captured by the camera is extracted at 25 frames per second, and the following processing is performed: adaptive histogram equalization to enhance image contrast, especially the details in shadow areas; Gaussian filter convolution kernel 5×5 to remove image noise; size normalization to scale the image to 640×640 pixels; frame sequence extraction, with each 16 frames forming a sample window, and adjacent windows overlapping by 8 frames to ensure temporal continuity.
[0051] like Figure 2 As shown, the deep neural network recognition module 3 is used to take a 16-frame continuous video image sequence of 640×640×3 as input. The spatial feature extraction module extracts the spatial features of each frame, and the temporal feature modeling module captures the temporal correlation between frames. The module outputs the classification results and confidence scores of 6 abnormal behaviors.
[0052] The spatial feature extraction module includes:
[0053] The backbone network consists of five CSP CrossStagePartial modules, including residual connections and cross-stage feature fusion. It performs multi-scale downsampling on the input image (640×640×3) with a stride of 2, outputting feature maps at three scales: shallow features (80×80×256) preserve details such as edges and textures, like human limb contours and elevator door gaps; mid-level features (40×40×512) contain local semantic information, like human poses and object shapes; and deep features (20×20×1024) contain global semantic information, like abstract features of "multiple people gathered" and "abnormal objects." The CSP module reduces computation while enhancing gradient propagation, avoiding the vanishing gradient problem common in deep networks.
[0054] like Figure 3 As shown, the Neck network adds a CBAM attention module to each feature fusion node, enabling the network to automatically focus on features more important to the current task. It contains two sub-modules: CBAM channel attention and CBAM spatial attention, with the specific structure as follows:
[0055] PANet feature fusion is used to combine shallow detail features with deep semantic features to solve the problem of recognizing small targets and complex behaviors.
[0056] CBAM channel attention is used to generate channel weights through global pooling and MLP, enhancing key feature channels such as the "human motion channel" and "metal object channel" while suppressing irrelevant channels. For example, when recognizing "violent destruction," the "object edge change" channel will be significantly enhanced.
[0057] CBAM Spatial Attention: It generates spatial weights through channel pooling and convolution to focus on regions of abnormal behavior, such as near elevator doors or where people fall, while weakening background areas. For example, when recognizing "pushing open a door," it highlights the area where the door edge contacts the hand.
[0058] The Head output layer consists of three convolutional layers: 1×1 convolution + BN + SiLU activation, corresponding to feature maps of three different scales. These layers compress the fused feature maps into a uniform dimension of 7×7×256, containing the target location bounding box, category, and confidence information, forming a spatial feature sequence of 16, 7, 7, 256. This provides standardized spatial feature input for the subsequent temporal feature modeling module, balancing computational cost and feature representation capability.
[0059] The temporal feature modeling module is used to capture the inter-frame dependencies of abnormal elevator behavior, such as "falling," which is an action sequence lasting 1-2 seconds. The specific structure is as follows:
[0060] The input layer is used to receive the 16-frame spatial feature sequence 16,7,7,256 output by the spatial feature extraction module, which is flattened into a temporal vector of 16,12544 dimensions, 7×7×256=12544;
[0061] The bidirectional LSTM layer consists of two stacked bidirectional LSTM layers: the first layer has 64 hidden units, and the second layer has 32 hidden units. Each layer contains a forward propagation path from frame 1 to frame 16 and a reverse propagation path from frame 16 back to frame 1. The first layer captures short-term temporal dependencies, such as changes in human posture in consecutive frames, like the process of falling from a standing position to a tilted position. The second layer captures long-term temporal dependencies, such as a lingering position lasting more than 10 seconds or a sequence of actions involving repeatedly scratching at a door.
[0062] Specifically, bidirectional propagation ensures that the network pays attention to the characteristics of both "past" and "future" frames simultaneously, avoiding information loss in one direction. For example, identifying "overload" requires combining the changes in the number of people in the preceding and following 5 seconds.
[0063] The dropout layer consists of a dropout layer with a probability of 0.5, which randomly deactivates some neurons to prevent overfitting. The fully connected layer outputs the probability distribution of six abnormal behaviors through the Softmax activation function, including: falling down, violent damage, prying open doors, prolonged stay, overloading, and leaving abnormal items behind.
[0064] Furthermore, for the aforementioned deep neural network recognition module 3, in application, monitoring videos of 100 elevators are collected, and 500,000 frames of images containing abnormal behavior are labeled, covering different lighting conditions and scenarios with 1-10 people. The training set is randomly cropped by ±10%, brightness adjusted by ±20%, and horizontally flipped. Using the early stopping method, if the accuracy of the validation set does not improve for 5 consecutive epochs, the model training is stopped to prevent overfitting, thereby obtaining the trained deep neural network recognition module 3 for elevator car abnormal behavior recognition.
[0065] The linkage alarm module 4 is used to determine the alarm level based on the type of abnormal behavior and trigger the corresponding response, as follows:
[0066] A Level 1 alarm corresponds to a minor abnormality and triggers a voice prompt inside the elevator car.
[0067] Level 2 alarm corresponds to a moderate anomaly, triggering an audible and visual alarm and notifying property management.
[0068] A Level 3 alarm corresponds to a serious anomaly, triggering audible and visual alarms, emergency elevator stop, and simultaneously notifying property management, maintenance center, and the public security platform.
[0069] This technical solution uses an STM32 microcontroller as its core, communicates with the elevator control system via an RS485 interface, and connects to the remote monitoring center via a 4G module. Its functions include:
[0070] Alarm Level Determination: When the confidence level of abnormal behavior is ≥85%, the alarm level is determined according to preset rules, such as violent damage corresponding to a Level 3 alarm. Alarm Execution: Controls the LED warning lights, buzzer, and voice module in the elevator car. For example, a Level 1 alarm plays "Please do not stay in the elevator car for an extended period of time," and a Level 2 / 3 alarm plays "An abnormality has occurred in the elevator, please remain calm." Commands are sent through the elevator control interface, such as sending the command "Emergency stop at the nearest floor and open the door" for a Level 3 alarm. Information Push: Alarm information, including the type of abnormality, time, elevator car location, and real-time video link, is sent to designated terminals via the 4G network, such as property monitoring computers, maintenance personnel mobile apps, and public security alarm platforms.
[0071] The remote monitoring center 5 communicates with the linkage alarm module 4 to receive and display alarm information, store video clips and alarm records, and remotely configure system parameters, such as identification thresholds and alarm level standards.
[0072] Specifically, it can be deployed in the property monitoring room and consists of a server, monitoring terminal, and management software. It uses a hybrid storage of MySQL and MongoDB, with MySQL storing alarm records and MongoDB storing video clips. The monitoring terminal displays alarm information in real time, pops up video windows, and supports manual snapshot and recording. Administrators can adjust the recognition threshold through the web interface, such as adjusting the confidence threshold of the third-level alarm from 85% to 90%, alarm level rules, and camera parameters.
[0073] According to an embodiment of the present invention, a method for identifying and triggering alarms for abnormal behavior in an elevator car is provided.
[0074] like Figure 4 As shown, the elevator car abnormal behavior identification and linkage alarm method according to an embodiment of the present invention includes the following steps:
[0075] Step S1: The camera of image acquisition module 1 acquires the video stream inside the car in real time and transmits it to preprocessing module 2 via Ethernet;
[0076] Step S2: The preprocessing module 2 extracts video frames at 25fps, performs enhancement, noise reduction, and normalization processing, and forms a sample window of 16 frames, which is then input into the deep neural network recognition module 3.
[0077] Step S3: The deep neural network recognition module 3 extracts features from each frame of the image in the sample window through the spatial feature extraction module and outputs a feature map containing spatial features; then the temporal feature modeling module performs time series modeling on the 16 frame feature maps and outputs the confidence scores of 6 abnormal behaviors; if the highest confidence score is ≥ the preset threshold, it is determined that there is a corresponding abnormal behavior and the result is output.
[0078] Step S4: The linkage alarm module 4 receives the identification result, determines the alarm level based on the anomaly type and confidence level, and triggers the corresponding response:
[0079] Level 1 alarm: If the person remains in the area for an extended period of time with a confidence level of 80%-90%, the voice module will play a prompt tone and no external alarm information will be sent.
[0080] Level 2 alarm, such as overload, with a confidence level ≥ 85%: activates LED warning lights and buzzers, and sends alarm information to the property monitoring terminal;
[0081] Level 3 alarm, such as forced entry, with a confidence level of ≥85%: activates LED warning lights, buzzers and voice alarms, sends an emergency stop command to the elevator control system, and sends alarm information to the property management, maintenance center and public security platform;
[0082] Step S5: The remote monitoring center 5 receives alarm information, displays it on the monitoring terminal, and stores relevant video clips and alarm records. Management personnel can view and handle the alarms.
[0083] In summary, by utilizing the above-described technical solution of the present invention, the following effects can be achieved:
[0084] 1. This invention effectively overcomes the limitations of traditional monitoring technologies by combining a spatial feature extraction module and a temporal feature modeling module to construct a deep neural network recognition module 3 that combines spatial feature extraction and temporal series modeling capabilities. The spatial feature extraction module introduces a CBAM attention module at the feature fusion node, enabling precise focusing on key areas and feature channels within the car related to abnormal behavior. This significantly enhances the ability to perceive subtle anomalies in complex scenarios. For example, even under conditions of changing lighting or object obstruction, it can clearly capture key information such as changes in posture when a person falls or details of hand movements during door-opening actions, avoiding the oversensitivity of traditional algorithms to environmental interference. The temporal feature modeling module, through temporal modeling of continuous frame sequences, deeply mines the dynamic evolution features of behavior, accurately distinguishing between statically similar but fundamentally different behaviors. For example, it clearly distinguishes between normal leaning and malicious door-opening, and between short-term stays and long-term lingering, fundamentally solving the problem that single-frame image recognition is insufficient to determine the nature of behavior.
[0085] Meanwhile, this invention forms a complete closed loop from anomaly identification to emergency response, significantly improving the intelligence level of elevator safety management. The system can automatically match corresponding alarm response strategies based on the severity of abnormal behavior. Through coordinated actions such as in-car audio-visual prompts, elevator operation status control, and multi-terminal information push, it achieves precise and efficient anomaly handling. Furthermore, the parameter configuration and data storage functions of the remote monitoring center not only facilitate managers in optimizing system performance according to actual scenarios but also provide reliable evidence for post-event traceability and responsibility determination. This reduces the workload of manual monitoring and, through proactive early warning and rapid response, minimizes the safety risks that abnormal behavior may cause, providing comprehensive technical support for elevator passenger safety and operation and maintenance management.
[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art, upon considering the disclosure in the specification and embodiments, will readily conceive of other embodiments of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.
[0087] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.
Claims
1. An elevator car abnormal behavior recognition and linkage alarm system, characterized in that, include: Image acquisition module (1) is used to acquire video image data inside the elevator car in real time; The preprocessing module (2) is electrically connected to the image acquisition module (1) and is used to preprocess the acquired video images, including enhancing contrast through adaptive histogram equalization, denoising with Gaussian filtering, standardizing the image size to 640×640 pixels, and extracting 16 frames in time sequence to form a sample window; The deep neural network recognition module (3) is connected to the preprocessing module (2) and is used to receive sample window data. After processing by the spatial feature extraction module and the temporal feature modeling module, it outputs the abnormal behavior type and confidence level. The spatial feature extraction module includes an improved PANet structure with a CBAM attention module, and the temporal feature modeling module uses a bidirectional LSTM network. The linkage alarm module (4) is connected to the deep neural network recognition module (3) and is used to determine the alarm level according to the type of abnormal behavior and confidence level, and trigger the corresponding response, including in-car audio and light prompts, elevator operation status control and pushing alarm information to multiple terminals; The remote monitoring center (5) is connected to the linkage alarm module (4) for receiving and displaying alarm information, storing video clips and alarm records, and supporting remote configuration of system parameters.
2. The elevator car abnormal behavior recognition and linkage alarm system according to claim 1, characterized in that, The image acquisition module (1) uses at least two high-definition cameras with a resolution of not less than 1080P and a frame rate of not less than 30fps, which are installed at the diagonal position on the top of the car, and work with a wide-angle lens to achieve no blind spots in the car.
3. The elevator car abnormal behavior recognition and linkage alarm system according to claim 1, characterized in that, The sample window of the preprocessing module (2) is extracted by overlapping adjacent windows by 8 frames to calibrate the temporal continuity.
4. The elevator car abnormal behavior recognition and linkage alarm system according to claim 1, characterized in that, The spatial feature extraction module includes: a Backbone network, a Neck network, and a Head output layer, wherein; The Backbone network adopts the CSPDarknet53 structure and outputs feature maps of three scales: 80×80×256, 40×40×512, and 20×20×1024. The Neck network adds a CBAM attention module after the PANet feature fusion node. The CBAM attention module includes a channel attention submodule and a spatial attention submodule. The Head output layer compresses the fused feature map into a 7×7×256-dimensional spatial feature sequence through three convolutional layers.
5. The elevator car abnormal behavior recognition and linkage alarm system according to claim 4, characterized in that, The channel attention submodule generates channel weights through global pooling and MLP, and the spatial attention submodule generates spatial weights through channel pooling and 7×7 convolution.
6. The elevator car abnormal behavior recognition and linkage alarm system according to claim 1, characterized in that, The temporal feature modeling module includes: an input layer, a bidirectional LSTM layer, and an output layer, wherein; The input layer is used to flatten the 16-frame spatial feature sequence into a temporal vector of (16, 12544) dimensions; The bidirectional LSTM layer contains a two-layer stacked structure. The first layer has 64 hidden units and the second layer has 32 hidden units. Each layer contains forward and backward propagation paths. The output layer outputs the probability distribution of six abnormal behaviors through a dropout layer and a fully connected layer. The abnormal behaviors include people falling, violent damage, prying open doors, prolonged stay, overloading, and leaving abnormal items behind.
7. The elevator car abnormal behavior identification and linkage alarm system and method according to claim 1, characterized in that, The alarm levels of the linkage alarm module (4) are divided into three levels, among which; A Level 1 alarm corresponds to a minor abnormality and triggers a voice prompt inside the elevator car. A level 2 alarm corresponds to a moderate anomaly, triggering an audible and visual alarm inside the elevator car and sending information to the property monitoring terminal. A Level 3 alarm corresponds to a serious anomaly, triggering an audible and visual alarm inside the elevator car, an emergency stop at the nearest floor with the doors opening, and simultaneously sending information to the property management, maintenance center, and public security platform.
8. A method for identifying and triggering alarms based on abnormal behavior of an elevator car, used in the method of the elevator car abnormal behavior identification and alarm system according to any one of claims 1-7, characterized in that, Includes the following steps: Step S1: The video stream inside the car is acquired in real time by the image acquisition module (1); Step S2, the preprocessing module (2) enhances and normalizes the video frames to form image samples, which are used as input data; Step S3, the deep neural network recognition module (3) extracts the spatial features of each frame through the spatial feature extraction module, and then captures the temporal correlation between frames through the temporal feature modeling module, and outputs the abnormal behavior type and confidence level; Step S4, the linkage alarm module (4) triggers the corresponding level of alarm response based on the identification result; Step S5: The remote monitoring center (5) receives, displays and stores alarm information.