Hydraulic lifting platform dangerous behavior identification method and system based on AI

Abstract

The invention relates to an AI-based method for identifying dangerous behaviors of a hydraulic lifting platform, which is characterized by comprising the following steps of: acquiring video image data of a working area of the hydraulic lifting platform in real time through a camera; preprocessing the collected video image data, including image scaling, normalization and data enhancement operations, so as to unify the data format and enhance the generalization ability of the model; inputting the preprocessed video image data into a target detection model based on deep learning, and determining a spatial relationship and behavior characteristics between target objects; and inputting the dangerous behavior feature vector into a dangerous behavior classification model based on a support vector machine. According to the system, multiple dangerous behaviors can be accurately identified based on deep learning, the identification accuracy of the system is greatly improved, the false alarm rate is remarkably reduced, and multiple complex dangerous scenes can be effectively covered.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of lifting platforms, in particular to an AI-based dangerous behavior recognition method and system for a hydraulic lifting platform. BACKGROUND

[0002] Hydraulic lifting platforms are widely used in many fields such as industrial production, construction, logistics and warehousing, and are used for cargo loading and unloading, equipment maintenance, high-altitude work, etc. However, there are many safety hazards in the operation of hydraulic lifting platforms, such as dangerous behaviors of operators climbing in violation of regulations, platform overloading, and approaching running parts, which can easily cause safety accidents and result in personnel casualties and property losses.

[0003] Traditional safety monitoring of hydraulic lifting platforms mainly relies on manual patrol and self-restraint of operators. Manual patrol has problems such as patrol interval and visual obstruction, and cannot monitor in real time and comprehensively; operators are prone to accidents when they are tired, careless or operate in violation of regulations. Some existing monitoring systems mainly focus on monitoring physical quantities such as platform position and speed, lack comprehensive perception of operator behavior and surrounding environment, and are difficult to identify dangerous behaviors and give early warnings in a timely manner.

[0004] In the prior art, there have been attempts to use ordinary cameras and simple image processing techniques for behavior recognition, but there are problems such as low recognition accuracy, high false positive rate, and poor adaptability to complex environments. For example, some systems can only recognize a limited number of dangerous behaviors, and the recognition of operator posture and behavior details is not accurate enough to meet the needs of actual applications. SUMMARY

[0005] The application provides an AI-based dangerous behavior recognition method for a hydraulic lifting platform, which solves the above problems in the prior art.

[0006] The technical solution of the application is as follows:

[0007] An AI-based dangerous behavior recognition method for a hydraulic lifting platform, characterized in that it comprises the following steps:

[0008] Real-time acquisition of video image data of the working area of the hydraulic lifting platform by a camera;

[0009] Preprocessing of the acquired video image data, including image scaling, normalization and data enhancement operations, to unify the data format and enhance the generalization ability of the model;

[0010] The pre-processed video image data is input into a deep learning-based target detection model, which at least includes a feature extraction layer, a region proposal network and a classification regression layer, for identifying target objects such as operators, various components of the hydraulic lifting platform and surrounding obstacles in the video image, and outputting the category and position information of each target object;

[0011] Based on the category and position information of the target objects, the spatial relationship and behavior characteristics between the target objects are determined, and a dangerous behavior feature vector including the proximity distance between the operator and the hydraulic lifting platform, the relative motion speed, the operator's posture and the motion state of the hydraulic lifting platform components, etc. is extracted;

[0012] The dangerous behavior feature vector is input into a support vector machine-based dangerous behavior classification model, which judges whether there is a dangerous behavior in the current video image according to the pre-trained classification rules, and outputs the type and degree of danger of the dangerous behavior;

[0013] If there is a dangerous behavior, the corresponding alarm and warning mechanism is triggered according to the type and degree of danger of the dangerous behavior, an information prompt containing the dangerous position, the dangerous type and the suggested measures is generated, and is displayed in real time on the monitoring terminal.

[0014] Preferably, before pre-processing the collected video image data, a step of correcting the distortion of the image collected by the camera is further included to eliminate the influence of camera lens distortion on image quality and improve the accuracy of subsequent target detection and behavior recognition.

[0015] Preferably, the deep learning-based target detection model adopts the Faster R-CNN architecture, and at least 5000 training pictures annotated with special components of the hydraulic lifting platform and typical dangerous scenes are added to the COCO dataset containing a large number of annotated target objects in the working scene of the hydraulic lifting platform, so as to migrate and optimize the model and improve the detection accuracy and recall rate of the model for specific targets of the hydraulic lifting platform.

[0016] Preferably, when extracting the dangerous behavior feature vector, a step of detecting the skeleton key points of the operator's posture is further included, the position coordinates of the main joints of the operator's body are determined based on the posture estimation algorithm, and whether the operator has dangerous posture behaviors such as climbing, leaning and illegal operation is further analyzed according to the angle change and relative position relationship between the joints, and the posture dangerous features are integrated into the dangerous behavior feature vector.

[0017] Preferably, the support vector machine-based dangerous behavior classification model adopts a radial basis function as a kernel function, and by performing feature extraction and labeling on collected dangerous behavior sample data of different types and different degrees, the SVM model is trained using a sequential minimal optimization algorithm to determine an optimal classification hyperplane, thereby realizing accurate classification of dangerous behavior and quantitative assessment of the degree of danger.

[0018] Preferably, the alarm and early warning mechanism includes sound alarms, light flashing alarms, and pop-up graphic prompt information on the monitoring terminal, and different levels of alarm and early warning methods are set according to different degrees of danger, wherein when the degree of danger is high, multiple alarm methods are triggered at the same time and are displayed in a prominent position on the monitoring terminal, ensuring that the on-site staff can timely perceive the danger and take corresponding measures.

[0019] An AI-based dangerous behavior recognition system for a hydraulic lifting platform, characterized by comprising:

[0020] An image acquisition module for acquiring video image data of the working area of the hydraulic lifting platform in real time through a camera;

[0021] A data preprocessing module for preprocessing the collected video image data, including image scaling, normalization, and data enhancement operations;

[0022] A target detection module for inputting the preprocessed video image data into a deep learning-based target detection model, identifying target objects such as operators, components of the hydraulic lifting platform, and surrounding obstacles in the video image, and outputting the category and position information of each target object;

[0023] A dangerous feature extraction module for determining the spatial relationship and behavior characteristics between target objects based on the category and position information of the target objects, and extracting a dangerous behavior feature vector including the proximity distance between the operator and the hydraulic lifting platform, the relative motion speed, the operator's posture, and the motion state of the hydraulic lifting platform components;

[0024] A dangerous behavior classification module for inputting the dangerous behavior feature vector into a support vector machine-based dangerous behavior classification model to determine whether there is a dangerous behavior in the current video image and output the type and degree of danger of the dangerous behavior;

[0025] An alarm and early warning module for triggering corresponding alarm and early warning mechanisms according to the type and degree of danger of the dangerous behavior if there is a dangerous behavior, generating information prompts containing the dangerous position, the dangerous type, and the recommended measures to be taken, and displaying them in real time on the monitoring terminal.

[0026] Preferably, an image distortion correction module is further included, which is arranged before the data preprocessing module to perform distortion correction processing on the image collected by the camera, so as to eliminate the influence of the camera lens distortion on the image quality.

[0027] Preferably, a pose estimation sub-module is further included in the dangerous feature extraction module, which is configured to detect the skeleton key points of the operator's pose, determine the position coordinates of the main joint points of the operator's body based on a pose estimation algorithm, analyze whether the operator has dangerous pose behaviors such as climbing, leaning and illegal operation, and integrate the pose dangerous features into the dangerous behavior feature vector.

[0028] Preferably, the alarm and warning mechanism in the alarm and warning module includes sound alarm, light flashing alarm, and pop-up of eye-catching text and picture prompt information on the monitoring terminal, and different levels of alarm and warning modes are set according to different dangerous levels, wherein when the dangerous level is high, multiple alarm modes are triggered at the same time and are displayed in a prominent position of the monitoring terminal.

[0029] In summary, the beneficial effects of the present application are:

[0030] The present application discloses an AI-based hydraulic lifting platform dangerous behavior recognition method and system:

[0031] (I) Accurate and comprehensive dangerous behavior recognition

[0032] The system uses a target detection model (Faster R-CNN) based on deep learning and an SVM dangerous behavior classification model, which can accurately recognize various dangerous behaviors such as illegal climbing, platform overload, and proximity to running parts. The target detection model is trained based on a large number of specific scene data of the hydraulic lifting platform extended from the COCO data set, which can accurately locate the operator, platform components and other targets; the SVM classification model accurately judges the behavior type and degree based on the dangerous behavior feature vector. Compared with the traditional simple image processing method, the recognition accuracy of the system is greatly improved, the false alarm rate is significantly reduced, and various complex dangerous scenes can be effectively covered.

[0033] (II) Fast and real-time alarm and warning

[0034] The system realizes real-time monitoring and rapid response, the camera collects images at a high frame rate, the data processing and model inference process is optimized, and the average response time from the occurrence of dangerous behavior to the system alarm is not more than 1.2 seconds, which is much faster than the traditional method. And according to the dangerous degree classification, different alarm and warning modes are triggered, and when the danger is high, sound, light and text and picture prompt are used for multi-mode warning at the same time, so as to ensure that the on-site personnel can timely perceive and handle the danger and effectively prevent the accident from expanding.

[0035] (III) High degree of intelligence and automation

[0036] Based on AI technology, this system automatically learns and adapts to different working environments and operator behavior patterns without manual intervention. It automatically updates and optimizes dangerous behavior recognition models, offering a high level of intelligence and automation. This reduces the cost and intensity of manual monitoring, improves safety monitoring efficiency and reliability, and achieves an intelligent upgrade for safety monitoring on hydraulic lift platforms.

[0037] (4) Reduce the risk of safety accidents and ensure the safety of personnel and equipment

[0038] By accurately identifying dangerous behaviors and quickly issuing alarms and warnings, operators and on-site staff can be effectively reminded to correct violations in a timely manner and take safety measures, thereby significantly reducing the risk of safety accidents during the operation of hydraulic lifting platforms, reducing casualties and equipment damage, improving the overall safety level of the work site, and providing reliable safety protection for industrial production and working environments.

[0039] (V) Improve work efficiency and management level

[0040] Timely identification and correction of dangerous behaviors, standardized operating procedures, and reduced equipment failures and downtime caused by illegal operations improve work efficiency. At the same time, the system automatically records and analyzes dangerous behavior data, providing managers with valuable information to help optimize safety management strategies and improve safety management levels.

[0041] (6) Strong adaptability, easy integration and expansion

[0042] The system utilizes a modular design with interoperable functional modules, offering excellent adaptability and scalability. It can be adapted to various models and specifications of hydraulic lift platforms to meet diverse needs. It can also be easily integrated with other safety monitoring systems, such as a factory's production safety management system, to enable information sharing and collaborative work, building a more comprehensive and robust safety monitoring system. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] Example

[0045] This embodiment discloses an AI-based method and system for identifying dangerous behaviors of hydraulic lifting platforms.

[0046] The AI-based hydraulic lifting platform dangerous behavior recognition system of the embodiment is applied to the safety monitoring of a hydraulic lifting platform in an industrial production site. The entire system is composed of an image acquisition module, a data preprocessing module, a target detection module, a dangerous feature extraction module, a dangerous behavior classification module, an alarm and early warning module, and a monitoring terminal.

[0047] The image acquisition module adopts multiple high-resolution industrial cameras, which are respectively installed at different positions of the hydraulic lifting platform to achieve all-around coverage of the working area. The cameras are connected to the data preprocessing module through a network interface to transmit the collected video image data in real time.

[0048] The data preprocessing module is deployed on a high-performance edge computing server, which is equipped with a high-performance GPU to accelerate image processing tasks. The data preprocessing module performs scaling, normalization, and data enhancement operations on the collected images, and simultaneously eliminates the influence of camera lens distortion on image quality by using an image distortion correction algorithm.

[0049] The target detection module runs a deep learning target detection model based on the Faster R-CNN architecture. The model is extended by adding 5000 training pictures labeled with special components of the hydraulic lifting platform and typical dangerous scenes on the basis of the COCO dataset. After migration training and optimization, the model can accurately identify target objects such as operators, components of the hydraulic lifting platform (such as lifting columns, platform panels, control handles, etc.), and surrounding obstacles in the video images, and output the category and position information (represented by bounding box coordinates) of each target object.

[0050] The dangerous feature extraction module calculates spatial and motion features such as the proximity distance and relative motion speed of the operator and the hydraulic lifting platform according to the category and position information of the target objects. Meanwhile, the skeleton key points of the operator's posture are detected by using a posture estimation algorithm to analyze whether there are dangerous posture behaviors such as climbing and leaning, and the corresponding posture dangerous features are extracted to form a dangerous behavior feature vector.

[0051] The dangerous behavior classification module adopts a support vector machine (SVM) model based on a radial basis function (RBF) kernel. By extracting and labeling features of a large number of different types and different degrees of dangerous behavior sample data collected, the SVM model is trained by using a sequential minimal optimization (SMO) algorithm to accurately determine whether there is a dangerous behavior in the current video image and output the type (such as illegal climbing, platform overload, and proximity to running components) and the dangerous degree (divided into low, medium, and high levels) of the dangerous behavior.

[0052] The alarm and early warning module triggers corresponding alarm and early warning mechanisms according to the type and degree of danger behavior. When the degree of danger is low, only a yellow warning prompt information is popped up on the monitoring terminal; when the degree of danger is medium, an orange warning information is popped up on the monitoring terminal, and a light flashing alarm is triggered; when the degree of danger is high, a sound alarm, a light flashing alarm and a red warning information are triggered at the same time, and the red warning information is popped up on a prominent position of the monitoring terminal, showing the dangerous position, the dangerous type and the suggested measures.

[0053] The monitoring terminal adopts an industrial-grade touch screen display and is installed in the on-site monitoring room for the safety supervisor to check the working state of the hydraulic lifting platform and the dangerous behavior recognition result in real time.

[0054] II. Detailed implementation of key modules

[0055] (I) Image acquisition module

[0056] 1. Camera selection: Four industrial cameras with a resolution of 1920x1080 pixels and a frame rate of 30fps are selected, which have wide dynamic range, low-light shooting and auto-focus functions, ensuring clear capture of images in the working area of the hydraulic lifting platform under different lighting conditions.

[0057] 2. Camera arrangement: Cameras are installed around and directly above the hydraulic lifting platform. The installation height of the cameras around the platform is about 3 meters, and the horizontal distance from the platform is 2-3 meters. The camera directly above the platform is installed 5 meters above the platform, achieving omnidirectional and dead-angle-free monitoring of the platform. The cameras are connected to the edge computing server through network cables, using PoE (Power over Ethernet) technology to simplify wiring and reduce construction costs.

[0058] (II) Data preprocessing module

[0059] 1. Image scaling: The collected images are uniformly scaled to 640x480 pixels to adapt to the input requirements of the target detection model, while reducing the amount of calculation and improving the processing speed.

[0060] 2. Normalization: The image pixel values are normalized, mapping the pixel values from [0, 255] to [0, 1] range, enhancing the model's adaptability to different lighting conditions and improving the stability of target detection.

[0061] 3. Data augmentation: Random rotation, flipping, cropping and color jittering data augmentation techniques are used to expand the training data set and improve the model's generalization ability, so that it can adapt to target detection scenarios under different angles, postures and lighting conditions.

[0062] 4. Distortion correction: Using camera calibration techniques, the intrinsic parameters and distortion parameters of the camera are obtained. The undistort function in the OpenCV image processing library is used to correct the distortion of the collected images, eliminate the image distortion caused by lens distortion, and improve the accuracy of target detection and behavior recognition.

[0063] (Three) Target detection module

[0064] 1. Faster R-CNN model architecture: The model mainly consists of a feature extraction layer, a region proposal network (RPN), and a classification regression layer. The feature extraction layer uses ResNet-50 as the backbone network to extract the deep features of the image; the RPN network generates a series of candidate region proposals and classifies each region proposal as foreground or background and performs boundary box regression; the classification regression layer further classifies the region proposals filtered by the RPN into target categories and fine-tunes the boundary box positions, outputting the final detection results.

[0065] 2. Model training and optimization:

[0066] Dataset preparation: Based on the COCO dataset, 5000 training images of hydraulic lifting platform working scenes are added, and target objects such as operators, hydraulic lifting platform components, and surrounding obstacles in the images are labeled. The labeling format uses the PascalVOC format, and the labeling content includes target categories and boundary box coordinates.

[0067] Data preprocessing: Random cropping, scaling, flipping, and other data augmentation operations are performed on the training data, and the images are normalized to unify the data format and range.

[0068] Model training: The stochastic gradient descent (SGD) optimization algorithm is used, with an initial learning rate of 0.001, a momentum of 0.9, a weight decay coefficient of 0.0001, a batch size of 8, and a total number of iterations of 100,000. During training, the model is saved every 10,000 iterations and validated on the validation set. The optimal model is selected based on the validation loss.

[0069] Model optimization: To improve the detection accuracy and recall rate of specific targets of the hydraulic lifting platform, the transfer learning technique is used, and the pre-trained ResNet-50 model is used as the initial weight to fine-tune the model. At the same time, the non-maximum suppression (NMS) algorithm is used for post-processing of the detection results to remove duplicate boundary boxes and improve the accuracy of the detection results.

[0070] (Four) Dangerous feature extraction module

[0071] 1. Spatial relationship and motion feature calculation: Based on the bounding box coordinates of the operator and hydraulic lifting platform components output by the target detection module, the proximity distance of the operator to the platform edge, lifting column and other components is calculated. The relative motion speed of the operator and the platform components is calculated using the target position change between adjacent frames. These spatial and motion features reflect the possible collision, extrusion and other dangerous situations between the operator and the platform.

[0072] 2. Pose estimation and dangerous pose analysis: The OpenPose pose estimation algorithm is used to detect the pose skeleton key points of the operator, obtaining the position coordinates of each major joint of the operator's body (such as head, shoulder, elbow, waist, knee, ankle, etc.). According to the angle change and relative position relationship between the joints, it is judged whether the operator has dangerous pose behaviors such as climbing (such as hands leaving the platform handrail, body climbing upwards), leaning (body center of gravity deviating towards the platform edge, distance from the platform edge less than the safety threshold). For example, when the operator's hand joint position is higher than the head joint position, and the distance between the waist and the platform edge is less than the set safety distance, it is determined as a climbing dangerous pose; when the operator's body center of gravity joint position deviates from the platform center by more than a certain threshold, and the distance from the platform edge is less than the safety distance, it is determined as a leaning dangerous pose. These pose dangerous features are integrated into the dangerous behavior feature vector, and the spatial and motion features are combined to form a complete dangerous behavior feature description.

[0073] (Five) Dangerous behavior classification module

[0074] 1. SVM model training:

[0075] Data collection and annotation: Collect a large number of different types of dangerous behavior sample data, including illegal climbing, platform overload, approaching running components, illegal operation control handle, etc., as well as normal operation behavior data. Extract the dangerous behavior feature vector for each sample data, and label the corresponding dangerous behavior type and dangerous degree.

[0076] Feature extraction: Perform feature normalization on the sample data, map each feature value in the feature vector to the range [-1, 1], and improve the training efficiency and stability of the SVM model.

[0077] Model training: Radial basis function (RBF) is used as the kernel function of SVM, and sequential minimal optimization (SMO) algorithm is used to train the SVM model. The kernel function parameters γ and the penalty parameter C are optimized by grid search method to maximize the classification accuracy and interval. During training, the data set is divided into training set and validation set, and cross-validation is used to evaluate the model performance to avoid overfitting phenomenon.

[0078] 2. Dangerous behavior classification and assessment: input the real-time extracted dangerous behavior feature vector into the trained SVM model, and the model outputs the type and danger level of the dangerous behavior. The danger level is divided into low, medium and high levels, among which the low danger level indicates that there is a certain safety hazard, but it will not immediately cause serious threat to personnel and equipment; the medium danger level indicates that there may be a serious dangerous situation, which needs to be handled in time; the high danger level indicates that there is a very high risk of danger, which may cause serious safety accidents.

[0079] (VI) Alarm and early warning module

[0080] 1. Alarm and early warning methods:

[0081] Sound alarm: install a high-pitched loudspeaker in the monitoring room, when the danger level is high, trigger the sound alarm, emit continuous alarm sound (frequency of 1 kHz, loudness of 80-90 decibels), remind the on-site workers and safety supervisors to pay attention to the dangerous situation.

[0082] Light flashing alarm: install multiple colored LED lights around the hydraulic lifting platform, when the danger level is medium or high, trigger the light flashing alarm. Among them, when the danger level is medium, the light flashing frequency is 2Hz, and the color is yellow; when the danger level is high, the light flashing frequency is 5Hz, and the color is red, which prompts different danger levels through different frequencies and colors of light flashing.

[0083] Monitoring terminal text and picture prompt: pop up a prominent text and picture prompt information on the monitoring terminal, the information content includes the dangerous position (represented by platform coordinates or area identifier), the dangerous type (such as illegal climbing, platform overload, etc.), the danger level and the suggested measures (such as immediately stopping the climbing behavior, evacuating the dangerous area, etc.). When the danger level is low, the background color of the prompt information is yellow, and the font color is black; when the danger level is medium, the background color is orange, and the font color is white; when the danger level is high, the background color is red, and the font color is white, and the prompt information is always displayed at the front end of the screen, ensuring that safety supervisors can check in time.

[0084] 2. Alarm and early warning mechanism control: the alarm and early warning module triggers the corresponding alarm and early warning methods according to the dangerous behavior type and danger level output by the dangerous behavior classification module, according to the preset alarm and early warning rules. The alarm and early warning rules are stored in the database of the edge computing server, which can be flexibly configured and adjusted according to actual needs. For example, for some specific dangerous behavior types (such as illegal operation control handle), even if the danger level is medium, sound alarm and light flashing alarm can be set to trigger at the same time to enhance the warning effect of this type of dangerous behavior.

[0085] III. System running process

[0086] (I) System initialization

[0087] 1. After the system power is turned on, the edge computing server starts and loads various functional modules, including the data preprocessing module, the target detection module, the dangerous feature extraction module, the dangerous behavior classification module, and the alarm and early warning module.

[0088] 2. The camera performs self-checking to check whether the image acquisition function is normal, and establishes a network connection with the edge computing server to start real-time transmission of video image data.

[0089] 3. The data preprocessing module performs distortion correction processing on the first frame of image collected, generates correction parameters, and takes the corrected image as a model input sample to initialize and calibrate the target detection module and the dangerous feature extraction module.

[0090] 4. The target detection module loads the trained Faster R-CNN model to complete the initialization and parameter loading of the model. The dangerous behavior classification module loads the trained SVM model to also complete the initialization operation. The alarm and early warning module initializes the alarm and early warning devices to check whether the sound alarm, light alarm, and display function of the monitoring terminal are normal.

[0091] 5. After the system completes the initialization, it enters the standby state and waits for the start monitoring instruction.

[0092] (II) Normal operation phase

[0093] 1. When the monitoring start instruction is received, the camera starts to collect video image data of the hydraulic lifting platform working area at a frame rate of 30 fps in real time, and transmits the image data to the data preprocessing module of the edge computing server through the network interface.

[0094] 2. The data preprocessing module sequentially performs scaling, normalization, data enhancement, and distortion correction processing on the collected image, and stores the processed image in the form of a tensor in the memory and as input to the target detection module.

[0095] 3. The target detection module receives the preprocessed image data, performs target detection through the Faster R-CNN model, and outputs the class and position information (bounding box coordinates) of target objects such as operators, hydraulic lifting platform components, and surrounding obstacles in the image. The detection result is stored in JSON format and sent to the dangerous feature extraction module.

[0096] 4. The dangerous feature extraction module calculates the proximity distance and relative motion speed between the operator and the hydraulic lifting platform components based on the target detection results, and uses the OpenPose pose estimation algorithm to detect the pose skeleton key points of the operator. Whether there is a dangerous pose behavior is analyzed, and the corresponding pose dangerous feature is extracted, and finally a dangerous behavior feature vector is formed. The feature vector is stored in an array form and transmitted to the dangerous behavior classification module.

[0097] 5. The dangerous behavior classification module inputs the dangerous behavior feature vector into the SVM model. The model judges whether there is a dangerous behavior and the type and degree of danger according to the trained classification rules. The classification results are stored in a dictionary form, containing information such as dangerous type and dangerous degree, and are sent to the alarm and warning module.

[0098] 6. The alarm and warning module triggers the corresponding alarm and warning mode according to the preset alarm and warning rules based on the classification results. For example, when it is detected that the operator has a rule-breaking climbing behavior and the danger degree is high, the sound alarm and the light flashing alarm are triggered immediately, and a red warning prompt information is popped up on the monitoring terminal, which shows the dangerous position, the dangerous type and the suggested measures in detail, reminding the on-site workers and safety supervisors to take timely action to avoid accidents.

[0099] 7. The monitoring terminal displays the working state of the hydraulic lifting platform in real time, including the real-time video picture collected by the camera, the target detection results (drawing target bounding box and category label on the video picture), the dangerous behavior recognition results (displayed in the form of text and picture), and the alarm and warning information. The safety supervisor can check the system running state through the monitoring terminal and perform remote control and management.

[0100] (Three) Alarm reset and system recovery

[0101] 1. When the dangerous behavior is eliminated, the alarm and warning module automatically detects the change of the dangerous state, and according to the preset reset rules, the alarm and warning state is released. The sound alarm stops, the light flashing alarm returns to the normal or off state, the warning prompt information on the monitoring terminal becomes a green "safe" prompt information, and it is displayed that the system has returned to normal operation.

[0102] 2. The system returns to the normal running state and continues to monitor the video image data of the working area of the hydraulic lifting platform in real time to identify and warn possible dangerous behaviors.

[0103] Four, system performance test and result

[0104] (I) Target detection performance test

[0105] 1. Test method: Precision, Recall, and Mean Average Precision (mAP) are used as evaluation indicators for target detection performance. 1000 pictures of hydraulic lifting platform working scenes are selected from the test data set, which cover different poses, angles, and lighting conditions of various target objects such as operators, hydraulic lifting platform components, and surrounding obstacles. The detection results of the target detection module on the test pictures are compared with the true values labeled by humans, and the precision and recall of each target class and the overall mAP value are calculated.

[0106] 2. Test results: After testing, the detection accuracy of the target detection module for operators is 92%, and the recall rate is 88%; the detection accuracy for hydraulic lifting platform components is 90%, and the recall rate is 85%; the detection accuracy for surrounding obstacles is 88%, and the recall rate is 83%. The overall mAP value reaches 87%, indicating that the target detection module can accurately identify various target objects in the hydraulic lifting platform working scene, meeting the requirements of actual application for target detection performance.

[0107] (II) Hazardous behavior classification performance test

[0108] 1. Test method: Accuracy and F1 score are used as evaluation indicators for hazardous behavior classification performance. 500 different types of hazardous behavior sample data and 500 normal operation behavior sample data are collected, and the hazardous behavior feature vectors of these sample data are input into the hazardous behavior classification module. The hazardous behavior type and dangerous degree judgment results output by the module are recorded and compared with the true hazardous behavior information labeled by humans to calculate the classification accuracy and F1 score.

[0109] 2. Test results: The classification accuracy of the hazardous behavior classification module reaches 91%, and the F1 score is 0.89. Among them, the classification accuracy for illegal climbing behavior is 95%, and the F1 score is 0.93; the classification accuracy for platform overload behavior is 88%, and the F1 score is 0.85; the classification accuracy for approaching running components behavior is 90%, and the F1 score is 0.88; the classification accuracy for illegal operation control handle behavior is 87%, and the F1 score is 0.84. The test results show that the hazardous behavior classification module can accurately judge the type and degree of dangerous behavior, providing a reliable basis for timely taking corresponding alarm and warning measures.

[0110] (III) Alarm and warning timeliness test

[0111] 1. Test method: By simulating different dangerous behavior scenes, record the response time from the occurrence of dangerous behavior to the triggering of alarm and warning by the system. Test the alarm and warning response time of low, medium and high three dangerous levels respectively, take the average value of 20 tests as the final response time result.

[0112] 2. Test results: The alarm and warning response time of low dangerous level is 1.2 seconds, the response time of medium dangerous level is 0.8 seconds, and the response time of high dangerous level is 0.5 seconds. These response times meet the requirements of the timeliness of alarm and warning in actual application, and can quickly remind the on-site workers and safety supervisors at the moment of dangerous behavior, and take timely measures to avoid accidents.

[0113] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An AI-based method for identifying dangerous behaviors of hydraulic lifting platforms, characterized in that: The following steps are involved: Collect video image data of the hydraulic lifting platform working area in real time through the camera; Preprocess the collected video image data, including image scaling, normalization, and data augmentation operations, to unify the data format and enhance the generalization ability of the model; The preprocessed video image data is input into a deep learning-based object detection model, which includes at least a feature extraction layer, a region proposal network, and a classification and regression layer. The model is used to identify target objects such as operators, components of the hydraulic lifting platform, and surrounding obstacles in the video image, and output the category and location information of each target object. Based on the category and location information of the target objects, the spatial relationship and behavioral characteristics between the target objects are determined, and the dangerous behavior feature vectors including the proximity distance between the operator and the hydraulic lifting platform, the relative movement speed, the operator's posture, and the movement status of the hydraulic lifting platform components are extracted; Inputting the dangerous behavior feature vector into a dangerous behavior classification model based on a support vector machine, the support vector machine dangerous behavior classification model determines whether there is a dangerous behavior in the current video image according to pre-trained classification rules, and outputs the type and degree of dangerous behavior; If there is any dangerous behavior, the corresponding alarm and early warning mechanism will be triggered according to the type and degree of dangerous behavior, and an information prompt including the dangerous location, danger type and recommended measures will be generated and displayed in real time on the monitoring terminal.

2. The AI-based dangerous behavior identification method for hydraulic lifting platforms according to claim 1 is characterized in that: Before preprocessing the collected video image data, a step of distortion correction is also included for the image collected by the camera to eliminate the impact of camera lens distortion on image quality and improve the accuracy of subsequent target detection and behavior recognition.

3. The AI-based dangerous behavior identification method for hydraulic lifting platforms according to claim 1 or 2 is characterized in that: The deep learning-based target detection model adopts the Faster R-CNN architecture. Based on the COCO dataset, which contains a large number of labeled target objects in hydraulic lift platform working scenes, it expands and adds at least 5,000 training images labeled with special components of hydraulic lift platforms and typical dangerous scenes. The model is transfer trained and optimized to improve the model's detection accuracy and recall rate for specific targets of hydraulic lift platforms.

4. The AI-based dangerous behavior identification method for hydraulic lifting platforms according to claim 1 is characterized in that: When extracting the dangerous behavior feature vector, it also includes the step of detecting the skeleton key points of the operator's posture, determining the position coordinates of the main joints of the operator's body based on the posture estimation algorithm, and further analyzing whether the operator has dangerous posture behaviors such as climbing, leaning, and illegal operations based on the angle changes and relative position relationships between the joints, and integrating the posture danger characteristics into the dangerous behavior feature vector.

5. The AI-based method for identifying dangerous behaviors of hydraulic lifting platforms according to claim 1 is characterized in that: The dangerous behavior classification model based on support vector machine adopts radial basis function as kernel function, extracts and labels features of collected dangerous behavior sample data of different types and degrees, and trains the SVM model using sequential minimum optimization algorithm to determine the optimal classification hyperplane, thereby achieving accurate classification of dangerous behaviors and quantitative assessment of the degree of danger.

6. The AI-based dangerous behavior identification method for hydraulic lifting platforms according to claim 1 is characterized in that: The alarm and early warning mechanism includes sound alarms, flashing light alarms, and eye-catching graphic and text prompts that pop up on the monitoring terminal. Different levels of alarm and early warning methods are set according to the degree of danger. When the degree of danger is high, multiple alarm methods are triggered at the same time and displayed in a prominent position on the monitoring terminal first, ensuring that on-site staff can detect the danger in time and take corresponding measures.

7. An AI-based dangerous behavior recognition system for hydraulic lifting platforms, characterized by: include: An image acquisition module is used to collect video image data of the working area of ​​the hydraulic lifting platform in real time through a camera; The data preprocessing module is used to preprocess the collected video image data, including image scaling, normalization and data enhancement operations; The target detection module is used to input the pre-processed video image data into the deep learning-based target detection model to identify target objects such as operators, various components of the hydraulic lifting platform, and surrounding obstacles in the video image, and output the category and location information of each target object; The hazard feature extraction module is used to determine the spatial relationship and behavioral characteristics between target objects based on their categories and location information, and extract the hazard behavior feature vector including the proximity distance between the operator and the hydraulic lift platform, the relative movement speed, the operator's posture, and the movement status of the hydraulic lift platform components; The dangerous behavior classification module is used to input the dangerous behavior feature vector into the dangerous behavior classification model based on the support vector machine, determine whether there is dangerous behavior in the current video image, and output the type and degree of dangerous behavior; The alarm and early warning module is used to trigger the corresponding alarm and early warning mechanism according to the type and degree of dangerous behavior if there is any dangerous behavior, generate information prompts including the dangerous location, danger type and recommended measures, and display them in real time on the monitoring terminal.

8. The AI-based dangerous behavior identification system for hydraulic lifting platforms according to claim 7 is characterized in that: It also includes an image distortion correction module, which performs distortion correction processing on the image collected by the camera before the data preprocessing module to eliminate the influence of camera lens distortion on image quality.

9. The AI-based dangerous behavior identification system for hydraulic lifting platforms according to claim 7 is characterized in that: The dangerous feature extraction module also includes a posture estimation submodule, which is used to detect the skeleton key points of the operator's posture, determine the position coordinates of the main joints of the operator's body based on the posture estimation algorithm, analyze whether the operator has dangerous posture behaviors such as climbing, leaning, and illegal operations, and integrate the posture danger features into the dangerous behavior feature vector.

10. The AI-based dangerous behavior identification system for hydraulic lifting platforms according to claim 7 is characterized in that: The alarm and warning mechanisms in the alarm and warning module include sound alarms, flashing light alarms, and eye-catching graphic and text prompts that pop up on the monitoring terminal. Different levels of alarm and warning methods are set according to the degree of danger. When the degree of danger is high, multiple alarm methods are triggered at the same time and displayed in a prominent position on the monitoring terminal first.