A method and system for intelligent monitoring of flight simulator training effectiveness

Abstract

The present application provides a kind of method and system for the intelligent monitoring of flight simulator training effectiveness, containing 5 steps, 1, identity data generation and training plan import;2, video acquisition and preprocessing;3, identity verification and training state monitoring;4, training duration statistics and data output 5, abnormal early warning and data archiving.The system includes user interaction layer, core processing layer and data storage layer.The present application realizes the identity of training personnel non-sensing verification, accurate statistics of effective training duration, real-time monitoring and duration record of off-duty state by fusing face recognition frame detection, wandering detection technology and low-delay network transmission, combined with the training plan imported in advance, improves the automation and precision level of flight simulator training monitoring, reduces labor cost, and guarantees the authenticity and standardization of training.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, computer vision and video surveillance technology. Specifically, it is a method and system that combines facial recognition and loitering detection technology with a low-latency network camera connection to realize the identity verification of flight simulator trainees, effective training time and off-duty status intelligent monitoring. Background Technology

[0002] Flight simulators are indoor flight simulation devices that simulate various environments and emergencies, playing a vital role in improving flying skills. Because they are conducted indoors, safety is guaranteed. Compared to civil aviation training, military flight simulation training is characterized by higher subject complexity, greater training intensity, stricter supervision requirements, and stronger confidentiality, placing far higher technical demands on the entire training process, high precision, and unmanned intervention monitoring than in general scenarios.

[0003] Currently, the personnel supervision aspect of flight simulator training in the military still generally adopts the traditional model of manual on-duty monitoring and post-training manual statistics: during training, dedicated personnel must be assigned to monitor the simulator cabin throughout the entire process, manually verifying the identities of trainees, recording training duration, and supervising their absence from duty; after training, training data is manually compiled and training records are filled out. This model is not only costly in terms of manpower and inefficient, but also has serious supervisory loopholes: manual monitoring is prone to errors in identity verification, omissions in recording absences, and deviations in duration statistics due to fatigue and negligence, making it impossible to accurately quantify effective training time; at the same time, manual intervention in monitoring can interfere with the pilots' training state, making it difficult to achieve seamless supervision throughout the entire process, and failing to meet the core requirements of "zero interference and full traceability" in military training.

[0004] Existing publicly available intelligent monitoring technologies are mostly designed for general civil aviation training and civilian security scenarios, and have three major flaws that make them unsuitable for the specific needs of military simulated flight training: First, they lack training plan-driven specialized monitoring logic, failing to achieve deep integration with military training outlines and plans, and thus unable to conduct targeted monitoring according to preset training subjects and time periods; second, identity verification and status monitoring technologies are disconnected, with general facial recognition only able to complete single identity verification, unable to achieve continuous facial tracking throughout training, and failing to combine with loitering detection technology to accurately determine off-duty status and duration statistics, making it difficult to prevent violations such as substitute training, unauthorized training, and mid-training absence; third, video transmission and detection lack real-time performance, with traditional monitoring systems experiencing high network latency and low frame extraction efficiency, failing to meet the military's requirements for real-time and low-latency monitoring, and unable to provide real-time early warnings of training anomalies.

[0005] In summary, existing technologies cannot solve the core technical pain points in military flight simulator training, namely "low efficiency of manual supervision, inapplicability of general monitoring, and disconnect between identity and status monitoring." There is an urgent need to develop an integrated intelligent monitoring method and system that is adapted to the actual combat training scenarios of the military, takes the training plan as the core, and integrates technologies such as low-latency streaming media transmission, adaptive adjustment of facial recognition accuracy, continuous facial tracking, and loitering detection. This system can achieve seamless identity verification, accurate statistics of effective training time, and full monitoring of off-duty status, providing technical support for the standardized and intelligent supervision of military flight simulator training. Summary of the Invention

[0006] The purpose of this invention is to address the problems of high reliance on manual labor, general monitoring methods not being relevant to specific scenarios, detection gaps, and high latency in existing flight simulator training monitoring. It proposes a method for intelligent monitoring of the effectiveness of flight simulator training, along with a corresponding system. This system enables seamless verification of trainee identities, accurate statistics of effective training time, real-time monitoring and recording of off-duty status, thereby improving the automation and accuracy of flight simulator training monitoring, reducing labor costs, and ensuring the authenticity and standardization of training.

[0007] The technical solution of this invention: A method for intelligent monitoring of the effectiveness of flight simulator training, comprising the following steps: Step 1: Identity Data Generation and Training Plan Import: Collect facial images of trainees in advance, extract facial feature vectors from each image using a facial feature extraction algorithm, and associate these feature vectors with the trainees' identity information to form an identity feature database; import the flight simulator training plan in advance, specifying the trainees' identity information, preset training duration, and monitoring periods; deploy network cameras and establish low-latency connections with the monitoring system, optimize transmission protocols and data processing flows, and ensure that video stream transmission latency is controlled within a preset range; Step 2, Video Acquisition and Preprocessing: The video stream of the flight simulator training area is acquired in real time through the deployed network cameras. The video stream is preprocessed by noise reduction and deblurring. Image frames are periodically extracted according to a preset frequency. Clear frames are selected for subsequent detection, and invalid frames with blurriness or severe occlusion are removed to ensure detection accuracy. Step 3, Identity Verification and Training Status Monitoring: Combining face recognition frame extraction detection and loitering detection, the entire training process is monitored seamlessly. The extracted image frames are input into the face recognition algorithm and compared with the target personnel features in the identity feature database to verify whether the current trainee is a planned participant. At the same time, through loitering detection, the trainee's dwelling status and movement trajectory within the monitored area are analyzed to determine whether there is any absence from duty. If absence from duty is detected, the timer automatically starts and stops after absence from duty ends, and the total absence time is recorded. Step 4. Training duration statistics and data output: Count the actual effective training duration of trainees in real time, deduct the total off-duty duration, and compare with the preset training duration to judge whether the standard is met; record the identity verification results, effective training duration, number of off-duty instances, and duration of each off-duty instance throughout the whole process, generate a complete training monitoring report after the training ends, and feed it back to the training management terminal synchronously; Step 5. Abnormal warning and data archiving: If abnormal conditions such as failed identity verification, unqualified effective training duration, and excessively long off-duty duration are detected, an alarm prompt is triggered; at the same time, all monitoring data, detection results, and training reports are classified and archived to facilitate subsequent query, traceability, and training management review.

[0008] As a preferred solution: the low-latency connection in Step 1 adopts the ZLMediaKit streaming media service, a ZLMediaKit streaming media server is deployed, the network camera establishes a stable connection with the ZLMediaKit streaming media server, the network camera enables the RTSP streaming push mode, and pushes the real-time collected video stream of the flight simulator training area to the ZLMediaKit streaming media server, the push parameters are configured as: video coding format H264, frame rate 25fps, bit rate 1024kbps, the monitoring system pulls the real-time video stream from the ZLMediaKit streaming media server, enables the low-latency streaming pull mode, configures the streaming pull cache time ≤ 100ms, and detects in real time the total delay of the video stream from collection by the network camera to reception by the monitoring system, the calculation formula is: , where, is the camera collection delay, is the video streaming push delay, is the streaming media cache delay, is the video streaming pull delay; when , the system automatically reduces the bit rate or increases the frame extraction frequency.

[0009] Specifically, preprocessing of the video stream in Step 2: For the real-time video stream pulled from ZLMediaKit, a 5×5 mean filter is used to smooth the picture and suppress Gaussian noise, the mean filter formula is: , if salt-and-pepper noise needs to be processed simultaneously, a median filter is superimposed, where F(x,y) is the pixel value of the original video frame, G(x,y) is the pixel value after noise reduction, a histogram equalization algorithm is used to correct the brightness and contrast of the video frame and improve the image definition; clear frame screening: for the extracted image frames, the frame definition value is calculated based on the image gray variance, and the definition threshold is set to , if the frame definition value is ≥ Th, it is determined as a valid frame for subsequent detection; if it is < Th, it is determined as an invalid frame and discarded directly, the calculation formula is: , where, M and N are the width and height of the image frame, The grayscale value at the (x, y) coordinates of the image frame. This represents the average grayscale value of the image frame.

[0010] Preferably, the facial feature vector in step 3 has 128 feature dimensions, and the trainer's identity information includes name, personnel number, and training plan number; at the same time, a feature update interface is reserved to dynamically update the facial feature vector according to the changes in the trainer's face.

[0011] Step 3: Identity verification and comparison. The calculation formula is as follows: , The facial feature vector of the extracted image frame. Given the facial feature vector of identity data in the identity feature database, if , The similarity threshold is dynamically adjusted. If an identity match is determined, the identity verification information is recorded, a unique face ID is assigned, and the information is synchronized to the loitering detection module. Initiate multi-frame fusion verification, set a 5-frame sliding window, calculate the average similarity of valid frames within the window, and if the average similarity still does not meet the standard, it is determined that the identity does not match and an abnormal signal is sent.

[0012] Calculation: , Assess the quality of the current video frame. The number of times the historical verification has passed is used. When Q < 0.7, the environment is considered poor. When N ≥ 3, the verification has passed continuously.

[0013] Preferably, step 3 involves continuous face tracking, assigning a unique UUID as a face ID to individuals who pass the initial verification, associating the face ID with the facial feature vector and initial position coordinates, storing it in the tracking cache, and synchronously transmitting it to the loitering detection module to achieve a one-to-one correspondence between identity, ID, and trajectory, and performing Kalman filter position prediction and tracking continuity maintenance.

[0014] Step 3: Implementation steps for loitering detection: a) Monitoring area calibration: Set the coordinate range of the monitoring area. , ( x 1, y 1) is the coordinate of the top left corner, ( x 2, y 2) The coordinates are the bottom right corner; at the same time, the tracking cache of the face recognition module is associated to realize the real-time correspondence between the personnel target and the face ID in the monitoring area; b) Loitering Detection Implementation: A target detection algorithm is used to locate personnel targets in valid sampled frames and real-time video streams, obtain the center coordinates (x, y) of the personnel target, and determine in real-time whether (x, y) is within the monitored area; a departure judgment time Th is set. leave=3s, if the target center coordinates of the personnel exceed the area for a duration ≥Th leave If the timeout period is determined to be absent from duty, the absence timeout will be triggered; if the duration is... <Th leave If the activity is deemed normal, the off-duty timer will not be triggered; at the same time, the start and end locations and times of the off-duty departure corresponding to the face ID will be recorded and linked to the tracking trajectory. c) Fusion Judgment Logic: Combining the facial recognition results, if the facial recognition verification passes and the loitering detection determines that the person is off duty, start recording the duration of absence. The time spent away from the post is calculated cumulatively; if facial recognition verification fails, regardless of the loitering detection result, the time spent away from the post will not be recorded, and an identity anomaly warning will be triggered directly; if the person returns to the monitored area, the time spent away from the post will be stopped, and the time spent away from the post will be recorded. Stored in the monitoring data storage module, total off-duty time n represents the number of times the person has left their post. If the identity of the returning person is inconsistent with that before leaving their post, an identity anomaly warning is triggered, the original ID's post-departure timer is terminated, and identity verification and tracking are restarted.

[0015] Step 4: Steps for calculating training duration statistics: a) Training time slot confirmation: Retrieve the preset training time slots for the current trainees from the training plan database. Preset training duration ; b) Actual training duration statistics: Real-time recording of the duration of successful identity verification. That is, the duration from the first successful identity verification to the end of the training period or until the identity verification fails; c) Calculation of effective training time: Effective training time The actual training time minus the total off-duty time is calculated using the following formula: ; d) Duration threshold determination: Set a threshold for achieving the target. ,like If the training duration is deemed to have exceeded the target; If the result is deemed substandard, an abnormal duration warning will be triggered.

[0016] A system for intelligent monitoring of the effectiveness of flight simulator training includes a user interaction layer, a core processing layer, and a data storage layer. The user interaction layer serves as the operation entry point for training administrators. It is used for importing training plans, configuring monitoring parameters, viewing real-time monitoring screens, previewing monitoring data, receiving abnormal warnings, and querying and exporting training monitoring reports. It collects operation instructions from administrators, transmits information such as training plans and parameter configurations to the core processing layer, and receives monitoring data and abnormal information fed back from the core processing layer. The core processing layer includes a ZLMediaKit streaming media processing module, a video preprocessing and frame extraction module, a face recognition module, a loitering detection module, and a duration statistics and anomaly warning module. The ZLMediaKit streaming media processing module contains a ZLMediaKit streaming media server, which establishes a low-latency connection with the network camera for pushing and pulling video streams, configuring streaming media transmission parameters, and enabling real-time reception and caching of surveillance video. The video preprocessing and frame extraction module receives the real-time video stream transmitted by the ZLMediaKit streaming media processing module, performs noise reduction, deblurring, and brightness correction on the video stream, and periodically extracts video frames according to a preset frequency, filtering valid frames and discarding invalid frames. The face recognition module receives the preprocessed valid frame images, calls a face recognition algorithm, and extracts the facial feature vectors of the people in the image frames. The system compares the data with the identity feature database in the data storage layer to determine whether the person is a planned trainee. The loitering detection module simultaneously receives the effective frame images after video preprocessing and the real-time video stream, locates the personnel target in the monitoring area, analyzes the personnel's movement trajectory, dwell time and position coordinates, and determines in real time whether the trainee is in the normal training position and whether the detection personnel have left their post. The duration statistics and anomaly warning module receives the training plan parameters, the identity verification results of the face recognition module, and the post-absence data of the loitering detection module, and calculates the effective training duration in real time. If any abnormal situation occurs, such as identity verification failure, insufficient effective training duration, or excessive post-absence time, an early warning is immediately triggered and the abnormal information is recorded simultaneously. After the training is completed, all monitoring data are integrated to generate a standardized training monitoring report, which is then transmitted to the data storage layer and the user interaction layer. The data storage layer comprises four parts: ① an identity feature database, storing facial feature vectors and corresponding identity information of trainees; ② a training plan database, storing imported flight simulator training plans; ③ a monitoring data storage module, storing pre-processed video streams, valid frame images, identity verification results, off-duty records, and valid training duration; and ④ a streaming media caching module, caching real-time video streams transmitted by the ZLMediaKit streaming media processing module. Simultaneously, an efficient index structure is employed for data querying and retrieval, supporting data traceability and review.

[0017] Furthermore, the face recognition module is also equipped with an accuracy adaptive adjustment unit to adapt to changes in lighting, changes in personnel posture, and partial occlusion in flight simulator training scenarios.

[0018] The beneficial effects of this invention are: reduced labor costs and improved monitoring efficiency: By using ZLMediaKit low-latency video push-pull streaming technology, video preprocessing and frame extraction technology, and face recognition and loitering detection fusion technology, fully automated and seamless monitoring is achieved without the need for dedicated personnel, completely solving the drawbacks of time-consuming and labor-intensive manual monitoring and significantly reducing the labor costs of training and management; at the same time, through multi-frame fusion verification of face recognition, dynamic threshold adjustment and continuous tracking technology, various errors such as identity verification errors, missed absences from duty, and time statistics deviations caused by human negligence are avoided, significantly improving monitoring efficiency and accuracy.

[0019] To achieve precise adaptation to specific scenarios and ensure the authenticity of training: Aligning with flight simulator training scenarios and centered on the training plan, the system uses facial recognition frame extraction detection (including accuracy adjustment and liveness detection) technology to accurately verify the identity of trainees. It also uses a combination of loitering detection and facial recognition technology (including continuous facial tracking and off-duty judgment logic) to monitor off-duty status. Finally, it uses effective training time statistics technology (combining identity verification results and total off-duty time) to accurately calculate effective training time, achieving integrated monitoring and effectively preventing issues such as identity fraud, proxy training, and inflated training durations, thus ensuring the authenticity and standardization of training.

[0020] Enhance the real-time performance and comprehensiveness of monitoring: By using ZLMediaKit low-latency video push-pull streaming technology (including latency detection and optimization, and automatic reconnection mechanism), video transmission latency is controlled within 100ms, ensuring no detection lag; through the deep integration of face recognition and loitering detection, combined with continuous face tracking (Kalman filtering + feature matching) technology, comprehensive coverage of trainee identity, off-duty status, and training duration is achieved. By setting off-duty judgment thresholds and calculating cumulative duration, accurate statistics of off-duty time are achieved, completely solving the problems of incomplete monitoring and detection lag in traditional monitoring.

[0021] Facilitating Refined Training Management: Through effective training duration statistics technology and anomaly warning mechanisms, standardized monitoring reports are generated, including identity verification results, off-duty information, and duration statistics. Combined with persistent storage technology in the data storage layer, monitoring data is queryable and traceable, facilitating training management review. At the same time, through functions such as facial recognition anomaly warnings and duration non-compliance warnings, training anomalies are detected in a timely manner, providing accurate data support for flight simulator training management and promoting the upgrading of training management towards refinement and standardization. Attached Figure Description

[0022] Figure 1 This is the system architecture diagram of the present invention. Detailed Implementation

[0023] This invention relates to a system for intelligent monitoring of the effectiveness of flight simulator training. The system comprises a user interaction layer, a core processing layer, and a data storage layer. These layers work together to achieve functions such as seamless monitoring of the entire flight simulator training process, identity verification, duration statistics, and anomaly warnings. The system relies on ZLMediaKit streaming media service, face recognition frame extraction detection, loitering detection, and face recognition fusion technology. The specific architecture and technical solution are as follows.

[0024] Specific architecture: User Interaction Layer: Provides operation entry for training managers, supporting functions such as importing training plans, configuring monitoring parameters (frame sampling frequency, latency threshold, off-duty judgment threshold, etc.), viewing real-time monitoring screens, previewing monitoring data, receiving anomaly alerts, and querying and exporting training monitoring reports; it is also responsible for collecting operation instructions from managers, transmitting information such as training plans and parameter configurations to the core processing layer, receiving monitoring data and anomaly information from the core processing layer, and synchronously displaying them to managers, thus achieving efficient human-computer interaction.

[0025] Core Processing Layer: As the core of the system, it is responsible for all data processing, analysis, and command execution. It mainly includes the ZLMediaKit streaming media processing module, video preprocessing and frame extraction module, face recognition module, loitering detection module, and duration statistics and anomaly warning module. These modules work collaboratively, with the following specific functions: 1. ZLMediaKit Streaming Media Processing Module: Responsible for establishing low-latency connections with network cameras, pushing and pulling video streams based on ZLMediaKit streaming media services, configuring streaming media transmission parameters, optimizing RTSP / RTP transmission protocols, reducing video stream transmission latency, and enabling real-time reception and caching of surveillance video to provide a stable, low-latency video source for subsequent preprocessing and frame extraction; at the same time, it automatically reconnects to abnormally interrupted video streams to ensure the continuity of monitoring.

[0026] 2. Video Preprocessing and Frame Extraction Module: Receives real-time video streams from the ZLMediaKit streaming media processing module, performs preprocessing operations such as noise reduction, deblurring, and brightness correction on the video stream to remove interference information and improve video clarity; periodically extracts video frames according to a preset frequency (configurable through the user interaction layer, such as 1 frame / second to 5 frames / second), filters valid frames using a frame clarity judgment algorithm (variance-based clarity assessment), and removes invalid frames that are blurry, severely occluded, or too dark / too bright, reducing the computational overhead of subsequent detection and ensuring detection accuracy.

[0027] 3. Face Recognition Module: This module receives pre-processed valid frame images, calls a face recognition algorithm to extract facial feature vectors from the image frames, compares them with the identity feature database in the data storage layer, calculates facial feature similarity, and sets a similarity threshold. When the similarity is greater than or equal to the threshold, the person is identified as a planned training participant, and identity verification is completed. If the similarity is less than the threshold, the person is identified as having an abnormal identity, triggering an anomaly signal to the anomaly warning module. To further improve face recognition accuracy and adapt to complex situations in flight simulator training scenarios such as changing lighting, diverse personnel postures (looking down, turning to view instruments), and partial occlusion, this module integrates an accuracy adaptive adjustment mechanism. It also combines with a loitering detection module to achieve continuous face tracking, ensuring the continuity of identity verification and personnel status monitoring.

[0028] 4. Loitering Detection Module: This module synchronously receives pre-processed valid frame images and real-time video streams. Using a target detection algorithm, it locates personnel targets within the monitored area (flight simulator operating area), analyzes their movement trajectory, dwell time, and location coordinates, and sets a departure threshold (personnel leaving the monitored area for more than 3 seconds is considered off-duty). It also determines in real-time whether trainees are in their normal training positions. If off-duty is detected, a departure signal is immediately sent to the duration statistics module to begin recording the departure time. Upon completion of the departure (person returning to the monitored area), a departure end signal is sent to stop timing and upload the departure time data. This module is deeply integrated with the face recognition module, receiving the unique face ID and tracking information from the face recognition module to achieve a one-to-one correspondence between "face ID and personnel trajectory." During tracking, it synchronously verifies personnel identity to avoid identity confusion in multi-person scenarios, ensuring consistency between departure determination, duration statistics, and identity information.

[0029] 5. Duration Statistics and Anomaly Warning Module: Receives training plan parameters, identity verification results from the face recognition module, and off-duty data from the loitering detection module, and calculates training duration in real time. Based on the preset training period in the training plan, it calculates the actual training duration, deducts the total off-duty time, and obtains the effective training duration. Simultaneously, it compares the effective training duration with the preset training duration to determine if the target has been met. If anomalies occur, such as failed identity verification, insufficient effective training duration, or excessive off-duty time (a single off-duty time exceeding 10 minutes), an early warning is immediately triggered (system message warning), and the anomaly information is recorded synchronously. After training, it integrates all monitoring data, generates a standardized training monitoring report, and transmits it to the data storage layer and user interaction layer.

[0030] Data storage layer: Responsible for persistent storage and efficient retrieval of all data, mainly including four database / storage modules: 1. Identity feature database, which stores the facial feature vectors of trainees and their corresponding identity information, providing a basis for comparison in face recognition; 2. Training plan database, which stores imported flight simulator training plans, including parameters such as trainees, preset training duration, and monitoring periods; 3. Monitoring data storage module, which stores core data such as pre-processed video streams, valid extracted frames, identity verification results, off-duty records, and effective training time; 4. Streaming media caching module: caches the real-time video stream transmitted by the ZLMediaKit streaming media processing module to avoid monitoring loss due to transmission interruption; at the same time, it adopts an efficient index structure to ensure the efficiency of data query and reading, and supports data traceability and review.

[0031] Technical solution: 1. Technology Implementation Process (Fully Intelligent Closed-Loop) This invention provides a method for intelligent monitoring of the effectiveness of flight simulator training. It constructs a five-step intelligent system encompassing video acquisition, preprocessing, identity verification, status monitoring, and duration statistics. Each step relies on dedicated algorithms and modules to achieve efficient collaboration, forming a complete monitoring loop from video transmission to data archiving. The specific process is as follows: S1: Low-latency video capture and push / pull streaming; S2: Video preprocessing and effective frame selection; S3: Facial recognition and continuous tracking throughout the entire process; S4: Fusion of loitering detection and face recognition; S5: Statistics on effective training time and determination of achievement.

[0032] 2. Specific implementation steps: (1) S1: Low-latency video acquisition and push-pull streaming This invention employs the ZLMediaKit streaming media service, with its core function being to achieve low-latency video stream transmission between network cameras and the monitoring system. This provides stable, real-time video input for all subsequent detection and analysis steps. The specific implementation is as follows: S11: Deploy the ZLMediaKit streaming media server, configure the core server parameters, including ports (RTSP port 554, HTTP port 80), buffer size, transmission protocol (RTSP protocol is preferred, UDP transmission mode is optimized), disable redundant functions, reduce server overhead, improve transmission efficiency, and provide hardware and software support for video push and pull streaming. S12: The network camera establishes a stable connection with the ZLMediaKit streaming media server. The camera enables RTSP streaming mode and pushes the real-time video stream of the flight simulator training area to the ZLMediaKit streaming media server. The push parameters are configured as follows: video encoding format H264, frame rate 25fps, bit rate 1024kbps. This reduces video compression loss and bandwidth usage while ensuring video stream quality. S13: The monitoring system pulls real-time video streams from the ZLMediaKit streaming media server through the API interface provided by ZLMediaKit, enables low-latency streaming mode, configures streaming buffer time ≤100ms, and sets up an automatic reconnection mechanism (reconnection interval 3 seconds, maximum number of reconnections 5) to avoid video stream interruption caused by network fluctuations and ensure the continuity of the video stream. S14: Video Stream Transmission Delay Detection and Optimization: Real-time detection of the total delay in video stream transmission from camera acquisition to monitoring system reception, calculated using the following formula: ,in, Due to camera acquisition delay, Due to video streaming delay, For streaming media caching latency, For video streaming delay; when At that time, the system automatically adjusts the bitrate (reduced to 512kbps) or the frame sampling frequency (increased to 5 frames / second) to ensure that the latency is controlled within the preset low range and to ensure the real-time performance of subsequent detection steps.

[0033] (2) S2: Video preprocessing and effective frame selection Video preprocessing and valid frame filtering follow the low-latency real-time video stream output by S1. By extracting frames and filtering valid frames, invalid data is removed and computational overhead is reduced. This is the foundation for ensuring the accuracy of subsequent face recognition and loitering detection. The specific implementation is as follows: S21: Video Preprocessing: For the real-time video stream pulled from ZLMediaKit, a 5×5 mean filter is used to smooth the image and suppress Gaussian noise; if salt-and-pepper noise needs to be processed simultaneously, a median filter is superimposed. The mean filter formula is as follows: Where F(x,y) is the original video frame pixel value, and G(x,y) is the denoised pixel value; at the same time, a histogram equalization algorithm is used to correct the brightness and contrast of the video frame, improve image clarity, and facilitate subsequent facial feature extraction. S22: Frame skipping implementation: A periodic frame skipping strategy is adopted. The frame skipping frequency can be configured through the user interaction layer (default 3 frames / second). The frame skipping logic is as follows: Set the frame skipping time interval. (f is the frame rate), starting from the first frame of the video stream, every [time range]... Extract one frame of image and store it in a temporary buffer; S23: Valid frame screening: For the extracted image frames, the frame sharpness value (based on image grayscale variance) is calculated, and the sharpness threshold is set to , if the frame sharpness value ≥ Th, the frame is determined as a valid frame for subsequent detection; if the frame sharpness value < Th, the frame is determined as an invalid frame and discarded directly, with the calculation formula being: , where, M 、 N are the width and height of the image frame, is the grayscale value of the (x,y) coordinate of the image frame, is the average grayscale value of the image frame.

[0034] (3) S3: Face recognition and continuous full-process tracking This step takes the valid image frames output from S2 as input, which is used to verify the identities of the trainees, while realizing continuous full-process face tracking, generating identity information and face IDs, ensuring that the trainees are personnel within the planned scope. Combined with the accuracy adjustment mechanism and continuous tracking requirements, the specific implementation is as follows: S31: Identity feature database construction: Collect facial images of trainees (3-5 images each for front face and side face) in advance, and extract the facial feature vector of each image through a facial feature extraction algorithm (k is the feature dimension, with a default value of 128 dimensions), associate the feature vectors with the trainee identity information (name, personnel number, training plan number) and store them in the identity feature database; at the same time, a feature update interface is reserved, which can dynamically update the feature vectors according to the facial changes of the trainees (makeup, hairstyle adjustment) to further improve the verification accuracy; S32: Facial feature extraction: For the screened valid frame-extracted images, the same facial feature extraction algorithm is adopted to extract the facial feature vector of the person in the image , during the extraction process, key facial regions (eyes, nose, mouth) are automatically located, and invalid features caused by facial occlusions (masks, hats) are eliminated; the extracted feature vectors are synchronously associated to a temporary tracking cache to provide data support for continuous face tracking; S33: Feature comparison and identity determination: Calculate the cosine similarity between the detected feature vector and the target personnel feature vector in the identity feature database , the similarity threshold is set to , the calculation formula is: , if ( is the dynamically adjusted similarity threshold), it is determined that the identity is matched, the identity verification passing information is recorded, a unique face ID is assigned and synchronized to the loitering detection module; if Initiate multi-frame fusion verification, set a 5-frame sliding window, calculate the average similarity of valid frames within the window, and if the average similarity still does not meet the standard, it is determined that the identity does not match and an abnormal signal is sent.

[0035] S34: Specific implementation of face recognition accuracy adjustment: ① Dynamic similarity threshold calculation: Calculates the quality score of the current video frame in real time. (Based on a comprehensive evaluation of clarity, light intensity, and degree of obstruction, with a value of 0-1), combined with the number of historical verification passes. The threshold is dynamically adjusted using the following formula: ,in, The environment was judged to be poor at that time. When the verification is passed continuously, the lower limit of the similarity threshold is set to 0.9 to avoid false judgments due to excessively low similarity. ② Multi-frame fusion verification: For the valid feature vectors of 5 frames within the sliding window, according to the middle frame... Weight 0.4, two adjacent frames , 0.2 each, edge frames , The fused feature vector is calculated by assigning weights of 0.1 to each feature vector. The formula is: ; ③ Liveness detection implementation: Extract facial dynamic features, detect blinking frequency (normal range 1-3 times / 10 seconds) and head posture change angle (within ±15°). If no dynamic change or fixed posture is detected, it is directly determined as non-liveness, triggering an identity abnormality warning.

[0036] S35: Specific implementation of continuous face tracking: ① Face ID-Trajectory Binding: A unique UUID is assigned as the face ID for each person who passes the initial verification. The ID is associated with the facial feature vector and the initial position coordinates, stored in the tracking cache, and synchronously transmitted to the loitering detection module to achieve a one-to-one correspondence between "identity-ID-trajectory". ② Kalman filter position prediction: based on the personnel position in the previous frame Movement speed Predict the location of people in the current frame The formula is: ,in, The frame interval (default 33ms, corresponding to 30fps) is used to predict the position to narrow the face detection range of the current frame and improve tracking efficiency; ③ Tracking continuity maintenance: If no face with the corresponding ID is detected in the current frame, the detection range is expanded based on the predicted position using Kalman filtering. If no face is detected for 3 consecutive frames, it is determined that the tracking has been lost, and a signal is sent to the loitering detection module to suspend the off-duty timer. After a face is detected again, the face ID is matched by feature comparison to resume tracking and off-duty timer, ensuring tracking continuity.

[0037] (4) S4: Fusion of loitering detection and face recognition This step takes the valid image frames output from S2, the face ID, identity verification results, and tracking information output from S3, and deeply integrates loitering detection and face recognition technologies to achieve monitoring and recording of personnel leaving their posts. The specific implementation is as follows: S41: Monitoring Area Marking During the system deployment phase, the flight simulator training operation area (monitoring area) is calibrated through the user interaction layer, and the coordinate range of the monitoring area is set. (( x 1, y 1) is the coordinate of the top left corner, ( x 2, y 2) The coordinates are the bottom right corner; at the same time, the tracking cache of the face recognition module is associated to realize the real-time correspondence between the personnel target and the face ID in the monitoring area.

[0038] S42: Implementation of loitering detection Using a target detection algorithm, personnel targets are located in valid sampled images and real-time video streams. The center coordinates (x, y) of the personnel targets are obtained, and it is determined in real time whether (x, y) is within the monitored area. A departure judgment time is set. If the target center coordinates of personnel exceed the area for a duration ≥ If the personnel are deemed to have left their post, the off-post timer will be triggered; if the personnel's target center coordinates exceed the duration of the Area, the timer will be triggered. If the activity is deemed normal (looking down, turning to the side), the off-duty timer will not be triggered; at the same time, the start and end positions and times of the off-duty time corresponding to the face ID will be recorded and linked to the tracking trajectory.

[0039] S43: Fusion Decision Logic Based on the facial recognition results, if the facial recognition verification passes and the loitering detection determines that the person has left their post, the duration of absence will be recorded. The time spent away from the post is calculated cumulatively. If facial recognition verification fails, regardless of the loitering detection result, the time spent away from the post will not be recorded, and an identity anomaly warning will be triggered directly. If the person returns to the monitored area (re-enters the Area at the central coordinates), the time spent away from the post will be stopped, and the time spent away from the post will be recorded. Stored in the monitoring data storage module, total off-duty time (n is the number of times the person has left their post); if the identity of the returning person is inconsistent with that before leaving their post, an identity anomaly warning will be triggered, the original ID's leave timer will be terminated, and identity verification and tracking will be restarted.

[0040] S44: Multi-person scene fusion adaptation For scenarios involving multiple people training simultaneously, face ID clustering and trajectory clustering are used to achieve separate tracking and departure determination for multiple individuals. All detected facial features within the monitored area are clustered, with features of the same person grouped into one category and bound to a unique face ID. Simultaneously, personnel trajectories are clustered to ensure that the same trajectory corresponds to a unique face ID, avoiding confusion between multiple identities. The departure status and duration of personnel with different IDs are statistically analyzed separately, without interference, further improving the applicability of fusion detection.

[0041] (5) S5: Statistics of effective training time and determination of achievement This step follows the identity verification results output by step S3 and the total absence time output by step S4. Combined with the pre-imported training plan, it completes the accurate statistics and qualification determination of effective training time. It is the final core output step of the entire monitoring method, and its specific implementation is as follows: S51: Training Period Confirmation Retrieve the preset training time slots for the current trainees from the training plan database. Preset training duration ; S52: Actual Training Duration Statistics Real-time recording of the duration of successful identity verification That is, the duration from the first successful identity verification to the end of the training period or until the identity verification fails; S53: Calculation of Effective Training Duration Effective training duration The actual training time minus the total off-duty time is calculated using the following formula: ; S54: Duration Standard Determination Set target threshold (That is, the effective training time is not less than 90% of the preset time), if If the training duration is deemed to have exceeded the target; If the result is deemed substandard, an abnormal duration warning will be triggered. Specific Implementation

[0042] 1. Taking the training monitoring scenario of a flight simulator at a flight academy training center as an example. (1) System deployment and configuration Deploy the ZLMediaKit streaming media server, configure RTSP port 554, buffer size 100MB, transmission protocol RTSP+UDP, and pull buffer time ≤100ms; Two network cameras (1080P resolution, 25fps) were deployed and installed in front of and to the side of the flight simulator operating area. They were connected to the ZLMediaKit server, RTSP streaming was enabled, and the video encoding format was H.264 with a bitrate of 1024kbps. Deploy a monitoring system, dividing it into a user interaction layer, a core processing layer, and a data storage layer. The core processing layer is configured with a frame extraction frequency of 3 frames / second, a face recognition similarity threshold of 0.9, an absence judgment time of 3 seconds, and a duration threshold of 90% of the preset duration. An identity feature database was constructed, and facial images (3 frontal and 2 side views) of 10 trainees were collected. 128-dimensional facial feature vectors were extracted and associated with the trainees' names, personnel numbers and corresponding training plans.

[0043] (2) Training plan introduction and preparation Import the training plan for the next day through the user interaction layer, select trainee Zhang San (personnel number 001), preset the training time period to 9:00-11:00, and preset the training duration. =120min, the monitoring period is the same as the training period; The system automatically reads Zhang San's facial feature vector to complete the monitoring preparation. The ZLMediaKit server connects with the camera and detects the video stream transmission delay. The initial delay is 85ms, which meets the preset requirement (≤100ms).

[0044] (3) Low-latency video transmission and preprocessing Training begins at 9:00 AM. The camera captures real-time video streams of the flight simulator's operating area and pushes them to the ZLMediaKit server via the RTSP protocol. The monitoring system retrieves the real-time video streams from the server, with transmission latency consistently between 80-95ms.

[0045] The system performs Gaussian filtering for noise reduction and histogram equalization preprocessing on the video stream, extracts image frames at a frequency of 3 frames per second, and each image frame is filtered for sharpness, with the proportion of effective frames reaching 92%, and invalid frames (blurred or occluded) are directly discarded.

[0046] (4) Identity verification, loitering detection and duration statistics Face recognition frame sampling detection: Extract facial feature vectors from valid frame images and compare them with Zhang San's feature vectors. If the similarity is between 0.92 and 0.98, the identity verification is deemed successful, and the identity verification result is recorded in real time. Loitering detection and facial recognition are integrated: The system monitors Zhang San's center coordinates in real time. At 9:30, Zhang San gets up and leaves the operating area. His center coordinates are outside the monitoring area for 4 seconds, which is considered as leaving his post, and the timer starts. At 9:35, Zhang San returns to the operating area, and the timer stops. The duration of this absence is 5 minutes. The system records the absence information and the total absence time. =5min; 10:10 Zhang San left his post again, lasting 3 minutes, and the total time away from his post was updated to 8 minutes; Effective training time statistics: actual training time =112 minutes (from 9:00 to 11:00, excluding two periods of absence from the work station), effective training time =112-8=104min, the preset training time is 120min, the threshold for meeting the standard is 108min. Since 104min<108min, the system triggers a warning that the training time has not met the standard.

[0047] (5) Anomaly warning and data output Training ended (11:00). The system integrated all data and generated a training monitoring report, which included information such as trainer Zhang San, identity verification pass rate of 100%, preset training duration of 120 minutes, actual training duration of 112 minutes, 2 absences from post, total absence time of 8 minutes, effective training time of 104 minutes, and time not meeting the standard. The system pushes monitoring reports to the user interaction layer, where administrators can view detailed monitoring data and play back video clips (cached in the streaming media cache module). At the same time, the data is automatically archived to the data storage layer for easy traceability later.

[0048] 2. Effect Comparison Comparing the traditional manual monitoring method with the method of this invention, under the same training scenario (10 trainees, 120 minutes of training per person), the results are shown in Table 1: Table 1 Comparison between traditional manual monitoring methods and the method of this invention

[0049] Manual monitoring requires two managers to be on duty throughout the process. Identity verification is prone to oversight (false judgment rate of about 8%), and the deviation of off-duty records is relatively large (average deviation ±2 minutes). Effective time statistics are time-consuming (about 15 minutes / person), resulting in high labor costs.

[0050] The method of this invention integrates three major functions—identity verification, effective duration statistics, and off-duty monitoring—with the training plan as its core. It achieves seamless monitoring throughout the entire process without requiring active cooperation from trainees, eliminating the need for dedicated personnel. The false positive rate for identity verification is reduced to below 1%, the deviation in off-duty records is ≤30s, effective duration statistics are automatically completed (real-time output), and video transmission latency is ≤100ms, ensuring strong real-time detection. Simultaneously, training monitoring data is traceable, improving the work efficiency of management personnel by over 80%, significantly reducing labor costs, and effectively ensuring the authenticity and standardization of training.

[0051] This invention solves the following technical problems:

[0052] Existing monitoring suffers from high reliance on manual labor and low efficiency: traditional manual monitoring is time-consuming and labor-intensive, and prone to verification errors and time statistics deviations. This invention achieves seamless monitoring throughout the entire process through automation technology, completely reducing reliance on manual labor and improving monitoring efficiency and accuracy.

[0053] The problem of mismatch between general monitoring and flight simulator training scenarios: Existing intelligent monitoring lacks specialized design and does not achieve targeted monitoring in conjunction with training plans. This invention is tailored to the training needs of flight simulators, achieving functional integration and scenario adaptability, and solving the pain point that the monitoring needs of specialized training scenarios cannot be met.

[0054] The problem of disconnect between identity verification and personnel status detection, and incomplete monitoring: Existing technologies mostly achieve identity verification or duration statistics only, and cannot simultaneously monitor off-duty status. This invention achieves comprehensive monitoring of identity, duration, and off-duty status by integrating two detection technologies.

[0055] The problem of high connection latency and poor real-time detection of network cameras: The problem of poor real-time detection: Existing monitoring systems suffer from high video transmission latency, which leads to detection delays. This invention optimizes low-latency transmission to ensure real-time detection and avoid detection errors and data deviations caused by latency.

[0056] Summary of the key technical points of this invention: 1. Low-latency streaming media push-pull integration technology: Adopting ZLMediaKit streaming media service, optimizing RTSP / RTP transmission protocol, configuring low-latency buffer parameters, realizing stable low-latency connection between network cameras and monitoring systems, supporting automatic reconnection of video streams, ensuring real-time transmission of monitoring video, and controlling the total video transmission latency within a preset range through latency detection and dynamic parameter adjustment, solving the problems of high latency and unstable connection in traditional monitoring, and is the core support for realizing real-time seamless monitoring.

[0057] 2. Face Recognition Frame Extraction and Detection Optimization Technology Based on Integrated Streaming Push and Pull Streams: Based on ZLMediaKit streaming media service, combined with video preprocessing (noise reduction, sharpness correction) and intelligent frame extraction strategies, effective video frames are extracted at a preset frequency, and invalid frames are removed to reduce computational overhead. Through facial feature extraction and similarity comparison algorithms, accurate verification of trainees' identities is achieved. Setting a reasonable similarity threshold balances verification accuracy and detection efficiency, which is the core technology to ensure the authenticity of trainees' identities. An innovative integrated adaptive adjustment mechanism for face recognition accuracy is adopted. Through multi-scale feature extraction, dynamic threshold adaptation, multi-frame fusion verification, and enhanced liveness detection, it adapts to the complex training scenarios of flight simulators and significantly reduces the false positive rate. At the same time, combined with continuous face tracking technology, a Kalman filter + feature matching strategy is used to achieve a unique binding between face ID and personnel trajectory, ensuring the continuity of identity verification and status monitoring throughout the process.

[0058] 3. Loitering Detection and Facial Recognition Fusion Technology: This innovative technology integrates loitering detection (personnel location and trajectory analysis) with facial recognition (identity verification). Combined with continuous facial tracking, it achieves a unique binding of "identity-ID-trajectory," sets scientific thresholds for leaving the post and timing logic, and enables accurate identification of trainees' off-duty status and real-time statistics of off-duty duration. Through secondary identity verification upon leaving and returning to the post, and clustering adaptation for multi-person scenarios, it prevents violations such as substitute work and personnel switching, and solves the technical pain points of disconnect between identity verification and status detection, and identity confusion in multi-person scenarios, further improving the accuracy and practicality of fusion detection.

[0059] 4. Intelligent statistical technology for effective training time based on training plan: Based on the pre-imported training plan, it automatically counts the actual training time and cumulative off-duty time, calculates the effective training time through formula, sets the time target threshold, realizes the time target judgment and abnormal warning, forming a closed loop of "plan-monitoring-statistics-judgment", synchronously records various core data and generates standardized reports, and solves the problems of large deviation and low efficiency of manual statistics.

[0060] Overall effects of the invention: 1. High real-time monitoring and stability: Through ZLMediaKit low-latency streaming media technology, the video transmission latency is controlled within 100ms, realizing real-time and imperceptible monitoring of the training process. The automatic reconnection mechanism of the video stream avoids monitoring interruption. Compared with traditional monitoring, the latency is reduced by more than 60%, and the monitoring stability is significantly improved, ensuring that situations such as leaving the post or abnormal identity can be detected in a timely manner.

[0061] 2. High accuracy and targeted monitoring: Face recognition frame sampling detection combined with preprocessing technology and an accuracy adaptive adjustment mechanism, through optimizations such as dynamic thresholds, multi-frame fusion, and liveness detection, reduces the false judgment rate of identity verification to below 0.5%, effectively resisting environmental interference and deception. The fusion design of loitering detection and face recognition, combined with continuous face tracking, accurately distinguishes between normal activities and off-duty behavior, achieving full binding of identity and trajectory, with off-duty record deviation ≤30s, no identity confusion in multi-person scenarios, and effectively avoiding misjudgment and missed judgment. The duration statistics are aligned with the training plan, accurately deducting off-duty time to ensure the accuracy of effective training time statistics, perfectly adapting to flight simulator training scenarios.

[0062] 3. Significantly reduce labor costs and improve management efficiency: The entire process is automated and seamless, requiring no dedicated personnel to be on duty. Compared with traditional manual monitoring, it can reduce the investment of management personnel by more than 80%. It automatically completes identity verification, duration statistics, abnormal warnings and report generation, avoiding the tedious process of manual statistics and recording, and greatly improving the automation and precision of flight simulator training management.

[0063] 4. Ensure the authenticity and standardization of training: Accurate identity verification will prevent violations such as substitute training and impersonation; monitoring off-duty personnel and statistics on effective training time will urge trainees to complete training according to plan and avoid training becoming a mere formality; all monitoring data is traceable and reviewable, providing reliable data support for training management, helping the standardization of flight training in the armed forces, and indirectly ensuring subsequent flight safety.

Claims

1. A method for intelligent monitoring of the effectiveness of flight simulator training, comprising the following steps: Step 1: Identity Data Generation and Training Plan Import: Collect facial images of trainees in advance, extract facial feature vectors from each image using a facial feature extraction algorithm, and associate the feature vectors with the trainees' identity information to form identity feature database. Import the flight simulator training plan in advance, clearly defining the trainees' identities, preset training duration, and monitoring periods; Deploy network cameras and establish low-latency connections with the monitoring system, optimize transmission protocols and data processing flows, and ensure that video stream transmission latency is controlled within a preset range; Step 2, Video Acquisition and Preprocessing: The video stream of the flight simulator training area is acquired in real time through the deployed network cameras. The video stream is preprocessed by noise reduction and deblurring. Image frames are periodically extracted according to a preset frequency. Clear frames are selected for subsequent detection, and invalid frames with blurriness or severe occlusion are removed to ensure detection accuracy. Step 3, Identity Verification and Training Status Monitoring: Combining face recognition frame extraction detection and loitering detection, the entire training process is monitored seamlessly. The extracted image frames are input into the face recognition algorithm and compared with the target personnel features in the identity feature database to verify whether the current trainee is a planned participant. At the same time, through loitering detection, the trainee's dwelling status and movement trajectory within the monitored area are analyzed to determine whether there is any absence from duty. If absence from duty is detected, the timer automatically starts and stops after absence from duty ends, and the total absence time is recorded. Step 4: Training Duration Statistics and Data Output: Real-time statistics of the actual effective training time of trainees, deducting the total off-duty time, comparing with the preset training time to determine whether the standard has been met; record the identity verification results, effective training time, number of off-duty times, and duration of each off-duty time throughout the process, generate a complete training monitoring report after the training ends, and synchronously feed it back to the training management terminal; Step 5, Anomaly Warning and Data Archiving: If anomalies are detected such as failed identity verification, insufficient effective training time, or excessive absence from duty, an early warning will be triggered. At the same time, all monitoring data, test results, and training reports are classified and archived to facilitate subsequent querying, tracing, and training management review.

2. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 1, characterized in that: Step 1's low-latency connection utilizes the ZLMediaKit streaming service. A ZLMediaKit streaming server is deployed, and the network camera establishes a stable connection with it. The network camera is set to RTSP streaming mode, pushing the real-time video stream from the flight simulator training area to the ZLMediaKit streaming server. Push parameters are configured as follows: H264 video encoding format, 25fps frame rate, and 1024kbps bitrate. The monitoring system retrieves the real-time video stream from the ZLMediaKit streaming server, enabling low-latency streaming mode with a buffer time of ≤100ms. The total latency from the network camera's capture to the monitoring system's reception is monitored in real-time and calculated using the following formula: ,in, Due to camera acquisition delay, Due to video streaming delay, For streaming media caching latency, For video streaming delay; when At this time, the system automatically reduces the bitrate or increases the frame skipping frequency.

3. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 1, characterized in that: Step 2: Video stream preprocessing: The real-time video stream pulled by ZLMediaKit is smoothed using a 5×5 mean filter to suppress Gaussian noise. The mean filter formula is: , If salt-and-pepper noise needs to be processed simultaneously, superimpose median filtering, where F(x,y) is the pixel value of the original video frame, and G(x,y) is the pixel value after noise reduction. The histogram equalization algorithm is used to correct the brightness and contrast of the video frame and improve the image definition; clear frame screening: for the extracted image frames, the frame definition value is calculated based on the image gray variance, and the definition threshold is set to , if the frame definition value is ≥ Th, it is determined as a valid frame for subsequent detection; if it is < Th, it is determined as an invalid frame and discarded directly, with the calculation formula being: , where, M and N are the width and height of the image frame, is the gray value at the (x,y) coordinate of the image frame, is the average gray value of the image frame.

4. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 1, characterized in that: Step 3: The facial feature vector has 128 feature dimensions. The trainer's identity information includes name, personnel number, and training plan number. At the same time, a feature update interface is reserved to dynamically update the facial feature vector based on changes in the trainer's face.

5. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 1, characterized in that: Step 3: Identity verification and comparison. The calculation formula is as follows: , This refers to the facial feature vector extracted from the image frame. Given the facial feature vector of identity data in the identity feature database, if , The dynamically adjusted similarity threshold is used to determine if an identity match is found, the identity verification information is recorded, a unique face ID is assigned, and the information is synchronized to the loitering detection module; if... Initiate multi-frame fusion verification, set a 5-frame sliding window, calculate the average similarity of valid frames within the window, and if the average similarity still does not meet the standard, it is determined that the identity does not match and an abnormal signal is sent. Calculation: , Assess the quality of the current video frame. The number of times the historical verification has passed is used. When Q < 0.7, the environment is considered poor. When N ≥ 3, the verification has passed continuously.

6. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 1, characterized in that: Step 3 involves continuous face tracking. A unique UUID is assigned as the face ID to the person who passes the initial verification. The face ID is associated with the facial feature vector and the initial position coordinates, stored in the tracking cache, and synchronously transmitted to the loitering detection module to achieve a one-to-one correspondence between identity, ID, and trajectory. Kalman filter position prediction and tracking continuity maintenance are then performed.

7. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 6, characterized in that: Step 3: Implementation steps for loitering detection: a) Monitoring area calibration: Set the coordinate range of the monitoring area. , ( x 1, y 1) is the coordinate of the top left corner, ( x 2, y 2) The coordinates are the bottom right corner; at the same time, the tracking cache of the face recognition module is associated to realize the real-time correspondence between the personnel target and the face ID in the monitoring area; b) Loitering Detection Implementation: A target detection algorithm is used to locate personnel targets in valid sampled frames and real-time video streams, obtain the center coordinates (x, y) of the personnel target, and determine in real-time whether (x, y) is within the monitored area; a departure judgment time Th is set. leave =3s, if the target center coordinates of the personnel exceed the area for a duration ≥Th leave If the timeout period is determined to be absent from duty, the absence timeout will be triggered; if the duration is... <Th leave If the activity is deemed normal, the off-duty timer will not be triggered; at the same time, the start and end locations and times of the off-duty departure corresponding to the face ID will be recorded and linked to the tracking trajectory. c) Fusion Judgment Logic: Combining the facial recognition results, if the facial recognition verification passes and the loitering detection determines that the person is off duty, start recording the duration of absence. The duration of absence from duty is calculated cumulatively; If facial recognition verification fails, regardless of the loitering detection result, the time spent away from the post will not be recorded, and an identity abnormality warning will be triggered directly; If personnel return to the monitored area, the off-duty timer will stop, and the duration of this off-duty period will be recorded. Stored in the monitoring data storage module, total absence time n represents the number of times the employee has left their post. If the identity of the returning personnel is inconsistent with that before leaving their post, an identity anomaly warning will be triggered, the original ID's post-departure timer will be terminated, and identity verification and tracking will be restarted.

8. The method for intelligent monitoring of the effectiveness of flight simulator training as described in claim 6, characterized in that: step 4. Steps for implementing training duration statistics: a) Training time slot confirmation: Retrieve the preset training time slots for the current trainees from the training plan database. Preset training duration ; b) Actual training duration statistics: Real-time recording of the duration of successful identity verification. That is, the duration from the first successful identity verification to the end of the training period or until the identity verification fails; c) Calculation of effective training time: Effective training time The actual training time minus the total off-duty time is calculated using the following formula: ; d) Duration threshold determination: Set a threshold for achieving the target. ,like If the training duration is deemed to have exceeded the target; If the result is deemed substandard, an abnormal duration warning will be triggered.

9. A system for intelligent monitoring of the effectiveness of flight simulator training, characterized in that: It includes a user interaction layer, a core processing layer, and a data storage layer; The user interaction layer serves as the operation entry point for training administrators. It is used for importing training plans, configuring monitoring parameters, viewing real-time monitoring screens, previewing monitoring data, receiving abnormal warnings, and querying and exporting training monitoring reports. It collects operation instructions from administrators, transmits information such as training plans and parameter configurations to the core processing layer, and receives monitoring data and abnormal information fed back from the core processing layer. The core processing layer includes a ZLMediaKit streaming media processing module, a video preprocessing and frame extraction module, a face recognition module, a loitering detection module, and a duration statistics and anomaly warning module. The ZLMediaKit streaming media processing module contains a ZLMediaKit streaming media server, which establishes a low-latency connection with the network camera for pushing and pulling video streams, configuring streaming media transmission parameters, and realizing real-time reception and caching of monitoring videos. The video preprocessing and frame extraction module receives the real-time video stream transmitted by the ZLMediaKit streaming media processing module, performs noise reduction, deblurring, and brightness correction on the video stream, and periodically extracts video frames according to a preset frequency, filtering valid frames and discarding invalid frames. The face recognition module receives the preprocessed valid frame images, calls the face recognition algorithm, extracts the facial feature vectors of the people in the image frames, compares them with the identity feature database in the data storage layer, and determines whether they are planned training personnel. The loitering detection module simultaneously receives the preprocessed valid frame images and the real-time video stream, locates personnel targets within the monitoring area, analyzes the personnel's movement trajectory, dwell time, and location coordinates, and determines in real time whether the training personnel are in the normal training position and whether the detection personnel have left their posts. The duration statistics and anomaly warning module receives training plan parameters, identity verification results from the face recognition module, and off-duty data from the loitering detection module, and calculates the effective training duration in real time. If any abnormal situation occurs, such as failure to pass identity verification, insufficient effective training time, or excessive absence from duty, an early warning will be triggered immediately, and the abnormal information will be recorded simultaneously. After training, all monitoring data is integrated to generate a standardized training monitoring report, which is then transmitted to the data storage layer and the user interaction layer. The data storage layer comprises four parts: ① an identity feature database, which stores the facial feature vectors of trainees and their corresponding identity information; ② a training plan database, which stores the imported flight simulator training plans; ③ Monitoring data storage module, which stores pre-processed video streams, valid frame images, identity verification results, off-duty records, and valid training time; ④ The streaming media caching module caches the real-time video stream transmitted by the ZLMediaKit streaming media processing module; it also uses an efficient index structure to perform data query and read, and supports data traceability and review.

10. The system for intelligent monitoring of the effectiveness of flight simulator training according to claim 9, characterized in that: The face recognition module is also equipped with an accuracy adaptive adjustment unit to adapt to changes in lighting, changes in personnel posture, and partial occlusion in flight simulator training scenarios.

Images