A monitoring system and method for a flight assurance node

Abstract

The application discloses a kind of flight guarantee node monitoring system and method, and monitoring system specifically includes: acquisition module obtains flight guarantee node strategy and the video sequence of each monitoring point;Behavior analysis module is through the target detection of fusing multi-path network and location attention, the analysis of video sequence, identifies multi-scale target;And, through the space perception function based on the category gap scale index to multi-scale target is corrected, obtains flight guarantee target and guarantee target node;Event construction module is based on the matching analysis of flight guarantee target, guarantee target node and flight guarantee node strategy, constructs flight guarantee node event. Flight guarantee target and corresponding guarantee target node are determined by target correction, to overcome the flight detection error caused by environmental change, with the aid of node judgment model, flight guarantee node event is constructed, realizes the real-time and accuracy of flight guarantee node monitoring.

Description

Technical Field

[0001] This invention belongs to the field of aviation data analysis technology, specifically relating to a monitoring system and method for flight support nodes. Background Technology

[0002] Airport aprons are crucial areas for aircraft ground support operations, and their operational safety and efficiency directly impact flight quality and aviation safety. Apron support involves the coordinated work of various personnel, vehicles, and support equipment, resulting in complex workflows and frequent critical incidents, placing high demands on real-time and accurate event data collection. Traditional flight support data collection relies primarily on manual entry and walkie-talkie notifications, making it difficult to guarantee information consistency and timeliness, easily leading to data delays or errors, and affecting flight scheduling and operational efficiency.

[0003] Furthermore, most current flight support monitoring systems rely on traditional fixed-rule algorithms or image recognition technology, using feature extraction and analysis of single-frame images to determine the status of a particular operation. While these methods can capture the static state of operational nodes, they cannot effectively identify dynamic changes and potential violations during the operational process. Especially in abnormal situations such as improper equipment placement or personnel shortages, traditional monitoring systems often struggle to make timely and accurate judgments, posing safety risks.

[0004] For example, patent CN117523500A discloses a monitoring system, method, and storage medium for flight support nodes. The monitoring system includes: a video acquisition module that collects real-time video streams from perspectives such as parking positions and aircraft doors within the flight support area; a positioning module that collects real-time positioning information related to personnel, aircraft, and ground moving targets within the flight support area; a real-time video analysis module that inputs real-time video frames from various perspectives at the current moment into a trained machine learning model to obtain the support node status, such as aircraft entry / exit status, door opening / closing status, and ground service status, corresponding to each perspective at the current moment; and calibrates the support node status based on the real-time positioning information collected at the current moment to obtain a high-precision support node status; and a data integration module that generates and reports an XML message of the aircraft at the current moment based on the high-precision support node status, flight identifier, and current time. This solution can effectively improve the accuracy and real-time performance of flight support node monitoring.

[0005] However, flight monitoring videos often contain multiple targets with large differences in scale, such as aircraft, ground equipment, and personnel. Using the existing technologies described above for target identification, localization, and node status judgment can easily lead to the inability to identify targets or to identify them incorrectly in harsh environments. It can also easily lead to inaccurate or even missing target localization, affecting the real-time performance and accuracy of flight support node monitoring.

[0006] Therefore, how to overcome flight detection errors caused by environmental changes and realize the automated collection, release and presentation of flight support node data to ensure the real-time and accuracy of flight support node monitoring is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the shortcomings of the existing technology, this invention provides a monitoring system and method for flight support nodes. The monitoring system specifically includes: a data acquisition module for acquiring flight support node strategies and video sequences of various monitoring points; a behavior analysis module for analyzing the video sequences by fusing multi-path network and positional attention target detection to identify multi-scale targets; and for correcting the multi-scale targets using a spatial perception function based on the category gap scaling index to obtain flight support targets and support target nodes; and an event construction module for constructing flight support node events based on matching analysis of flight support targets, support target nodes, and flight support node strategies.

[0008] By integrating multi-path network and positional attention target detection, video sequences from various monitoring points are analyzed to identify targets at multiple scales. Spatial perception functions are then used to correct these targets, thus determining flight support targets and overcoming flight detection errors caused by environmental changes. Furthermore, continuous time step analysis is used to further determine the support target nodes corresponding to the flight support targets. By leveraging a node judgment model, flight support node events are constructed, expanding the application environment of flight support node monitoring and achieving real-time and accurate monitoring of flight support nodes.

[0009] In a first aspect, the present invention provides a monitoring system for flight support nodes, specifically comprising:

[0010] The acquisition module is used to acquire flight support node strategies and video sequences from various monitoring points;

[0011] The behavior analysis module is used to analyze video sequences and identify targets at multiple scales by fusing multi-path networks and positional attention for target detection.

[0012] Furthermore, by using a spatial perception function based on the category gap scaling index to correct multi-scale targets, flight support targets and support target nodes are obtained;

[0013] The event construction module is used to construct flight support node events based on the matching analysis of flight support targets, support target nodes, and flight support node strategies.

[0014] Furthermore, the video sequences from each monitoring point are obtained through the following steps:

[0015] Collect initial video data from preset monitoring points;

[0016] Based on the time interval of the event trigger, extract the key video streams from the initial video data to form multiple video clips;

[0017] The various video clips are combined to form a video sequence for each monitoring point.

[0018] Furthermore, by fusing multi-path networks and positional attention for target detection, the analysis of video sequences identifies targets at multiple scales, specifically including:

[0019] Read the video sequences from each monitoring point to obtain multiple video frame images;

[0020] The identification range is given based on the target area of ​​flight support;

[0021] By identifying the range, each video frame image is cropped to obtain the key images of the video frame;

[0022] By using a multi-scale feature extraction pathway, image targets at different scales are extracted from key images of video frames to obtain image features at different scales.

[0023] By employing a multi-feature fusion matching strategy that combines location enhancement, image features at different scales are fused and analyzed, and image output vectors are provided for each image target.

[0024] Furthermore, a multi-feature fusion matching strategy incorporating location enhancement is used to perform fusion analysis on image features at different scales, providing image output vectors for each image target, specifically including:

[0025] By combining scale weights, image features of different scales in the key images of each video frame are fused to obtain fused features;

[0026] Based on the feature location of the fused features in the video frame image, and combined with the position weight, the fused features are enhanced to form the enhanced feature vector of each image target;

[0027] Based on a pre-trained target detection head, target localization and judgment are performed on the enhanced feature vectors of each image target, and the image output vector of the image target in each video frame is given.

[0028] Furthermore, the image output vector includes the spatial location parameters of the image target, the target category, and the detection confidence.

[0029] By refining multi-scale targets using a spatial perception function based on the category gap scaling index, flight support targets and target nodes are obtained, specifically including:

[0030] Based on the target category of the image object and the number of target categories, a classification calibration function is set;

[0031] By scaling the classification calibration function and combining it with the gap enhancement intensity, a confidence correction function is determined to correct the detection confidence of the image target.

[0032] Based on the corrected detection confidence, flight support targets are determined from the image output vector of the image targets;

[0033] Based on the target area of ​​flight support, the image output vectors of each flight support target are judged, and the judgment result is given;

[0034] Based on continuous time steps, the results of each judgment are merged to form a time state sequence;

[0035] Based on statistical analysis of time state sequences, the motion state of flight support targets is determined, and the corresponding support target nodes are given.

[0036] The classification calibration function is specifically expressed as follows:

[0037]

[0038] Where Cal() is the classification calibration function, Let N be the detection confidence level before correction for target e under target category number c, where 1, ..., C are all target category numbers, A is the number of target categories, bg is the background category number, and N is the target category number. c β represents the number of samples corresponding to the target class number c in the training dataset, and β is the logarithmic basis.

[0039] Furthermore, the confidence correction function is obtained through the following steps:

[0040] By analyzing the distribution of the number of each target category in the training dataset, the class distribution gap of each target category is determined;

[0041] Based on the class distribution gapness, and combined with different window sizes, multi-scale analysis is performed to determine the class gapness scaling index for each target class.

[0042] Scale mapping is performed on the image output vector to obtain the scale mapping parameters for each target category;

[0043] Based on the scale mapping parameters and combined with the standardized category gap scaling index, the spatial perception calibration parameters are determined through the spatial perception function.

[0044] The confidence correction function is constructed by normalizing the spatial perception calibration parameters of different target categories.

[0045] Furthermore, the category gap scaling index is specifically expressed as:

[0046]

[0047] in, The category gap scaling index corresponding to the target category number c. The category distribution gap size corresponding to the target category number c, N c Let c be the number of samples corresponding to the target category number c in the training dataset, O() is the order of magnitude function, and r is the window side length.

[0048] Furthermore, the spatial perception function is specifically expressed as:

[0049]

[0050] Where S() is the spatial perception function and Cal() is the classification calibration function. Let σ() be the detection confidence level before correction for flight support target e under target category number c, and let σ() be the scaling function. Let c be the standardized category void scaling index, α be the void enhancement intensity, 1,…,C be all the target category numbers, and bg be the background category number.

[0051] Furthermore, the confidence correction function is specifically expressed as:

[0052]

[0053] in, The corrected detection confidence level for flight support target e under target category number c. Let S() be the detection confidence level of flight support target e under target category number c before correction, where S() is the spatial perception function and Cal() is the classification calibration function.

[0054] Furthermore, based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed, specifically including:

[0055] Obtain flight support targets, support target nodes, and flight support node strategies. The flight support node strategies include standard support node events corresponding to each flight support target.

[0056] Based on a pre-built node judgment model, a matching analysis is performed on the flight support target and flight support node strategy at the target node.

[0057] Based on the results of the matching analysis, flight support node events are constructed.

[0058] Furthermore, the node judgment model specifically includes: behavior unit, probability calculation unit, and rule unit;

[0059] The behavioral unit is used to analyze the relationship between the flight support target and the standard support node events based on the flight support target and the support target node, and to obtain the current node status.

[0060] The probability calculation unit is used to calculate the probability distribution of the current node state based on the probability density function of the target node and the current node state, and to give the probability distribution likelihood value of the target node.

[0061] The rule unit is used to judge the probability distribution likelihood value of the target node through a preset logical expression relationship, and give the matching result of the flight support target between the target node and the flight support node.

[0062] Furthermore, the probability density function is a Gaussian mixture model, obtained through the following steps:

[0063] Based on each protection node, a Gaussian mixture model is used to model the state distribution of the protection nodes to obtain an initial distribution model;

[0064] Calculate the posterior responsibility degree corresponding to the Gaussian component based on the Gaussian component of each guarantee node;

[0065] By combining the preset learning rate and the posterior responsibility corresponding to the Gaussian components, the model parameters in the initial distribution model are updated to determine the Gaussian mixture model. The model parameters include the component weights, center positions, and covariance matrices of the Gaussian components.

[0066] Secondly, the present invention also provides a method for monitoring flight support nodes, employing the flight support node monitoring system described above, specifically including the following steps:

[0067] Obtain flight support node strategies and video sequences from each monitoring point;

[0068] By integrating multi-path network and positional attention for target detection, the analysis of video sequences identifies targets at multiple scales.

[0069] Furthermore, by using a spatial perception function based on the category gap scaling index to correct multi-scale targets, flight support targets and support target nodes are obtained;

[0070] Based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed.

[0071] The monitoring system and method for flight support nodes provided by this invention have at least the following beneficial effects:

[0072] (1) By integrating multi-path network and positional attention target detection, the video sequences of each monitoring point are analyzed to give multi-scale targets. The multi-scale targets are corrected by spatial perception function to determine flight support targets, overcome flight detection errors caused by environmental changes, and further determine the support target nodes corresponding to the flight support targets through continuous time step analysis and judgment. With the help of the node judgment model, flight support node events are constructed, expanding the application environment of flight support node monitoring and realizing the real-time and accuracy of flight support node monitoring.

[0073] (2) Based on multi-path feature extraction, a position attention module is introduced to model the spatial position information of the target in the key image of the video frame. By learning the spatial relationship between the target area and the background area, the key area features are weighted and enhanced, thereby improving the positioning accuracy and detection robustness of the target in complex backgrounds. Attached Figure Description

[0074] Figure 1 This is a structural diagram of a monitoring system for flight support nodes provided in an embodiment of the present invention;

[0075] Figure 2 A flowchart for identifying multi-scale targets provided in an embodiment of the present invention;

[0076] Figure 3 A flowchart providing an image output vector is provided for embodiments of the present invention;

[0077] Figure 4 A flowchart for determining flight support targets and target nodes provided in an embodiment of the present invention;

[0078] Figure 5 This is a flowchart of constructing a confidence correction function provided in an embodiment of the present invention;

[0079] Figure 6 An architecture diagram of target detection and target correction provided in an embodiment of the present invention;

[0080] Figure 7 A flowchart for constructing flight support node events provided in an embodiment of the present invention;

[0081] Figure 8 This is a flowchart illustrating a method for monitoring flight support nodes provided in an embodiment of the present invention.

[0082] Among them, 201 is the data acquisition module; 202 is the behavior analysis module; and 203 is the event construction module. Detailed Implementation

[0083] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0084] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0085] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0086] Existing methods for collecting apron support events mainly include manual inspection, rule-based triggering mechanisms, and target detection and event recognition methods based on video analytics. Among these, automated detection technology based on computer vision is gradually becoming the mainstream solution, but it still faces several technical challenges in practical applications. In apron scenarios, there are numerous target categories with significantly varying frequencies of occurrence. Some key support targets and abnormal events have relatively few samples, exhibiting a clear long-tail distribution characteristic. This leads to insufficient accuracy and stability of existing models for detecting fewer sample categories, making them prone to missed or false detections. Existing systems typically use fixed rules or predefined logic to trigger event judgments based on detection results, failing to effectively consider the uncertainty of detection results and state transition relationships. This makes them susceptible to false alarms or omissions due to detection fluctuations, occlusion, or environmental changes, affecting the reliability and practicality of the apron support system.

[0087] The targets in apron operations vary considerably in terms of spatial scale and temporal duration. Some targets exhibit ultra-large or ultra-small scale characteristics, and their state changes show strong temporal correlation. Existing methods mostly rely on detection results within a single frame or short period of time for judgment, lacking modeling of the target state evolution process, making it difficult to continuously and stably distinguish complex support events.

[0088] Based on this, the present invention provides a monitoring system for flight support nodes, specifically including: a data acquisition module 201 for acquiring flight support node strategies and video sequences of various monitoring points; a behavior analysis module 202 for analyzing the video sequences by fusing multi-path network and positional attention target detection to identify multi-scale targets; and for correcting multi-scale targets using a spatial perception function based on the category gap scaling index to obtain flight support targets and support target nodes; and an event construction module 203 for constructing flight support node events based on matching analysis of flight support targets, support target nodes, and flight support node strategies. By using multi-scale target detection that fuses multi-path network and positional attention to provide multiple image targets, and combining this with a spatial perception function based on the category gap scaling index to correct multi-scale targets, the system determines flight support targets, overcomes flight detection errors caused by environmental changes, and further determines the support target nodes corresponding to the flight support targets through continuous time step analysis and judgment. By using a node judgment model, the system constructs flight support node events, expanding the application environment of flight support node monitoring and achieving real-time and accurate monitoring of flight support nodes.

[0089] The monitoring system for flight support nodes based on intelligent video analytics deploys high-definition cameras in key areas of the airport and utilizes artificial intelligence and video image recognition technology to conduct real-time monitoring, event detection, identification, and analysis of support nodes during the ground support process.

[0090] In addition, the monitoring system for flight support nodes can collect and analyze data from support nodes in real time, and automatically record the time of key operations such as aircraft positioning, wheel chock operation, jet bridge docking, refueling, and catering, reducing human intervention and ensuring efficient execution of operations.

[0091] like Figure 1 As shown, this embodiment of the invention provides a monitoring system for flight support nodes, specifically including a data acquisition module 201, a behavior analysis module 202, and an event construction module 203. The data acquisition module 201, behavior analysis module 202, and event construction module 203 are communicatively connected. Specifically, the data acquisition module 201 acquires flight support node strategies and video sequences of various monitoring points; the behavior analysis module 202 analyzes the video sequences by fusing multi-path network and positional attention target detection to identify multi-scale targets; and, by using a spatial perception function based on the category gap scaling index, it corrects the multi-scale targets to obtain flight support targets and support target nodes; the event construction module 203 constructs flight support node events based on the matching analysis of flight support targets, support target nodes, and flight support node strategies.

[0092] Specifically, the video sequences of each monitoring point are obtained through the following steps:

[0093] Collect initial video data from preset monitoring points;

[0094] Based on the time interval of the event trigger, extract the key video streams from the initial video data to form multiple video clips;

[0095] The various video clips are combined to form a video sequence for each monitoring point.

[0096] In one specific implementation, during airport apron operations, fixed or pan-tilt-zoom (PTZ) video surveillance equipment deployed at various monitoring points in the apron area continuously captures video of the entire flight support operation process, obtaining initial video data. Based on preset operation area configurations and camera parameters, the captured video stream is analyzed in real time. Key video streams are extracted from the video stream according to preset event triggering time intervals, forming multiple video segments. The video segments from each monitoring point are then aggregated using time frames to form a video sequence for each monitoring point. It is understood that different flight support node events will be triggered at different monitoring points and at different time intervals. To improve the efficiency of video data analysis, only key video streams related to triggering flight support node events are extracted. For example, at a certain monitoring point, a flight support node event is triggered every 60 minutes, and this triggering process lasts for 10 minutes. Therefore, for that monitoring point, ten minutes of key video stream are extracted from the initial video data every 60 minutes.

[0097] Furthermore, by fusing multi-path networks and positional attention for target detection, and analyzing video sequences, multi-scale targets are identified, referencing... Figure 2 Specifically, it includes:

[0098] Read the video sequences from each monitoring point to obtain multiple video frame images;

[0099] The identification range is given based on the target area of ​​flight support;

[0100] By identifying the range, each video frame image is cropped to obtain the key images of the video frame;

[0101] By using a multi-scale feature extraction pathway, image targets at different scales are extracted from key images of video frames to obtain image features at different scales.

[0102] By combining a location-enhanced multi-feature fusion matching strategy, image features at different scales are fused and analyzed, and image output vectors for each image target are given, thus completing the recognition of multi-scale targets.

[0103] In one specific implementation, video sequences from various monitoring points are read and images are extracted to obtain multiple video frame images. Based on the differences between each monitoring point and the target area for flight support, a recognition range corresponding to each monitoring point is determined. Using this recognition range, each video frame image is cropped to obtain key images. For example, based on the camera's perspective (i.e., different monitoring points) and a pre-configured apron operation area (i.e., the target area for flight support), the video frame images are cropped to obtain key images. This means only image information within the target area for flight support is retained, while irrelevant background areas are removed. The obtained key images are numbered and stored in chronological order and associated with the camera numbers corresponding to the monitoring points, forming an image sequence for each monitoring point, used for subsequent target element detection and key event identification.

[0104] Furthermore, by employing a multi-feature fusion matching strategy combined with location enhancement, image features at different scales are fused and analyzed to provide image output vectors for each image target, referring to... Figure 3 Specifically, it includes:

[0105] By combining scale weights, image features of different scales in the key images of each video frame are fused to obtain fused features;

[0106] Based on the feature location of the fused features in the video frame image, and combined with the position weight, the fused features are enhanced to form the enhanced feature vector of each image target;

[0107] Based on a pre-trained target detection head, target localization and judgment are performed on the enhanced feature vectors of each image target, and the image output vector of the image target in each video frame is given.

[0108] In one specific implementation, a multi-scale feature extraction pathway is first used to extract image targets of different scales from key images of video frames, resulting in image features of different scales. Specifically, the multi-scale feature extraction pathway employs multiple parallel feature extraction paths based on target scale differences. Each pathway corresponds to a different receptive field range, used to extract multi-level feature representations of small-scale, medium-scale, and large-scale targets, resulting in multiple image features. Then, feature fusion is used to align and fuse the image features output from each pathway, forming a unified multi-scale feature representation, resulting in fused features.

[0109] In a specific example, in an apron support scenario, various image targets exhibit significant scale differences. For instance, ground crew and wheel chocks are small-scale targets, catering trucks, fuel trucks, and tractor units are medium-scale targets, and jet bridges and aircraft are large-scale targets. Due to the inadequacy of single-scale features, a multi-scale feature extraction approach is employed. The fused features are specifically represented as follows:

[0110]

[0111] Among them, F multi To fuse features, k is the feature extraction pathway number, K is the number of feature extraction pathways, and α k F represents the scale weight corresponding to feature extraction path k. k For the image features corresponding to feature extraction path k, P k () represents the k-th feature extraction path, I represents the key image of the video frame, Backbone() represents the shared backbone network, and F represents the intermediate features.

[0112] By inputting the intermediate feature F into K feature extraction pathways at different scales, the corresponding image features are obtained and fused. Different feature pathways employ different convolutional kernel sizes, different downsampling rates, and different receptive field structures. Multiple feature extraction pathways are constructed in parallel with different receptive fields, enabling simultaneous modeling of target features at different scales.

[0113] After obtaining the fused features, position enhancement is performed on the fused features based on their feature positions in the video frame images, forming enhanced feature vectors for each image target, specifically represented as follows:

[0114]

[0115] Among them, F enhanced To enhance the feature vector, F multi For fusion features, F att (i,j) represents the enhanced feature corresponding to the feature position (i,j) in the fused features, and Attention() is the positional attention module. (i,j),(m,n) V represents the correlation between any feature position (i,j) and all other feature positions (m,n), i.e., the position weight. m,n Q is the value vector corresponding to the feature position (m,n). i,j K is the query vector corresponding to the feature position (i,j). m,n This is the key vector corresponding to the feature position (m,n).

[0116] The positional attention module weighted aggregates the contextual information of each spatial location in the fused feature, enabling the enhanced feature vector to adaptively enhance based on global spatial relationships, thereby improving detection and state discrimination capabilities in complex scenes. Specifically, for any feature location (i,j) in the fused feature, the positional attention module calculates its correlation with all other feature locations (m,n), adaptively aggregating global contextual information to enhance the representation of key region features. The positional attention module is used to model the correlations between different feature locations on the fused feature.

[0117] After obtaining the enhanced feature vector, it is input into a pre-trained target detection head for target localization and category discrimination, giving the corresponding image output vector, specifically represented as follows:

[0118]

[0119] Among them, F enhanced To enhance the feature vectors, D() is a pre-trained target detection head. For the spatial location parameters of the target in the image, For the target category of the image target, The confidence level for the detection of targets in the image.

[0120] During the data collection process for flight support events, the multi-scale feature extraction pathway was adjusted and improved based on the scale differences of targets on the apron. By introducing adaptive attention weights (i.e., scale weights corresponding to each feature extraction pathway), the detailed features of small-scale targets (such as ground staff) are enhanced during feature fusion, preventing them from being overwhelmed by the features of large-scale targets (such as aircraft). Simultaneously, combined with event sequence logic, a temporal context module is integrated into the target detection head to correlate the multi-scale detection results of consecutive frames, accurately identifying the sequence of support events. For example, detecting aircraft parking first, then triggering precise tracking of medium-scale vehicles and small-scale personnel, thereby improving the overall accuracy and real-time performance of event data collection.

[0121] The training process of the object detection head includes the following steps:

[0122] First, the historical target images in the training dataset are preprocessed to form standardized input images for training. The preprocessing includes format conversion, size normalization, and quality screening to remove images with severe occlusion, blurriness, or invalid targets.

[0123] The standardized input image is labeled to generate annotation information, which includes the target category and target bounding box of each image target.

[0124] In a specific example, historical target images and their corresponding annotation information, acquisition time, and monitoring point numbers are stored to provide a standardized data foundation for subsequent target detection model training, multi-scale feature learning, and event state discrimination.

[0125] The model is trained based on previously collected and labeled historical target images. Supervised learning is employed during training, using manually labeled target categories and bounding boxes as supervision signals. The classification branch minimizes the classification loss L. cls The regression branch minimizes the location regression loss L. regThe overall optimization objective is expressed as: L=L cls +λL reg , where λ is the loss weight coefficient, used to balance classification and positioning accuracy.

[0126] To enhance the model's robustness under different viewpoints and scales, stochastic scaling and image enhancement strategies are introduced during the training phase, enabling the model to adapt to the significant differences in target scale in tarmac scenes. After multiple rounds of iterative training, a target detection head for target feature detection is obtained. Through stochastic scaling and image enhancement strategies, each batch of standardized input images undergoes a series of random "modifications." For scaling, a multi-scale training strategy is employed, randomly selecting a new resolution (e.g., dynamically varying between 320 and 960 pixels) as the uniform input size for this batch. This allows the same aircraft to be presented with drastically different absolute pixel sizes in different batches, forcing the target detection head to abandon its dependence on a fixed scale and focus on learning the essential features of the target. Simultaneously, basic image enhancement is performed, including randomly adjusting the brightness and contrast of the standardized input images to simulate different lighting and weather conditions, and applying small rotations and flips to simulate diverse camera angles.

[0127] Furthermore, the image output vector includes the spatial location parameters of the image target, the target category, and the detection confidence.

[0128] By refining multi-scale targets using a spatial perception function based on the category gap scaling index, flight support targets and target nodes are obtained, with reference to... Figure 4 Specifically, it includes:

[0129] Based on the target category of the image object and the number of target categories, a classification calibration function is set;

[0130] By scaling the classification calibration function and combining it with the gap enhancement intensity, a confidence correction function is determined to correct the detection confidence of the image target.

[0131] Based on the corrected detection confidence, flight support targets are determined from the image output vector of the image targets;

[0132] Based on the target area of ​​flight support, the image output vectors of each flight support target are judged, and the judgment result is given;

[0133] Based on continuous time steps, the results of each judgment are merged to form a time state sequence;

[0134] Based on statistical analysis of the time state sequence, the motion state of the flight support target is determined, and the corresponding support target node is given.

[0135] Specifically, the confidence correction function, refer to Figure 5 This can be obtained through the following steps:

[0136] By analyzing the distribution of the number of each target category in the training dataset, the class distribution gap of each target category is determined;

[0137] Based on the class distribution gapness, and combined with different window sizes, multi-scale analysis is performed to determine the class gapness scaling index for each target class.

[0138] Scale mapping is performed on the image output vector to obtain the scale mapping parameters for each target category;

[0139] Based on the scale mapping parameters and combined with the standardized category gap scaling index, the spatial perception calibration parameters are determined through the spatial perception function.

[0140] The confidence correction function is constructed by normalizing the spatial perception calibration parameters of different target categories.

[0141] In one specific implementation, after training the target detection head, the target detection head is used to detect objects on the training dataset, and outputs a logits vector corresponding to the category of the historical target images. Here, C represents the maximum value of the target category number, providing the basic input for subsequent confidence calibration and screening. The category logits vector includes each target category number and the detection confidence corresponding to that target category number.

[0142] First, the object detection head performs joint prediction on image features at different scales to obtain the class logits vectors of each image object. To alleviate the long-tailed class prediction bias of the object detection head, the class logits vectors output by the object detection head are subjected to a porosity metric and calibrated.

[0143] Based on the category logits vectors corresponding to each historical target image in the training dataset, the category distribution gaps are calculated for each target category. The category distribution gaps corresponding to target category number c are specifically expressed as follows:

[0144]

[0145] in, M represents the gapness of the category distribution corresponding to the target category number c. c (r) represents the random variable that falls into the target category number c within the sliding window when the window side length is r. Var() is the variance function, and Mean() is the expectation function. As the window side length increases, the logarithmic growth rate of the class distribution gaps with scale tends to a stable value. This stable value is used to characterize the intensity of spatial non-uniformity of the classes.

[0146] The categorical gap scaling index is specifically expressed as:

[0147]

[0148] in, The category gap scaling index corresponding to the target category number c. The category distribution gap size corresponding to the target category number c, N c Let c be the number of samples corresponding to the target category number c in the training dataset, and O() is the order of magnitude function.

[0149] Since the number of samples in each target category in the training dataset is limited, using excessively fine-scale grid partitioning will lead to statistical instability and biased porosity estimation. Therefore, in multi-scale analysis, the maximum number of grids in the spatial partitioning scale is limited to the order of the square root of the number of samples in each category. Within this scale range, the relationship between the category porosity scaling exponent and scale is modeled in the logarithmic domain, by applying (log r, log A linear fit is used to approximately estimate the asymptotic slope of the class gap scaling index on a logarithmic scale, thus obtaining a stable and reliable class gap scaling index under finite sample conditions. When the number of samples is less than 4 and a linear fit cannot be performed, the class gap scaling index is set to 1.

[0150] Let the set of category gap scaling indices for all target categories be . The standardized category gap scaling index is specifically expressed as:

[0151]

[0152] in, The standardized category gap scaling index corresponding to the target category number c. The category gap scaling index corresponding to the target category number c. The maximum value in the gap scaling index corresponding to the target category number c. The minimum value in the gap scaling index corresponding to the target category number c.

[0153] For the foreground category, prior category information obtained from statistics of the training dataset is introduced to perform translational calibration on the category logits vector. The classification calibration function is specifically expressed as follows:

[0154]

[0155] Where Cal() is the classification calibration function, Let N be the detection confidence level before correction for target e under target category number c, where 1, ..., C are all target category numbers, A is the number of target categories, bg is the background category number, and N is the target category number. c β represents the number of samples corresponding to target class number c in the training dataset, and β is the logarithmic basis. β is a hyperparameter, which can be set according to the actual situation. β can take values ​​of 2, the natural constant e, or 10. In this example, β=10. The target class numbers are 1, ..., C. In this example, the maximum value C of the target class number is equal to the number of target classes A.

[0156] The classification calibration function is used to eliminate the overall bias caused by class sample size imbalance and to perform class-related scaling adjustments.

[0157] The confidence correction function is specifically expressed as follows:

[0158]

[0159] in, The corrected detection confidence level for flight support target e under target category number c. Let S() be the detection confidence level of flight support target e under target category number c before correction, S() be the spatial perception function, through which spatial perception calibration parameters can be calculated, and σ() be the scale mapping function, used to scale the category logits vector. The standardized category gap scaling index is given to the target category number c. α is the gap enhancement intensity. α>0 is used to control the enhancement intensity of the category gap scaling index on the tail category. The background category is not calibrated or modulated.

[0160] The category gap scaling index is used to characterize the spatial sparsity differences among different target categories. By modulating the gapness of the category logits vector, it effectively alleviates the confidence suppression problem of long-tail categories during the classification stage.

[0161] In a specific example, there are three target categories: tractor, fuel truck, and aircraft. Tractor is a rare class, while fuel truck and aircraft are frequent classes. The standardized class gap scaling index [1.1, 1.7, 1.9] can be calculated using the training dataset. For a target category of tractor, the class logits vector (detection confidence) is [3.9, 4.5, 0.2, 0.9], indicating a detection confidence of 3.9 for tractor, 4.5 for fuel truck, 0.2 for aircraft, and 0.9 for background. In this case, the probability of fuel truck is slightly higher. Although the probability of tractor is not low, due to unfamiliarity with tractor, the probability of fuel truck is still considered higher. After calibration using the classification calibration function, the class logits vector is [5.0, 5.3, 0.7, 0.9]. After further correction using the confidence correction function, the class logits vector is [0.5, 0.4, 0.03, 0.07]. The target detection head will correctly calibrate the target category of the target within the bounding box as a tractor.

[0162] Reference Figure 6 , Figure 6 To provide a structural framework for target detection and multi-scale target correction that integrates multi-path networks and positional attention, this invention performs feature transformation on key images of video frames in a video sequence to obtain intermediate features. Then, it extracts and fuses image features at different scales through a multi-scale feature extraction pathway, resulting in fused features. These fused features are then spatially aggregated and enhanced by a positional attention module to obtain an enhanced feature vector. Finally, a target detection head is used for localization and classification to provide an image output vector, thus achieving target detection that integrates multi-path networks and positional attention. Furthermore, to improve the accuracy of identifying flight support targets, the image output vector undergoes further target correction. The final flight support target is determined based on the corrected detection confidence level, providing an accurate data foundation for the construction of subsequent flight support node events.

[0163] Specifically, the support target nodes corresponding to each flight support target are determined, including:

[0164] Based on the target area of ​​flight support, the image output vectors of each flight support target are judged, and the judgment result is given;

[0165] Based on continuous time steps, the results of each judgment are merged to form a time state sequence;

[0166] Based on statistical analysis of the time state sequence, the motion state of the flight support target is determined, and the corresponding support target node is given.

[0167] After identifying flight support targets from various image targets, to address the issues of temporal continuity and instantaneous false detection interference in the state values ​​of flight support targets during apron operations, a time-state sequence is used to determine the state of each flight support target. For the spatial relationship between the image output vector at consecutive time steps and the predefined target region, a decision function generates the corresponding state value, specifically expressed as follows:

[0168]

[0169] Among them, f state Let s be the decision function. e (t) represents the state value of flight support target e at time t, which can also be understood as the judgment result obtained by flight support target e at time t. e For the target area of ​​flight support corresponding to flight support target e, d e (t) represents the image output vector of the flight support target e at time t. The image output vector includes the spatial location parameters of the image target, the target category, and the detection confidence. s1 represents a stable state that satisfies the spatial constraints, and s2 represents an unstable state that does not satisfy the spatial constraints. That is, if the center coordinates of the flight support target are within the corresponding target area and the detection confidence is higher than the confidence threshold, the state value of the flight support target is considered to be s1, and the judgment result is that the flight support target is a stable state that satisfies the spatial constraints at time t. Conversely, if the center coordinates of the flight support target are not within the corresponding target area or the detection confidence is not higher than the confidence threshold, the state value of the flight support target is considered to be s2, and the judgment result is that the flight support target is an unstable state that does not satisfy the spatial constraints at time t.

[0170] The status values ​​are written sequentially to the time status queue. Each flight support target has a corresponding time status queue, specifically represented as follows:

[0171]

[0172] in, This indicates a queue update operation, with the queue length set to L to control the historical time window, Q. e (t) represents the time state sequence of flight support target e at time step t. By analyzing the statistical characteristics of the state values ​​in the time state queue and combining them with the time consistency rule, it is determined whether the flight support target has entered a stable operating state, specifically expressed as:

[0173]

[0174] in, The motion state of flight support target e is used to trigger the output of flight support node events, f temporal () represents the time consistency rule, Q eThis is the time state sequence of flight support target e. When the motion state of the flight support target is stable, the corresponding support target node is given.

[0175] In a specific example, since the time state queue contains the state values ​​of the flight support target, assuming the queue length is 70, if the flight support target is a catering truck, and the first 35 states in the queue are s2 and the last 35 states are s1, then it can be assumed that the catering truck has started preparing meals, and the time t corresponding to the 36th state value can be determined as the support target node. If the first 30 states are s2, the last 30 states are s2, and the middle 10 states are s1, then it can be assumed that the catering truck has only passed by and has not started preparing meals, and the support target node cannot be determined.

[0176] Specifically, based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed, referring to... Figure 7 Specifically, it includes:

[0177] Obtain flight support targets, support target nodes, and flight support node strategies. The flight support node strategies include standard support node events corresponding to each flight support target.

[0178] Based on a pre-built node judgment model, a matching analysis is performed on the flight support target and flight support node strategy at the target node.

[0179] Based on the results of the matching analysis, flight support node events are constructed.

[0180] Furthermore, the node judgment model specifically includes: behavior unit, probability calculation unit, and rule unit;

[0181] The behavioral unit is used to analyze the relationship between the flight support target and the standard support node events based on the flight support target and the support target node, and to obtain the current node status.

[0182] The probability calculation unit is used to calculate the probability distribution of the current node state based on the probability density function of the target node and the current node state, and to give the probability distribution likelihood value of the target node.

[0183] The rule unit is used to judge the probability distribution likelihood value of the target node through a preset logical expression relationship, and give the matching result of the flight support target between the target node and the flight support node.

[0184] In one specific implementation, each type of flight support objective (such as vehicles, equipment, or operational elements) in the flight support process is considered as an independent decision node. For node v, its corresponding state feature vector is extracted from the continuous time series to obtain the current node state, specifically represented as follows:

[0185]

[0186] in, Let v be the current node state at time t. Let t be the time interval between node v and the standard support node event (such as flight arrival). This represents the spatial distance between the flight support target corresponding to node v at time t and the reference location (such as the gate or fuselage area) in the standard support node event. Let v be the confidence feature of the flight support target corresponding to node v at time t.

[0187] For node v, a Gaussian mixture model is used to model its state distribution, and its probability density function is defined as:

[0188]

[0189] Where p(x) v Let K be the probability density of node v. v π is the number of Gaussian components at node v. v,k Let μ be the component weight of the k-th Gaussian component. v,k Let ∑ be the mean vector of the k-th Gaussian component. v,k Let be the covariance matrix of the k-th Gaussian component.

[0190] When node v generates a new state observation at time t At that time, the probability distribution of the current node state is calculated, and the likelihood value of the probability distribution of the current support node is given. Based on the relationship between the likelihood value of the probability distribution and a preset threshold (i.e., the preset logical expression relationship), it is determined whether the current state meets the event establishment condition, and the matching result of the flight support target in the flight support node and the corresponding standard support node event in the flight support node strategy is given, specifically expressed as follows:

[0191]

[0192] in, This represents the matching result between the flight support target and the standard support node event for node v at time t, ensuring the support for the target node. Let be the likelihood value of the probability distribution corresponding to node v at time t, which guarantees the target node. For node v to generate new state observations at time t, τ vThe preset threshold is set based on the standard support node events corresponding to different flight support targets.

[0193] Through the above process, it is possible to effectively distinguish between stable event states that conform to historical statistical patterns and short-term, occasional, and atypical states, thus avoiding the direct triggering of event judgments based solely on detection results.

[0194] Furthermore, the probability density function is a Gaussian mixture model, obtained through the following steps:

[0195] Based on each protection node, a Gaussian mixture model is used to model the state distribution of the protection nodes to obtain an initial distribution model;

[0196] Calculate the posterior responsibility degree corresponding to the Gaussian component based on the Gaussian component of each guarantee node;

[0197] By combining the preset learning rate and the posterior responsibility corresponding to the Gaussian components, the model parameters in the initial distribution model are updated to determine the Gaussian mixture model. The model parameters include the component weights, center positions, and covariance matrices of the Gaussian components.

[0198] In one specific implementation, considering that the operational process of flight support nodes may change over time, an online update strategy is introduced to enable the initial distribution model to adapt. When the state is determined to be a valid event state, the current observation samples are used to update the model parameters.

[0199] Based on the foregoing description, the initial distribution model can be obtained, specifically expressed as:

[0200]

[0201] For the k-th Gaussian component of node v, the posterior responsibility is specifically expressed as:

[0202]

[0203] in, Let K be the posterior responsibility of the k-th Gaussian component of node v at time t. v π is the number of Gaussian components at node v. v,k Let μ be the component weight of the k-th Gaussian component. v,k Let be the mean vector of the k-th Gaussian component, representing the center position of the Gaussian component, ∑ v,k Let be the covariance matrix of the k-th Gaussian component. Generate new state observations for node v at time t.

[0204] Posterior responsibility is used to calculate the degree to which the current node's state belongs among the various Gaussian components. Posterior responsibility γ v,kThis represents the relative probability that the current observation state is generated by the k-th Gaussian component, providing a weighting basis for subsequent model parameter updates.

[0205] Based on this, a learning rate α0 is introduced to update the model parameters online:

[0206]

[0207] in, Let α0 be the component weight of the k-th Gaussian component of node v at time (t+1), and let α0 be the learning rate. Let V be the component weight of the k-th Gaussian component of node v at time t. Let be the posterior responsibility of the k-th Gaussian component at time t. Let be the mean vector of the k-th Gaussian component of node v at time (t+1). Let be the mean vector of the k-th Gaussian component of node v at time t. Let be the covariance matrix of the k-th Gaussian component of node v at time (t+1). Let v be the covariance matrix of the k-th Gaussian component of node v at time t. T This is the transpose of the matrix.

[0208] By introducing a learning rate α0, the weights of each component in the Gaussian mixture model are adjusted online, allowing new observation samples to gradually influence the model parameters. This ensures the model can adapt to new data without drastic changes due to short-term fluctuations. Through posterior responsibility weighting, the central positions of the Gaussian components are updated, gradually incorporating new effective state features into the original mean, enabling the model to reflect long-term trends in the operational process. Online updates to the covariance describe the dynamic adjustment process of the node state distribution, characterizing the temporal and spatial fluctuation range of node states, allowing the model to adapt to changes under different operational rhythms and environmental conditions. Through this online update mechanism, the Gaussian mixture model can gradually adapt to changes in state distribution under different monitoring locations, different flights, or different operational rhythms.

[0209] Reference Figure 8 This invention provides a method for monitoring flight support nodes, which specifically includes the following steps:

[0210] Obtain flight support node strategies and video sequences from each monitoring point;

[0211] By integrating multi-path network and positional attention for target detection, the analysis of video sequences identifies targets at multiple scales.

[0212] Furthermore, by using a spatial perception function based on the category gap scaling index to correct multi-scale targets, flight support targets and support target nodes are obtained;

[0213] Based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed.

[0214] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described steps can be referred to the corresponding process in the aforementioned system embodiments, and will not be repeated here.

[0215] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and variations of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A monitoring system for flight support nodes, characterized in that, Specifically, it includes: The acquisition module is used to acquire flight support node strategies and video sequences from various monitoring points; The behavior analysis module is used to analyze video sequences and identify targets at multiple scales by fusing multi-path networks and positional attention for target detection. Specifically, it includes: reading video sequences from various monitoring points to obtain multiple video frame images; determining the identification range based on the target area of ​​flight support; cropping each video frame image based on the identification range to obtain key images of the video frames; extracting image targets of different scales from the key images of the video frames through a multi-scale feature extraction pathway to obtain image features of different scales; performing fusion analysis on image features of different scales using a multi-feature fusion matching strategy combined with positional enhancement to provide image output vectors for each image target. Specifically, it includes: fusing image features of different scales in each key image of the video frame with scale weights to obtain fused features; enhancing the fused features based on their feature positions in the video frame images with positional weights to form enhanced feature vectors for each image target; and performing target localization judgment on the enhanced feature vectors of each image target based on a pre-trained target detection head to provide image output vectors for the image targets in each video frame image. The image output vectors include the spatial location parameters of the image target, the target category, and the detection confidence. Furthermore, by using a spatial perception function based on the category gap scaling index to correct multi-scale targets, flight support targets and support target nodes are obtained. Specifically, this includes: setting a classification calibration function based on the target category of the image target and the number of target categories, where the classification calibration function is used to eliminate the overall bias caused by the imbalance of category sample size and to adjust the category scale; the gap enhancement intensity is used to control the enhancement intensity of the category gap scaling index for tail categories; the category gap scaling index is used to characterize the spatial sparsity differences of different target categories; by scaling the classification calibration function and combining it with the gap enhancement intensity, a confidence correction function is determined to correct the detection confidence of the image target; based on the corrected detection confidence, the flight support target is determined from the image output vector of the image target; combined with the target area of ​​flight support, the image output vector of each flight support target is judged, and the judgment result is given; based on continuous time steps, the judgment results are fused to form a time state sequence; based on the statistical analysis of the time state sequence, the motion state of the flight support target is judged, and the support target node corresponding to the flight support target is given. The event construction module is used to construct flight support node events based on the matching analysis of flight support targets, support target nodes, and flight support node strategies.

2. The monitoring system for flight support nodes as described in claim 1, characterized in that, The video sequences from each monitoring point are obtained through the following steps: Collect initial video data from preset monitoring points; Based on the time interval of the event trigger, extract the key video streams from the initial video data to form multiple video clips; The various video clips are combined to form a video sequence for each monitoring point.

3. The monitoring system for flight support nodes as described in claim 1, characterized in that, The classification calibration function is specifically expressed as follows: ； Where Cal() is the classification calibration function, Let N be the detection confidence level before correction for target e under target category number c, where 1, ..., C are all target category numbers, A is the number of target categories, bg is the background category number, and N is the target category number. c β represents the number of samples corresponding to the target class number c in the training dataset, and β is the logarithmic basis.

4. The monitoring system for flight support nodes as described in claim 1, characterized in that, The confidence correction function is obtained through the following steps: By analyzing the distribution of the number of each target category in the training dataset, the class distribution gap of each target category is determined; Based on the class distribution gapness, and combined with different window sizes, multi-scale analysis is performed to determine the class gapness scaling index for each target class. Scale mapping is performed on the image output vector to obtain the scale mapping parameters for each target category; Based on the scale mapping parameters and combined with the standardized category gap scaling index, the spatial perception calibration parameters are determined through the spatial perception function. The confidence correction function is constructed by normalizing the spatial perception calibration parameters of different target categories.

5. The monitoring system for flight support nodes as described in claim 1, characterized in that, Based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed, specifically including: Obtain flight support targets, support target nodes, and flight support node strategies. The flight support node strategies include standard support node events corresponding to each flight support target. Based on a pre-built node judgment model, a matching analysis is performed on the flight support target and flight support node strategy at the target node. Based on the results of the matching analysis, flight support node events are constructed.

6. The monitoring system for flight support nodes as described in claim 5, characterized in that, The node judgment model specifically includes: behavior unit, probability calculation unit, and rule unit; The behavioral unit is used to analyze the relationship between the flight support target and the standard support node events based on the flight support target and the support target node, and to obtain the current node status. The probability calculation unit is used to calculate the probability distribution of the current node state based on the probability density function of the target node and the current node state, and to give the probability distribution likelihood value of the target node. The rule unit is used to judge the probability distribution likelihood value of the target node through a preset logical expression relationship, and give the matching result of the flight support target between the target node and the flight support node.

7. The monitoring system for flight support nodes as described in claim 6, characterized in that, The probability density function is a Gaussian mixture model, obtained through the following steps: Based on each protection node, a Gaussian mixture model is used to model the state distribution of the protection nodes to obtain an initial distribution model; Calculate the posterior responsibility degree corresponding to the Gaussian component based on the Gaussian component of each guarantee node; By combining the preset learning rate and the posterior responsibility corresponding to the Gaussian components, the model parameters in the initial distribution model are updated to determine the Gaussian mixture model. The model parameters include the component weights, center positions, and covariance matrices of the Gaussian components.

8. A method for monitoring flight support nodes, characterized in that, The monitoring system for flight support nodes as described in any one of claims 1-7 specifically includes the following steps: Obtain flight support node strategies and video sequences from each monitoring point; By integrating multi-path network and positional attention for target detection, this method analyzes video sequences to identify targets at multiple scales. Specifically, it involves: reading video sequences from various monitoring points to obtain multiple video frame images; defining the identification range based on the target area of ​​flight support; cropping each video frame image based on the identification range to obtain key images; extracting image targets of different scales from the key images of the video frames using a multi-scale feature extraction pathway to obtain image features of different scales; and using a multi-feature fusion matching strategy combined with positional enhancement to perform fusion analysis on the image features of different scales, providing image output vectors for each image target. Specifically, this includes: fusing image features of different scales in each key image of the video frame with scale weights to obtain fused features; enhancing the fused features based on their feature positions in the video frame images with positional weights to form enhanced feature vectors for each image target; and using a pre-trained target detection head to perform target localization judgment on the enhanced feature vectors of each image target, providing image output vectors for the image targets in each video frame image. The image output vectors include the spatial location parameters of the image target, the target category, and the detection confidence. Furthermore, by using a spatial perception function based on the category gap scaling index to correct multi-scale targets, flight support targets and support target nodes are obtained. Specifically, this includes: setting a classification calibration function based on the target category of the image target and the number of target categories, where the classification calibration function is used to eliminate the overall bias caused by the imbalance of category sample size and to adjust the category scale; the gap enhancement intensity is used to control the enhancement intensity of the category gap scaling index for tail categories; the category gap scaling index is used to characterize the spatial sparsity differences of different target categories; by scaling the classification calibration function and combining it with the gap enhancement intensity, a confidence correction function is determined to correct the detection confidence of the image target; based on the corrected detection confidence, the flight support target is determined from the image output vector of the image target; combined with the target area of ​​flight support, the image output vector of each flight support target is judged, and the judgment result is given; based on continuous time steps, the judgment results are fused to form a time state sequence; based on the statistical analysis of the time state sequence, the motion state of the flight support target is judged, and the support target node corresponding to the flight support target is given. Based on the matching analysis of flight support objectives, target nodes, and flight support node strategies, flight support node events are constructed.

Images