Self-adaptive optimization airport activity relevance assessment method

Abstract

The invention discloses a self-adaptive optimization airport activity relevance calculation method, and belongs to the technical field of video processing. The method comprises the following steps: collecting video data containing airport scene activities, and labeling the airport scene video data; preprocessing the video and generating a data set; constructing a feature extraction network based on a convolutional neural network; constructing an activity mapping module, and designing an activity prediction result uncertainty representation method based on a quality function; constructing an activity relevance representation method based on Renyi divergence; constructing an autonomous optimizer, and performing adaptive optimization on a prediction result based on a simulated annealing algorithm to complete adaptive optimization of airport activity relevance; and designing an activity relevance fusion calculation method based on the Dempster-Shafer evidence theory. According to the method, on the premise of using extremely small model calculation amount, evaluation of relevance between different airport activities is greatly improved, and then the accuracy of activity evaluation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of video processing, and particularly to a method for evaluating the relevance of airport activities with autonomous adaptation and optimization. Background Art

[0002] With the rapid development of the air transportation industry, airports, as the core hubs of air transportation activities, involve a large number of complex activity collaborations in their daily operations, such as aircraft ground taxiing, apron support, baggage handling, cargo transportation, and passenger boarding and alighting. There are complex correlations among these activities. By effectively monitoring and analyzing the correlations of airport surface activities, important decision-making support can be provided for airport management, further improving airport operation efficiency, reducing operation risks, and lowering management costs. However, there are still many challenges in the current airport surface activity correlation calculation technology, especially in terms of activity diversification, complex surface environment, and inefficient correlation evaluation, making it difficult to meet the actual needs.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the wide application of deep learning technology, new solutions have been provided for video processing and activity recognition. The Convolutional Neural Network (CNN) has been widely used in feature extraction and object detection due to its outstanding performance in the field of image and video analysis. However, in airport scenarios, due to the complexity of the activity video scenario and the requirements for robustness, there are still deficiencies in relying solely on the CNN model to achieve efficient activity correlation calculation. On the one hand, the huge data volume and model calculation overhead increase the difficulty of actual deployment; on the other hand, the prediction results of deep learning models usually lack the quantification and representation of uncertainty, making it difficult to fully reflect the reliability of the prediction results and the measurement of activity correlations in complex scenarios.

[0004] In addition to the limitations of feature extraction technology, the current mainstream activity correlation calculation methods also mostly focus on simple correlation analysis, such as simple evaluation of activities based on linear regression or statistical methods. These methods often rely on predefined rules or specific scenario bases and are difficult to adapt to the complex dynamic correlation characteristics in the airport surface. Moreover, with the increasing demand for data fusion from multiple sensors in the surface, the correlation fusion of multi-modal data also poses higher requirements on traditional methods. In terms of optimization calculation, modern meta-heuristic optimization methods - such as the Simulated Annealing Algorithm (SA) and genetic algorithms - have shown excellent performance in many complex optimization problems. Due to its strong global optimization ability and simple implementation process, the simulated annealing algorithm has potential application value in the fields of video processing and correlation optimization. However, there is currently little research effectively combining the simulated annealing algorithm with deep learning methods to solve the adaptive optimization problem of activity correlations.

[0005] In addition, due to its unique uncertainty quantification ability, the Dempster-Shafer evidence theory has also become a popular research direction for dealing with multi-source information fusion in recent years. The Dempster-Shafer theory has significant advantages in scenarios with high uncertainty and sensor data fusion, especially in complex environments such as airports filled with dynamic information and multi-source data. However, the current research on the fusion calculation of activity relevance based on the Dempster-Shafer theory is not sufficient, especially in using it for the dynamic relevance assessment of activities, and there is still much room for improvement. Summary of the Invention

[0006] In order to solve the above technical problems existing in the prior art, the present invention provides a method for calculating the relevance of airport activities with autonomous adaptation and optimization, aiming to improve the accuracy and reliability of the calculation of airport activity relevance and provide more accurate decision-making support for airport surface traffic control.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A method for evaluating the relevance of airport activities with autonomous adaptation and optimization, the method comprising the following steps:

[0009] Step 1: Collect video data containing airport surface activities, and label the airport activities in the video data, including surface targets, activity start time, end time, and activity categories, to obtain the labeled airport surface video data;

[0010] Step 2: Perform data preprocessing on the labeled airport surface video data to adapt it to the input layer of the adopted feature extraction module;

[0011] Step 3: Construct a feature extraction module based on a convolutional neural network, and input the video data preprocessed in Step 2 into the feature extraction module for feature extraction to obtain activity features related to airport surface targets;

[0012] Step 4: Construct an activity mapping module based on a feedforward neural network, and represent the uncertainty of the activity prediction results through a mass function to generate the activity probability distributions of several activities, and obtain the mass function of each activity;

[0013] Step 5: Calculate the distribution difference between pairwise mass functions using Renyi divergence, and construct an n*n-dimensional divergence matrix based on the Renyi divergence between the mass functions of all activities to represent activity relevance; where n represents the number of activities;

[0014] Step 6: Construct a divergence loss function based on Renyi divergence, and adaptively optimize the constructed divergence loss function based on the simulated annealing algorithm to learn a set of weighted coefficients representing the weights of the quality functions of each activity; and weight the different activity probability distributions in Step 4 based on this set of weighted coefficients to obtain a corrected activity probability distribution;

[0015] Step 7: Use Dempster-Shafer evidence theory to perform information fusion on the corrected activity probability distribution obtained in Step 6 and the activity probability distribution obtained in Step 4 to generate a final activity correlation result;

[0016] The present invention first uses a convolutional neural network and a feedforward neural network to predict the activities of airport surface targets, and introduces a quality function to assign the activity probability distribution that each target may have. Then, by introducing Renyi divergence to calculate the activity differences between pairwise targets, a divergence loss function is constructed to measure the divergence between different target activities. Secondly, the simulated annealing algorithm is used to adaptively optimize the divergence loss function to obtain the optimal weights of different quality functions, and then all the original quality functions are integrated to obtain a new weighted quality function. Finally, based on Dempster-Shafer evidence theory, information fusion of different activities is performed to calculate the correlation between different activities.

[0017] Further, the data preprocessing in Step 2 specifically includes:

[0018] Randomly extract frames from the labeled airport surface video data according to the set frame rate (which can be set to 1-30 FPS) to obtain a number of image frame sequences;

[0019] Normalize the resolution of the extracted image frame sequences, and then reconstruct the image sizes of all images into a specified size.

[0020] In addition, during the training process of the neural network, the data preprocessing may also include data augmentation processing of the training data, such as random flipping and mosaic augmentation, etc.;

[0021] Further, the feature extraction module constructed in Step 3 is a feature extraction module based on the ResNet network structure to obtain N-dimensional activity features, where N is a preset value.

[0022] Further, the activity mapping module constructed in Step 4 is composed of a cascade of at least three layers of linear layers, the dimensions of its respective linear layers decrease layer by layer, and the dimension of the last linear layer of the activity mapping module is 2 n , and the image activity output corresponding to each frame of video is a quality function m i , where the subscript i is the activity identifier.

[0023] Further, in step 4, the uncertainty representation of the activity prediction result by the mass function is specifically as follows:

[0024] m i : m({A1}) = a1, m({A2}) = a2, …,

[0025] m({A1, A2}) = a n+1 m({A1, A3}) = a n+2 , …,

[0026]

[0027] where m() represents the probability assignment function for the activity event, {A} represents an activity event, and its subscript represents the activity index, that is, {A j} represents the jth activity event, i = 1, 2, n; a represents the probability assignment value of each probability assignment function m(), and for each mass function m i , it calculates the corresponding probability assignment values for different combinations of n activities respectively, that is, a p (p ∈ {1, 2, …, 2 n}), that is, the subscript of the probability assignment value a is the activity combination index of n activities, and for each mass function m i satisfies ∑a p = 1.

[0028] Further, in step 5, the expression of the Renyi divergence between the mass functions of activities is:

[0029]

[0030] where α represents the preset Renyi coefficient, {A i} and {A j} represent the ith and jth activity events, m p (A i ), m q (A j ) respectively represent the probability assignment values of the activities of activity p and activity q with respect to their corresponding activity events A i , A j .

[0031] Further, in step 6, the difference loss function is specifically:

[0032]

[0033] where D() represents the Renyi divergence between two mass functions, M represents the weighted sum of the mass functions of all activities, that is ω i represents the mass function mi weight.

[0034] Furthermore, the simulated annealing algorithm is used to optimize the constructed difference loss function L dif , and the weight combination w1, w2,..., w under the optimal solution is obtained n ; finally, based on this optimal weight combination, the n mass functions m obtained in step 4 are weighted and summed i to obtain the reconstructed mass function M. In step 6 of the present invention, based on the constructed difference loss function and adaptive loss, the optimal set of weighting coefficients ω i obtained by optimization, and the reconstructed mass function M obtained by weighted summation together constitute the self-optimization method of step 6.

[0035] Further, step 7 specifically includes the following steps:

[0036] Step 7.1: Construct the Dempster-Shafer evidence fusion formula, and its expression is as follows:

[0037]

[0038] Among them, represents the new mass function after the fusion of two mass functions, {A i} and {A j} represent two active events, and K represents the conflict coefficient between the two active events, and its expression is as follows:

[0039]

[0040] Step 7.2: Use the Dempster-Shafer evidence theory to calculate the activity correlation between different activities, and its calculation expression is as follows

[0041]

[0042] Among them, m p represents the activity probability distribution of activity p (i.e., the mass function of activity p), m q represents the activity probability distribution of activity q, M represents the weighted sum of all generated activity distributions, and finally, the result of represents the final activity correlation result.

[0043] The technical solution provided by the present invention at least brings the following beneficial effects:

[0044] The present invention constructs a feature extraction network based on a convolutional neural network, which can efficiently extract key features of airport surface activities, solving the problem of insufficient adaptability of traditional methods to complex scenarios; the designed uncertainty representation method based on the quality function and the Renyi divergence correlation representation method can effectively quantify the uncertainty of activity prediction results, improving the accuracy and robustness of activity correlation calculation; the autonomous optimizer using the simulated annealing algorithm enables the model to achieve adaptive optimization in different scenarios, reducing the computational amount while improving the efficiency of correlation calculation; the correlation fusion calculation method based on the Dempster-Shafer evidence theory makes full use of the redundancy and complementarity of multi-source data, significantly improving the comprehensive evaluation ability of activity correlation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of the present invention will be described in detail and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described by referring to the accompanying drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation to the present invention.

[0046] Figure 1 It is a schematic flowchart of an autonomous adaptive optimization method for airport activity correlation evaluation provided by an embodiment of the present invention;

[0047] Figure 2 It is a schematic structural diagram of the process of an autonomous adaptive optimization method for airport activity correlation evaluation provided by an embodiment of the present invention;

[0048] Figure 3 It is a data preprocessing flowchart in an embodiment of the present invention;

[0049] Figure 4 It is a structural diagram of a feature extraction module in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of the present invention will be described in detail and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described by referring to the accompanying drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation to the present invention.

[0051] In view of the deficiencies of the prior art, the embodiments of the present invention propose a method for calculating the relevance of airport activities with autonomous adaptation and optimization. By adopting the feature extraction technology based on CNN combined with an autonomous optimization algorithm, further integrating the uncertainty quantification technology and Dempster-Shafer theory, it will help to improve the accuracy and reliability of calculating the relevance of airport activities. In particular, it optimizes the consumption of computing resources, realizes a lightweight and efficient model design, which can not only improve the real-time performance of the algorithm, but also provide more accurate decision-making support for airport managers, and is of great significance for improving airport operation efficiency.

[0052] In one embodiment, referring to Figure 1 and Figure 2 , the specific implementation of a method for calculating the relevance of airport activities with autonomous adaptation and optimization provided by the embodiments of the present invention includes:

[0053] Step 1, Airport surface video data collection and annotation:

[0054] Collect video data containing airport surface activities, and annotate the airport activities in the video data, including surface targets, activity start times, end times, and activity categories, to obtain the original video data set;

[0055] Step 2, Data preprocessing and data set division:

[0056] Divide the annotated airport surface video data according to a certain ratio, and preprocess the data to construct training and test samples;

[0057] Step 3, Construct a feature extraction network for extracting activity features:

[0058] Based on the convolutional neural network, construct a feature extraction module to extract features from the data in the original video data set, and obtain activity features related to airport surface targets;

[0059] Step 4, Construct an activity mapping module, which involves a method for representing the uncertainty of activity prediction results based on the mass function:

[0060] Construct an activity mapping module based on the feedforward neural network, and represent the uncertainty of the activity prediction results through the mass function;

[0061] Step 5, Construct a method for representing the relevance of activities based on Renyi divergence:

[0062] Construct a method for representing the relevance of activities based on Renyi divergence, and construct a divergence matrix for the activity prediction distribution generated in step 4 to represent the activity relevance.

[0063] Step 6, Construct an autonomous optimization method for the relevance of activities based on the simulated annealing algorithm:

[0064] Construct a differential loss function, adaptively optimize the loss function based on the simulated annealing algorithm, and use the divergence matrix obtained in step 5. Through the simulated annealing algorithm, learn a set of weighted coefficients to weight different activity probability distributions in step 4 to obtain a corrected activity probability distribution.

[0065] Step 7: Design a method for calculating activity relevance based on Dempster-Shafer evidence theory:

[0066] Use Dempster-Shafer evidence theory to perform information fusion on the corrected activity probability distribution and the original probability distribution obtained in step 6 to generate the final activity relevance result;

[0067] In one embodiment, in the embodiments of the present invention, each step is specifically as follows:

[0068] In step 1: Use airport video acquisition equipment to collect activity video data of common targets on the airport scene, and label the targets and target activities in time sequence through manual and annotation software to obtain an original video data set; specifically including target types, target activity types, target activity start times, and target activity end times. Among them, the time accuracy of the activity start time and end time annotation is in milliseconds.

[0069] In step 2: Preprocess the original video images. In the embodiments of the present invention, the collected original video is first randomly divided into a training set and a test set according to a certain ratio for constructing training samples and test samples. In this embodiment, the collected video data is divided into a training set and a test set with a ratio of 1:1 as the basic data set for model development. Specifically, each video in the divided video data set is frame-extracted at a random frame rate of 1 to 30 FPS to generate a corresponding video frame sequence. Each video obtains a number of RGB image frame sequences through random frame extraction, so that each video samples high-quality image frames with a uniform distribution. Subsequently, the resolution of all the extracted RGB image frame sequences is uniformly adjusted to 640×640 to eliminate the influence of the difference in the original video resolution on subsequent processing. After the resolution is unified, all the pictures are further reconstructed into a specification of 256×256 to adapt to the computing requirements of subsequent neural network processing. In terms of data augmentation, for the image frame sequences in the training set, a series of additional data augmentation operations are performed, including random flipping (horizontal or vertical direction) and mosaic data augmentation. These operations can effectively enhance the model's adaptability to the diversity of input data and alleviate the overfitting problem. At the same time, only the unified 256×256 resolution reconstruction is performed on the image frames in the test set to maintain the consistency of its input format with the training set to ensure the fairness and consistency of the test results. The detailed processing steps are as Figure 3 shown.

[0070] In step 3: A single-stream feature extraction module is constructed based on the ResNet network structure to extract features from the preprocessed image sequence. The feature extraction module consists of 1 convolutional layer module and 4 residual modules, and high-dimensional features are extracted through layer-by-layer convolution and feature mapping. For each frame of RGB image in the input image sequence, after being processed by the feature extraction module, a feature vector f of size N is output, where N = 2048, representing the high-dimensional feature representation of the image. The specific process is as follows: Each frame of image I preprocessed in step 2 is input into the ResNet50 network (whose structure is as Figure 4 shown), and after convolutional operation, batch normalization, activation function, and global average pooling layer, the corresponding feature vector f ∈ R N . In the entire image sequence, after the feature vectors of each frame of image are input in sequence, they are concatenated into a feature matrix, which is used as the input of the subsequent activity correlation calculation module.

[0071] That is, in this embodiment, referring to Figure 4 , the network structure of the adopted feature extraction module is specifically:

[0072] The input image sequence first passes through a 7*7 convolutional layer, a normalization layer, and a Relu mapping of the activation function, and then is sent into the max pooling layer. After passing through two stacked first bottleneck modules and two stacked second bottleneck modules in sequence, the obtained output feature map is then sent into an average pooling layer and finally through a fully connected layer to obtain the image features output by the feature extraction module, that is, the feature vector f. Among them, the first bottleneck module is stacked by several BottleBlock layers, and the second bottleneck module is stacked by several BottleNeck layers. For example, the two first bottleneck modules are stacked by 3 layers and 4 layers of BottleBlock layers respectively. And the two second bottleneck modules are stacked by 5 layers and 3 layers of BottleBlock layers respectively. Among them, the structure of the BottleBlock layer is: two 3*3 convolutional layers, and a shortcut connection is adopted between the input of the first 3*3 convolutional layer and the output of the second 3*3 convolutional layer; and the BottleNeck layer sequentially includes: a 1*1 convolutional layer, a 3*3 convolutional layer, a 1*1 convolutional layer, and a shortcut connection is adopted between the input of the first 1*1 convolutional layer and the output of the second 1*1 convolutional layer, specifically as Figure 4 shown.

[0073] In step 4: f is input into the activity mapping module based on the feedforward neural network. This module consists of at least 3 linear layers with equal input and output dimensions and size N. Here, the structure of the linear layer is: a fully connected layer and a ReLU layer. The output of the module is the probability distribution of several activities (i.e., the mass function), and its expression is:

[0074] mi : m({A1}) = a1, m({A2}) = a2, …,

[0075] m({A1, A2}) = a n+1 ,m({A1, A3}) = a n+2 ,…,

[0076]

[0077] where m() represents the probability assignment function for activity events. Any m() represents the probability distribution of all possible activities for a target. {A} represents a certain activity event, expressed as a set, and a represents the probability assignment value of the activity, satisfying ∑a i = 1, m i represents one of the mass functions output by the activity mapping module, a i represents the probability assignment value of the i-th activity, n represents the number of activities. For example, m({A1}) represents the probability assignment value corresponding to the 1st activity, and m({A1, A1}) represents the probability assignment value corresponding to the 1st and 2nd activities.

[0078] In step 5: Construct a method for representing activity relevance based on Renyi divergence. Build a divergence matrix for the activity prediction distribution (i.e., the probability assignment values of all possible activities) generated in step 4 to represent activity relevance. Its expression is:

[0079]

[0080] where α represents the preset Renyi coefficient, {A i} and {A j} represent a certain activity event. Then, use Renyi divergence to calculate the distribution difference between pairwise mass functions and build a divergence matrix. The expression is as follows:

[0081]

[0082] where D(m i , m j ) represents the Renyi divergence between the mass functions m i and m j .

[0083] In step 6: First, construct a difference loss function to represent the target to be optimized. Its expression is:

[0084]

[0085] where M represents the weighted sum of the n mass functions output in step 4. Its expression is:

[0086] M = w1m1 + w2m2 + … + w n m n

[0087] where w i represents the weight of the mass function m i .

[0088] Then, use the simulated annealing algorithm to adaptively optimize the constructed difference loss function L dif to obtain the weight combination w1, w2, …, w n under the optimal solution. Then, according to the obtained optimal weights, calculate the weighted mass function M of all activities.

[0089] In step 7: Use the Dempster-Shafer evidence fusion formula to fuse the information of different activities to be processed to calculate their activity relevance. First, construct the Dempster-Shafer evidence fusion formula, and its expression is as follows:

[0090]

[0091] where represents the mass fusion function, {A i} and {A j} represent two activity events, and K represents the conflict coefficient between the two activity events, and its expression is as follows:

[0092]

[0093] Next, use the Dempster-Shafer evidence theory to calculate the activity relevance between different activities, and its calculation expression is as follows:

[0094]

[0095] where represents the new mass function after the fusion of two mass functions, m p represents the activity probability distribution of target p, m q represents the activity probability distribution of target q, and M represents the weighted sum of all activity distributions generated in step 6. Finally, the result of can be used to represent the final activity relevance result, as shown in Figure 2 .

[0096] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0097] In addition, the descriptions such as "first", "second", etc. are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first", "second", etc. can explicitly or implicitly include at least one of such features.

[0098] Any process or method description shown in a flowchart or described in other ways in this specification can be understood as representing a module, segment, or part of code including one or more executable instructions for implementing a customized logic function or process, and the scope of the preferred embodiments of the present invention includes additional implementations, where the functions can be executed in a way that is not in the order shown or discussed, including in a substantially simultaneous manner according to the involved functions or in the reverse order, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0099] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware as in another embodiment, any one of the following techniques well known in the art or a combination of them can be used: discrete logic circuits having logic gate circuits for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0100] Those of ordinary skill in the art of this technology can understand that all or part of the steps carried by the method for implementing the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the application, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An airport activity relevance assessment method with autonomous adaptation and optimization, characterized in that, Including the following steps: Step 1: Collect video data containing airport surface activities, and label the airport activities in the video data, including surface targets, activity start time, end time, and activity category, to obtain the labeled airport surface video data; Step 2: Perform data preprocessing on the labeled airport surface video data to adapt to the input layer of the adopted feature extraction module; Step 3: Construct a feature extraction module based on a convolutional neural network, and input the video data preprocessed in Step 2 into the feature extraction module for feature extraction to obtain activity features related to airport surface targets; Step 4: Construct an activity mapping module based on a feedforward neural network, and represent the uncertainty of the activity prediction result through a quality function to generate the activity probability distribution of several activities, and obtain the quality function of each activity; Step 5: Use Renyi divergence to calculate the distribution difference between pairwise quality functions, and construct an n*n-dimensional divergence matrix based on the Renyi divergence between the quality functions of all activities to represent activity correlation; where n represents the number of activities; Step 6: Construct a difference loss function based on Renyi divergence, and adaptively optimize the constructed difference loss function based on the simulated annealing algorithm to learn a set of weighted coefficients representing the weights of the quality functions of each activity; and weight the different activity probability distributions in Step 4 based on this set of weighted coefficients to obtain a corrected activity probability distribution; Step 7: Use Dempster-Shafer evidence theory to perform information fusion on the corrected activity probability distribution obtained in Step 6 and the activity probability distribution obtained in Step 4 to generate the final activity correlation result.

2. The method according to claim 1, characterized in that, The data preprocessing in Step 2 specifically includes: Randomly extract frames from the labeled airport surface video data according to the set frame rate to obtain several image frame sequences; Normalize the resolution of the extracted image frame sequences, and then reconstruct the image size of all images into a specified size.

3. The method according to claim 2, wherein The set frame rate is 1 - 30 FPS.

4. The method according to claim 1, wherein The feature extraction module constructed in Step 3 is a feature extraction module based on the ResNet network structure to obtain N-dimensional activity features, where N is a preset value.

5. The method according to claim 1, wherein The activity mapping module constructed in step 4 is composed of a cascade of at least three layers of linear layers, the dimensions of each linear layer decreasing layer by layer, and the dimension of the last linear layer of the activity mapping module is 2 n , and the image activity output corresponding to each frame of video is a quality function m i , where the subscript i is the activity identifier.

6. The method according to claim 5, wherein In Step 4, the uncertainty representation of the activity prediction result through the quality function is specifically: m i : m({A1}) = a1, m({A2}) = a2, …, m({A1,A2}) = a n+1 , m({A1,A3}) = a n+2 , … Among them, m() represents the probability assignment function for activity events, {A} represents an activity event, and its subscript represents the activity index. For each mass function m i , it calculates the corresponding probability assignment values for different combinations of n activities respectively, representing the probability assignment values of each probability assignment function m(), and its subscript is the activity combination index of n activities. And for each mass function m i satisfies ∑a p = 1.

7. The method according to claim 6, characterized in that, In Step 5, the expression for Renyi divergence representing the quality functions of any two activities is: Among them, α represents the preset Renyi coefficient, {A i} and {A j} represent the i-th and j-th active events, m p (A i ), m q (A j ) respectively represent the probability assignment values of activity p and activity q with respect to their corresponding active events A i , A j .

8. The method according to claim 1, wherein In Step 6, the difference loss function is specifically: where D() represents the Renyi divergence between two mass functions, and M represents the weighted sum of all active mass functions, that is ω i represents the weight of the mass function m i .

9. The method according to claim 8, wherein Step 7 specifically includes the following steps: Step 7.1: Construct a Dempster-Shafer evidence fusion formula, and its expression is as follows: Among them, represents the new mass function after the fusion of two mass functions, {A i} and {A j} represent two activity events, and K represents the conflict coefficient between the two activity events. Its expression is as follows: Step 7.2: Use Dempster-Shafer evidence theory to calculate the activity relevance between any two activities, and its calculation expression is: where m p represents the mass function of activity p, and m q represents the mass function of activity q.

Images