Airport runway light anomaly detection and positioning method based on aerial video analysis

Abstract

The invention discloses an airport runway light anomaly detection and positioning method based on aerial video analysis. The method is characterized by comprising the following steps: S1, inputting a runway light sample data set into a YOLOV8 target detector for target detection training; s2, sequentially shooting video images through the unmanned aerial vehicle, inputting the video images into a YOLOV8 target detector to carry out target detection, and carrying out target data association by using a ByteTrack algorithm to obtain a target tracking trajectory set; s3, extracting coordinates of target trajectory points in the target tracking trajectory set, performing motion linear equation fitting, and obtaining a straight slope and a Y-axis intercept; s4, using a Gaussian mixture model to perform unsupervised clustering processing on the target trajectory points according to the motion features, and obtaining a left line lamp target trajectory set, a middle line lamp target trajectory set and a right line lamp target trajectory set; and carrying out runway light abnormity judgment, and outputting a target and position prompt. According to the invention, light abnormity of the runway center line lamp and the sideline lamp can be rapidly detected, and technical support is provided for airport guarantee.

Description

Technical Field

[0001] The present invention relates to the field of airport runway operation and maintenance, and particularly to a method for detecting and locating abnormal airport runway lights based on aerial video analysis. Background Art

[0002] In the modern aviation field, ensuring the operation safety of airports is of utmost importance. The safety detection of the airport runway lighting system, especially the safety detection of airport runway lights, is an important part of ensuring the safety of aircraft takeoff and landing. Generally, airport runway lights refer to the lighting devices distributed on both sides and in the middle of the straight runway of the airport, and are used to guide aircraft to take off and land correctly.

[0003] Traditional methods for detecting abnormal airport runway lights mainly rely on manual inspections. Staff need to check along the runway at specific time intervals to observe whether the lights are working properly. This method not only consumes a large amount of manpower and time, but also increases the difficulty and danger of manual inspections at night or under adverse weather conditions. With the development of technology, some automatic detection methods based on sensors and monitoring devices have gradually been applied. For example, using optoelectronic sensors to detect abnormalities in light brightness and color changes. However, the installation and maintenance costs of these hardware devices such as sensors are relatively high, and a single sensor can only monitor a limited area. To achieve full coverage of the entire airport runway lights, a large number of sensors are required, increasing the complexity and cost of the system.

[0004] With the development of aerial photography technology, aerial photography technology has also been gradually applied in the field of civil aviation safety detection. However, in the aspect of airport runway light detection, the method of using aerial video for analyzing abnormal lights is still in a technical blank state. Even if runway light images are obtained through aerial photography, there is no relevant technology in image processing and detection and recognition processing, and accurate recognition of runway lights cannot be achieved (especially in complex lighting conditions and background environments, it is difficult to accurately distinguish normal lights from abnormal lights, and it is even more difficult to accurately locate the position of abnormal lights). Therefore, the aerial detection of airport runway lights is the current direction of technological development. How to achieve technical problems such as target recognition, target tracking detection, abnormal light detection, and abnormal light location of airport runway lights is the current technical difficulty in research and development, and also restricts the technical expansion of unmanned aerial vehicle (UAV) aerial photography technology in the field of civil aviation runway detection. Summary of the Invention

[0005] The object of the present invention is to solve the technical problems pointed out in the background art, and to provide an abnormal detection and positioning method for airport runway lights based on aerial video analysis. Based on the runway light trajectory analysis of the aerial video, the motion trajectory is fitted and the light trajectories with similar motion characteristics are calculated. At the same time, GMM clustering analysis is carried out to obtain the light sets of the left line, the middle line and the right line of the runway respectively. Further analyze the trajectory information of each set (light count, spacing determination, brightness determination) to identify abnormal runway lights and locate the precise positions.

[0006] The object of the present invention is achieved by the following technical solutions:

[0007] An abnormal detection and positioning method for airport runway lights based on aerial video analysis, the method comprising:

[0008] S1. Construct a runway light sample data set, and input the runway light sample data set into the YOLOV8 object detector for object detection training;

[0009] S2. Use a drone to sequentially capture video images of the airport runway light area with the runway center line as the flight reference direction, input them into the YOLOV8 object detector for object detection, and use the ByteTrack algorithm for object data association to obtain a set of target tracking trajectories;

[0010] S3. Based on the horizontal direction of the video image as the X-axis and the vertical direction as the Y-axis, construct a coordinate system, extract the coordinates of the target trajectory points in the set of target tracking trajectories for motion linear equation fitting, and obtain the straight line slope and the Y-axis intercept, and use the straight line slope and the Y-axis intercept as motion features;

[0011] S4. Use the Gaussian mixture model to perform unsupervised clustering processing on the target trajectory points according to the motion features and divide them into four categories. The four categories are the left line lights, the middle line lights, the right line lights, and other lights. Eliminate the set of other light candidate categories to obtain the left line light target trajectory set, the middle line light target trajectory set, and the right line light target trajectory set;

[0012] S5. Calculate the distance between adjacent two runway lights in the video image of the video frame for the left line light target trajectory set, the middle line light target trajectory set, and the right line light target trajectory set respectively. If the distance is greater than the light interval threshold, there is a completely unlit runway light between the adjacent two runway lights. Mark the unlit position in the middle of the adjacent two runway lights and output correspondingly;

[0013] S6. Statistically calculate the historical brightness average value of the left - side line light target trajectory set, the middle - side line light target trajectory set, and the right - side line light target trajectory set according to the runway light target. If the historical brightness average value is less than the low - brightness mean threshold, it is determined that the runway light target brightness is too low and a low - brightness prompt is output accordingly; if the historical brightness average value is greater than the high - brightness mean threshold, it is determined that the runway light target brightness is too high and a high - brightness prompt is output accordingly.

[0014] To better implement the present invention, in step S5, extract the video frame position where the middle - position marker of two adjacent runway lights first appears not to be lit and output the position marker not lit; in step S6, extract the runway light target at the video frame position where the target brightness is too low first appears and output a low - brightness prompt, and extract the runway light target at the video frame position where the target brightness is too high first appears and output a high - brightness prompt.

[0015] Preferably, the runway light video images of the runway light sample data set are taken by an unmanned aerial vehicle with the runway center line as the flight reference direction, and the runway light video images are processed with runway light label annotation.

[0016] Preferably, the YOLOV8 target detector of the present invention performs frame - by - frame runway light target detection and annotation on the video image stream F(t), where t represents the sequential number of the video frame, and the ByteTrack algorithm performs data association and target tracking based on the targets in the front and rear frames being runway lights.

[0017] Preferably, the motion linear equation fitted for the coordinates (x, y) of the target trajectory point m is: y = a m ·x + b m where a m is the slope of the straight line and b m is the Y - axis intercept; the sum of the squared errors of the trajectory points to the fitted straight line is set as the optimization objective function E(i), x m represents the X - axis coordinate after the motion linear equation fitting, y is the y - axis coordinate before fitting, and M represents the total number of trajectory points for target data association.

[0018] Preferably, the least - squares method is used to solve the optimal fitting parameters a m and b m，优 b m，优 = y 均 - a m，优 ×x 均 where x 均 represents the mean value of the abscissas of the target trajectory points and y 均 represents the mean value of the ordinates of the target trajectory points.

[0019] Preferably, the Gaussian mixture model of the present invention includes the following methods in unsupervised clustering processing:

[0020] S41. Take the motion features of the target trajectory points as two-dimensional data points in the clustering process. The Gaussian mixture model determines that the probability density function of each data point S(i) is:

[0021] where K is the number of Gaussian components, and π k is the weight parameter of the k-th Gaussian component, N(S(i)|μ k , ∑ k ) is the probability density function of the k Gaussian distributions of the data point S(i), μ k is the mean vector parameter of the Gaussian distribution, ∑ k is the covariance matrix parameter of the Gaussian distribution, and θ is the vector parameter composed of all parameters;

[0022] S42. Iterative optimization. For each data point S(i), calculate the posterior probability generated by each Gaussian component:

[0023] r ik is the probability that the data point S(i) belongs to the k-th Gaussian component, and z i is the latent variable corresponding to the data point S(i);

[0024] S43. Re-estimate the mean, covariance matrix, and mixing coefficient of each Gaussian component to maximize the likelihood function of the entire model. The specific update process is as follows:

[0025]

[0026] where M represents the total number of trajectory points associated with the target data;

[0027] S44. Loop and execute the above steps until the parameters of the Gaussian mixture model converge or the number of loops reaches the preset threshold;

[0028] S45. Cluster the data points based on the optimal parameters of the Gaussian mixture model, and cluster the data points S(i) into the corresponding categories respectively.

[0029] Preferably, the parameter convergence is whether the change mean value before and after the update of the parameter μ k is less than the threshold thr_mu, and the preset threshold set for the number of loops is the threshold thr_loop.

[0030] Preferably, when performing clustering processing, the Gaussian mixture model calculates the mean slope of the data points included in each clustering set, takes the category set with the largest absolute value of the mean slope as the center line light set $Z_{mid}$, takes the category set with a mean slope greater than 0 as the left line light set $Z_{left}$, and takes the category set with a mean slope less than 0 as the right line light set $Z_{right}$.

[0031] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0032] (1) Based on the runway light trajectory analysis of the aerial video, the present invention fits the motion trajectory and calculates the light trajectories with similar motion characteristics. At the same time, GMM clustering analysis is performed to obtain the light sets of the left line, middle line, and right line of the runway respectively, and the trajectory information (light count, spacing determination, brightness determination) of each set is further analyzed to identify abnormal runway lights and locate the precise positions.

[0033] (2) Through the light trajectory analysis in the aerial video, the present invention realizes the accurate identification and precise positioning of abnormal lights on the left line, middle line, and right line of the airport runway; the present invention relies on a single unmanned aerial vehicle for regular runway light inspection, which has the characteristics of low cost, fast deployment, high flexibility, and strong scalability, helps the large-scale promotion of the method, and effectively improves the digital level of the airport.

[0034] (3) The present invention can effectively identify the light anomalies (abnormal light brightness, non - lighting) existing in the center line lights and side line lights (including left line lights and right line lights) of the airport runway in the aerial video, effectively saving the labor inspection cost and hardware deployment cost, and providing strong technical support for the safety guarantee of airport flights. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the method flow chart of the method for detecting and locating abnormal airport runway lights of the present invention;

[0036] Figure 2 is the schematic diagram of runway light recognition in the 4989th frame in the embodiment;

[0037] Figure 3 is the schematic diagram of runway light recognition in the 5700th frame in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The present invention will be further described in detail below in conjunction with the embodiments:

[0039] Embodiment

[0040] As Figure 1 shown, a method for detecting and locating abnormal airport runway lights based on aerial video analysis includes:

[0041] S1. Construct a runway lighting sample dataset and input the runway lighting sample dataset into the YOLOV8 object detector for object detection training. The runway lighting video images in the runway lighting sample dataset are taken by a drone with the runway center line as the flight reference direction (the flight direction is from the start end of the runway to the end of the runway at a constant speed), and the preferred height of the aerial photography route is 20 meters. The runway lighting video images are processed with runway lighting label annotation. The runway lighting video images in the runway lighting sample dataset need to cover various weather production scenarios and different lighting states of all lamp groups (including the too dark and too bright states based on brightness grading). When performing runway lighting label annotation processing, a further preferred technical solution is: label all normal runway lighting areas as light_normal, label the runway lighting areas with too low brightness as light_low, label the runway lighting areas with too high brightness as light_high, and perform separate annotations for too low and too high brightness. Subsequently, preliminary identification and judgment of too low and too high brightness can be carried out first, and then comparison and judgment between the historical brightness average value and the threshold can be performed.

[0042] S2. Use a drone to sequentially capture video images of the airport runway lighting area with the runway center line as the flight reference direction (the flight direction is from the start end of the runway to the end of the runway at a constant speed, and the preferred height of the aerial photography route is 20 meters; the video images are collected at intervals of 1 second for each image frame) and input them into the YOLOV8 object detector for object detection and use the ByteTrack algorithm for object data association to obtain a set of object tracking trajectories.

[0043] The YOLOV8 object detector performs frame-by-frame runway light object detection (see Figure 2 , for example, the object recognition and object number recognition results of a frame of video image) and annotation on the video image stream F(t). t represents the sequential number of the video frame. The ByteTrack algorithm performs data association and object tracking based on the runway light objects in the front and rear frames. The YOLOV8 object detector performs object recognition on the front and rear frames, and the ByteTrack algorithm associates the same runway light object. The YOLOV8 object detector and the ByteTrack algorithm jointly achieve object tracking. As Figure 3 shown, in this embodiment, the YOLOV8 object detector performs frame-by-frame runway light object detection and numbering. The YOLOV8 object detector gives a number to a newly emerging object, and the subsequent same object follows the same object number. The ByteTrack algorithm associates the same runway light object.

[0044] S3. A coordinate system is constructed with the horizontal direction of the video image as the X-axis and the vertical direction as the Y-axis. The coordinates of the target trajectory points in the target tracking trajectory set are extracted to fit a motion linear equation, and the slope of the straight line and the Y-axis intercept are obtained, which are used as motion features. The runway lights include the left line lights, the middle line lights, and the right line lights, and they show historical regularity in the video image of the UAV. Taking the left line lights as an example, each left line light is on the left side of the video image, and the historical trajectory of the same left line light target is a straight line located on the left side of the video image. Therefore, the straight line feature of the historical trajectory of the runway light target can be used for feature extraction and runway light type recognition. Since there are often other light sources (such as vehicles, warning lights, light sources of distant buildings and towers, other types of navigation lights, etc.) on both sides of the airport runway, these lights will be identified and labeled as targets, and the historical trajectories of these targets (defined as other lights) are substantially different from the runway light features and need to be excluded first. Through the creative analysis of a large number of aerial videos of runway lights, it is found that the motion directions (slopes) and the motion positions (offsets) in the image of the runway lights located on the left line, the middle line, and the right line have a high degree of similarity. Therefore, the present invention creatively uses motion features for recognition and other light exclusion processing.

[0045] In some embodiments, the method for fitting the motion linear equation of the target trajectory point m is as follows: The motion linear equation corresponding to the coordinates (x, y) of the target trajectory point m is fitted as: y = a m ·x + b m , where a m is the slope of the straight line, and b m is the Y-axis intercept; the sum of the squares of the errors from the trajectory points to the fitted straight line is set as the optimization objective function E(i), x m represents the X-axis coordinate after fitting the motion linear equation (based on the coordinate system constructed with the horizontal direction of the video image as the X-axis and the vertical direction as the Y-axis, and the X-axis coordinate is the X-axis coordinate data of the coordinate system), y is the y-axis coordinate before fitting, and M represents the total number of trajectory points associated with the target data. The least squares method is used to solve the optimal fitting parameters a m , opt and b m , opt, b m，优 = y 均 - a m，优 × x 均 , where x 均 represents the mean value of the abscissas of the target trajectory points, and y 均 represents the mean value of the ordinates of the target trajectory points. Thus, the a m , b m of the target at each trajectory point can be obtained, as well as the optimal a m，优 and b m，优 after fitting at the historical trajectory points of the same target., the runway lights' targets show regularity, and the sum of their squared errors is significantly small (preferably, a threshold can be set). At the same time, a m，优 and b m，优 (For those that cannot be fitted by other lights, they can be directly excluded). Based on this, other lights can be excluded.

[0046] S4. Use the Gaussian mixture model to perform unsupervised clustering on the target trajectory points according to their motion characteristics and divide them into four categories, namely the left line lights, the middle line lights, the right line lights, and other lights. Eliminate the set of candidate categories of other lights to obtain the left line light target trajectory set, the middle line light target trajectory set, and the right line light target trajectory set.

[0047] In some embodiments, the Gaussian mixture model includes the following methods in the unsupervised clustering process:

[0048] S41. Take the motion characteristics of the target trajectory points as two-dimensional data points in the clustering process. The Gaussian mixture model determines the probability density function of each data point S(i) as:

[0049] where K is the number of Gaussian components (set to 4 in this embodiment), π k is the weight parameter of the k-th Gaussian component (satisfying and π k ≥0), N(S(i)|μ k , ∑ k ) is the probability density function of the k Gaussian distributions of the data point S(i), μ k is the mean vector parameter of the Gaussian distribution, ∑ k is the covariance matrix parameter of the Gaussian distribution, and θ is the vector parameter composed of all the parameters. At the beginning of the processing of the Gaussian mixture model, the above parameters are all randomly initialized.

[0050] S42. Iterative optimization. For each data point S(i), calculate the posterior probability generated by each Gaussian component:

[0051] r ik is the probability that the data point S(i) belongs to the k-th Gaussian component, and z i is the latent variable corresponding to the data point S(i);

[0052] S43. Re-estimate the mean, covariance matrix, and mixing coefficient of each Gaussian component to maximize the likelihood function of the entire model. The specific update process is as follows:

[0053]

[0054] where M represents the total number of trajectory points associated with the target data.

[0055] S44. Repeat the above steps until the parameters of the Gaussian mixture model converge or the number of loops reaches a preset threshold. Parameter convergence means that the change mean value of the parameter μ before and after update is less than the threshold thr_m (set to 0.01 in this embodiment), and the preset threshold set for the number of loops is the threshold thr_loop (set to 500 in this embodiment). k Whether the change mean value before and after update is less than the threshold thr_m (set to 0.01 in this embodiment), and the preset threshold set for the number of loops is the threshold thr_loop (set to 500 in this embodiment).

[0056] S45. Cluster the data points based on the optimal parameters of the Gaussian mixture model, and cluster the data points S(i) into the corresponding categories respectively. After multiple iterations until convergence, according to the posterior probability of each data point belonging to each Gaussian component, the data points are assigned to the Gaussian component with the largest posterior probability to complete the clustering. The number of lights in the categories of the left line, the middle line, and the right line is much larger than the number of lights in other positions (obviously, the number of lights in the categories of the left line, the middle line, and the right line are the three categories with the largest number). Therefore, it is determined that the set with significantly fewer data points is the set where other lights are located, and this set is deleted (here, further deletion processing of other lights is performed); in this embodiment, after the Gaussian mixture model clustering is completed, all clustering sets are sorted in descending order of the number of data points, and the top three clustering sets in the sorting are retained (as the category sets), and the remaining clustering sets are all deleted. In some embodiments, when the Gaussian mixture model performs clustering processing, it calculates the slope mean value of the data points included in each clustering set, and takes the category set with the largest absolute value of the slope mean value as the middle line light set Z_mid, takes the category set with the slope mean value greater than 0 as the left line light set Z_left, and takes the category set with the slope mean value less than 0 as the right line light set Z_right.

[0057] S5. Calculate the distance between two adjacent runway lights in the video image of the video frame for the left line light target trajectory set, the middle line light target trajectory set, and the right line light target trajectory set respectively. If the distance is greater than the light interval threshold, there is a completely unlit runway light between the two adjacent runway lights, and mark it as unlit at the middle position between the two adjacent runway lights and output it correspondingly (thus locating and identifying the position of the unlit runway light target). In some embodiments, extract the video frame position where the unlit mark first appears at the middle position between two adjacent runway lights and perform the unlit position mark output.

[0058] S6. Statistically calculate the historical brightness average value of the left-side line light target trajectory set, the middle-line light target trajectory set, and the right-side line light target trajectory set according to the runway light target. If the historical brightness average value is less than the low brightness average threshold thr_low, it is determined that the brightness of the runway light target is too low and a low brightness prompt is output accordingly. If the historical brightness average value is greater than the high brightness average threshold thr_high, it is determined that the brightness of the runway light target is too high and a high brightness prompt is output accordingly. In some embodiments, the runway light target at the video frame position where the target brightness is too low first appears is extracted and a low brightness prompt is output (thereby locating and identifying the position of the runway light target with too low brightness), and the runway light target at the video frame position where the target brightness is too high first appears is extracted and a high brightness prompt is output (thereby locating and identifying the position of the runway light target with too high brightness). The present invention can effectively identify the abnormal lighting problems (abnormal lighting brightness, non-lighting) of the center line lights and side line lights (including left-side line lights and right-side line lights) on the airport runway in the aerial video, effectively saving the labor inspection cost and the hardware deployment cost, and providing a strong technical support for the safety guarantee of airport flights.

[0059] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting and locating abnormalities of airport runway lights based on aerial video analysis, characterized in that: The methods include: S1. Build a runway light sample dataset and input the runway light sample dataset into the YOLOV8 target detector for target detection training; S2, using the drone to take the runway centerline as the flight reference direction to sequentially shoot video images of the airport runway lighting area, input them into the YOLOV8 target detector for target detection, and use the ByteTrack algorithm to associate target data and obtain a target tracking trajectory set; S3, constructing a coordinate system based on the horizontal direction of the video image as the X-axis and the vertical direction as the Y-axis, extracting the coordinates of the target trajectory points in the target tracking trajectory set to fit the motion linear equation and obtain the straight line slope and Y-axis intercept, and using the straight line slope and Y-axis intercept as motion features; S4. Use a Gaussian mixture model to perform unsupervised clustering on the target trajectory points according to the motion characteristics and divide them into four categories, namely, left sideline lights, middle sideline lights, right sideline lights, and other lights. Eliminate the set of candidate categories of other lights to obtain the left sideline light target trajectory set, the middle sideline light target trajectory set, and the right sideline light target trajectory set; S5, respectively calculating the distance between two adjacent runway lights in the video image of the video frame for the left runway light target trajectory set, the middle runway light target trajectory set, and the right runway light target trajectory set. If the distance is greater than the light interval threshold, there is a completely unlit runway light between the two adjacent runway lights, and the position between the two adjacent runway lights is marked as unlit and output accordingly; S6. Perform historical brightness average statistics on the left side light target trajectory set, the middle side light target trajectory set, and the right side light target trajectory set according to the runway light target. If the historical brightness average is less than the low brightness average threshold, it is determined that the runway light target brightness is too low and a corresponding output brightness is too low prompt; if the historical brightness average is greater than the high brightness average threshold, it is determined that the runway light target brightness is too high and a corresponding output brightness is too high prompt.

2. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1 is characterized in that: In step S5, the video frame position where the position mark between two adjacent runway lights is not illuminated is extracted and the position mark is not illuminated output; in step S6, the runway light target where the target brightness is too low and the video frame position where the target brightness is too high is extracted and the video frame position where the target brightness is too high is output.

3. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1 is characterized in that: The runway light video images of the runway light sample data set are taken by a drone with the runway centerline as the flight reference direction, and the runway light video images are annotated with runway light labels.

4. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1 is characterized in that: The YOLOV8 target detector detects and annotates the runway light targets frame by frame in the video image stream F(t), where t represents the sequential number of the video frame. The ByteTrack algorithm performs data association and target tracking based on the runway lights in the previous and next frames.

5. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1 is characterized in that: The linear equation of motion corresponding to the coordinates (x, y) of the target trajectory point m is fitted as: y = a m ·x+b m , a m is the slope of the straight line, b m is the Y-axis intercept; the sum of squared errors from the trajectory point to the fitting line is set as the optimization objective function E(i), x m represents the x-axis coordinate after the linear equation of motion is fitted, y represents the y-axis coordinate before fitting, and M represents the total number of trajectory points associated with the target data.

6. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 5 is characterized in that: Use the least squares method to solve the optimal fitting parameter a for the optimization objective function m , U and B m ,excellent b m , good = y 均 -a m , Excellent 均 , where x 均 Represents the mean value of the horizontal coordinate of the target trajectory point, y 均 Represents the mean ordinate of the target trajectory points.

7. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1 is characterized in that: Gaussian mixture models include the following methods in unsupervised clustering: S41, taking the motion features of the target trajectory points as the two-dimensional data points of the clustering process, the Gaussian mixture model determines the probability density function of each data point S(i) as: Where K is the number of Gaussian components, π k is the weight parameter of the kth Gaussian component, N(S(i)|μ k ,∑ k ) is the k Gaussian distribution probability density function of the data point S(i), μ k is the mean vector parameter of the Gaussian distribution, ∑ k is the covariance matrix parameter of Gaussian distribution, θ is the vector parameter composed of all parameters; S42, iterative optimization, for each data point S(i), calculate the posterior probability generated by each Gaussian component: r ik is the probability that the data point S(i) belongs to the kth Gaussian component, z i is the hidden variable corresponding to the data point S(i); S43, re-estimate the mean, covariance matrix and mixing coefficient of each Gaussian component to maximize the likelihood function of the entire model. The specific updating process is: Where M represents the total number of trajectory points associated with the target data; S44, executing the above steps repeatedly until the parameters of the Gaussian mixture model converge or the number of cycles reaches a preset threshold; S45. Cluster the data points based on the optimal parameters of the Gaussian mixture model, and cluster the data points S(i) into corresponding categories.

8. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 7 is characterized in that: The parameter converges to the parameter μ k Whether the mean of the changes before and after the update is less than the threshold thr_mu, the preset threshold set by the number of cycles is the threshold thr_loop.

9. The method for detecting and locating abnormal airport runway lighting based on aerial video analysis according to claim 1, characterized in that: During clustering, the Gaussian mixture model calculates the mean slope of the data points contained in each cluster set, and takes the category set with the largest absolute value of the mean slope as the center line light set Z_mid, the category set with a mean slope greater than 0 as the left line light set Z_left, and the category set with a mean slope less than 0 as the right line light set Z_right.

Images