A method for fast capture of sparse point clouds by aircraft in low visibility

Abstract

The present invention discloses a method for quickly capturing aircraft with sparse point clouds in low visibility, and the steps are as follows: obtaining continuous point cloud data of a laser radar at an airport berth position, and preprocessing each frame of point cloud; density clustering, obtaining a nose cluster; judging the target position of an aircraft that enters the laser radar coverage range according to the relative position transformation of the coordinates of the center point of the nose cluster and the intersection of the taxi line and the stop line; performing density clustering for aircraft berthed to a designated position, obtaining a complete aircraft point cloud set, and performing intrusion detection. The present invention adaptively determines the point cloud resolution based on actual application scenarios, selects points at the position with the minimum number of neighbors to calculate the distance between two points, and uses the statistical average value as the cluster point cloud resolution.

Description

Technical Field

[0001] The invention belongs to the technical field of laser radar scanning, and in particular relates to a method for quickly capturing sparse point clouds by aircraft under low visibility. Background Art

[0002] LiDAR is less affected by environmental factors, but the scanned point cloud is sparser than the image. In low-visibility weather such as rain and fog, the point cloud is severely blocked, and the aircraft themselves will block each other, and the point cloud clusters of the same target in the point cloud are discrete. Therefore, the output of LiDAR is sparse 3D point cloud data, and the scanned object is incomplete, which poses a great obstacle to target detection and recognition and target size measurement.

[0003] A variety of aircraft types can be docked at each parking stand at the airport. The accurate parking position of the aircraft is when the front wheels are pressed on the intersection of the parking line and the taxiway of the corresponding aircraft type. When an aircraft enters a parking stand, it needs to go through three stages: turning and straightening (the fuselage is parallel to the taxiway) / deceleration / stopping. If the aircraft stops / slides before the parking line or does not slide to the stop line, it exceeds a certain range and the jet bridge cannot be docked. It must be towed to the correct position by a tractor. At the beginning of the parking guidance, manual assistance is used to improve the parking accuracy and reduce manpower consumption. Mature parking guidance replaces manual guidance, reduces manpower consumption, and reduces airport operating costs.

[0004] The preliminary work to realize parking guidance is to quickly obtain the complete cluster of aircraft that are about to park, determine whether the moving target on the scene is the aircraft that is about to park at the designated location, prevent the parking guidance system from being activated by mistake, distinguish the aircraft from other moving targets on the scene, and realize intrusion detection.

[0005] The existing technology uses the point cloud scanned by the lidar to perform plane segmentation and background removal, and uses the point cloud deep learning network to train the model for target detection and recognition. The point cloud obtained by the radar is sparse, and the deep learning network for point cloud target detection and recognition has a complex network model. This method is suitable for various target detections in point clouds, but it is not targeted at relatively special and specific situations such as airports. This method has high requirements on the machine, time-consuming model training, time-consuming detection and recognition, and low accuracy, especially in rainy and foggy weather, the accuracy is reduced.

[0006] The disadvantages of the existing technology are that the shape of the object obtained by LiDAR scanning is incomplete and the point cloud is sparse, the deep learning model is large, the training and calibration of the data set is time-consuming, the machine requirements are high and time-consuming, and the accuracy is not high. Although this method is widely used, it is rarely combined with actual scenarios and is not very practical. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for rapid capture of sparse point cloud aircraft under low visibility, so as to solve the problems in the prior art that the object shape obtained by lidar scanning in airport scenes is incomplete and the point cloud is sparse, the deep learning model is large in size, the training and calibration data set is complex and time-consuming, and the detection accuracy is low.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] The invention provides a method for quickly capturing sparse point clouds of aircraft in low visibility, comprising the following steps:

[0010] 1) Obtain continuous point cloud data of the laser radar at the airport berth location and pre-process each frame of point cloud;

[0011] 2) Density clustering to obtain machine head clusters;

[0012] 3) According to the relative position transformation of the center coordinates of the nose cluster and the intersection of the taxi line and the stop line, the target position of the aircraft entering the laser radar coverage is determined;

[0013] 4) For aircraft berthing at designated locations, density clustering is performed to obtain a complete set of aircraft point clouds and perform intrusion detection.

[0014] Furthermore, the step 1) specifically includes:

[0015] 11) Based on the area where the aircraft moves, remove points that do not belong to the area;

[0016] 12) performing plane segmentation on the acquired original point cloud frame by frame to remove the ground point set in the point cloud;

[0017] 13) Perform moving target detection based on continuous frames to obtain a point cloud containing only moving targets.

[0018] Furthermore, the step 13) is specifically as follows: based on the octree algorithm (Octree), spatial transformation detection of previous and next frames is performed, the aircraft enters the radar scanning range, spatial transformation detection is performed to obtain the moving target point cloud, when the number of points in the moving target point cloud of multiple consecutive frames is on an increasing trend, the coordinates of the highest point in the cluster meet the aircraft motion trajectory, the spatial transformation point cloud is saved based on time, and a moving target point cloud set based on a time series is obtained.

[0019] Furthermore, the step 2) specifically includes:

[0020] Based on the moving target point cloud set obtained, the neighborhood radius is initialized as the distance between the two farthest points of the guide vehicle cluster, the minimum number of neighbors is the number of points of the person cluster point in the point cloud, and the spatial transformation point cloud is density clustered to obtain the average height of the highest point of the cluster and the point with the minimum number of neighbors near the highest point in two consecutive frames. If its movement direction and location meet the aircraft motion trajectory, it is judged to belong to the aircraft, and the cluster contains the aircraft nose. The resolution of the cluster point cloud is calculated as the neighborhood radius of the next frame, the cluster points are saved, and the coordinates of the cluster center point are calculated and saved; the coordinates of the cluster center point for the next clustering are predicted according to the aircraft motion trajectory and Kalman filtering, and the cluster (previous frame) is drawn. The clusters obtained by clustering (including the cluster of the aircraft nose, etc.) are located in the area, and the next frame clustering is performed in this area, the aircraft clusters are saved, the coordinates of the cluster center points are calculated and saved, the cluster point cloud resolution is calculated, the area is updated based on the aircraft motion trajectory, and clustering is iterated according to the above method (clustering in the new area, calculating the cluster resolution according to the clustering result clustering, and using it as the clustering parameter of the next frame to update the cluster area); when the cluster resolution difference between the current and the next frames is less than the threshold, or the aircraft motion direction is determined based on the change in the coordinates of the cluster center points, the cluster resolution is no longer iterated, and clustering is continued with a fixed neighborhood radius. At this time, it is determined that the clustered cluster is the aircraft nose cluster.

[0021] Furthermore, the system state equation of the Kalman filter is: k =Ax k-1 +Bu k-1 +w k-1 ;

[0022] The observation equation of Kalman filter: z k =Hx k +v k ;

[0023] Among them, X k represents the system state matrix, A is the state transfer matrix, B is the control input matrix, W k-1 is the process noise; Z k Represents the observed value of the state matrix, that is, the measured value, H is the state observation matrix, V k is the measurement noise;

[0024] Kalman filter time update equation:

[0025]

[0026] in, and Respectively represent the posterior state estimates at time k and time k-1, that is, the most accurate estimates; represents the prior state estimate at time k, and predicts the result at time k based on the best estimate at the previous time; u k is the effect of the outside world on the system at time k; Q is the prediction noise covariance matrix; P k-1 and P k They represent the posterior estimated covariance at time k-1 and time k, respectively, that is, and The covariance of is the prior estimated covariance at time k, that is The covariance of

[0027] Kalman filter state update equation:

[0028]

[0029]

[0030]

[0031] Among them, K k is the filter gain matrix, R is the measurement noise covariance matrix, Represents the residual between the actual observation and the prediction, which is corrected by the Kalman gain prior to obtain the posterior, a matrix with I being 1. For a single model and a single measurement, I=1.

[0032] Furthermore, the step 3) specifically includes:

[0033] Assume that the taxiline direction is the y-axis direction, the stop line direction is the x-axis direction, the intersection of the taxiline and the stop line is set to D(x0, y0), and the coordinates of the center point of the nose cluster are (x, y). If the distance (abs(y, y0)) between the center point of the nose cluster and the intersection point D of the taxiline and the stop line in the taxiline direction is continuously decreasing, it indicates that the aircraft is approaching the monitored berth; if the distance (abs(y, y0)) in the taxiline direction increases or the cluster that meets the aircraft height characteristics is lost, it indicates that the aircraft is stopped at the berth opposite to the monitored berth; if the distance between the center point of the nose cluster and the intersection point D in the taxiline direction is basically unchanged, and the distance between the center point of the nose cluster and the intersection point D in the stop line direction first decreases and then increases, it indicates passing by.

[0034] Furthermore, the step 4) specifically includes:

[0035] When the aircraft turns, it starts to cluster its complete fuselage, initializes the neighborhood radius to the length of a single wing, saves the neighborhood radius of each frame, and the initial minimum number of neighbors is the average number of points in the radar scanned character clusters; the position clustering point cloud when the aircraft turns can obtain the complete fuselage and other moving targets that may exist outside the fuselage; according to the nose position of the previous frame, a cluster that is determined to include the fuselage is obtained, and the cluster point cloud resolution is calculated based on the cluster. According to the aircraft parking trajectory, the motion area including the cluster is delineated, and the next frame clustering is continued. Each time, the cluster area of ​​the next frame is delineated based on the aircraft parking trajectory on the basis of the cluster of the previous frame, and the cluster of the previous frame is divided into The resolution is used as the neighborhood radius of the next frame. When clustering the next frame, N clusters are obtained, N>1, and there are three possibilities: the first is that the aircraft cluster at the berth is segmented; the second is that the aircraft cluster is complete and not segmented, and the remaining N-1 clusters are other moving targets; the third is that the aircraft cluster is segmented and other moving target clusters are detected at the same time; the next frame is clustered with the same parameters to obtain the clusters and cluster center points of three consecutive frames, and the multi-target tracking method based on trajectory prediction is used to track the moving target. If the tracking is lost, the relative positions of the multiple clusters clustered and the nose cluster are used to assist in judging the corresponding relationship between the clusters. If it is still impossible to judge, the method is used The point cloud registration method is used to calculate the matching point cloud clusters (the point cloud registration also calculates the rotation and translation matrix of the registered point cloud), obtains the correspondence between multiple clusters in two consecutive frames, and judges the above three possibilities based on the motion direction and motion angle of the corresponding clusters; if the motion direction and motion angle of each pair of clusters are close, it indicates that the aircraft cluster is segmented and there are no other moving targets. The binary search method is used to iterate upward to calculate the neighborhood radius (the neighborhood radius r1 of the clustering result is only one cluster, the neighborhood radius r2 of multiple clusters at the cluster, and the binary search method is iterated upward The neighborhood radius r=(r1+r2) / 2) until clustering is achieved, and the most suitable neighborhood radius for this frame is obtained; if there is a cluster pair in the direction of motion, If the direction or angle is different from the movement direction and angle of the nose cluster, there is a moving target other than the aircraft, an alarm is issued, and whether the aircraft is segmented is determined based on the number of clusters. If only one cluster satisfies the aircraft parking trajectory, the aircraft cluster is not segmented, and the neighborhood radius continues to be iterated; if multiple clusters satisfy the aircraft motion trajectory, the clusters are segmented, and the binary search method is used to iterate the neighborhood radius upward until only one cluster satisfies the aircraft motion trajectory in this frame, and the iteration is stopped; through the above method, the aircraft in the actual scene is obtained, and the aircraft is distinguished from other moving targets on the scene. When other moving targets on the scene are in the aircraft parking space, they are segmented to achieve intrusion detection.

[0036] Furthermore, the step 4) specifically includes: the cluster movement direction is the center points of the two corresponding clusters, which are respectively recorded as: c1 and c2, the slope of the vector formed by the two points, the cluster movement angle is the angle between the vector α formed by the two points c1 and c2 and the unit vector β of the glide line, and the angle is recorded as θ; the cosine angle of the three-dimensional coordinates of the space is used for comparison, when the angle is acute, cosθ>0, when the angle is obtuse, cosθ<0, the angle between the vector α and the vector β, Vector α=(x1,y1,z1), vector β=(x2,y2,z2), α*β=x1x2+y1y2+z1z2, When the aircraft turns into the parking space, the cluster movement angle starts to decrease until it becomes parallel to the taxi line, and the directions and angles of cluster pairs of the same object are similar.

[0037] Beneficial effects of the present invention:

[0038] (1) The present invention aims at the situation where the airport scene is single and the number of background points or static target points is much larger than the number of moving target points. The present invention fully combines the actual scene situation to perform motion area division and plane segmentation, obtain the clusters of moving targets in the area, and save the spatial change point cloud according to the time series, so as to solve the problem of time-consuming processing and inaccurate moving target segmentation due to too many invalid points.

[0039] (2) The present invention solves the problem of clustering sparse point clouds of irregular structures by traditional clustering methods, where the same target is divided into multiple points or other targets and irregular structure targets are clustered into one cluster. It adopts density clustering and distinguishes them from other targets, initializes the neighborhood radius kernel and the minimum number of neighbors based on the actual scene, and iterates the neighborhood radius based on the cluster resolution. It realizes intrusion detection according to the movement direction and movement angle of the split cluster pairs.

[0040] (3) The present invention adaptively determines the point cloud resolution based on the actual application scenario, selects the point with the minimum number of neighbors to calculate the distance between the two points, and uses the statistical average value as the cluster point cloud resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0042] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0043] Reference Figure 1 As shown, a method for quickly capturing sparse point clouds of aircraft in low visibility of the present invention comprises the following steps:

[0044] 1) Obtain continuous point cloud data of the laser radar at the airport berth location and pre-process each frame of point cloud;

[0045] Specifically, the step 1) includes:

[0046] 11) Based on the area where the aircraft moves, remove points that do not belong to the area;

[0047] 12) performing plane segmentation on the acquired original point cloud frame by frame to remove the ground point set in the point cloud;

[0048] 13) Perform moving target detection based on continuous frames to obtain a point cloud containing only moving targets.

[0049] Among them, the step 13) is specifically as follows: based on the octree algorithm (Octree), spatial transformation detection of previous and next frames is performed, the aircraft enters the radar scanning range, spatial transformation detection is performed to obtain the moving target point cloud, when the number of points in the moving target point cloud of multiple consecutive frames is on an increasing trend, the coordinates of the highest point in the cluster meet the aircraft motion trajectory, the spatial transformation point cloud is saved based on time, and a moving target point cloud set based on a time series is obtained.

[0050] 2) Density clustering to obtain machine head clusters;

[0051] Specifically, the step 2) includes:

[0052] Based on the moving target point cloud set obtained, the neighborhood radius is initialized as the distance between the two farthest points of the guide vehicle cluster, the minimum number of neighbors is the number of points of the person cluster point in the point cloud, and the spatial transformation point cloud is density clustered to obtain the average height of the highest point of the cluster and the point with the minimum number of neighbors near the highest point in two consecutive frames. If its movement direction and location meet the aircraft motion trajectory, it is judged to belong to the aircraft, and the cluster contains the aircraft nose. The resolution of the cluster point cloud is calculated as the neighborhood radius of the next frame, the cluster points are saved, and the coordinates of the cluster center point are calculated and saved; the coordinates of the cluster center point for the next clustering are predicted according to the aircraft motion trajectory and Kalman filtering, and the cluster (previous frame) is drawn. The clusters obtained by clustering (including the cluster of the aircraft nose, etc.) are located in the area, and the next frame clustering is performed in this area, the aircraft clusters are saved, the coordinates of the cluster center points are calculated and saved, the cluster point cloud resolution is calculated, the area is updated based on the aircraft motion trajectory, and clustering is iterated according to the above method (clustering in the new area, calculating the cluster resolution according to the clustering result clustering, and using it as the clustering parameter of the next frame to update the cluster area); when the cluster resolution difference between the current and the next frames is less than the threshold, or the aircraft motion direction is determined based on the change in the coordinates of the cluster center points, the cluster resolution is no longer iterated, and clustering is continued with a fixed neighborhood radius. At this time, it is determined that the clustered cluster is the aircraft nose cluster.

[0053] In this example, the system state equation of the Kalman filter is:k =Ax k-1 +Bu k-1 +w k-1 ;

[0054] The observation equation of Kalman filter: z k =Hx k +v k ;

[0055] Among them, X k represents the system state matrix, A is the state transfer matrix, B is the control input matrix, W k-1 is the process noise; Z k Represents the observed value of the state matrix, that is, the measured value, H is the state observation matrix, V k is the measurement noise;

[0056] Kalman filter time update equation:

[0057]

[0058] in, and Respectively represent the posterior state estimates at time k and time k-1, that is, the most accurate estimates; represents the prior state estimate at time k, and predicts the result at time k based on the best estimate at the previous time; u k is the effect of the outside world on the system at time k; Q is the prediction noise covariance matrix; P k-1 and P k They represent the posterior estimated covariance at time k-1 and time k, respectively, that is, and The covariance of is the prior estimated covariance at time k, that is The covariance of

[0059] Kalman filter state update equation:

[0060]

[0061]

[0062]

[0063] Among them, K k is the filter gain matrix, R is the measurement noise covariance matrix, Represents the residual between the actual observation and the prediction, which is corrected by the Kalman gain prior (prediction) to obtain the posterior, a matrix with I being 1. For a single model and a single measurement, I=1.

[0064] 3) According to the relative position transformation of the center coordinates of the nose cluster and the intersection of the taxi line and the stop line, the target position of the aircraft entering the laser radar coverage is determined;

[0065] Specifically, the step 3) includes:

[0066] Assume that the taxiline direction is the y-axis direction, the stop line direction is the x-axis direction, the intersection of the taxiline and the stop line is set to D(x0, y0), and the coordinates of the center point of the nose cluster are (x, y). If the distance (abs(y, y0)) between the center point of the nose cluster and the intersection point D of the taxiline and the stop line in the taxiline direction is continuously decreasing, it indicates that the aircraft is approaching the monitored berth; if the distance (abs(y, y0)) in the taxiline direction increases or the cluster that meets the aircraft height characteristics is lost, it indicates that the aircraft is stopped at the berth opposite to the monitored berth; if the distance between the center point of the nose cluster and the intersection point D in the taxiline direction is basically unchanged, and the distance between the center point of the nose cluster and the intersection point D in the stop line direction first decreases and then increases, it indicates passing by.

[0067] 4) For aircraft berthing at designated locations, density clustering is performed to obtain a complete aircraft point cloud set and perform intrusion detection;

[0068] When the aircraft turns, it starts to cluster its complete fuselage, initializes the neighborhood radius to the length of a single wing, saves the neighborhood radius of each frame, and the initial minimum number of neighbors is the average number of points in the radar scanned character clusters; the position clustering point cloud when the aircraft turns can obtain the complete fuselage and other moving targets that may exist outside the fuselage; according to the nose position of the previous frame, a cluster that is determined to include the fuselage is obtained, and the cluster point cloud resolution is calculated based on the cluster. According to the aircraft parking trajectory, the motion area including the cluster is delineated, and the next frame clustering is continued. Each time, the cluster area of ​​the next frame is delineated based on the aircraft parking trajectory on the basis of the cluster of the previous frame, and the cluster of the previous frame is divided into The resolution is used as the neighborhood radius of the next frame. When clustering the next frame, N clusters are obtained, N>1, and there are three possibilities: the first is that the aircraft cluster at the berth is segmented; the second is that the aircraft cluster is complete and not segmented, and the remaining N-1 clusters are other moving targets; the third is that the aircraft cluster is segmented and other moving target clusters are detected at the same time; the next frame is clustered with the same parameters to obtain the clusters and cluster center points of three consecutive frames, and the multi-target tracking method based on trajectory prediction is used to track the moving target. If the tracking is lost, the relative positions of the multiple clusters clustered and the nose cluster are used to assist in judging the corresponding relationship between the clusters. If it is still impossible to judge, the method is used The point cloud registration method is used to calculate the matching point cloud clusters (the point cloud registration also calculates the rotation and translation matrix of the registered point cloud), obtains the correspondence between multiple clusters in two consecutive frames, and judges the above three possibilities based on the motion direction and motion angle of the corresponding clusters; if the motion direction and motion angle of each pair of clusters are close, it indicates that the aircraft cluster is segmented and there are no other moving targets. The binary search method is used to iterate upward to calculate the neighborhood radius (the neighborhood radius r1 of the clustering result is only one cluster, the neighborhood radius r2 of multiple clusters at the cluster, and the binary search method is iterated upward The neighborhood radius r=(r1+r2) / 2) until clustering is achieved, and the most suitable neighborhood radius for this frame is obtained; if there is a cluster pair in the direction of motion, If the direction or angle is different from the movement direction and angle of the nose cluster, there is a moving target other than the aircraft, an alarm is issued, and whether the aircraft is segmented is determined based on the number of clusters. If only one cluster satisfies the aircraft parking trajectory, the aircraft cluster is not segmented, and the neighborhood radius continues to be iterated; if multiple clusters satisfy the aircraft motion trajectory, the clusters are segmented, and the binary search method is used to iterate the neighborhood radius upward until only one cluster satisfies the aircraft motion trajectory in this frame, and the iteration is stopped; through the above method, the aircraft in the actual scene is obtained, and the aircraft is distinguished from other moving targets on the scene. When other moving targets on the scene are in the aircraft parking space, they are segmented to achieve intrusion detection.

[0069] Specifically, the step 4) further includes: the cluster movement direction is the center points of the two corresponding clusters, which are respectively recorded as: c1 and c2, the slope of the vector formed by the two points, the cluster movement angle is the angle between the vector α formed by the two points c1 and c2 and the unit vector β of the glide line, and the angle is recorded as θ; the cosine angle of the three-dimensional coordinates of the space is used for comparison, when the angle is acute, cosθ>0, when the angle is obtuse, cosθ<0, the angle between the vector α and the vector β, Vector α=(x1,y1,z1), vector β=(x2,y2,z2), α*β=x1x2+y1y2+z1z2, When the aircraft turns into the parking space, the cluster movement angle starts to decrease until it becomes parallel to the taxi line, and the directions and angles of cluster pairs of the same object are similar.

[0070] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principle of the present invention. These improvements should also be regarded as the protection scope of the present invention.

Claims

1. A method for rapidly capturing sparse point clouds of aircraft in low visibility, characterized in that: Here are the steps: 1) Obtain continuous point cloud data of the laser radar at the airport berth location and pre-process each frame of point cloud; 2) Density clustering to obtain machine head clusters; 3) According to the relative position transformation of the center coordinates of the nose cluster and the intersection of the taxi line and the stop line, the target position of the aircraft entering the laser radar coverage is determined; 4) For aircraft berthing at designated locations, density clustering is performed to obtain a complete set of aircraft point clouds and perform intrusion detection; The step 2) specifically includes: Based on the moving target point cloud set obtained as described, the neighborhood radius is initialized as the distance between the two farthest points of the guide vehicle cluster, the minimum number of neighbors is the number of points in the person cluster point in the point cloud, and the spatial transformation point cloud is density clustered to obtain the average height of the highest point of the cluster and the point with the minimum number of neighbors near the highest point in two consecutive frames. If its movement direction and location meet the aircraft motion trajectory, it is judged to belong to the aircraft, and the cluster contains the aircraft nose. The resolution of the cluster point cloud is calculated as the neighborhood radius of the next frame, the cluster points are saved, and the coordinates of the cluster center point are calculated and saved; according to the aircraft motion trajectory and Kalman filtering to predict the coordinates of the center point of the next clustering, delineate the area including the cluster, cluster the next frame in the area, save the aircraft cluster, calculate and save the coordinates of the center point of the cluster, calculate the point cloud resolution of the cluster, update the area based on the aircraft motion trajectory, and iterate clustering according to the above method; when the difference in cluster resolution between the previous and next frames is less than a threshold, or the direction of aircraft motion is determined based on the change in the coordinates of the center point of the cluster, the cluster resolution is no longer iterated, and clustering is continued with a fixed neighborhood radius, and at this time, it is determined that the cluster is the aircraft nose cluster; The step 4) specifically includes: When the aircraft turns, the complete fuselage is clustered, and the neighborhood radius is initialized to the length of a single wing. The neighborhood radius of each frame is saved, and the initial minimum number of neighbors is the average number of points in the human cluster scanned by the radar. The position clustering point cloud of the aircraft when turning can obtain the complete fuselage and other moving targets that may exist outside the fuselage. According to the nose position of the previous frame, a cluster that is determined to include the fuselage is obtained, and the cluster point cloud resolution is calculated based on the cluster. According to the aircraft parking trajectory, the motion area including the cluster is delineated, and the next frame clustering is continued. Each time, the clustering is based on the previous frame cluster according to the aircraft parking trajectory. The clustering area of ​​the next frame is delineated by the trajectory, and the cluster resolution of the previous frame is used as the neighborhood radius of the next frame. N clusters are obtained by clustering the next frame, N>1, and there are three possibilities: the first is that the aircraft cluster of the berth is segmented; the second is that the aircraft cluster is complete and not segmented, and the remaining N-1 clusters are other moving targets; the third is that the aircraft cluster is segmented and other moving target clusters are detected at the same time; the next frame is clustered with the same parameters to obtain the clusters and cluster center points of three consecutive frames, and the multi-target tracking method based on trajectory prediction is used to track the moving target. If the tracking is lost, the clustered The relative positions of multiple clusters and the nose cluster are used to assist in determining the corresponding relationship between the clusters. If it is still impossible to determine, the point cloud registration method is used to calculate the matching point cloud clusters, and the corresponding relationship between multiple clusters in two consecutive frames is obtained. The above three possibilities are determined based on the movement direction and movement angle of the corresponding clusters; if the movement direction and movement angle of each pair of clusters are close, it indicates that the aircraft cluster is segmented and there are no other moving targets. The neighborhood radius is calculated iteratively upward using the binary search method until it is clustered into one cluster, and the most suitable neighborhood radius for this frame is obtained; if the movement direction or angle of a cluster pair is different from that of the nose cluster, , then there is a moving target that is not an aircraft, an alarm is issued and it is determined whether the aircraft is segmented according to the number of clusters. If only one cluster satisfies the aircraft parking trajectory, the aircraft cluster is not segmented, and the neighborhood radius continues to be iterated; if multiple clusters satisfy the aircraft motion trajectory, the clusters are segmented, and the binary search method is used to iterate the neighborhood radius upward until only one cluster satisfies the aircraft motion trajectory in this frame, and the iteration is stopped; through the above method, the aircraft in the actual scene is obtained, and the aircraft is distinguished from other moving targets on the scene. When other moving targets on the scene are in the aircraft parking space, they are segmented to achieve intrusion detection.

2. The method for rapidly capturing sparse point clouds of aircraft in low visibility conditions according to claim 1, characterized in that: The step 1) specifically includes: 11) Based on the area where the aircraft moves, remove points that do not belong to the area; 12) performing plane segmentation on the acquired original point cloud frame by frame to remove the ground point set in the point cloud; 13) Perform moving target detection based on continuous frames to obtain a point cloud containing only moving targets.

3. The method for rapidly capturing sparse point clouds of aircraft in low visibility conditions according to claim 2, characterized in that: The step 13) is specifically as follows: based on the octree algorithm, spatial transformation detection of previous and next frames is performed, the aircraft enters the radar scanning range, and spatial transformation detection is performed to obtain the moving target point cloud. When the number of points in the moving target point cloud of multiple consecutive frames is on an increasing trend, the coordinates of the highest point in the cluster meet the aircraft motion trajectory, and the spatial transformation point cloud is saved based on time to obtain a moving target point cloud set based on a time series.

4. The method for rapidly capturing sparse point clouds from aircraft in low visibility conditions according to claim 1, characterized in that: The system state equation of the Kalman filter is: k =Ax k-1 +Bu k-1 +w k-1 ; The observation equation of Kalman filter: z k =Hx k +v k ; Among them, X k represents the system state matrix, A is the state transfer matrix, B is the control input matrix, W k-1 is the process noise; Z k Represents the observed value of the state matrix, that is, the measured value, H is the state observation matrix, V k is the measurement noise; Kalman filter time update equation: in, and Respectively represent the posterior state estimates at time k and time k-1, that is, the most accurate estimates; represents the prior state estimate at time k, and predicts the result at time k based on the best estimate at the previous time; u k is the effect of the outside world on the system at time k; Q is the prediction noise covariance matrix; P k-1 and P k They represent the posterior estimated covariance at time k-1 and time k, respectively. and The covariance of is the prior estimated covariance at time k, that is The covariance of Kalman filter state update equation: Among them, K k is the filter gain matrix, R is the measurement noise covariance matrix, Represents the residual between the actual observation and the prediction, which is corrected by the Kalman gain prior to obtain the posterior, a matrix with I being 1. For a single model and a single measurement, I=1.

5. The method for rapid capture of sparse point clouds by aircraft in low visibility according to claim 1, characterized in that: The step 3) specifically includes: Assume that the taxiline direction is the y-axis direction, the stop line direction is the x-axis direction, the intersection of the taxiline and the stop line is D(x0,y0), and the coordinates of the center point of the nose cluster are (x,y). If the distance (abs(y,y0)) between the center point of the nose cluster and the intersection point D of the taxiline and the stop line in the taxiline direction is decreasing, it means that the aircraft is approaching the monitored berth; if the distance (abs(y,y0)) in the taxiline direction increases or the cluster that meets the aircraft height characteristics is lost, it means that the aircraft is stopped at the berth opposite to the monitored berth; if the distance between the center point of the nose cluster and the intersection point D in the taxiline direction is basically unchanged, and the distance between the center point of the nose cluster and the intersection point D in the stop line direction first decreases and then increases, it means passing by.

6. The method for rapid capture of sparse point clouds by aircraft in low visibility according to claim 1, characterized in that: The step 4) specifically includes: the cluster movement direction is the center points of the two corresponding clusters, which are respectively recorded as: c1 and c2, the slope of the vector formed by the two points, the cluster movement angle is the angle between the vector α formed by the two points c1 and c2 and the unit vector β of the glide line, and the angle is recorded as θ; the cosine angle of the three-dimensional spatial coordinates is used for comparison, when the angle is acute, cosθ>0, when the angle is obtuse, cosθ<0, the angle between the vector α and the vector β, Vector α = (x1, y1, z1), vector β = (x2, y2, z2), α*β=x1x2+y1y2+z1z2, When the aircraft turns into the parking space, the cluster motion angle starts to decrease until it becomes parallel to the taxi line, and the directions and angles of cluster pairs of the same object are similar.

Images