Non-urban scene inter-strip point cloud registration method and device based on multi-scale point features

Abstract

The application discloses a method and device for inter-strip point cloud registration in non-urban scene based on multi-scale point features, and belongs to the technical field of unmanned aerial vehicle (UAV) laser radar point cloud data processing. In order to solve the problem that existing point cloud registration methods are difficult to extract structured features in non-urban scenes, multi-scale point features are extracted from UAV laser radar point cloud data, and key point pairing is performed according to the point features; random sample consensus (RANSAC) is used to select paired points and calculate point cloud transformation parameters, including a rotation matrix and a translation vector, so as to perform point cloud transformation on point clouds to be registered and realize point cloud registration. The method can be applied to point cloud data collected in non-urban scenes, does not need to extract structured features, is easy to implement, and eliminates spatial errors between point clouds of different strips. The application is suitable for point cloud registration of a laser radar.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle laser radar point cloud data processing, and particularly relates to a non-urban scene inter-strip point cloud registration method and system, a storage medium and equipment. BACKGROUND

[0002] Laser radar is an integrated system including a laser scanner, a global positioning system and an inertial measurement unit. Due to its high angular resolution and strong anti-interference capability, the laser radar has been widely applied to remote sensing, surveying and mapping, environmental protection and other fields. The laser radar installed on an unmanned aerial vehicle platform has the advantages of low cost, high point density and convenient data acquisition, which has attracted extensive attention of researchers. Due to the limited field of view, the unmanned aerial vehicle laser radar data acquisition usually adopts a multi-strip mode, and multiple strips are required to cover the entire observation scene. Due to factors such as laser ranging error, mirror angle scale error, installation error and system error, the same ground object point in adjacent strips usually has different spatial coordinates, thereby causing spatial differences. In actual application, spatial differences are inevitable, and direct use of original data will result in a large deviation of the result.

[0003] In order to eliminate spatial differences, point cloud registration processing is required. However, the existing point cloud registration methods are mostly concentrated in urban areas, and line features or surface features are extracted in structural elements such as buildings. However, in non-urban scenes, there is a lack of such ground objects with structural elements, and the corresponding point cloud data is also disordered, and it is difficult to extract line features and surface structures by traditional methods. SUMMARY

[0004] The application is to solve the problem that the existing point cloud registration method is difficult to extract structured features in non-urban scenes.

[0005] The non-urban scene inter-strip point cloud registration method based on multi-scale point features comprises the following steps:

[0006] Step 1, two point clouds to be registered are denoted as point cloud 1 and point cloud 2, and the outliers of point cloud 1 and point cloud 2 are removed, and then the overlapping region of point cloud 1 and point cloud 2 is selected;

[0007] Step 2, selecting uniformly distributed key points in the overlapping region selected from point cloud 1 and point cloud 2, selecting key point set P1 in point cloud 1, and selecting points within the r k neighborhood range centered on the key points of point cloud 1 as the key point set P2 of point cloud 2;

[0008] Step 3, extracting multi-scale point features of the key points, and the specific process comprises the following steps:

[0009] Step 3.1, take the point p0 in the point cloud 1 as the center of the sphere, divide its spherical neighborhood into several spherical shells according to a set of successively increasing radii, which is expressed as follows:

[0010] ps m ={p|R m <r≤R m+1 ,m=1,…,M}

[0011] wherein, ps m is the point set contained in the mth spherical shell, r is the Euclidean distance between the neighborhood point p and the center point p0 of the sphere, R m and R m+1 are the inner radius and outer radius of the mth spherical shell, and M represents the number of spherical shells;

[0012] Step 3.2, for each point set ps m contained in the spherical shell, construct two planes to satisfy the following two constraints: ① all points ps m in the spherical shell are located between the two planes; ② the two planes are parallel and the distance is minimum; the two constraints are expressed as follows:

[0013]

[0014] s.t.||w||2=1

[0015] wherein, ||·|| ∞ represents the infinite norm, and ||·||2 represents the two norm; and are the normal direction of the two planes and the intercept after projection to the normal direction;

[0016] Step 3.3, after constructing the above two planes, calculate the distance between the two planes and the distance from the center p0 of the sphere to the two planes and

[0017] Then concatenate the three distances to generate the feature vector and redefine wherein is the maximum value between and ; and are calculated as follows:

[0018]

[0019] wherein, |·| represents taking the absolute value, and according to the M different scales of the spherical shells with p0 as the center, the multi-scale feature vector is obtained through the above process, which is as follows:

[0020] f1 = [d 1 ,d 23 ] T

[0021]

[0022]

[0023] wherein (·) T denotes matrix transposition, is a multi-scale point feature of point p0;

[0024] Step 3.4, multi-scale point feature extraction is performed on all points in the point cloud 1 key point set P1 and the point cloud 2 key point set P2;

[0025] Step 4, based on the multi-scale feature vectors d 1 and d 23 , feature distances of the key points in the point cloud 1 and the point cloud 2 are determined, and then the point with the smallest feature distance is selected as the matching point; the random sample consensus algorithm is used for key point matching, the transformation parameters are calculated using the matching points, and then the point cloud to be registered and the transformation parameters are substituted into the point cloud transformation model to realize point cloud registration.

[0026] Further, the process of outlier elimination for the point cloud 1 and the point cloud 2 includes the following steps:

[0027] For each point p in each point cloud, the number N of other points in the neighborhood range with p as the center and r o as the radius is counted, and N is compared with the set threshold value N threshold , and for the point with a neighborhood point number less than the threshold value, the point is eliminated as an outlier, and each point in the point cloud 1 and the point cloud 2 is subjected to neighborhood point counting and threshold value comparison to complete outlier elimination.

[0028] Further, the overlapping region of the point cloud 1 and the point cloud 2 is as follows:

[0029] overlap = {p(x, y, z) | x min ≤ x ≤ x max , y min ≤ y ≤ y max}

[0030] x min = max{min(X1), min(X2)}, x max = min{max(X1), max(X2)}

[0031] y min = max{min(Y1), min(Y2)}, y max= min{max(Y1), max(Y2)}

[0032] wherein overlap represents the overlapping region, p(x, y, z) is a point in the point cloud, X1, Y1, Z1 and X2, Y2, Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

[0033] Further, based on the multi-scale feature vector d 1 and d 23 , the process of determining the feature distance of the key points in point cloud 1 and point cloud 2 includes the following steps:

[0034] For the extracted multi-scale point features, the feature distance between a point p1 in the key point set P1 of point cloud 1 and several key points p n in the key point set P2 of point cloud 2 is calculated; for the convenience of expression, d 1 is denoted as a, and d 23 is denoted as b; the a and b corresponding to the point p1 in the key point set P1 are recorded as a1 and b1, respectively, and the a and b corresponding to the point p n in the key point set P2 are recorded as a n and b n , respectively; the calculation process of the feature distance is as follows:

[0035] [Δa n ] m = |a n ] m - [a1] m |

[0036] [Δb n ] m = |b n ] m - [b1] m |

[0037] wherein [·] m represents the mth element of the vector;

[0038] Then, binary processing is performed, as follows:

[0039]

[0040]

[0041]

[0042] wherein sign is a sign operator, and mid represents the median; and the mth element in e M,m is equal to 1, and other elements are equal to 0;

[0043] Finally, the feature distance Δf1 is obtained n :

[0044] Δf1 n =∑(Δa n +Δb n )

[0045] where Δa n is a vector composed of the binarized [Δa n ] m , and Δb n is a vector composed of the binarized [Δb n ] m .

[0046] The non-urban scene inter-strip point cloud registration system based on multi-scale point features comprises:

[0047] An outlier elimination module: two point clouds to be registered are denoted as point cloud 1 and point cloud 2, and outlier elimination is performed on the point cloud 1 and the point cloud 2;

[0048] An overlapping area selection module: used for selecting an overlapping area of the point cloud 1 and the point cloud 2;

[0049] A key point selection module: selecting uniformly distributed key points in the overlapping area selected from the point cloud 1 and the point cloud 2, selecting a key point set P1 from the point cloud 1, and selecting, as a key point set P2 of the point cloud 2, points within a r k neighborhood range centered at the key points of the point cloud 1;

[0050] A multi-scale point feature extraction module: extracting multi-scale point features of the key points; the process of extracting the multi-scale point features of the key points comprises the following steps:

[0051] Step 3.1: taking a point p0 in the point cloud 1 as a spherical center, dividing a spherical neighborhood thereof into a plurality of spherical shells according to a set of successively increasing radii, and using the following formula:

[0052] ps m ={p|R m <r≤R m+1 ,m=1,…,M}

[0053] where ps m is a neighborhood point set contained in the mth spherical shell, r is the Euclidean distance between a neighborhood point p and the spherical center point p0, R m and R m+1 are the inner radius and the outer radius of the mth spherical shell, and M represents the number of the spherical shells;

[0054] Step 3.2: for each point set ps m, two planes are constructed to satisfy the following two constraints: ① all points ps m in the sphere are located between the two planes; ② the two planes are parallel and the distance is minimum; the two constraints are expressed as follows:

[0055]

[0056] s.t.||w||2=1

[0057] wherein, ||·|| ∞ represents the infinite norm, and ||·||2 represents the two norm; and are the normal direction of the two planes and the intercept after projection to the normal direction;

[0058] Step 3.3, after constructing the above two planes, the distance between the two planes and the distance from the sphere center p0 to the two planes and

[0059] Then concatenate the three distances to generate a feature vector and redefine wherein is the maximum value between and ; and are calculated as follows:

[0060]

[0061] wherein, |·| represents taking the absolute value, and according to the M different scale sphere shells with p0 as the sphere center, the multi-scale feature vector is obtained through the above process, as follows:

[0062] f1=[d 1 ,d 23 ] T

[0063]

[0064]

[0065] wherein, (·) T represents matrix transposition, is the multi-scale point feature of the point p0;

[0066] Step 3.4, multi-scale point feature extraction is performed on the point cloud 1 key point set P1 and the point cloud 2 key point set P2;

[0067] Feature distance calculation module: based on the multi-scale feature vector d1 and d 23 Determine the feature distance between key points in point cloud 1 and point cloud 2;

[0068] Point cloud registration module: Select the point with the smallest feature distance as the pairing point, use the random sampling consensus algorithm to pair key points, use the pairing points to calculate transformation parameters, and then substitute the point cloud to be registered and the transformation parameters into the point cloud transformation model to achieve point cloud registration.

[0069] Furthermore, the outlier removal module's processing steps include the following:

[0070] For each point p in the point cloud, the sphere r centered at p is calculated. o Let N be the number of other points within the radius's neighborhood, and then compare N with a set threshold N. threshold The comparison is performed, and points with fewer than a threshold in their neighborhood are identified as outliers and removed. The neighborhood count and threshold comparison are performed on each point in point cloud 1 and point cloud 2 to complete the outlier removal.

[0071] Furthermore, the processing steps of the overlapping region selection module include the following:

[0072] overlap={p(x,y,z)|x min ≤x≤x max ,y min ≤y≤y max}

[0073] x min =max{min(X1),min(X2)},x max =min{max(X1),max(X2)}

[0074] y min =max{min(Y1),min(Y2)},y max =min{max(Y1),max(Y2)}

[0075] Wherein, overlap represents the overlapping region, p(x,y,z) is a point in the point cloud, and X1,Y1,Z1 and X2,Y2,Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

[0076] Furthermore, the processing procedure of the feature distance calculation module includes the following steps:

[0077] For the extracted multi-scale point features, calculate a point p1 in the keypoint set P1 of point cloud 1 and several keypoints p in the keypoint set P2 of point cloud 2. n The characteristic distance between them; for ease of expression, d 1 Represented as a, d23 are denoted as b; a and b corresponding to p1 in the key point set P1 are denoted as a1 and b1 respectively, and a and b corresponding to p2 in the key point set P2 are denoted as a2 and b2 respectively. n are denoted as b; a and b corresponding to p1 in the key point set P1 are denoted as a1 and b1 respectively, and a and b corresponding to p2 in the key point set P2 are denoted as a2 and b2 respectively. n , b n ; the calculation process of the feature distance is as follows:

[0078] [Δa n ] m =|[a n ] m -[a1] m |

[0079] [Δb n ] m =|[b n ] m -[b1] m |

[0080] wherein [·] m represents the mth element of a vector;

[0081] Then, binarization processing is performed, as follows:

[0082]

[0083]

[0084]

[0085] wherein sign is a sign operator, and mid represents a median; and the mth element in e M,m is equal to 1, and other elements are equal to 0;

[0086] Finally, the feature distance Δf1 n is obtained:

[0087] Δf1 n =∑(Δa n +Δb n )

[0088] wherein Δa n is a vector composed of binarization-processed [Δa n ] m , and Δb n is a vector composed of binarization-processed [Δb n ] m .

[0089] A computer storage medium, the storage medium has at least one instruction stored therein, the at least one instruction is loaded by a processor and runs the multi-scale point feature based non-urban market scene inter-strip point cloud registration system.

[0090] A multi-scale point feature based non-urban market scene inter-strip point cloud registration device, the device comprises a processor and a memory, the memory has at least one instruction stored therein, the at least one instruction is loaded by the processor and runs the multi-scale point feature based non-urban market scene inter-strip point cloud registration system.

[0091] Compared with the prior art, the present application has the following advantages:

[0092] The purpose of the present application is to propose a multi-scale point feature based point cloud registration method for point cloud data collected by unmanned aerial vehicle laser radar in non-urban market scene, to perfect the existing point cloud registration method for non-urban market scene, to extract multi-scale point features in unmanned aerial vehicle laser radar point cloud data, and to pair key points according to the point features, to select paired points and calculate point cloud transformation parameters including rotation matrix and translation vector by using random sample consensus algorithm, and to perform point cloud transformation on the point cloud to be registered to realize point cloud registration. The method can be applied to point cloud data collected in non-urban market scene, does not need to extract structured features, is easy to implement, and eliminates the spatial error between different strip point clouds.

[0093] In order to verify the performance of the algorithm proposed in the present application, experiments are carried out on the point cloud data collected by unmanned aerial vehicle laser radar in non-urban market scene, and the experimental results verify the effectiveness of the multi-scale point feature based non-urban market scene inter-strip point cloud registration method proposed in the present application. BRIEF DESCRIPTION OF DRAWINGS

[0094] Figure 1 It is a flowchart of multi-scale point feature based non-urban market scene inter-strip point cloud registration.

[0095] Figure 2 It is a multi-scale point feature diagram.

[0096] Figure 3 It is original data.

[0097] Figure 4 It is point cloud registration data. DETAILED DESCRIPTION

[0098] The purpose of the present application is to propose a point cloud registration method based on multi-scale point features for point cloud data collected by unmanned aerial vehicle laser radar in non-urban market scenarios, to improve the existing point cloud registration method for non-urban market scenarios, to extract multi-scale point features in unmanned aerial vehicle laser radar point cloud data, and to pair key points according to the point features, to select paired points and calculate point cloud transformation parameters including rotation matrix and translation vector by using the random sample consensus algorithm, and to perform point cloud transformation on the point clouds to be registered to realize point cloud registration. The method can be applied to point cloud data collected in non-urban market scenarios, does not need to extract structured features, is easy to implement, and eliminates the spatial error between different flight strips of point clouds. The specific embodiments will be described in detail below.

[0099] Specific embodiment one: combined with Figure 1 The present embodiment is described,

[0100] The present embodiment is a point cloud registration method for non-urban market scenarios between flight strips based on multi-scale point features, including the following steps:

[0101] Step 1, pre-process the point cloud data to be registered, including outlier removal and overlap area selection; the specific process includes the following steps:

[0102] For unmanned aerial vehicle laser radar, there is noise in the point cloud acquisition process, and first, outlier removal processing is needed:

[0103] The two point clouds to be registered are denoted as point cloud 1 and point cloud 2;

[0104] For each point p in the point cloud, count the number N of other points in the neighborhood range with p as the center and r o as the radius, and compare N with the set threshold value N threshold , and for the points with a neighborhood point number less than the threshold value, they will be outliers and will be removed. The neighborhood point statistics and threshold comparison are performed for each point in point cloud 1 and point cloud 2, and the outlier removal is completed.

[0105] In order to ensure that the key points and paired points are located in the overlap area of point cloud 1 and point cloud 2, the overlap area overlap needs to be selected:

[0106] overlap={p(x,y,z)|x min ≤x≤x max ,y min ≤y≤y max}

[0107] x min =max{min(X1),min(X2)},x max =min{max(X1),max(X2)}

[0108] y min =max{min(Y1),min(Y2)},y max =min{max(Y1),max(Y2)}

[0109] Where p(x,y,z) is a point in the point cloud, and X1,Y1,Z1 and X2,Y2,Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

[0110] Step 2: Select uniformly distributed key points in the overlapping region of point cloud 1 and point cloud 2; the set of key points P1 selected in point cloud 1 should be uniformly distributed in the overlapping region, and select key points r in point cloud 2 centered on the key points of point cloud 1. k Points within the neighborhood are used as the key point set P2 of point cloud 2.

[0111] Step 3: Establish a multi-scale point feature model and extract multi-scale point features of key points; the specific process includes the following steps:

[0112] Step 3.1: Taking point p0 in point cloud 1 as the center of the sphere, divide its spherical neighborhood into several spherical shells according to a set of successively increasing radii, expressed by the following formula:

[0113] ps m ={p|R m <r≤R m+1 ,m=1,…,M}

[0114] Among them, ps m Let R be the set of neighborhood points contained within the m-th spherical shell, and let r be the Euclidean distance between neighborhood point p and the center point p0 of the sphere. m and R m+1 Let R be the inner and outer radii of the m-th spherical shell. m+1 =R m +t, where t represents the thickness of each spherical shell and M represents the number of spherical shells.

[0115] Step 3.2: For each spherical shell containing the point set ps m Construct two planes that satisfy the following two constraints: ① All points ps inside the spherical shell m Both are located between two planes; ② The two planes are parallel and the distance between them is minimized. The two constraints are expressed as follows:

[0116]

[0117] st||w||2=1

[0118] in, For a vector consisting entirely of 1s, ||·|| ∞ Let ||·||2 denote the infinity norm, and let ||·||2 denote the 2-norm. and are the normal direction of the two planes and the intercepts after projection to the normal direction.

[0119] Step 3.3, after constructing the above two planes, calculate three distances: the distance between the two planes and the distance from the sphere center p0 to the two planes and Then concatenate the three distances to generate a feature vector Considering d m elements are not independent of each other, redefine where is the maximum value between and . and are calculated as follows:

[0120]

[0121] where |·| represents taking the absolute value, and according to the M different scales of the spherical shell with p0 as the center, the multi-scale feature vector is obtained by the above process, as follows:

[0122] f1 = [d 1 , d 23 ] T

[0123]

[0124]

[0125] where (·) T represents matrix transposition, is the multi-scale point feature of point p0.

[0126] The obtained multi-scale point feature is shown in Figure 2 .

[0127] Step 3.4, perform the above multi-scale point feature extraction on all points in the point cloud 1 key point set P1 and the point cloud 2 key point set P2.

[0128] Step 4, use the random sample consensus algorithm to pair the key points, calculate the transformation parameters using the paired points, then substitute the to-be-registered point cloud and the transformation parameters into the point cloud transformation model to realize point cloud registration; the specific process includes the following steps:

[0129] Step 4.1, for the extracted multi-scale point features, calculate the feature distance between a point p1 in the point cloud 1 key point set P1 and several key points p n in the point cloud 2 key point set P2. For ease of expression, d1 denoted as a, d 23 denoted as b; a, b corresponding to p1 in P1 are denoted as a1, b1, and a, b corresponding to p2 in P2 are denoted as a2, b2 n denoted as a, d n , b n ; the calculation process of the feature distance is as follows:

[0130] [Δa n ] m =|[a n ] m -[a1] m |

[0131] [Δb n ] m =|[b n ] m -[b1] m |

[0132] wherein, [·] m represents the mth element of the vector;

[0133] Then, binary processing is performed, as follows:

[0134]

[0135]

[0136]

[0137] wherein, sign is a sign operator, and mid represents a median; and the mth element in e M,m is equal to 1, and other elements are equal to 0.

[0138] Finally, the feature distance Δf1 n is obtained:

[0139] Δf1 n =∑(Δa n +Δb n )

[0140] wherein, Δa n is a vector composed of [Δa n ] m after binary processing, and Δb n is a vector composed of [Δb n ] m after binary processing;

[0141] Step 4.2, after obtaining the multi-scale point feature distance, the point with the minimum feature distance is selected as the matching point. For several matching points, the random sample consensus (RANSAC) algorithm is used to calculate the point cloud transformation parameters, including the rotation matrix and the translation vector Then, the rotation matrix and the translation vector are used to perform point cloud transformation on the point cloud 2, and the point cloud registration is completed.

[0142] The experimental results of the registration of the application are shown in Figure 3 and Figure 4 , wherein Figure 3 is the original data (wherein ① and ② are local magnification effects of two different point clouds), and Figure 4 is the data after point cloud registration (wherein ① and ② are local magnification effects of two different point clouds). The evaluation index of the registration of the application is shown in Table 1.

[0143] Table 1 Quantitative evaluation index

[0144]

[0145] Specific implementation method two:

[0147] The embodiment is a multi-scale point feature-based point cloud registration system for non-urban scene flight strips, which comprises an outlier elimination module, an overlapping area selection module, a key point selection module, a multi-scale point feature extraction module, a feature distance calculation module and a point cloud registration module; wherein

[0148] The outlier elimination module: two point clouds to be registered are denoted as point cloud 1 and point cloud 2, and outlier elimination is performed on point cloud 1 and point cloud 2. The processing process of the outlier elimination module comprises the following steps:

[0149] For each point p in the point cloud, the number N of other points in the neighborhood range with p as the center and r o as the radius is counted, and N is compared with the set threshold value N threshold . For the points with a neighborhood point number less than the threshold value, the points are regarded as outliers and are eliminated. The neighborhood point number and the threshold value are compared for each point in point cloud 1 and point cloud 2, and the outlier elimination is completed.

[0150] The overlapping area selection module: used for selecting the overlapping area of point cloud 1 and point cloud 2; the processing process of the overlapping area selection module comprises the following steps:

[0151] overlap={p(x,y,z)|x min ≤x≤x max ,y min ≤y≤y max}

[0152] x min = max{min(X1), min(X2)}, x max = min{max(X1), max(X2)}

[0153] y min = max{min(Y1), min(Y2)}, y max = min{max(Y1), max(Y2)}

[0154] wherein overlap represents the overlapping region, p(x, y, z) is a point in the point cloud, X1, Y1, Z1 and X2, Y2, Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

[0155] Key point selection module: selecting uniformly distributed key points in the overlapping region selected from point cloud 1 and point cloud 2, selecting a key point set P1 from point cloud 1, and selecting points within the r k neighborhood range centered on the key points of point cloud 1 as the key point set P2 of point cloud 2.

[0156] Multi-scale point feature extraction module: extracting multi-scale point features of the key points; the process of extracting multi-scale point features of the key points includes the following steps:

[0157] Step 3.1, taking the point p0 in point cloud 1 as the center of a sphere, dividing its spherical neighborhood into several spherical shells according to a set of successively increasing radii, which is expressed as follows:

[0158] ps m = {p | R m < r ≤ R m+1 , m = 1, …, M}

[0159] wherein ps m is the neighborhood point set contained in the mth spherical shell, r is the Euclidean distance between the neighborhood point p and the center point p0 of the sphere, R m and R m+1 are the inner radius and outer radius of the mth spherical shell, and M represents the number of spherical shells;

[0160] Step 3.2, for each point set ps m contained in each spherical shell, two planes are constructed to satisfy the following two constraints: ① all points ps m in the spherical shell are located between the two planes; ② the two planes are parallel and have the minimum distance; the two constraints are expressed as follows:

[0161]

[0162] s.t. ||w||2 = 1

[0163] in, ||·|| ∞ represents the infinity norm, and ||·||2 represents the 2-norm; and It is the normal direction of the two planes and the intercept after projection onto the normal direction;

[0164] Step 3.3: After constructing the two planes described above, calculate the distance between the two planes. and the distance from the center of the sphere p0 to the two planes and

[0165] Then, these three distances are concatenated to generate a feature vector. And redefine in for and The maximum value between; and The calculation is as follows:

[0166]

[0167] Where |·| represents taking the absolute value, and based on M spherical shells of different scales with p0 as the center, the multi-scale feature vector is obtained through the above process, as shown in the following formula:

[0168] f1 = [d 1 ,d 23 ] T

[0169]

[0170]

[0171] in,(·) T Indicates matrix transpose. Multi-scale point features for point p0;

[0172] Step 3.4: Perform multi-scale point feature extraction on all key point sets P1 of point cloud 1 and P2 of point cloud 2.

[0173] Feature distance calculation module: based on multi-scale feature vector d 1 and d 23 Determine the feature distance between key points in point cloud 1 and point cloud 2; the feature distance calculation module's processing includes the following steps:

[0174] For the extracted multi-scale point features, calculate a point p1 in the keypoint set P1 of point cloud 1 and several keypoints p in the keypoint set P2 of point cloud 2. nbetween the features; for ease of expression, let d 1 denote a, d 23 denote b; let a1 and b1 denote a and b corresponding to the point p1 in the key point set P1, and let a2 and b2 denote a and b corresponding to the point p2 in the key point set P2 n . n n ; the calculation process of the feature distance is as follows:

[0175] [Δa n ] m = |[a n ] m -[a1] m |

[0176] [Δb n ] m = |[b n ] m -[b2] m |

[0177] wherein [·] m denotes the mth element of a vector;

[0178] Then, binary processing is performed, as follows:

[0179]

[0180]

[0181]

[0182] wherein sign is a sign operator, and mid represents a median; and the mth element in e M,m is equal to 1, and other elements are equal to 0;

[0183] Finally, the feature distance Δf1 n is obtained:

[0184] Δf1 n =∑(Δa n +Δb n )

[0185] wherein Δa n is a vector composed of [Δa n ] m after binary processing, and Δb n is a vector composed of [Δb n ] m after binary processing.

[0186] ​The point cloud registration module: the point with the minimum feature distance is selected as the matching point, the random sample consensus algorithm is used for key point matching, the transformation parameters are calculated by using the matching points, then the point cloud to be registered and the transformation parameters are substituted into the point cloud transformation model, and the point cloud registration is realized. THREE SPECIFIC EMBODIMENTS

[0188] The embodiment is a computer storage medium, and the storage medium stores at least one instruction, the at least one instruction is loaded by a processor and runs the point cloud registration system between flight strips in a non-urban market scene based on multi-scale point features.

[0189] It should be understood that the instructions include a computer program product, software or computerized method corresponding to any method described in the present application; the instructions can be used to program a computer system or other electronic device. The computer storage medium can include a readable medium having instructions stored thereon, and can include but is not limited to a magnetic storage medium, an optical storage medium, a magneto-optical storage medium, a read-only memory (ROM), a random access memory (RAM), an erasable programmable memory (such as an EPROM and an EEPROM), and a flash memory layer, or other types of media suitable for storing electronic instructions. FOUR SPECIFIC EMBODIMENTS

[0191] The embodiment is a point cloud registration device based on multi-scale point features in a non-urban market scene, and the device includes a processor and a memory. It should be understood that any device described in the present application includes a processor and a memory, and the device can also include other units, modules, etc. that display, interact, process, control, etc. through signals or instructions, and other functions;

[0192] The storage medium stores at least one instruction, and the at least one instruction is loaded by the processor and runs the point cloud registration system between flight strips in a non-urban market scene based on multi-scale point features.

[0193] The above examples of the present application are only used to illustrate the calculation model and calculation process of the present application, and are not limited to the embodiments of the present application. For those skilled in the art, other different forms of changes or variations can be made on the basis of the above description, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application still fall within the protection scope of the present application.

Claims

1. A point cloud registration method for non-urban scene flight strips based on multi-scale point features, characterized in that, Includes the following steps: Step 1: Label the two point clouds to be registered as point cloud 1 and point cloud 2 respectively. Remove outliers from point cloud 1 and point cloud 2, and then select the overlapping area of ​​point cloud 1 and point cloud 2. Step 2: Select uniformly distributed key points in the overlapping area of ​​point cloud 1 and point cloud 2. Select a key point set P1 in point cloud 1, and select a key point r in point cloud 2 centered on the key points in point cloud 1. k Points within the neighborhood are used as the key point set P2 of point cloud 2; Step 3: Extract multi-scale point features of key points. The specific process includes the following steps: Step 3.1: Taking point p0 in point cloud 1 as the center of the sphere, divide its spherical neighborhood into several spherical shells according to a set of successively increasing radii, expressed by the following formula: p.s. m ={p|R m <r≤R m+1 ,m=1,…,M} Among them, ps m Let R be the set of neighborhood points contained within the m-th spherical shell, and let r be the Euclidean distance between neighborhood point p and the center point p0 of the sphere. m and R m+1 Let M be the inner and outer radii of the m-th spherical shell, and M represent the number of spherical shells. Step 3.2: For each spherical shell containing the point set ps m Construct two planes that satisfy the following two constraints: ① All points ps inside the spherical shell m Both are located between two planes; ② The two planes are parallel and have the minimum distance; the two constraints are expressed as follows: st||w||2=1 in, ||·|| ∞ represents the infinity norm, and ||·||2 represents the 2-norm; and It is the normal direction of the two planes and the intercept after projection onto the normal direction; Step 3.3: After constructing the two planes described above, calculate the distance between the two planes. and the distance from the center of the sphere p0 to the two planes and Then, these three distances are concatenated to generate a feature vector. And redefine in for and The maximum value between; and The calculation is as follows: Where |·| represents taking the absolute value, and based on M spherical shells of different scales with p0 as the center, the multi-scale feature vector is obtained through the above process, as shown in the following formula: f1=[d 1 ,d 23 ] T in,(·) T Indicates matrix transpose. Multi-scale point features for point p0; Step 3.4: Perform multi-scale point feature extraction on all keypoint sets P1 of point cloud 1 and P2 of point cloud 2; Step 4: Based on multi-scale feature vector d 1 and d 23 The feature distances of key points in point cloud 1 and point cloud 2 are determined, and the point with the smallest feature distance is selected as the pairing point. The key point pairing is performed using the random sampling consensus algorithm, and the transformation parameters are calculated using the pairing points. Then, the point cloud to be registered and the transformation parameters are substituted into the point cloud transformation model to achieve point cloud registration.

2. The point cloud registration method for non-urban scene flight strips based on multi-scale point features according to claim 1, characterized in that, The outlier removal process for point cloud 1 and point cloud 2 includes the following steps: For each point p in the point cloud, the sphere r centered at p is calculated. o Let N be the number of other points within the radius's neighborhood, and then compare N with a set threshold N. threshold The comparison is performed, and points with fewer than a threshold in their neighborhood are identified as outliers and removed. The neighborhood count and threshold comparison are performed on each point in point cloud 1 and point cloud 2 to complete the outlier removal.

3. The point cloud registration method for non-urban scene flight strips based on multi-scale point features according to claim 2, characterized in that, The overlapping areas of point cloud 1 and point cloud 2 are as follows: overlap={p(x,y,z)|x min ≤x≤x max ,y min ≤y≤y max } x min <max{min(X1),min(X2)},x max <min{max(X1),max(X2)} y min =max{min(Y1),min(Y2)},y max =min{max(Y1),max(Y2)} Wherein, overlap represents the overlapping region, p(x,y,z) is a point in the point cloud, and X1,Y1,Z1 and X2,Y2,Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

4. The point cloud registration method for non-urban scene flight strips based on multi-scale point features according to claim 1, 2, or 3, characterized in that, Based on multi-scale feature vector d 1 and d 23 The process of determining the feature distance of key points in point cloud 1 and point cloud 2 includes the following steps: For the extracted multi-scale point features, calculate a point p1 in the keypoint set P1 of point cloud 1 and several keypoints p in the keypoint set P2 of point cloud 2. n The characteristic distance between them; for ease of expression, d 1 Represented as a, d 23 Let b be the key point; let a1 and b1 be the points corresponding to point p1 in key point set P1, respectively; and let a1 and b1 be the points corresponding to point p1 in key point set P2. n Let a and b be denoted as a respectively. n b n The calculation process for the feature distance is as follows: [Δa n ] m =|[a n ] m -[a1] m | [Δb n ] m =|[b n ] m -[b1] m | in,[·] m This represents the m-th element of the vector; Then, binarization is performed, as shown in the following formula: Where sign is the sign operator and mid represents taking the median; And e M,m The m-th element in the array is equal to 1, and all other elements are equal to 0; Finally, the feature distance Δf1 is obtained. n : Δf1 n =∑(Δa n +Δb n ) Where, Δa n [Δa] after binarization n ] m The vector formed, Δb n [Δb] after binarization n ] m The vector formed.

5. A point cloud registration system for non-urban scene flight strips based on multi-scale point features, characterized in that, include: Outlier Removal Module: The two point clouds to be registered are denoted as point cloud 1 and point cloud 2, and outlier removal is performed on point cloud 1 and point cloud 2. Overlapping region selection module: used to select the overlapping region between point cloud 1 and point cloud 2; Key point selection module: Selects uniformly distributed key points in the overlapping area of ​​point cloud 1 and point cloud 2. A key point set P1 is selected in point cloud 1, and a key point r is selected in point cloud 2 centered on the key points in point cloud 1. k Points within the neighborhood are used as the key point set P2 of point cloud 2; Multi-scale point feature extraction module: Extracts multi-scale point features of key points; the process of extracting multi-scale point features of key points includes the following steps: Step 3.1: Taking point p0 in point cloud 1 as the center of the sphere, divide its spherical neighborhood into several spherical shells according to a set of successively increasing radii, expressed by the following formula: p.s. m ={p|R m <r≤R m+1 ,m=1,…,M} Among them, ps m Let R be the set of neighborhood points contained within the m-th spherical shell, and let r be the Euclidean distance between neighborhood point p and the center point p0 of the sphere. m and R m+1 Let M be the inner and outer radii of the m-th spherical shell, and M represent the number of spherical shells. Step 3.2: For each spherical shell containing the point set ps m Construct two planes that satisfy the following two constraints: ① All points ps inside the spherical shell m Both are located between two planes; ② The two planes are parallel and have the minimum distance; the two constraints are expressed as follows: st||w||2=1 in, ||·|| ∞ represents the infinity norm, and ||·||2 represents the 2-norm; and It is the normal direction of the two planes and the intercept after projection onto the normal direction; Step 3.3: After constructing the two planes described above, calculate the distance between the two planes. and the distance from the center of the sphere p0 to the two planes and Then, these three distances are concatenated to generate a feature vector. And redefine in for and The maximum value between; and The calculation is as follows: Where |·| represents taking the absolute value, and based on M spherical shells of different scales with p0 as the center, the multi-scale feature vector is obtained through the above process, as shown in the following formula: f1=[d 1 ,d 23 ] T in,(·) T Indicates matrix transpose. Multi-scale point features for point p0; Step 3.4: Perform multi-scale point feature extraction on all keypoint sets P1 of point cloud 1 and P2 of point cloud 2; Feature distance calculation module: based on multi-scale feature vector d 1 and d 23 Determine the feature distance between key points in point cloud 1 and point cloud 2; Point cloud registration module: Select the point with the smallest feature distance as the pairing point, use the random sampling consensus algorithm to pair key points, use the pairing points to calculate transformation parameters, and then substitute the point cloud to be registered and the transformation parameters into the point cloud transformation model to achieve point cloud registration.

6. The point cloud registration system for non-urban scene flight strips based on multi-scale point features according to claim 5, characterized in that, The outlier removal module's processing steps include the following: For each point p in the point cloud, the sphere r centered at p is calculated. o Let N be the number of other points within the radius's neighborhood, and then compare N with a set threshold N. threshold The comparison is performed, and points with fewer than a threshold in their neighborhood are identified as outliers and removed. The neighborhood count and threshold comparison are performed on each point in point cloud 1 and point cloud 2 to complete the outlier removal.

7. The point cloud registration system for non-urban scene flight strips based on multi-scale point features according to claim 6, characterized in that, The processing steps of the overlapping region selection module include the following: overlap={p(x,y,z)|x min ≤x≤x max ,y min ≤y≤y max } x min <max{min(X1),min(X2)},x max <min{max(X1),max(X2)} y min =max{min(Y1),min(Y2)},y max =min{max(Y1),max(Y2)} Wherein, overlap represents the overlapping region, p(x,y,z) is a point in the point cloud, and X1,Y1,Z1 and X2,Y2,Z2 represent the three-dimensional coordinates of all points in point cloud 1 and point cloud 2, respectively.

8. The point cloud registration system for non-urban scene flight strips based on multi-scale point features according to claim 5, 6, or 7, characterized in that, The feature distance calculation module's processing steps include the following: For the extracted multi-scale point features, calculate a point p1 in the keypoint set P1 of point cloud 1 and several keypoints p in the keypoint set P2 of point cloud 2. n The characteristic distance between them; for ease of expression, d 1 Represented as a, d 23 Let b be the key point; let a1 and b1 be the points corresponding to point p1 in key point set P1, respectively; and let a1 and b1 be the points corresponding to point p1 in key point set P2. n Let a and b be denoted as a respectively. n b n The calculation process for the feature distance is as follows: [Δa n ] m =|[a n ] m -[a1] m | [Δb n ] m =|[b n ] m -[b1] m | in,[·] m This represents the m-th element of the vector; Then, binarization is performed, as shown in the following formula: Where sign is the sign operator and mid represents taking the median; And e M,m The m-th element in the array is equal to 1, and all other elements are equal to 0; Finally, the feature distance Δf1 is obtained. n : Δf1 n =∑(Δa n +Δb n ) Where, Δa n [Δa] after binarization n ] m The vector formed, Δb n [Δb] after binarization n ] m The vector formed.

9. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor as described in any one of claims 5 to 8, for the non-urban scene inter-strip point cloud registration system based on multi-scale point features.

10. A point cloud registration device for non-urban scene flight strips based on multi-scale point features, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, the at least one instruction being loaded by the processor and executed as described in any one of claims 5 to 8, the non-urban scene inter-strip point cloud registration system based on multi-scale point features.

Images