A two-stage key region semantic map construction method based on Hough transform

Abstract

The present invention discloses a two-stage key area semantic map construction method based on Hough transform. Aiming at the problem that existing UAV environment modeling involves a large amount of calculation and cannot always guarantee the real-time processing of actual ranging images on currently available airborne computing devices, a two-stage linear plane extraction method is proposed. In the first stage, the three-dimensional space of the original point cloud is decomposed into a two-dimensional plane subspace, and the search problem of the plane object in the three-dimensional space is converted into the search problem of the straight line segment in the two-dimensional subspace (scanning plane), thereby significantly reducing the amount of calculation by reducing the dimensions of the search object and the search space; the second stage also does not involve a large amount of calculation, because the search for the plane object is not based on the original point cloud, but on the straight line segments extracted in the first stage, and the number of these straight line segments is much less than the number of points in the original image, which further reduces the amount of calculation, thereby improving the autonomous task execution efficiency of the UAV in the industrial city environment.

Description

Technical Field

[0001] The invention belongs to the technical field of unmanned aerial vehicle environment modeling, and in particular relates to the design of a two-stage key area semantic map construction method based on Hough transform. Background Art

[0002] Drones are increasingly in demand in industrial urban environments, which are characterized by shielded areas that significantly limit the use of traditional remote control and navigation devices. In this context, robotic devices with higher autonomy are more promising, capable of operating independently for a period of time without artificial (e.g. satellite navigation) and natural (e.g. geomagnetic field) navigation signals. When building these autonomous control systems, key tasks are the modeling of the external environment and the navigation system. Solving these tasks in real time on an onboard computer based on onboard sensor data enables goal-oriented planning and execution of safe motion trajectories without operator intervention, which is the basis for autonomous operation. For autonomous systems, the most common and preferred onboard sensors are 2D and 3D laser sensors, which are able to acquire geometric information of the external environment directly in the field of view in the form of point clouds. The raw point cloud data generated by laser sensors is usually voluminous and therefore not always suitable for onboard control systems. Therefore, how to compress the raw data without losing geometric and navigation information becomes a key issue. A more promising solution is to represent these raw odometry images as a collection of geometric linear primitives (such as planes, lines, angles, etc.), which are the main components of industrial urban environments, such as roads, sidewalks, fences, walls, floors, steps, and other parts of industrial urban infrastructure, buildings, and rooms.

[0003] Many studies are currently devoted to solving the problem of extracting linear objects from point clouds. The most famous of these are RANSAC (Random Sample Consensus) type algorithms, which extract a subset of points belonging to a specific plane in a point cloud through an iterative method to meet a given tolerance requirement. RANSAC type algorithms are robust to noise in the input data, but as the number of points in the point cloud increases and the number of planes increases, the algorithm will run slower. In addition, algorithms based on local normal calculation are also widely used. These algorithms extract planes by calculating the neighborhood of each point and then merging points with similar normals. However, this type of algorithm is sensitive to sensor errors because the local normal of each point is calculated based on the neighborhood points, and errors in the measured coordinates can lead to significant errors in the calculation of local normals. There are also some methods for extracting linear objects. Global methods for 3D image plane segmentation based on energy function minimization, especially 3D image plane segmentation based on graph cut methods, and interactive unstructured point cloud segmentation based on graph cut methods are proposed, in which the cost function is obtained by the Euclidean distance of the points in the point cloud.

[0004] Practical experience shows that the solution to extract linear objects from point clouds involves a lot of calculations, which does not always guarantee real-time processing of actual ranging images on currently available airborne computing devices. Therefore, it is particularly important to create and study algorithms and software and hardware tools that can effectively solve this problem. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of large data volume, slow processing speed and poor real-time performance faced by existing UAV systems in industrial and environmental applications when performing environmental modeling. A two-stage key area semantic map construction method based on Hough transform is proposed, which can effectively extract key information from a large amount of data, thereby constructing an accurate and detailed environmental model.

[0006] The technical solution of the present invention is: a two-stage key area semantic map construction method based on Hough transform, comprising the following steps:

[0007] S1. Collect stereo ranging images of the external environment of the UAV.

[0008] S2. Layer the stereo ranging image according to different scanning angles to obtain a multi-layer scanning plane.

[0009] S3. Use two-dimensional Hough transform to extract line segments in each scanning plane to obtain a line segment subset.

[0010] S4. Re-cluster the incorrectly clustered line segments in each line segment subset to obtain a point subset for each layer of the scanning plane.

[0011] S5. Construct a plane equation based on a subset of points on each layer of the scanned plane, and then construct a semantic map of the key area.

[0012] S6. Determine the displacement of the UAV and incrementally update the semantic map.

[0013] Furthermore, step S1 includes the following sub-steps:

[0014] S11. Use the 3D laser sensor on the drone to collect the original point cloud data of the drone's external environment.

[0015] S12, filtering the original point cloud data to obtain a stereo ranging image of the external environment of the UAV.

[0016] Further, step S2 includes the following sub-steps:

[0017] S21. Establish the spatial coordinate system of the space where the drone is located. The axis rotates to obtain different scanning angles.

[0018] S22, layering the stereo ranging image according to different scanning angles.

[0019] S23, projecting the layered point cloud into a binary plane image to obtain a multi-layer scanning plane.

[0020] Further, step S3 includes the following sub-steps:

[0021] S31, using two-dimensional Hough transform to extract line segments in each scanning plane, and counting the extracted line segments and The intersection point on the axis.

[0022] S32, construct an accumulation function, whose value is The number of line segments that have the same intersection point on the axis.

[0023] S33, according to the value of the accumulation function, The local maximum of the intersection histogram is extracted from the local interval on the axis. The line segments with the same intersection point on the axis are clustered to obtain line segment subsets.

[0024] Furthermore, the two-dimensional Hough transform formula in step S31 is:

[0025] ;

[0026] in represents the vertical distance from the origin to the line segment, Represents the vertical line The angle between the axes, Represents the coordinates in the scan plane.

[0027] Further, step S4 includes the following sub-steps:

[0028] S41. Count the line segments in each line segment subset and The intersection of the planes.

[0029] S42, will The line segments whose intersection points of the planes are on the same straight line are re-clustered to obtain line segments belonging to the same scanning plane.

[0030] S43. Obtain a point subset of the layer of scanning plane according to line segments belonging to the same scanning plane.

[0031] Further, step S5 includes the following sub-steps:

[0032] S51. Randomly select three points from a point subset belonging to the same scanning plane to construct an initial plane equation.

[0033] S52. Using the solution of the initial plane equation as the initial value, construct a least squares equation, the expectation of which is that the distance from a point on the same plane to the plane is the minimum.

[0034] S53. Solve the least squares equation to obtain the precise plane equation of the key area, and add corresponding plane labels to the point cloud data of the key area through the precise plane equation of the key area to construct a semantic map of the key area.

[0035] Further, step S6 includes the following sub-steps:

[0036] S61, matching each layer of scanned plane with the plane in the existing database, and obtaining List of corresponding planes in the local and global coordinate systems of the drone:

[0037] ;

[0038] in Indicates The unit normal vector of the plane in the local coordinate system of the drone, Indicates The unit normal vector of a plane in the global coordinate system is Represents the direction vector of any point on the plane in the local coordinate system of the drone, Represents the direction vector of any point on the plane in the global coordinate system, Indicates The distance from the plane to the origin of the local coordinate system of the drone, Indicates The distance from the plane to the origin of the global coordinate system.

[0039] S62. Construct an overdetermined system of nonlinear equations based on a list:

[0040] ;

[0041] in Indicates The vector from the plane to the origin of the local coordinate system of the drone, Represents the displacement vector between the local coordinate system of the drone and the global coordinate system, Represents the rotation matrix between the local coordinate system of the drone and the global coordinate system. .

[0042] S63. Solve the nonlinear equations to obtain the displacement of the UAV.

[0043] S64. Determine a new plane according to the displacement of the UAV, and incrementally update the semantic map according to the new plane.

[0044] The beneficial effects of the present invention are as follows: the present invention aims at the problem that the existing UAV environment modeling involves a large amount of calculations and cannot always guarantee the real-time processing of the actual ranging image on the currently available airborne computing equipment, and proposes a two-stage linear plane extraction method. In the first stage, the three-dimensional space of the original point cloud is decomposed into a two-dimensional plane subspace, and the search problem of the plane object in the three-dimensional space is converted into the search problem of the straight line segment in the two-dimensional subspace (scanning plane), thereby significantly reducing the amount of calculation by reducing the dimensions of the search object and the search space; the second stage also does not involve a large amount of calculations, because the search for the plane object is not based on the original point cloud, but on the straight line segments extracted in the first stage, and the number of these straight line segments is far less than the number of points in the original image. The present invention can convert a large amount of original information from the ranging sensor into a compact geometric and semantic environment, thereby improving the efficiency of autonomous task execution of UAVs in industrial urban environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The figure shows a flow chart of a two-stage key area semantic map construction method based on Hough transform provided by an embodiment of the present invention.

[0046] Figure 2 Shown is a schematic diagram of original point cloud data provided by an embodiment of the present invention.

[0047] Figure 3 FIG. 1 is a schematic diagram of line segment extraction using two-dimensional Hough transform provided by an embodiment of the present invention.

[0048] Figure 4 The figure is a schematic diagram of the mapping relationship between the image space and the parameter space under the two-dimensional Hough transform provided by an embodiment of the present invention.

[0049] Figure 5 It is a schematic diagram showing the principle of clustering line segment subsets in the same scanning plane and eliminating erroneous clustering results provided by an embodiment of the present invention.

[0050] Figure 6 Shown is a distribution histogram of the intersection of a line segment subset and a coordinate axis provided by an embodiment of the present invention.

[0051] Figure 7 The figure shows the line segments and the Schematic diagram of the straight lines where the planes intersect.

[0052] Figure 8 Shown is a schematic diagram of a double-stage plane extraction result provided by an embodiment of the present invention.

[0053] Fig. 9 The figure is a schematic diagram of the mapping relationship between the image space and the parameter space under the three-dimensional Hough transform provided by an embodiment of the present invention.

[0054] Fig.10 It is a schematic diagram showing matching of the scan plane provided by the embodiment of the present invention with the plane in the existing database.

[0055] Fig.11 Shown is a schematic diagram of plane equation transformation provided by an embodiment of the present invention.

[0056] Fig.12 Shown are a photograph and an extracted plan view of the first spatial region provided in an embodiment of the present invention.

[0057] Fig.13 Shown are a photograph and an extracted plan view of a second spatial region provided in an embodiment of the present invention.

[0058] Fig.14 It is a schematic diagram of the calculation result of the relative position of the sensor in the working area provided by the embodiment of the present invention.

[0059] Fig.15 FIG. 4 is a schematic diagram of a semantic map construction result in a complex indoor situation provided by an embodiment of the present invention.

[0060] Fig.16 Shown is a schematic diagram of a semantic map construction result in an outdoor large scene environment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0062] The embodiment of the present invention provides a two-stage key area semantic map construction method based on Hough transform, such as Figure 1 As shown, the following steps S1 to S6 are included:

[0063] S1. Collect stereo ranging images of the external environment of the UAV.

[0064] Step S1 includes the following sub-steps S11-S12:

[0065] S11, collect the original point cloud data of the external environment of the drone by carrying a 3D laser sensor on the drone, such as Figure 2 shown.

[0066] S12, filtering the original point cloud data to obtain a stereo ranging image of the external environment of the UAV.

[0067] S2. Layer the stereo ranging image according to different scanning angles to obtain a multi-layer scanning plane.

[0068] Step S2 includes the following sub-steps S21 to S23:

[0069] S21. Establish the spatial coordinate system of the space where the drone is located. The axis rotates to obtain different scanning angles.

[0070] S22, layering the stereo ranging image according to different scanning angles.

[0071] S23, projecting the layered point cloud into a binary plane image to obtain a multi-layer scanning plane.

[0072] S3. Use two-dimensional Hough transform to extract line segments in each scanning plane to obtain a line segment subset.

[0073] Step S3 includes the following sub-steps S31 to S33:

[0074] S31, using two-dimensional Hough transform to extract line segments in each scanning plane, and counting the extracted line segments and The intersection point on the axis.

[0075] In the embodiment of the present invention, Figure 3 As shown in Figure 1, two-dimensional Hough transform is used to extract line segments. The two-dimensional Hough transform is designed to find straight lines in binary contour plane images. Figure 4 As shown, the 2D Hough transform is based on using a 2D parameter space in which a straight line represented by a normal is searched:

[0076] ;

[0077] in represents the vertical distance from the origin to the line segment, Represents the vertical line The angle between the axes, represents the coordinates in the scan plane. For each point in the image , we can find a set of points in the parameter space , which correspond to all possible passing points straight line.

[0078] gather is the point in parameter space The response represents the vector-based Point on the circle with diameter The polar coordinate set of (through the point and The line is perpendicular to the vector , because the angle is a right angle, being the angle inscribed in a circle based on its diameter).

[0079] If we perform the Hough transform on all points of the original discrete image and calculate the response value in the discrete parameter space, we can construct the discrete function , the discrete function is The value at is equal to the original space Points on the corresponding straight line .

[0080] S32, construct an accumulation function, whose value is The number of line segments that have the same intersection point on the axis.

[0081] S33, according to the value of the accumulation function, The local maximum of the intersection histogram is extracted from the local interval on the axis. The line segments with the same intersection point on the axis are clustered to obtain line segment subsets.

[0082] In the embodiment of the present invention, the size of the local interval is determined by the measurement error of the sensor and the discreteness of the parameter space. If there are multiple straight line segments in the image, the accumulation function will have multiple local maxima at the points corresponding to the existing straight line segments. Therefore, to detect all straight line segments in the original image, it is only necessary to find all the significant local maxima of the accumulation function.

[0083] S4. Re-cluster the incorrectly clustered line segments in each line segment subset to obtain a point subset for each layer of the scanning plane.

[0084] Step S4 includes the following sub-steps S41 to S43:

[0085] S41. Count the line segments in each line segment subset and The intersection of the planes.

[0086] S42, will The line segments whose intersection points of the planes are on the same straight line are re-clustered to obtain line segments belonging to the same scanning plane.

[0087] S43. Obtain a point subset of the layer of scanning plane according to line segments belonging to the same scanning plane.

[0088] Due to the different planes at different positions and angles in space, the line segments they contain are The axes may have the same intersection points, so the line segment subset extracted in step S3 contains the line segments that are incorrectly clustered. The line segments that are incorrectly clustered need to be re-clustered in step S4. The intersection of the planes should be on the same straight line, so it will be After the line segments whose intersection points are on the same straight line are correctly clustered, the line segments that were incorrectly clustered in step S3 are re-clustered.

[0089] In the embodiment of the present invention, steps S3-S4 are a two-stage plane extraction process.

[0090] like Figure 5 As shown, according to the elevation angle Scan. is a planar object in the sensor's field of view, , and Corresponding elevation angles (The angle value is 0, so Figure 5 Not shown in and The scanning plane ( , and If the plane object Not parallel to axis, then the line segments belonging to the object ( , , ), that is, the line segment extracted in the first stage of the algorithm, will be in the plane and The intersection of the axes If the plane object Not parallel to axis, then the line segments belonging to the object ( , , ), that is, the line segment extracted in the first stage of the algorithm, will be in the plane The point in ( , , ) intersect, these points belong to the plane and The straight line formed by the intersection of the planes (if the plane object and The axis is parallel, then there is no point cloud that belongs to the object. point).

[0091] In the embodiment of the present invention, Figure 6 Shows the line segments selected in the first stage of the algorithm and Histogram of the axis intersections, where the peak of the histogram corresponds to the plane , and , Represents line segments and The distribution of axis intersection points.

[0092] In the embodiment of the present invention, Figure 7 Shows that line segments belong to the same plane The straight line where the planes intersect.

[0093] In the embodiment of the present invention, under the condition of pitch angle scanning, a plane object extraction method is used to obtain a plane object. (0.393194x + 0.636345y + 0.663675z - 0.876937 = 0), (0.42971x +0.657433y + 0.618976z - 1.5562 = 0) and (-0.89452x + 0.155785y + 0.419006z -0.918841 = 0), in the case of azimuth scanning, the plane is also extracted (0.172041x - 0.752091y +0.636209z - 1.50634 = 0). The final result of the two-stage plane extraction is as follows: Figure 8 shown.

[0094] The two-stage plane extraction method based on two-dimensional Hough transform in the embodiment of the present invention is compared and evaluated with the effect of three-dimensional Hough transform. A set of coordinates in is The transformation is based on the three-dimensional parameter space , in which the plane represented by its normal is searched:

[0095] ;

[0096] in represents the vertical distance from the origin to the plane, Represents the vertical line The angle between the axes, Represents the vertical line The angle between the axes. Coordinates from Transformation from space to parameter space The rule is as follows: Consider the point Any plane , if we construct a perpendicular line from the origin to the plane , and record its length as , relative to and The angles of the axes are denoted as and , then the plane It is expressed by the above plane equation. is a right angle ( Is a plane The vertical line, line segment belongs to the plane), and based on Vector ,point Belongs to vector is a sphere with a diameter of . Therefore, from Transformation from space to parameter space The process boils down to generating a series of vectors from the origin to the vector On a sphere with a diameter of Fig. 9 Each such vector passes through the point The normal of the plane is defined in the form of the above plane equation. Therefore, when using the three-dimensional transformation method to extract a planar object from a point cloud, the calculation amount Depends on the number of similar operations that generate for each point in the point cloud a set of vectors ending up on the corresponding sphere:

[0097] ;

[0098] in represents the index of the point in the 3D image, represents the total number of points in the three-dimensional image, Indicated by The radius of the sphere generated by the points, represents the spherical discretization step size.

[0099] In the method proposed in the embodiment of the present invention, the computational effort in the first stage is determined by the number of synchronous operations to generate multiple vectors for each point in the point cloud, the endpoints of these vectors are located on the corresponding circles, and the computational effort is It can be expressed as:

[0100] ;

[0101] in represents the index of the point in the two-dimensional image, represents the total number of points in the two-dimensional image, Indicated by The radius of the circle generated by the points.

[0102] In the second stage of the embodiment of the present invention, for each line segment extracted in the first stage, the intersection points of the line segments with the two scanning axes and the projection plane perpendicular to the scanning axes are calculated; in addition, the local maximum values ​​in the histogram of the intersection points with the scanning axes are found, and the lines corresponding to these local maximum values ​​are extracted from the corresponding intersection points in the projection plane. Since the number of line segments extracted in the first stage is much less than the number of points in the original three-dimensional image, the amount of calculation in the second stage is much smaller than that in the first stage.

[0103] Using the same original data on the same computer, the processing time for plane semantic segmentation using the 3D Hough transform is 27 seconds, while the overall running time using the two-stage plane extraction method proposed in the embodiment of the present invention is 0.277 seconds, of which only 0.016 seconds is used for the second stage. Because the computational complexity of the 3D Hough transform is proportional to the total area of ​​the sphere generated by all the original point clouds, and the computational complexity of the two-stage plane extraction method proposed in the embodiment of the present invention is proportional to the total length of the circumference generated by all the original point clouds, this makes the execution speed of the embodiment of the present invention two orders of magnitude higher in the experiment discussed.

[0104] S5. Construct a plane equation based on a subset of points on each layer of the scanned plane, and then construct a semantic map of the key area.

[0105] Step S5 includes the following sub-steps S51 to S53:

[0106] S51. Randomly select three points from a point subset belonging to the same scanning plane to construct an initial plane equation.

[0107] S52. Using the solution of the initial plane equation as the initial value, construct a least squares equation, the expectation of which is that the distance from a point on the same plane to the plane is the minimum.

[0108] Ideally, the distance from a point on the same plane to the plane should be 0. However, in the actual measurement process, there are problems such as measurement noise and calculation deviation, which make it impossible to find a plane passing through all points. Therefore, we can only find an approximate solution so that the model can achieve the best fit at each observation point.

[0109] S53. Solve the least squares equation to obtain the precise plane equation of the key area, and add corresponding plane labels to the point cloud data of the key area through the precise plane equation of the key area to construct a semantic map of the key area.

[0110] S6. Determine the displacement of the UAV and incrementally update the semantic map.

[0111] In order to realize the autonomous function of the UAV, it is necessary to build a semantic map of inter-frame changes by incrementally updating the semantic map.

[0112] Step S6 includes the following sub-steps S61 to S64:

[0113] S61. Match each layer of scanned plane with the planes in the existing database to obtain a list of M pairs of planes corresponding to each other in the local coordinate system and the global coordinate system of the drone.

[0114] like Fig.10As shown, the paired planes are searched in the database based on possible motion information, which can be roughly determined based on the data of the onboard sensors. The relative rotation can be calculated from the data of the onboard gyroscope, while the relative displacement can be calculated from the odometer.

[0115] The equation of the plane in the local coordinate system of the drone (ЛСК) and its approximate position in the global coordinate system (ГСК) are known, which is estimated by the linear displacement vector and the generalized rotation matrix estimate (This matrix defines the orientation of the local coordinate system relative to the global coordinate system) Given, the equation of the plane in the global coordinate system can be determined.

[0116] For the Plane introduction vector , and Its collinear vectors , which in general may be the same as the vector The directions are inconsistent. It can be determined that ,in, Representation vector The estimated value of Representation vector Then, taking into account the assumed motion, the equation of the plane in the global coordinate system has the following form: ,in, .

[0117] Then, by adjusting the parameters of the generated plane, a matching plane can be found near it. The result is List of corresponding planes in the local and global coordinate systems of the drone:

[0118] ;

[0119] in Indicates The unit normal vector of the plane in the local coordinate system of the drone, Indicates The unit normal vector of a plane in the global coordinate system is Represents the direction vector of any point on the plane in the local coordinate system of the drone, Represents the direction vector of any point on the plane in the global coordinate system, Indicates The distance from the plane to the origin of the local coordinate system of the drone, Indicates The distance from the plane to the origin of the global coordinate system. Represents the unit normal vector The estimated value of Indicates distance The estimated value of .

[0120] At this stage, multiple alternative plane pairs may be found from the database, or none may be found at all. False matches can be screened out by a simple comparative analysis after all resulting planes have been calculated in a subsequent step.

[0121] S62. Construct an overdetermined system of nonlinear equations based on the list.

[0122] In order to determine all the motion parameters, at least three pairs of planes are required. For a plane, in the local coordinate system and the global coordinate system, such as Fig.11 As shown, the following equation holds: as well as

[0123] So, for For corresponding planes, we can obtain a nonlinear system of equations that is generally overdetermined:

[0124] ;

[0125] in Indicates The vector from the plane to the origin of the local coordinate system of the drone, Represents the displacement vector between the local coordinate system of the drone and the global coordinate system, represents the rotation matrix between the local coordinate system of the drone and the global coordinate system, and Represents the rotation matrix between the global coordinate system and the local coordinate system of the drone, .

[0126] S63. Solve the nonlinear equations to obtain the displacement of the UAV.

[0127] A generalized rotation matrix is ​​represented as a sequence of rotations around three axes ,in , and Represents the rotation sequence around the Z axis, Y axis and X axis respectively, , and Represent the rotation angles around the Z axis, Y axis, and X axis respectively. Then, the rotation angle that best satisfies the equation group is , and , and the offset , which can be found by optimization methods. The found solution must meet the given accuracy requirements for all plane pairs, thereby screening out wrong alternative plane pairs.

[0128] S64. Determine a new plane according to the displacement of the UAV, and incrementally update the semantic map according to the new plane.

[0129] After the displacement is determined, the database and external environment model can be updated and refined based on the parameters of the newly extracted plane. New planes are directly added to the database, and existing planes are refined by averaging the new data, which helps reduce the impact of noise in the sensor raw data.

[0130] In the examples of the present invention, Fig.12 and Fig.13 The photographic images of two spatial region segments and their corresponding colored point clouds are obtained from two different positions of the sensor. It is assumed that the local coordinate system and the global coordinate system of the initial position of the first spatial region coincide. In both ranging images, the corresponding planes are marked, and their equations are listed in Table 1.

[0131] Table 1

[0132]

[0133] An initial approximation for the displacement is By solving the corresponding nonlinear equations, the true displacement of the sensor is obtained as: The calculation result of the relative position of the sensor in the working area is as follows: Fig.14 shown.

[0134] The performance and efficiency of the two-stage key area semantic map construction method proposed in the embodiment of the present invention have been further verified on different sensor data sets. Fig.15 As shown, the method of the present invention can still robustly construct a linear semantic map in a complex indoor environment with interference from multiple nonlinear objects.

[0135] In the embodiment of the present invention, Fig.16 As shown, in a large outdoor scene environment, the original 3D data information is very sparse, but the method of the present invention can still effectively predict key scenes and construct a semantic map.

[0136] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A two-stage key area semantic map construction method based on Hough transform, characterized in that: The following steps are involved: S1, collect stereo ranging images of the external environment of the UAV; S2, layering the stereo ranging image according to different scanning angles to obtain a multi-layer scanning plane; S3, extracting line segments using two-dimensional Hough transform in each scanning plane to obtain a line segment subset; S4, re-clustering the incorrectly clustered line segments in each line segment subset to obtain a point subset of each scanning plane; S5, constructing a plane equation based on a subset of points on each layer of the scanned plane, and then constructing a semantic map of the key area; S6, determine the displacement of the UAV and incrementally update the semantic map; The step S2 comprises the following sub-steps: S21. Establish the spatial coordinate system of the space where the drone is located. Axis rotation to obtain different scanning angles; S22, layering the stereo ranging image according to different scanning angles; S23, projecting the layered point cloud into a binary plane image to obtain a multi-layer scanning plane; The step S3 comprises the following sub-steps: S31, using two-dimensional Hough transform to extract line segments in each scanning plane, and counting the extracted line segments and The intersection points on the axis; S32, construct an accumulation function, whose value is The number of line segments that have the same intersection point on the axis; S33, according to the value of the accumulation function, The local maximum of the intersection histogram is extracted from the local interval on the axis. The line segments with the same intersection point on the axis are clustered to obtain a line segment subset; The two-dimensional Hough transform formula in step S31 is: ； in represents the vertical distance from the origin to the line segment, Represents the vertical line The angle between the axes, represents the coordinates in the scan plane; The step S4 comprises the following sub-steps: S41. Count the line segments in each line segment subset and Intersection of planes; S42, will The line segments whose intersection points of the planes are on the same straight line are re-clustered to obtain line segments belonging to the same scanning plane; S43, obtaining a point subset of the layer of scanning plane according to line segments belonging to the same scanning plane; The step S6 comprises the following sub-steps: S61, matching each layer of scanned plane with the plane in the existing database, and obtaining List of corresponding planes in the local and global coordinate systems of the drone: ； in Indicates The unit normal vector of the plane in the local coordinate system of the drone, Indicates The unit normal vector of a plane in the global coordinate system is Represents the direction vector of any point on the plane in the local coordinate system of the drone, Represents the direction vector of any point on the plane in the global coordinate system, Indicates The distance from the plane to the origin of the local coordinate system of the drone, Indicates The distance from the plane to the origin of the global coordinate system; S62, constructing an overdetermined nonlinear system of equations according to the list: ； in Indicates The vector from the plane to the origin of the local coordinate system of the drone, Represents the displacement vector between the local coordinate system of the drone and the global coordinate system, Represents the rotation matrix between the local coordinate system of the drone and the global coordinate system. ; S63, solving the nonlinear equation group to obtain the displacement of the UAV; S64. Determine a new plane according to the displacement of the UAV, and incrementally update the semantic map according to the new plane.

2. The method for constructing a two-stage key area semantic map based on Hough transform according to claim 1 is characterized in that: The step S1 comprises the following sub-steps: S11, collect the original point cloud data of the external environment of the drone by carrying a 3D laser sensor on the drone; S12, filtering the original point cloud data to obtain a stereo ranging image of the external environment of the UAV.

3. The method for constructing a two-stage key area semantic map based on Hough transform according to claim 1, characterized in that: The step S5 comprises the following sub-steps: S51, randomly selecting three points from a point subset belonging to the same scanning plane, and constructing an initial plane equation; S52, using the solution of the initial plane equation as an initial value to construct a least squares equation, wherein the expectation of the least squares equation is that the distance from a point on the same plane to the plane is the minimum; S53. Solve the least squares equation to obtain the precise plane equation of the key area, and add corresponding plane labels to the point cloud data of the key area through the precise plane equation of the key area to construct a semantic map of the key area.

Images