Vehicle target detection method and device, electronic equipment and computer readable medium

Abstract

The application provides a target detection method and device for a vehicle, electronic equipment and a computer readable medium, the target detection method comprising: obtaining main point cloud data from a main radar, obtaining left point cloud data from a left blind spot filling radar, obtaining right point cloud data from a right blind spot filling radar, and fusing any point in the left and right point cloud data into the main point cloud data; performing down-sampling processing on the main point cloud data, and then deleting all point clouds in all object boxes in each point cloud frame in the main point cloud data; deleting all point clouds in all object boxes in each point cloud frame in the main point cloud data; performing plane gridding processing on the main point cloud data, extracting non-ground point clouds and non-ground plane grids based on grid features; performing clustering processing on the non-ground point clouds, and removing grids in the clustering object boxes from the non-ground plane grids to obtain irregular object grids that cannot be clustered. Thus, the point cloud data can be subjected to target detection processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and in particular to a target detection method and device for a vehicle, an electronic device and a computer readable medium. BACKGROUND

[0002] With the advancement and development of technology, intelligent driving has been widely applied. One basic process of intelligent driving is as follows: a radar (e.g., a laser radar, etc.) obtains distance, speed and angle of objects in the environment by emitting laser pulses and measuring their return time. When the laser pulses encounter objects, part of the light is reflected back to the radar receiver to form so-called “point cloud” data. Each point represents a reflection point on the surface of an object and contains information such as its three-dimensional coordinates and reflection intensity; then, the point cloud data is processed and target detection (e.g., obstacle detection) is performed; then, map construction and navigation are performed, and path planning and decision-making are performed.

[0003] It can be understood that target detection on point cloud data is a very key step. SUMMARY

[0004] The present application relates to the technical field of automatic driving, and in particular to a target detection method and device for a vehicle, an electronic device and a computer readable medium.

[0005] To achieve one of the above-mentioned purposes, an embodiment of the present application provides a target detection method for a vehicle, wherein the vehicle is provided with a main radar, a left blind spot filling radar and a right blind spot filling radar; the method comprises the following steps: obtaining main point cloud data from the main radar, obtaining left point cloud data from the left blind spot filling radar, and obtaining right point cloud data from the right blind spot filling radar; the interval of the acquisition time between any two point cloud data in the main point cloud data, the left point cloud data and the right point cloud data is less than a preset threshold; a first coordinate transformation formula is generated for calibrating the left point cloud data to the main point cloud data, and a second coordinate transformation formula is generated for calibrating the right point cloud data to the main point cloud data; target detection processing is performed on the main point cloud data, the left point cloud data and the right point cloud data, and a plurality of object frames and a detection score of an object in each object frame are detected from the main point cloud data, the left point cloud data and the right point cloud data, respectively; for any point cloud frame Pic M in the main point cloud data, the following processing is performed: selecting a point cloud frame Pic M closest in time from the left point cloud data, selecting a point cloud frame Pic L closest in time from the right point cloud data, and selecting a point cloud frame Pic M closest in time from the left point cloud data and the right point cloud data; if any object frame Frame R in the point cloud frame Pic L is determined to be a false positive, the object frame Frame L is removed from the point cloud frame Pic MFrame of objects ML When matching, use the object frame. L Update the object frame ML The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M Frame of objects MR When matching, use the object frame. R Update the object frame MR The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M All object bounding boxes do not match, but the object bounding box Frame R With point cloud frame Pic L Frame of objects LR When matching, the object bounding box Frame is transformed based on the first coordinate transformation formula. LR Transform into a bounding box Frame' LR , set the object frame' LR Add to point cloud frame Pic M In the middle, use the object frame. R and object frame LR Update the object bounding box Frame' LR The detection score is calculated; the main point cloud data is downsampled, and then all point clouds in the object boxes of all point cloud frames in the main point cloud data are deleted; the main point cloud data is rasterized, and non-ground point clouds and non-ground planar grids are extracted based on grid features; the non-ground point clouds are clustered, and the grids in the clustered object boxes are removed from the non-ground planar grids to obtain irregular object grids that cannot be clustered.

[0006] As a further improvement to one embodiment of the present invention, the "first coordinate transformation formula for generating the left point cloud data calibration to the main point cloud data" specifically includes: the first coordinate transformation formula for the left point cloud data calibration to the main point cloud data is: x′ L =x L ·cosθ L -y L ·sinθ L +T xL y′ L =x L ·sinθ L +y L ·cosθ L +T yL , z′L = z L + T zL ; wherein, the coordinates (x L , y L , z L ) are three-dimensional coordinates of any point pot1 in the left point cloud data, the coordinates (x' L , y' L , z' L ) are the projection coordinates of the point pot1 in the main point cloud data, the translation transformation of the main radar relative to the left blind-filling radar is (T xL , T yL , T zL ), and the angle of the main radar relative to the left blind-filling radar counterclockwise around the z-axis is θ L ; the “generate a second coordinate transformation formula of the right point cloud data to the main point cloud data” specifically comprises: the second coordinate transformation formula of the right point cloud data to the main point cloud data is x' R = x R cos θ R - y R sin θ R + T xR , y' R = x R sin θ R + y R cos θ R + T yR , z' R = z R + T zR ; wherein, the coordinates (x R , y R , z R ) are three-dimensional coordinates of any point pot2 in the right point cloud data, the coordinates (x' R , y' R , z' R ) are the projection coordinates of the point pot2 in the main point cloud data, the translation transformation of the main radar relative to the right blind-filling radar is (T xR , T yR , T zR ), and the angle of the main radar relative to the right blind-filling radar counterclockwise around the z-axis is θ R .

[0007] As a further improvement of an embodiment of the present application, the “perform target detection processing on the main point cloud data, the left point cloud data and the right point cloud data” specifically comprises: performing target detection processing on the main point cloud data, the left point cloud data and the right point cloud data based on a deep learning algorithm.

[0008] As a further improvement to one embodiment of the present invention, the step of "performing target detection processing on the main point cloud data, left point cloud data, and right point cloud data, and detecting several object boxes from the main point cloud data, left point cloud data, and right point cloud data respectively, and the detection score of the object in each object box" specifically includes: performing target detection processing on the main point cloud data, left point cloud data, and right point cloud data, and detecting several object boxes from the main point cloud data, left point cloud data, and right point cloud data respectively, and the detection score of the object in each object box, and setting an initial value of zero for each object box in the main point cloud data; the step of "using the object box Frame" specifically includes: performing target detection processing on the main point cloud data, left point cloud data, and right point cloud data, and detecting several object boxes from the main point cloud data, left point cloud data, and right point cloud data respectively, and the detection score of the object in each object box; and setting an initial value of zero for a score modification value for each object box in the main point cloud data. L Update the object frame ML The "detection score" specifically includes: the object bounding box (Frame). ML The score modification value increases the object frame. L The detection score; the "using object frame" R Update the object frame MR The "detection score" specifically includes: the object bounding box (Frame). MR The score modification value increases the object frame. R The detection score; the "using object frame" R and object frame LR Update the object bounding box Frame' LR The detection score specifically includes: the object bounding box Frame' LR The score modification value increases the object frame. R The detection score is then used to add the object bounding box (Frame). LR The detection score, the object bounding box Frame′ LR The detection score is zero; the processing also includes: if the point cloud frame Pic M If the score modification value of any object box F is not zero, then Score = α * detection score of object box F + β * score modification value of object box F, where α > 0 and β > 0.

[0009] As a further improvement of one embodiment of the present invention, the step of "deleting all point clouds in the object frames of any point cloud frame in the main point cloud data" specifically includes: performing the following operations on any point pot3 and any object frame Frame in all point cloud frames of the main point cloud data: obtaining the four coordinates (x1, y1), (x2, y2), (x3, y3) and (x4, y4) of the object frame Frame, and the coordinates (x, y) of point pot3, wherein the directed distance from the coordinates (x, y) to the four sides of the object frame Frame is d. i =(xx) i )·(y i％4+1 -yi ) · (x i ) · (x i％4+1 -x i ), i = 1, 2, 3, 4; when the first condition or the second condition is met, the point pot3 is deleted; the first condition is d1≥0 and d2≥0 and d3≥0 and d4≥0, and the second condition is d1≤0 and d2≤0 and d3≤0 and d4≤0.

[0010] As a further improvement of an embodiment of the application, the "plane rasterization processing on the main point cloud data" specifically comprises: obtaining the range and resolution of the xy plane raster in the main point cloud data, creating a plurality of rasters on a two-dimensional plane based on the range and resolution; obtaining a plurality of point clouds corresponding to each raster, each raster intersects with a plurality of corresponding point clouds; the "extracting non-ground point cloud and non-ground plane raster based on raster features" specifically comprises: calculating the average height, maximum height, minimum height and normal vector of the point cloud in each raster; obtaining a gradient threshold, marking the raster as a ground raster; for the blind area raster around the ego vehicle, directly marking it as a ground raster; using a breadth-first search algorithm, starting from the raster where the ego vehicle is located to search the surrounding rasters, for adjacent rasters, marking according to the height change and normal consistency; if the height change > gradient threshold, mark it as an obstacle raster; generating map information according to the ground raster and the obstacle raster, the map information at least includes: raster coordinates, height information, ground point cloud and non-ground point cloud; then, using the index of the obstacle raster to extract the non-ground raster.

[0011] As a further improvement of the embodiment of the present application, the "clustering processing on the non-ground point cloud" specifically comprises: using the Euclidean clustering DBSCAN algorithm, combining Kd-Tree accelerated search for the nearest neighbor points in the non-ground point cloud, after completing the DBSCAN clustering, extracting the points in the same cluster to generate a plurality of clustered point clouds; for each clustered point cloud, the following processing is performed: obtaining the minimum value ZMin and the maximum value ZMax of the Z-axis coordinates of the midpoint of the clustered point cloud, obtaining the height h of the cluster h = ZMax-ZMin, projecting the clustered point cloud onto the XOY plane to obtain a projected point set, generating a minimum convex polygon enclosing the projected point set, for each edge of the minimum convex polygon, calculating the sum of distances from all vertices of the minimum convex polygon to the edge, obtaining the edge Edge1 with the minimum sum of distances, projecting all vertices of the minimum convex polygon onto the edge Edge1 to obtain two points p1 and p2 with the farthest projection distance, obtaining the slope of the edge Edge2 composed of the points p1 and p2, and finding the point Point in the point set of the minimum convex polygon that is farthest from the edge Edge2, generating the edge Edge3 passing through the point Point and parallel to the edge Edge2, projecting the points p1 and p2 onto the edge Edge3 to obtain two points p3 and p4, generating a minimum circumscribed rectangle with the points p1, p2, p3 and p4 as vertices, determining the length l, the width w and the diagonal of the minimum circumscribed rectangle by calculating the distances between the four vertices, wherein the slope of the long side relative to the X-axis in the forward direction of the automobile is the orientation angle θ of the minimum circumscribed rectangle, and the center point coordinates (x M ,y M ) of the minimum circumscribed rectangle are obtained from the coordinates of the four vertices of the minimum circumscribed rectangle. A three-dimensional clustered object frame is generated, the center point coordinates of the three-dimensional clustered object frame are (x M ,y M ,z M ,), the orientation angle θ, the length l, the width w and the height h are the orientation angle, the length, the width and the height on the z-axis of the three-dimensional clustered object frame respectively; the "removing the grids in the clustered object frame from the non-ground plane grid to obtain irregular object grids that cannot be clustered" specifically comprises: the coordinates of the four vertices of the rotated rectangle of the three-dimensional clustered object frame are (x1, y1), (x2, y2), (x3, y3) and (x4, y4), for each point in the non-ground grid, the following processing is performed: obtaining the directed distance of the point to the four edges of the rotated rectangle, and deleting the point when it is determined that the point is in the rotated matrix based on the cross product calculation method.

[0012] The embodiment of the application also provides a target detection device of an automobile, wherein the automobile is provided with a main radar, a left blind area filling radar and a right blind area filling radar; the device comprises the following modules: a data acquisition module, which is used for acquiring main point cloud data from the main radar, acquiring left point cloud data from the left blind area filling radar and acquiring right point cloud data from the right blind area filling radar, the interval of the acquisition time between any two point cloud data in the main point cloud data, the left point cloud data and the right point cloud data is less than a preset threshold, a first coordinate transformation formula of the left point cloud data calibrated to the main point cloud data is generated, and a second coordinate transformation formula of the right point cloud data calibrated to the main point cloud data is generated; target detection processing is performed on the main point cloud data, the left point cloud data and the right point cloud data, and a plurality of object frames and the detection scores of objects in each object frame are detected from the main point cloud data, the left point cloud data and the right point cloud data respectively; an alignment module is used for performing the following processing on any point cloud frame Pic M in the main point cloud data: selecting a point cloud frame Pic M closest in time from the left point cloud data, L selecting a point cloud frame Pic M closest in time from the right point cloud data, R if any object frame Frame L in the point cloud frame Pic L matches an object frame Frame M in the point cloud frame Pic ML , updating the detection score of the object frame Frame L using the object frame Frame ML , if any object frame Frame R in the point cloud frame Pic R matches an object frame Frame M in the point cloud frame Pic MR , updating the detection score of the object frame Frame R using the object frame Frame MR , and if any object frame Frame R in the point cloud frame Pic R does not match all object frames in the point cloud frame Pic M , but the object frame Frame R matches an object frame Frame L in the point cloud frame Pic LR , transforming the object frame Frame LR into an object frame Frame′ LR based on the first coordinate transformation formula, adding the object frame Frame′ LR to the point cloud frame Pic M , and updating the detection score of the object frame Frame RAnd the object frame Frame LR Update the object frame Frame' LR The detection score of the object frame; the processing module is used for downsampling processing on the main point cloud data, and then deleting all points in the object frame in all point cloud frames in the main point cloud data; the main point cloud data is subjected to plane rasterization processing, and non-ground point clouds and non-ground plane grids are extracted based on grid features; the non-ground point clouds are subjected to clustering processing, and the grids in the clustered object frame are removed from the non-ground plane grids, so as to obtain irregular object grids that cannot be clustered.

[0013] The embodiment of the application further provides an electronic device, comprising: one or more processors; a memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, the one or more processors implement the target detection method according to the above.

[0014] The embodiment of the application further provides a computer readable medium having a computer program stored thereon, when the program is executed by a processor, the target detection method according to the above is implemented.

[0015] Compared with the prior art, the technical effects of the application are as follows: 1. Multi-radar information integration: the main radar, left and right blind filling radars, GPS and IMU devices are mounted on the autonomous vehicle, the point clouds of the left and right blind filling radars are calibrated to the main radar coordinate system, the objects of the three radars at similar time points are detected through deep learning, then the recognition scores of the left and right blind filling radars are integrated into the main radar result, and the accuracy of the deep learning detection target is improved; 2. Solve the problem of the direction of the clustered object frame: by clustering the point clouds and searching the longest edge in the clustered point clouds, the direction of the object can be more accurately determined, and the problem that the direction of the axis-aligned bounding box (AABB) or the principal component analysis (PCA) frame is greatly different from the actual object direction in the traditional method is solved; 3. Diversity of perception information: deep learning, point cloud clustering and rasterization algorithms are comprehensively used, the deep learning detection object frame is responsible for perceiving common objects, the clustered object frame is used for perceiving the first three types of objects and other types of objects that are not recognized, and the plane grid is used for perceiving irregular objects that cannot be clustered, and the diverse perception information improves the adaptability of the system in a complex environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the target detection method of the automobile in the embodiment of the application;

[0017] Figure 2 is an experimental result graph of the target detection method of the automobile in the embodiment of the application;

[0018] Figure 3 andFigure 4 is the schematic diagram of the target detection method of the automobile in the embodiment of the present application. DETAILED DESCRIPTION

[0019] The following description and drawings are illustrative of the specific embodiments herein and are not intended to limit the scope of the application. Various aspects of the embodiments can be included or substituted into other embodiments. The scope of the embodiments herein includes the whole scope of the claims and all available equivalents of the claims. In this document, the terms "first", "second", and so on are used only to distinguish one element from another, and do not require or imply any actual relationship or order between the elements. In fact, the first element can also be called the second element, and vice versa. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a structure, device or apparatus including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such structure, device or apparatus. Without more limitations, the element defined by the statement "including a" does not exclude the existence of other identical elements in the structure, device or apparatus including the element. Various embodiments herein are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between various embodiments can be referred to each other.

[0020] The terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like in this document indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of description herein and simplification of description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description herein, unless otherwise specified and limited, the terms "mount", "connect", "connection" should be understood broadly, for example, it can be a mechanical connection or an electrical connection, it can be a communication between two elements inside, it can be directly connected, or indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0021] The embodiment one of the present application provides a target detection method of a vehicle, wherein the vehicle is provided with a main radar, a left blind area filling radar and a right blind area filling radar. Optionally, the vehicle is provided with a main radar, a left blind area filling radar and a right blind area filling radar; the main radar is arranged at the front part of the roof of the vehicle, the left blind area filling radar is located at the left side of the main radar, the right blind area filling radar is located at the right side of the main radar, the main radar, the left blind area filling radar and the right blind area filling radar are located on the same horizontal plane, and the main radar, the left blind area filling radar and the right blind area filling radar are located on the same straight line, and the straight line is perpendicular to the straight running direction of the vehicle; as shown in the figure, the method comprises the following steps: Figure 1 and Figure 2 The method comprises the following steps:

[0022] Step 101: obtaining main point cloud data from the main radar, obtaining left point cloud data from the left blind area filling radar, and obtaining right point cloud data from the right blind area filling radar; the interval of the acquisition time between any two point cloud data in the main point cloud data, the left point cloud data and the right point cloud data is less than a preset threshold; a first coordinate transformation formula of the left point cloud data calibrated to the main point cloud data is generated, and a second coordinate transformation formula of the right point cloud data calibrated to the main point cloud data is generated; target detection processing is performed on the main point cloud data, the left point cloud data and the right point cloud data, and a plurality of object frames and the detection scores of the objects in each object frame are detected from the main point cloud data, the left point cloud data and the right point cloud data, respectively; here, after the main point cloud data, the left point cloud data and the right point cloud data are obtained, noise removal, data filtering, data scaling and the like can be performed on the three point cloud data to improve the quality and accuracy of the data.

[0023] Step 102: for any point cloud frame Pic M in the main point cloud data, the following processing is performed: selecting the point cloud frame Pic M closest in time to the point cloud frame Pic L from the left point cloud data, and selecting the point cloud frame Pic M closest in time to the point cloud frame Pic R from the right point cloud data; if any object frame Frame L in the point cloud frame Pic L matches the object frame Frame M in the point cloud frame Pic ML , the detection score of the object frame Frame L is updated using the object frame Frame ML ; if any object frame Frame R in the point cloud frame Pic R matches the object frame Frame M in the point cloud frame Pic MR , the detection score of the object frame Frame Rupdating the detection score of the object frame Frame MR ; if any object frame Frame R in the point cloud frame Pic R does not match all object frames in the point cloud frame Pic M , but the object frame Frame R matches the object frame Frame L in the point cloud frame Pic LR , the object frame Frame LR is transformed into the object frame Frame′ LR based on the first coordinate transformation formula, the object frame Frame′ LR is added to the point cloud frame Pic M , the detection score of the object frame Frame′ R is updated using the object frame Frame LR and the object frame Frame LR ;

[0024] Here, target detection processing is performed on the main point cloud data, the left point cloud data and the right point cloud data, and the specific attributes of the object frame include: the time stamp time_stamp of the point cloud frame to which the object frame belongs, the object type type, the object frame center point coordinates (x, y, z), the object frame orientation angle θ, the length l, width w and height h of the object frame, and the detection score typescore of the object.

[0025] The operation "selecting the point cloud frame Pic M closest in time to the point cloud frame Pic L from the left point cloud data" can be understood as: selecting the point cloud frame Pic M closest in time stamp time_stamp to the point cloud frame Pic L from the left point cloud data, and in the left point cloud data, the value of |time_stamp of point cloud frame Pic M - time_stamp of point cloud frame Pic L | is the smallest. Similarly, the operation "selecting the point cloud frame Pic M closest in time to the point cloud frame Pic R from the right point cloud data" can be understood as: selecting the point cloud frame Pic M closest in time stamp time_stamp to the point cloud frame Pic R from the right point cloud data, and in the right point cloud data, the value of |time_stamp of point cloud frame Pic M - time_stamp of point cloud frame Pic R | is the smallest.

[0026] Here, the object box can be understood as a rotated rectangle with height information ignored and an orientation angle. In the experiment, the inventor used the RotatedRect class of OpenCV to represent a rotated rectangle, which consists of a center point (x, y), length l, width w, and rotation angle θ. Then, two RotatedRect objects were defined, rect1 for the left and right blind radar detected object box and rect2 for the main radar detected object box. The rotatedRectangleIntersection function was used with rect1 and rect2 as parameters. This function accepts two rotated rectangles as input and returns a vector containing the intersection points, which can calculate the intersection area of the two rotated rectangles. Then, the intersection area result was stored in the vector intersectingRegion. If intersectingRegion is empty, it means the rectangles do not intersect.

[0027] When determining whether two object boxes match, the matching degree of the two object boxes is calculated through the intersection area of the rotated rectangle. The contourArea function of OpenCV was used, which accepts a vector containing intersection points as input and returns the area of the intersection region. If the area of the intersection region is > 0, it means that the match is found, i.e., the two object boxes have overlapping areas. Then, the area of the intersection region is found to be the largest, which is the best matching object box.

[0028] Step 103: The main point cloud data is down-sampled, and then all point cloud frames in the main point cloud data are deleted from all object boxes in the point cloud frames. The main point cloud data is processed by plane gridding, and non-ground point clouds and non-ground plane grids are extracted based on grid features. The non-ground point clouds are clustered, and the grids in the clustered object boxes are removed from the non-ground plane grids to obtain irregular object grids that cannot be clustered.

[0029] Here, first, the main point cloud data is down-sampled to reduce the density of the main point cloud data. Then, for any point and any object box in all point cloud frames in the main point cloud data, it is determined whether the point is located in the object box. If it is located, the point is deleted.

[0030] Three-dimensional point cloud coordinates are acquired from the main radar sensor, the range and resolution of the xy plane grid are defined, the range is determined by the autonomous driving map region, and the resolution determines the size of the grid. Higher resolution means more detailed grid. Then, according to the defined range and resolution, a grid data structure is created on the two-dimensional plane, each grid will store the point cloud information intersected with it, and for each point, its three-dimensional coordinates are mapped to the grid. Then, search for the features of the grid to extract non-ground point clouds and non-ground grids.

[0031] In the autonomous driving system, the key of perception cannot be ignored, and the laser radar as an important part of the perception system, its accuracy directly affects the planning and control accuracy of autonomous driving. The traditional method often uses axis-aligned bounding box (AABB) or principal component analysis (PCA) box to box the object point cloud, but the direction of the bounding box obtained by this method is greatly different from the actual object orientation. This not only affects the accurate positioning of the object, but also brings great difficulty to the subsequent object tracking. In addition, the traditional method performs poorly in object classification, making it difficult to accurately distinguish different types of objects, limiting the applicability of the system in complex environments.

[0032] In the embodiment, the "first coordinate transformation formula of generating left point cloud data calibration to main point cloud data" specifically includes: the first coordinate transformation formula of left point cloud data calibration to main point cloud data is: x' L =x L ·cosθ L -y L ·sinθ L +T xL , y' L =x L ·sinθ L +y L ·cosθ L +T yL , z' L =z L +T zL ; wherein the coordinates (x L ,y L ,z L ) are the three-dimensional coordinates of any point pot1 in the left point cloud data, the coordinates (x' L ,y' L ,z' L ) are the projection coordinates of the point pot1 in the main point cloud data, the translation transformation of the main radar relative to the left blind filling radar is (T xL ,T yL ,T zL ), and the angle of the main radar relative to the left blind filling radar around the z-axis counterclockwise is θ L ;

[0033] The "generate the right point cloud data calibration to the main point cloud data second coordinate transformation formula" specifically includes: the right point cloud data

[0034] The second coordinate transformation formula calibrated to the main point cloud data is: x' R =x R ·cosθ R -y R ·sinθ R +T xR , y' R =x R ·sinθ R +y R ·cosθ R +T yR , z' R =z R +T zR ; wherein the coordinates (x R , y R , z R ) are the three-dimensional coordinates of any point pot2 in the right point cloud data, the coordinates (x' R , y' R , z' R ) are the projection coordinates of the point pot2 in the main point cloud data, the translation transformation of the main radar relative to the right blind filling radar is (T xR , T yR , T zR ), and the angle of the main radar relative to the right blind filling radar counterclockwise around the z-axis is θ R .

[0035] In this embodiment, the "perform target detection processing on the main point cloud data, the left point cloud data and the right point cloud data" specifically includes: performing target detection processing on the main point cloud data, the left point cloud data and the right point cloud data based on a deep learning algorithm.

[0036] In this embodiment, the "perform target detection processing on the main point cloud data, the left point cloud data and the right point cloud data, and detect a plurality of object frames and the detection scores of objects in each object frame from the main point cloud data, the left point cloud data and the right point cloud data respectively" specifically includes: performing target detection processing on the main point cloud data, the left point cloud data and the right point cloud data, and detecting a plurality of object frames and the detection scores of objects in each object frame from the main point cloud data, the left point cloud data and the right point cloud data respectively, and setting a score modification value with an initial value of zero for each object frame in the main point cloud data.

[0037] The "update the detection scores of the object frames Frame L using the object frames Frame ML " specifically includes: updating the detection scores of the object frames FrameML The score modification value increases the object frame. L The test score;

[0038] The "use of object frame" R Update the object frame MR The "detection score" specifically includes: the object bounding box (Frame). MR The score modification value increases the object frame. R The test score;

[0039] The "use of object frame" R and object frame LR Update the object bounding box Frame' LR The detection score specifically includes: the object bounding box Frame' LR The score modification value increases the object frame. R The detection score is then used to add the object bounding box (Frame). LR The detection score, the object bounding box Frame′ LR The test score was zero;

[0040] The processing also includes: if the point cloud frame Pic M If the score modification value of any object box F is not zero, then Score = α * detection score of object box F + β * score modification value of object box F, where α > 0 and β > 0.

[0041] In this embodiment, the following operations are performed on any point pot3 and any object frame in all point cloud frames of the main point cloud data: obtain the four coordinates (x1, y1), (x2, y2), (x3, y3), and (x4, y4) of the object frame, and the coordinates (x, y) of point pot3, where the directed distance from coordinates (x, y) to the four sides of the object frame is d. i =(xx) i )·(y i％4+1 -y i )-(yy i )·(x i％4+1 -x i ), i = 1, 2, 3, 4; when the first condition or the second condition is met, delete point pot3; the first condition is d1≥0 and d2≥0 and d3≥0 and d4≥0, and the second condition is d1≤0 and d2≤0 and d3≤0 and d4≤0.

[0042] Here, the operation "determine whether the point is located within the object frame" specifically includes: obtaining the four coordinates of the object frame (x1, y1), (x2, y2), (x3, y3), and (x4, y4), and the coordinates of the point (x, y). The four sides of the object frame are Line1, Line2, Line3, and Line4. The coordinates of the two endpoints of side Line1 are (x1, y1) and (x2, y2), the coordinates of the two endpoints of side Line2 are (x2, y2) and (x3, y3), the coordinates of the two endpoints of side Line3 are (x3, y3) and (x4, y4), and the coordinates of the two endpoints of side Line4 are (x4, y4) and (x1, y1). The coordinates (x, y) are then used to determine the distance from the point to the edge of the object frame. i The directed distance d i =(xx) i )·(y i％4+1 -y i )-(yy i )·(x i％4+1 -x i ), i = 1, 2, 3, 4; the point is located in the object frame when either of the following two conditions is met, otherwise the point is not located in the object frame; the two conditions are: first condition, d1≥0 and d2≥0 and d3≥0 and d4≥0; second condition, d1≤0 and d2≤0 and d3≤0 and d4≤0.

[0043] In this embodiment, the "planar rasterization processing of the main point cloud data" specifically includes: obtaining the range and resolution of the xy plane raster in the main point cloud data; creating multiple rasters on a two-dimensional plane based on the range and resolution; obtaining several point clouds corresponding to each raster, with each raster intersecting with the corresponding several point clouds;

[0044] The "extraction of non-ground point clouds and non-ground planar grids based on grid features" specifically includes: calculating the average height, maximum height, minimum height, and normal vector for the point cloud within each grid; obtaining a gradient threshold and marking the grid as a ground grid; directly marking the grids in the blind spot around the vehicle as ground grids; using a breadth-first search algorithm, searching for surrounding grids starting from the grid where the vehicle is located, and marking adjacent grids based on height change and normal consistency; if the height change > gradient threshold, marking it as an obstacle grid; generating map information based on the ground grids and obstacle grids, the map information including at least: grid coordinates, height information, ground point clouds, and non-ground point clouds; and then extracting non-ground grids using the index of the obstacle grids.

[0045] Here, the calculation of the normal vector can employ a normal estimator such as a nearest neighbor normal estimator, etc.; the gradient threshold can be set according to the ground slope.

[0046] In this embodiment, as shown in Figure 4 , the "clustering processing on the non-ground point cloud" specifically includes: using the Euclidean clustering DBSCAN algorithm, combining Kd-Tree accelerated search for the nearest neighbor points in the non-ground point cloud, after completing the DBSCAN clustering, extracting the points in the same cluster to generate a plurality of clustered point clouds; for each clustered point cloud, the following processing is performed: obtaining the minimum value ZMin and the maximum value ZMax of the Z-axis coordinates of the midpoint of the clustered point cloud, obtaining the height h of the cluster h = ZMax-ZMin, projecting the clustered point cloud onto the XOY plane to obtain a projection point set, generating a minimum convex polygon enclosing the projection point set, for each edge of the minimum convex polygon, calculating the sum of the distances from all vertices of the minimum convex polygon to the edge, obtaining the edge Edge1 with the minimum sum of distances, projecting all vertices of the minimum convex polygon onto the edge Edge1 to obtain two points p1 and p2 with the farthest projection distance, obtaining the slope of the edge Edge2 composed of the points p1 and p2, and finding the point Point in the point set of the minimum convex polygon that is farthest from the edge Edge2, generating the edge Edge3 passing through the point Point and parallel to the edge Edge2, projecting the points p1 and p2 onto the edge Edge3 to obtain two points p3 and p4, generating a minimum circumscribed rectangle with the points p1, p2, p3 and p4 as vertices, determining the length l, width w and hypotenuse of the minimum circumscribed rectangle by calculating the distances between the four vertices, wherein the slope of the long side relative to the X-axis in the forward direction of the car is the orientation angle θ of the minimum circumscribed rectangle, and the center point coordinates (x M ,y M ) of the minimum circumscribed rectangle are obtained from the coordinates of the four vertices of the minimum circumscribed rectangle. A three-dimensional clustered object frame is generated, the center point coordinates of the three-dimensional clustered object frame are (x M ,y M ,z M ,), the orientation angle θ, the length l, the width w and the height h are the orientation angle, the length, the width and the height on the z-axis of the three-dimensional clustered object frame respectively.

[0047] Here, when generating the projection point set, the point with the minimum y i coordinate value in the point set can be selected as the starting point P0, and the point set is sorted according to the angle from P0 from small to large, specifically, the angle between the connecting line and the Y-axis is calculated, if the angle between the connecting line and the Y-axis is in the clockwise direction, a negative value is used to represent, if the angle between the connecting line and the Y-axis is in the counterclockwise direction, a positive value is used to represent, and the projection point set is arranged from small to large according to the angle. As shown in Figure 3 .

[0048] Here, when generating the smallest convex polygon enclosing the set of projection points, the second point in the sorted set of projection points can be connected with P0 to obtain a first edge, then the third point is connected with the second point to obtain a second edge, and it is determined whether the angle formed by the two edges is greater than 180°, if yes, the second point is removed, and the first point and the third point are connected and the third point and the fourth point are connected, and then it is determined, if no, the second point is retained, the second point is a convex point, and the fourth point is connected with the third point, and then it is determined with the previous edge. Each point is considered in turn, and finally the point set of the convex polygon is obtained. In addition, it can also be expressed that the angle between the adjacent two edges of the sorted point set is determined, if the angle is greater than 180°, then the point common to the two edges is a convex point, and the two edges can enclose the set of projection points. For example, when it is determined that the angle between the edge P0P1 and the edge P1P2 is less than 180°, the point P1 is retained, when it is determined that the angle between the edge P0P1 and the edge P2P3 is greater than 180°, the point P2 is an enclosed inner point, the point P2 is discarded, and it is determined whether the angle between the edge P1P3 and the edge P3P4 is greater than 180°, and so on.

[0049] The "removing the grid in the clustered object frame from the non-ground plane grid to obtain an irregular object grid that cannot be clustered" specifically includes that the coordinates of the four vertices of the rotated rectangle of the three-dimensional clustered object frame are (x1, y1), (x2, y2), (x3, y3) and (x4, y4), and for each point in the non-ground grid, the following processing is performed: obtaining the directed distance of the point to the four edges of the rotated rectangle, and deleting the point when it is determined that the point is in the rotated matrix based on the cross product calculation method.

[0050] Embodiment two of the present application provides a target detection device of a vehicle, wherein the vehicle is provided with a main radar, a left blind filling radar and a right blind filling radar; and the device comprises the following modules:

[0051] A data acquisition module is configured to acquire main point cloud data from the main radar, left point cloud data from the left blind filling radar, and right point cloud data from the right blind filling radar, wherein the interval of the acquisition time between any two point cloud data in the main point cloud data, the left point cloud data and the right point cloud data is less than a preset threshold, a first coordinate transformation formula for calibrating the left point cloud data to the main point cloud data is generated, and a second coordinate transformation formula for calibrating the right point cloud data to the main point cloud data is generated; target detection processing is performed on the main point cloud data, the left point cloud data and the right point cloud data, and a plurality of object frames and the detection scores of the objects in each object frame are detected from the main point cloud data, the left point cloud data and the right point cloud data, respectively.

[0052] An alignment module is configured to align any point cloud frame Pic MAll are processed as follows: Select the point cloud frame Pic from the left point cloud data. M Pic, the point cloud frame that is closest in time L Select the point cloud frame Pic from the right point cloud data. M Pic, the point cloud frame that is closest in time R If the point cloud frame Pic L Any object frame in L With point cloud frame Pic M Frame of objects ML When matching, use the object frame. L Update the object frame ML The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M Frame of objects MR When matching, use the object frame. R Update the object frame MR The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M All object bounding boxes do not match, but the object bounding box Frame R With point cloud frame Pic L Frame of objects LR When matching, the object bounding box Frame is transformed based on the first coordinate transformation formula. LR Transform into a bounding box Frame' LR , set the object frame' LR Add to point cloud frame Pic M In the middle, use the object frame. R and object frame LR Update the object bounding box Frame' LR The test score;

[0053] The processing module is used to downsample the main point cloud data, and then delete the point clouds in all object frames of all point cloud frames in the main point cloud data; perform planar rasterization processing on the main point cloud data, and extract non-ground point clouds and non-ground planar raster based on raster features; perform clustering processing on the non-ground point clouds, and remove the raster in the clustered object frames from the non-ground planar raster to obtain irregular object raster that cannot be clustered.

[0054] Embodiment three of the present application provides an electronic device, comprising: one or more processors; a memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, the one or more processors implement the target detection method in embodiment one.

[0055] Embodiment four of the present application provides a computer readable medium having a computer program stored thereon, when the program is executed by a processor, the target detection method in embodiment one is implemented.

[0056] It should be noted that although the above describes each step in a specific order, it does not mean that each step must be performed in the above specific order, in fact, some of the steps can be performed concurrently, or even change the order, as long as the required function can be realized.

[0057] The present application can be a system, a method, and / or a computer program product. The computer program product can include a readable storage medium having stored thereon instructions that, when executed by a processor, implement the various aspects of the present application.

[0058] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch cards or punched tape, and any suitable combination of the above. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se.

[0059] The above has described various embodiments of the present application, the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical application or technical improvement in the market, or to enable other ordinary skilled in the art to understand the embodiments disclosed herein.

Claims

1. A target detection method for a vehicle, wherein the vehicle is equipped with a main radar, a left blind-spot radar, and a right blind-spot radar; characterized in that, Includes the following steps: The system acquires main point cloud data from the main radar, left point cloud data from the left blind-filling radar, and right point cloud data from the right blind-filling radar. The time interval between any two point cloud data in the main point cloud data, left point cloud data, and right point cloud data is less than a preset threshold. A first coordinate transformation formula for calibrating the left point cloud data to the main point cloud data is generated, and a second coordinate transformation formula for calibrating the right point cloud data to the main point cloud data is generated. Object detection processing is performed on the main point cloud data, left point cloud data, and right point cloud data. Several bounding boxes are detected from the main point cloud data, left point cloud data, and right point cloud data respectively, and the detection score of the object in each bounding box is obtained. For any point cloud frame Pic in the main point cloud data M All are processed as follows: Select the point cloud frame Pic from the left point cloud data. M Pic, the point cloud frame that is closest in time L Select the point cloud frame Pic from the right point cloud data. M Pic, the point cloud frame that is closest in time R If the point cloud frame Pic L Any object frame in L With point cloud frame Pic M Frame of objects ML When matching, use the object frame. L Update the object frame ML The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M Frame of objects MR When matching, use the object frame. R Update the object frame MR The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M All object bounding boxes do not match, but the object bounding box Frame R With point cloud frame Pic L Frame of objects LR When matching, the object bounding box Frame is transformed based on the first coordinate transformation formula. LR Transform into a bounding box Frame' LR , set the object frame' LR Add to point cloud frame Pic M In the middle, use the object frame. R and object frame LR Update the object bounding box Frame' LR The test score; The main point cloud data is downsampled. Then, for all point cloud frames in the main point cloud data, the point clouds in all object boxes of the point cloud frames are deleted. The main point cloud data is then rasterized, and non-ground point clouds and non-ground planar grids are extracted based on grid features. The non-ground point clouds are then clustered, and the grids in the clustered object boxes are removed from the non-ground planar grids to obtain irregular object grids that cannot be clustered.

2. The target detection method according to claim 1, characterized in that, The "first coordinate transformation formula for generating the left point cloud data calibration to the main point cloud data" specifically includes: the first coordinate transformation formula for the left point cloud data calibration to the main point cloud data is: x′ L =x L ·cosθ L -y L ·sinθ L +T xL y′ L =x L ·sinθ L +y L ·cosθ L +T yL , z′ L =z L +T zL ; where the coordinates (x L y L , z L Let (x') be the 3D coordinates of any point pot1 in the left point cloud data, with coordinates (x') L y′ L , z′ L Let be the projected coordinates of point pot1 in the main point cloud data. The translation transformation of the main radar relative to the left blind-filling radar is (T) xL T yL T zL The main radar rotates counterclockwise around the z-axis at an angle θ relative to the left blind-filling radar. L ; The "generating the second coordinate transformation formula from right point cloud data calibration to main point cloud data" specifically includes: the second coordinate transformation formula from right point cloud data calibration to main point cloud data is: x′ R =x R ·cosα R -y R ·sinθ R +T xR y′ R =x R ·sinθ R +y R ·cosθ R +T yR , z′ R =z R +T zR ; where the coordinates (x R ,y R , z R Let (x′) be the 3D coordinates of any point pot2 in the right point cloud data. R y′ R , z′ R Let be the projected coordinates of point pot2 in the main point cloud data. The translation transformation of the main radar relative to the right blind-filling radar is (T). xR T yR T zR The main radar rotates counterclockwise around the z-axis at an angle θ relative to the right blind-filling radar. R .

3. The target detection method according to claim 1, characterized in that, The phrase "performing target detection processing on the main point cloud data, left point cloud data, and right point cloud data" specifically includes: Based on deep learning algorithms, target detection processing is performed on the main point cloud data, left point cloud data, and right point cloud data.

4. The target detection method according to claim 1, characterized in that, The phrase "performs target detection processing on the main point cloud data, left point cloud data, and right point cloud data, and detects several bounding boxes from the main point cloud data, left point cloud data, and right point cloud data respectively, as well as the detection score of the object in each bounding box" specifically includes: performing target detection processing on the main point cloud data, left point cloud data, and right point cloud data, and detecting several bounding boxes from the main point cloud data, left point cloud data, and right point cloud data respectively, as well as the detection score of the object in each bounding box, and setting an initial value of zero for each bounding box in the main point cloud data; The phrase "using an object frame" L Update the object frame ML The "detection score" specifically includes: the object bounding box (Frame). ML The score modification value increases the object frame. L The test score; The phrase "using an object frame" R Update the object frame MR The "detection score" specifically includes: the object bounding box (Frame). MR The score modification value increases the object frame. R The test score; The phrase "using an object frame" R and object frame LR Update the object bounding box Frame' LR The detection score specifically includes: the object bounding box Frame' LR The score modification value increases the object frame. R The detection score is then used to add the object bounding box (Frame). LR The detection score, the object bounding box Frame′ LR The test score was zero; The processing also includes: if the point cloud frame Pic M If the score modification value of any object box F is not zero, then Score = a * detection score of object box F + β * score modification value of object box F, where α > 0 and β > 0.

5. The target detection method according to claim 1, characterized in that, The phrase "delete all point clouds within the bounding boxes of any point cloud frame in the main point cloud data" specifically includes: For any point pot3 and any object bounding box Frame in all point cloud frames of the main point cloud data, the following operations are performed: obtain the four coordinates (x1, y1), (x2, y2), (x3, y3), and (x4, y4) of the object bounding box Frame, and the coordinates (x, y) of point pot3, where the directed distance from coordinates (x, y) to the four sides of the object bounding box Frame is d. i =(xx) i )·(y i％4+1 -y i )-(yy i )·(x i％4+1 -x i ), i = 1, 2, 3, 4; when the first condition or the second condition is met, delete point pot3; the first condition is d1≥0 and d2≥0 and d3≥0 and d4≥0, and the second condition is d1≤0 and d2≤0 and d3≤0 and d4≤0.

6. The target detection method according to claim 1, characterized in that, The "planar rasterization processing of the main point cloud data" specifically includes: obtaining the range and resolution of the xy plane raster in the main point cloud data; creating multiple rasters on a two-dimensional plane based on the range and resolution; obtaining several point clouds corresponding to each raster, with each raster intersecting with the corresponding several point clouds; The "extraction of non-ground point clouds and non-ground planar grids based on grid features" specifically includes: For each grid cell, the average height, maximum height, minimum height, and normal vector of the point cloud are calculated; a gradient threshold is obtained, and the grid cell is marked as a ground grid cell; for blind spot grid cells around the vehicle, they are directly marked as ground grid cells; a breadth-first search algorithm is used to search for surrounding grid cells starting from the grid cell where the vehicle is located, and adjacent grid cells are marked according to height change and normal consistency; if the height change is greater than the gradient threshold, it is marked as an obstacle grid cell; map information is generated based on the ground grid cells and obstacle grid cells, and the map information includes at least: grid coordinates, height information, ground point cloud, and non-ground point cloud; then, the non-ground grid cells are extracted using the index of the obstacle grid cells.

7. The target detection method according to claim 1, characterized in that, The "clustering of non-ground point clouds" specifically includes: using the Euclidean clustering algorithm DBSCAN, combined with Kd-Tree to accelerate the search for nearest neighbor points in the non-ground point cloud; after completing DBSCAN clustering, extracting points within the same cluster to generate several clustered point clouds; for each clustered point cloud, the following processing is performed: obtaining the minimum Z-axis coordinate ZMin and maximum ZMax of the midpoint of the clustered point cloud, obtaining the cluster height h = ZMax - ZMin; projecting the clustered point cloud onto the XOY plane to obtain a projection point set; generating a minimum convex polygon surrounding the projection point set; for each edge of the minimum convex polygon, calculating the sum of distances from all vertices of the minimum convex polygon to the edge, obtaining the edge Edge1 with the smallest sum of distances; and then processing the minimum convex polygon... All vertices are projected onto edge Edge1. The two points p1 and p2 with the furthest projection distance are obtained. The slope of edge Edge2 formed by points p1 and p2 is obtained. The point Point farthest from edge Edge2 is found in the point set of the minimum convex polygon. An edge Edge3 passing through point Point and parallel to edge Edge2 is generated. Points p1 and p2 are projected onto edge Edge3 to obtain two points p3 and p4. Points p1, p2, p3, and p4 are used as vertices to generate the minimum bounding rectangle. The length l, width w, and hypotenuse of the minimum bounding rectangle are determined by calculating the distance between these four vertices. The slope of the longer side relative to the X-axis (positive direction of the car) is the orientation angle θ of the minimum bounding rectangle. The coordinates of the center point (x, y, y) are obtained from the coordinates of the four vertices of the minimum bounding rectangle. M y M ), Generate a 3D clustered object bounding box, wherein the center point coordinates of the 3D clustered object bounding box are (X... M y M Z M ,), orientation angle θ, length l, width w and height h are the orientation angle, length, width and height on the z-axis of the three-dimensional clustering object box, respectively; The phrase "removing grid cells from the non-ground plane grid to obtain irregular object grid cells that cannot be clustered" specifically includes: the coordinates of the four vertices of the rotation rectangle of the three-dimensional clustering object frame are (x1, y1), (x2, y2), (x3, y3), and (x4, y4), respectively. For each point in the non-ground grid, the following processing is performed: the directed distance from the point to the four sides of the rotation rectangle is obtained; when the point is determined to be in the rotation matrix based on the cross product calculation method, the point is deleted.

8. A target detection device for a vehicle, wherein the vehicle is equipped with a main radar, a left blind-spot radar, and a right blind-spot radar; characterized in that, Includes the following modules: The data acquisition module is used to acquire main point cloud data from the main radar, left point cloud data from the left blind-filling radar, and right point cloud data from the right blind-filling radar. The time interval between any two point cloud data in the main point cloud data, left point cloud data, and right point cloud data is less than a preset threshold. The module generates a first coordinate transformation formula from left point cloud data to main point cloud data and a second coordinate transformation formula from right point cloud data to main point cloud data. Object detection processing is performed on the main point cloud data, left point cloud data, and right point cloud data. Several bounding boxes are detected from the main point cloud data, left point cloud data, and right point cloud data respectively, and the detection score of the object in each bounding box is obtained. The alignment module is used to align any point cloud frame (Pic) in the main point cloud data. M All are processed as follows: Select the point cloud frame Pic from the left point cloud data. M Pic, the point cloud frame that is closest in time L Select the point cloud frame Pic from the right point cloud data. M Pic, the point cloud frame that is closest in time R If the point cloud frame Pic L Any object frame in L With point cloud frame Pic M Frame of objects ML When matching, use the object frame. L Update the object frame ML The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M Frame of objects MR When matching, use the object frame. R Update the object frame MR The detection score; if the point cloud frame Pic R Any object frame in R With point cloud frame Pic M All object bounding boxes do not match, but the object bounding box Frame R With point cloud frame Pic L Frame of objects LR When matching, the object bounding box FrameLR is transformed into the object bounding box Frame′ based on the first coordinate transformation formula. LR , set the object frame' LR Add to point cloud frame Pic M In the middle, use the object frame. R and object frame LR Update the object bounding box Frame' LR The test score; The processing module is used to downsample the main point cloud data, and then delete the point clouds in all object frames of all point cloud frames in the main point cloud data; perform planar rasterization processing on the main point cloud data, and extract non-ground point clouds and non-ground planar raster based on raster features; perform clustering processing on the non-ground point clouds, and remove the raster in the clustered object frames from the non-ground planar raster to obtain irregular object raster that cannot be clustered.

9. An electronic device, characterized in that, include: One or more processors; A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement the target detection method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, It stores a computer program that, when executed by a processor, implements the target detection method according to any one of claims 1 to 7.

Images