Out-of-order target identifying and positioning method based on binocular laser three-dimensional scanning imaging

Abstract

The invention relates to the technical field of robot visual guidance, in particular to an out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging, which comprises the following steps of: performing laser scanning on an object in a target area, and synchronously acquiring laser stripe images by adopting a binocular camera to obtain a target area image sequence; extracting a laser stripe center line of the target area image sequence to obtain auxiliary features; performing three-dimensional point cloud reconstruction on the auxiliary features to obtain original point cloud data; compensating the original point cloud data to obtain fused point cloud data; on the basis of a spherical multi-view sampling and positioning algorithm framework, off-line model point cloud features are constructed, and object recognition and pose matching are carried out on fused point cloud data; and visualizing an object identification and pose matching result. The method has the characteristics of stable identification and positioning results, high applicability in complex scenes, high identification accuracy and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of robot vision guidance, and in particular to a method for identifying and positioning out-of-order targets based on binocular laser three-dimensional scanning imaging. Background Art

[0002] Machine vision technology is a non-contact, high-precision, and highly flexible sensing method used to help robots automatically grasp scattered objects in industrial production. Through vision systems, robots can perceive the posture, position, and shape of objects in their environment in real time, providing accurate data support for subsequent grasping, effectively overcoming the limitations of traditional industrial robots when handling out-of-order objects.

[0003] Prior art vision-based methods for robotic grasping target recognition and localization can be categorized as either two-dimensional (2D) or three-dimensional (3D) vision-based. 2D image analysis suffers from unstable recognition and localization results and limited applicability in complex scenes due to the complex structure of target objects, single-color, weakly textured surfaces, and the potential for multiple objects to overlap. 3D vision systems can provide spatial position information and geometric shape, mitigating the impact of occlusion on recognition accuracy. However, in highly disordered scenes, 3D vision data can suffer from increased measurement errors or localized information loss due to factors such as the complex stacking of parts, severe occlusion, and varying ambient lighting, thus impacting grasping stability and accuracy. In industrial scenarios, grasping randomly stacked parts in disorder still faces numerous challenges, such as weak or no texture on the objects being grasped, severe ambient lighting interference, and the diversity of 3D surface morphology. These factors leave significant room for improvement in target recognition accuracy and grasping precision. Furthermore, existing methods lack effective solutions for occlusion or partial missing components.

[0004] Therefore, there is an urgent need for a robot out-of-order grasping target recognition and positioning method with stable recognition and positioning results, strong applicability in complex scenes, and high recognition accuracy. Summary of the Invention

[0005] In view of this, the present application provides a method for identifying and locating out-of-order targets based on binocular laser 3D scanning imaging to solve the above problems; the method includes:

[0006] Step 1: Use a laser transmitter to perform laser scanning on the object in the target area, and use a binocular camera to synchronously capture laser stripe images to obtain a target area image sequence;

[0007] Step 2: extracting the center line of the laser stripe in the target area image sequence to obtain auxiliary features, wherein the auxiliary features include the coordinates of the center of gravity of the light stripe;

[0008] Step 3: Based on the auxiliary features, the three-dimensional point cloud is reconstructed by combining the binocular camera joint calibration and the laser triangulation principle to obtain the original point cloud data;

[0009] Step 4: Use a blind area 3D compensation method based on a laser plane model to compensate the original point cloud data to obtain fused point cloud data;

[0010] Step 5: Based on the spherical multi-view sampling and positioning algorithm framework, build an offline model point cloud feature library;

[0011] Step 6: Based on the offline model point cloud feature library, perform object recognition and pose matching on the fused point cloud data;

[0012] Step 7: Visualize the results of object recognition and pose matching.

[0013] The beneficial effects of the present invention include:

[0014] By combining the texture characteristics and grayscale distribution information of the stripes, the threshold is dynamically adjusted to optimize the extraction effect of the stripe center;

[0015] The point positions in the blind spots are completed and calculated by the principle of multi-view line-surface intersection, and the initial pose deviation compensation strategy is combined to achieve fast multi-source point cloud registration, which significantly improves the integrity and efficiency of 3D reconstruction and enhances adaptability in complex scenes.

[0016] By adopting the dynamic farthest point sampling method to downsample the point cloud, uniform distribution sampling is achieved, and the global shape characteristics of the point cloud are retained while avoiding the loss of local details;

[0017] By building an offline model point cloud feature library and applying a multi-view HPR operator for visibility judgment, the interference of invisible points is eliminated and a sufficiently accurate feature library is established, thereby enhancing the stability of recognition and positioning results. In addition, by adopting voxel-based pose verification and a top-down grasping strategy, as well as the ICP fine registration algorithm to optimize the pose estimation of the initial matching results, the stability of recognition and positioning results is further ensured.

[0018] It provides a new design idea for the design of out-of-order target recognition and positioning methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the process of the out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging in the present invention;

[0020] Figure 2 Schematic diagram of performing laser stripe centerline extraction processing on a target area image sequence in an embodiment of the present invention;

[0021] Figure 3 Schematic diagram of the principle of 3D point cloud reconstruction in an embodiment of the present invention;

[0022] Figure 4 Schematic diagram of a three-dimensional blind area compensation method based on a laser plane model in an embodiment of the present invention;

[0023] Figure 5 Schematic diagram of step 5 and step 6 in an embodiment of the present invention;

[0024] Figure 6 Schematic diagram of the geometric relationship between viewpoint distribution and object point cloud in an embodiment of the present invention;

[0025] Figure 7 Schematic diagram of the offline construction process of the global model hash table in an embodiment of the present invention;

[0026] Figure 8 Schematic diagram of accumulating votes for possible target postures based on the Hough voting mechanism in an embodiment of the present invention;

[0027] Figure 9 Schematic diagram for visualizing the results of recognition and posture estimation in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the objectives, technical solutions, features and advantages of the present invention more clearly understood, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0029] This embodiment includes a method for identifying and locating out-of-order targets based on binocular laser three-dimensional scanning imaging, such as Figure 1 Shown, including:

[0030] Step 1: Use a laser transmitter to perform laser scanning on the object in the target area, and use a binocular camera to synchronously capture laser stripe images to obtain a target area image sequence.

[0031] Step 2: Extract the center line of the laser stripe in the target area image sequence to obtain auxiliary features, which include the coordinates of the center of gravity of the light stripe. Specifically, the results of the laser stripe center line extraction process in this embodiment are as follows: Figure 2 As shown in the figure, the left side is the target area image, and the red line on the right side represents the auxiliary feature, that is, the center of the laser stripe. The laser stripe centerline extraction process includes:

[0032] S201, using the Otsu threshold method, to perform adaptive region extraction on the target region image sequence to obtain a region of interest.

[0033] Specifically, the laser light band and the background area are separated, and the minimum bounding rectangle is calculated based on the laser stripe profile. The minimum bounding rectangle is expanded, and the expanded area and the rectangular area are defined as the region of interest.

[0034] The Otsu threshold method divides the image into foreground and background by finding the only optimal segmentation threshold among all possible grayscale thresholds, which correspond to the laser light band area and the non-laser area respectively, and maximizes the inter-class variance of the two types of pixels after division, thereby achieving the best segmentation effect. The formula for the inter-class variance is:

[0035]

[0036] in, represents the inter-class variance, ω0(t) and ω1(t) represent the pixel ratio of the background and the pixel ratio of the foreground, μ0(t) and μ1(t) represent the grayscale mean of the background and the grayscale mean of the foreground. Traverse all possible thresholds t and select the one that makes the inter-class variance The largest threshold is taken as the optimal segmentation threshold.

[0037] S202 , using a grayscale centroid method with a threshold to extract the center of the stripes in the region of interest, and obtaining the centroid coordinates of the light stripes.

[0038] Specifically, to improve extraction accuracy, the present invention dynamically adjusts the threshold based on the texture characteristics and grayscale distribution of the stripes, thereby optimizing the extraction of the stripe center. The grayscale centroid method with thresholding considers the grayscale distribution of the light streak and the grayscale values ​​of the pixels within the cross section as weights. The centroid coordinates of the pixels within the cross section are calculated based on the weights to determine the center position of the light streak. The calculation formula is:

[0039]

[0040] Among them, Gravity represents the coordinates of the center of gravity of the light stripe, g i Represents the grayscale value of the i-th pixel, μ i represents the coordinates of the i-th pixel point, A and B represent the boundaries of the light streak area to be calculated, and the method for determining A and B includes: setting a grayscale threshold Th, and checking pixel by pixel along the cross-sectional direction of the light streak starting from the edge of the image. The coordinates of the first pixel point whose grayscale value is not less than the grayscale threshold Th are A, and the coordinates of the pixel point before the first pixel point whose grayscale value is less than the grayscale threshold Th are B.

[0041] S203 , using the Sobel operator to perform gradient calculation on the center-of-gravity coordinates of the light stripes to obtain a direction field.

[0042] S204: Calculate the grayscale centroid of the image sequence based on the direction field to obtain auxiliary features.

[0043] Specifically, the assist feature represents the center of the laser stripe.

[0044] Step 3: Based on the auxiliary features, the three-dimensional point cloud is reconstructed by combining the binocular camera joint calibration and the laser triangulation principle to obtain the original point cloud data; specifically, the principle of three-dimensional point cloud reconstruction is as follows: Figure 3 Shown, including:

[0045] S301, decompose the binocular camera into two coplanar and parallel virtual cameras, and construct an image aligned in a plane; the image includes the left image P l With the right image P r , alignment is co-row alignment or co-column alignment.

[0046] S302, using the center of gravity coordinates of the light stripes as the features of binocular vision stereo matching, calculate P l With P r The parallax between points of the same name.

[0047] Specifically, the same-name point refers to the left image P in the binocular vision system. l and right image P r Two pixels corresponding to the same spatial point on the left and right images. Because the binocular vision system simulates the visual mechanism of the human eye, using two cameras to capture the same scene from different angles, object points in the same space are projected to different positions on the two images. The two projected points are geometrically corresponding, meaning they represent the imaging points of the same spatial entity from different perspectives. Pixels corresponding to the same spatial point on the left and right images are called homonymous points.

[0048] S303: Obtain calibration parameters of the binocular camera, and determine the plane coordinates (X, Y) of the feature points according to the calibration parameters.

[0049] Specifically, after imaging, the binocular camera is calibrated by the camera system, including camera calibration, line laser light plane calibration, and displacement platform calibration. Camera calibration is used to correct image distortion and establish an accurate mapping relationship between image pixel coordinates and real-world coordinates. Line laser light plane calibration is used to determine the general equation of the laser light plane. Displacement platform calibration is used to determine the displacement direction of the laser light strip images in adjacent frames. The spatial coordinate system is the global spatial coordinate system.

[0050] S304. Calculate the depth value of the feature point using the laser triangulation principle and parallax. The plane coordinates of the feature point and the depth value constitute the original point cloud data. The formula for the laser triangulation principle is:

[0051]

[0052] Where Z represents the depth value, f represents the focal length of the binocular camera, and the binocular camera is built by two cameras with the same internal parameters, f = f l =f r , that is, the focal lengths of the two cameras are the same, b represents the baseline length, d represents the parallax, and x l with x r The x, y, and z coordinates represent the horizontal coordinates of the feature points captured by the two cameras. The (X, Y, Z) form the location information of the feature points, and the location information of all the feature points constitutes the original point cloud data.

[0053] Step 4: Use a blind area three-dimensional compensation method based on a laser plane model to compensate the original point cloud data to obtain fused point cloud data.

[0054] The blind spot 3D compensation method based on the laser plane model includes: using the multi-view line-plane intersection principle to complete the calculation of the blind spot points, and combining the initial posture deviation compensation strategy to achieve fast multi-source point cloud registration. The principle is as follows: Figure 4 As shown in the figure, the blind spot 3D compensation method based on the laser plane model can significantly improve the integrity and efficiency of 3D reconstruction and enhance the adaptability of the method in complex scenes. Specifically:

[0055] S401. Establish a spatial equation of the laser scanning plane.

[0056] Specifically, in traditional binocular laser vision systems, if one camera fails to capture the laser stripes due to occlusion, reflection, or an inappropriate angle, it will be impossible to calculate accurate depth information for that area. In this case, the area where depth data was not collected will have missing data, resulting in holes or errors in the 3D reconstruction.

[0057] The present invention solves the three-dimensional information of the blind area through the spatial equation of the laser scanning plane. During measurement, if one of the cameras fails to successfully capture the laser stripe due to occlusion, since the center of the laser stripe is located in the laser light plane, the laser stripe center captured by the other camera is obtained and the camera calibration parameters are used to calculate the X and Y values ​​of the laser stripe center feature in space. The X and Y values ​​are substituted into the spatial equation of the laser scanning plane to obtain the Z value, i.e., the depth information, to complete the recovery of the missing three-dimensional information. The formula of the spatial equation of the laser scanning plane in this embodiment is:

[0058] ax+by+cz+d=0

[0059] Among them, (a, b, c) is the unit normal vector, (x, y, z) represents the restored three-dimensional information, and d is a constant term.

[0060] S402: Establish a local world coordinate system based on the position information of the binocular camera.

[0061] Specifically, the local world coordinate system is used for subsequent point cloud transformation and registration based on the binocular camera setup.

[0062] S403: Perform coordinate transformation on the original point cloud data to obtain left and right point cloud information. The formula used is:

[0063]

[0064] in, and Represents the left and right point cloud information, T left and T right They represent the transformation matrix obtained by global calibration of the left image and the transformation matrix obtained by global calibration of the right image, PCL left With PCL right Respectively represent the original point cloud data obtained based on the two cameras in the binocular camera.

[0065] S404: In the local world coordinate system, based on the line-plane intersection principle, the left and right point cloud information are compensated to restore the missing three-dimensional information and obtain the completed point cloud data.

[0066] S405 , in the local world coordinate system, using the ICP (Iterative Closest Point) algorithm to align and optimize the original point cloud data and the completed point cloud data to obtain fused point cloud data.

[0067] Step 5: Based on the spherical multi-view sampling and positioning algorithm framework, build an offline model point cloud feature library. Specifically, build an offline model point cloud feature library such as Figure 5 As shown in the offline training, it includes:

[0068] S501: Construct a model point cloud, perform spherical multi-view sampling on the model point cloud, and obtain a uniform sampling perspective.

[0069] Specifically, the model point cloud is used for offline training. In this embodiment, the model point cloud is a model of the target object in .STL format obtained by modeling with SolidWorks, and then the .STL format model is converted into a point cloud data format using the PCL (Point Cloud Library) point cloud library. Figure 6 It shows the geometric relationship between the distribution of 200 viewpoints and the object point cloud when the spherical radius is 0.65 meters.

[0070] S502: For each sampling viewpoint, use the HPR (Hidden Point Removal) operator to perform visibility analysis on the fused point cloud data, filter out occluded points, and retain observable points in the viewpoint direction. The observable points constitute a visible point set.

[0071] Specifically, the HPR operator is used to perform visibility analysis on the fused point cloud data, including: mapping the point cloud to a coordinate system relative to the viewpoint, filtering out the obscured points through inverse convex hull analysis, and retaining the remaining points to form a visible point set.

[0072] S503 , selecting spatially uniformly distributed reference points from the visible point set, and calculating point pair feature descriptors of the reference points.

[0073] Specifically, the model point cloud after applying the multi-view HPR operator still contains a large amount of redundant data, which not only increases storage and computational overhead, but also affects the efficiency of subsequent feature calculation and matching. To achieve uniformly distributed sampling, preserve the global shape characteristics of the point cloud, and avoid the loss of local details, the present invention uses the farthest point sampling method to downsample the point cloud when selecting spatially uniformly distributed reference points.

[0074] Furthermore, the target number of sampling points K for farthest point sampling in the present invention is dynamically set according to the diameter D of the point cloud, and the formula is K=[D / a], where a is the sampling resolution, indicating the maximum distance between adjacent sampling points.

[0075] S504: Store the point-to-point feature descriptors under all sampling perspectives into the hash feature library to complete the offline construction of the global model hash table.

[0076] Specifically, the offline construction process of the global model hash table is as follows: Figure 7 As shown in the figure, (m1, m2), (m3, m4), (m5, m6) represent key point pairs, F represents the point pair feature descriptor under all sampling perspectives, and A represents the global model hash table.

[0077] Step 6: Based on the offline model point cloud feature library, perform object recognition and pose matching on the fused point cloud data. Specifically, object recognition and pose matching are as follows: Figure 5 As shown in the online matching, the scene point cloud is the fused point cloud data, including:

[0078] S601 , pre-processing the fused point cloud data to extract key points; the pre-processing includes: noise reduction, downsampling, and background separation processing.

[0079] S602. Calculate the PPF (Point-Pair Feature) of the key point, query the descriptor in the global model hash table, quickly match the PPF with the descriptor, and solve the pose transformation relationship between the scene object and the model through the corresponding point-pair feature to obtain the possible pose of the object.

[0080] Specifically, the formula of point pair feature PPF is:

[0081] f(m1,m2)=(║d║,∠(n1,d),∠(n2,d),∠(n1,n2))

[0082] Among them, f(m1,m2) represents the point pair feature, m1 and m2 represent the key point pair, d represents the point pair vector, ∠ represents the angle operation, and n1 and n2 represent the normal vectors of the key point pair.

[0083] S603: Based on the Hough voting mechanism, votes are accumulated on the possible positions of the object.

[0084] Specifically, if Figure 8 As shown, (s r ,s i ) represents the scene point pair, F S (s r ,s i ) means (s r ,s i )’s point pair features, (m r , m i ) and (m′ r , m′ i ) represents different model point pairs. ① represents the point pairs located on the target object (s r ,s i );② represents the point pair feature F S (s r ,s i ) as an index, and find similar model point pair features in the hash table of the global model constructed in the offline stage. The embodiment of the present invention constructs a two-dimensional accumulator array to count the number of votes when the scene point pairs are matched with the model point pairs. The number of rows of the two-dimensional accumulator is equal to the number of key points of the model point cloud, and the number of columns is equal to the discrete number of rotation angles α; ③ represents the calculation of the rotation angle α between each group of scene and model point pairs; ④ represents the discretization of the rotation angle between each group of scene and model point pairs and the addition of 1 to the corresponding position of the two-dimensional accumulator array, traversing all point pairs in the scene, and finally further setting a vote threshold in the accumulator array, considering only the local coordinates corresponding to the peak value whose votes exceed a certain threshold, to ensure that the initial 6D poses of multiple instances are obtained in the scene.

[0085] S604: Using voxel-based posture verification and a top-down grasping strategy, incorrect postures are eliminated, and the priority of object grasping is determined, with the object at the top being given a high priority, to complete the initial matching. The voxel-based posture verification includes:

[0086]

[0087] Among them, [] represents the rounding operation, (x si ,y si,z si ) represents the scene point, (x int ,y int ,z int ) represents the minimum value of all points in the scene point cloud on the three coordinate axes.

[0088] S605: Use the ICP fine registration algorithm to optimize the pose estimation of the initial matching result to complete the pose matching.

[0089] Specifically, the objective function L(T) of the ICP fine registration algorithm is:

[0090] L(T)=∑║T·p i -q i ║ 2

[0091] Where T represents the rigid transformation matrix, p i represents the model point set, q i Represents the corresponding point set of the target scene.

[0092] Step 7: Visualize the results of object recognition and pose matching.

[0093] Specifically, if Figure 9 As shown in the figure, the blue point cloud represents the accurate recognition and positioning candidate results, and the green point cloud represents the optimal 6D grasping target outputted in the end.

[0094] Finally, it should be noted that the above only describes some embodiments of the present invention. For those skilled in the art, it is conceivable that various changes, modifications, substitutions and deformations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents, and the above-mentioned actions should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying and locating out-of-order targets based on binocular laser three-dimensional scanning imaging, characterized in that: include: Step 1: Use a laser transmitter to perform laser scanning on the object in the target area, and use a binocular camera to synchronously capture laser stripe images to obtain a target area image sequence; Step 2: extracting the center line of the laser stripe in the target area image sequence to obtain auxiliary features, wherein the auxiliary features include the coordinates of the center of gravity of the light stripe; Step 3: Based on the auxiliary features, the three-dimensional point cloud is reconstructed by combining the binocular camera joint calibration and the laser triangulation principle to obtain the original point cloud data; Step 4: Use a blind area 3D compensation method based on a laser plane model to compensate the original point cloud data to obtain fused point cloud data; Step 5: Based on the spherical multi-view sampling and positioning algorithm framework, build an offline model point cloud feature library; Step 6: Based on the offline model point cloud feature library, perform object recognition and pose matching on the fused point cloud data; Step 7: Visualize the results of object recognition and pose matching.

2. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 1 is characterized in that: Extracting the laser stripe centerline of the target area image sequence includes: S201, using the Otsu threshold method, performing adaptive region extraction on the target region image sequence to obtain a region of interest; S202, using a grayscale centroid method with a threshold to extract the center of the light streak in the region of interest to obtain the centroid coordinates of the light streak; S203, using the Sobel operator to calculate the gradient of the center of gravity coordinates of the light stripes to obtain a direction field; S204: Calculate the grayscale centroid of the image sequence based on the direction field to obtain auxiliary features.

3. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 2 is characterized in that: The grayscale centroid method with threshold includes: taking the grayscale distribution of the light streak and the grayscale value of the pixel in the cross section as weights, and calculating the centroid coordinates of the pixels in the cross section according to the weights. The calculation formula is: Among them, Gravity represents the coordinates of the center of gravity of the light stripe, g i Represents the grayscale value of the i-th pixel, μ i represents the coordinates of the i-th pixel point, A and B represent the boundaries of the light streak area to be calculated, and the method for determining A and B includes: setting a grayscale threshold Th, and checking pixel by pixel along the cross-sectional direction of the light streak starting from the edge of the image. The coordinates of the first pixel point whose grayscale value is not less than the grayscale threshold Th are A, and the coordinates of the pixel point before the first pixel point whose grayscale value is less than the grayscale threshold Th are B.

4. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 1 is characterized in that: 3D point cloud reconstruction includes: S301, decomposing a binocular camera into two coplanar and parallel virtual cameras, and constructing an image aligned in a plane; the image includes a left image and a right image, and the alignment is co-row alignment or co-column alignment; S302, using the centroid coordinates of the light streaks as features for binocular stereo matching, and calculating the disparity between the same-name points in the left and right images; S303, obtaining calibration parameters of the binocular camera, and determining the plane coordinates of the feature points according to the calibration parameters; S304 , calculating the depth value of the feature point through the laser triangulation principle and parallax; the plane coordinates of the feature point and the depth value constitute the original point cloud data.

5. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 1 is characterized in that: The three-dimensional blind zone compensation method based on the laser plane model includes: S401, establishing a spatial equation of the laser scanning plane; S402, establishing a local world coordinate system based on the position information of the binocular camera; S403, performing coordinate transformation on the original point cloud data to obtain left and right point cloud information; S404: In the local world coordinate system, the left and right point cloud information are compensated based on the spatial equation of the laser scanning plane to restore the missing three-dimensional information and obtain the completed point cloud data; S405. In the local world coordinate system, the ICP algorithm is used to align and optimize the original point cloud data and the completed point cloud data to obtain fused point cloud data.

6. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 1 is characterized in that: Building an offline model point cloud feature library includes: S501, constructing a model point cloud, performing spherical multi-view sampling on the model point cloud to obtain a uniform sampling perspective; S502: For each sampling viewpoint, perform visibility analysis on the fused point cloud data using a hidden point elimination operator, filter out occluded points, and retain observable points in the viewpoint direction. The observable points constitute a visible point set. S503, selecting spatially uniformly distributed reference points from the visible point set, and calculating point pair feature descriptors of the reference points; S504: Store the point-to-point feature descriptors under all sampling perspectives into the hash feature library to complete the offline construction of the global model hash table.

7. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 6 is characterized in that: The selection of reference points uniformly distributed in space includes: downsampling the visible point set using the farthest point sampling method, and the target sampling point number K of the farthest point sampling is dynamically set according to the diameter D of the point cloud. The formula is K = [D / a], where a is the sampling resolution, which represents the maximum distance between adjacent sampling points.

8. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 1 is characterized in that: Object recognition and pose matching include: S601, pre-processing the fused point cloud data to extract key points; S602: Calculate the point-pair features of the key points, query the descriptors in the global model hash table, quickly match the point-pair features with the descriptors, and solve the pose transformation relationship between the scene object and the model based on the matching results to obtain the possible pose of the object; S603: Based on the Hough voting mechanism, votes are accumulated for the possible positions of the object. S604: Using voxel-based pose verification and a top-down grasping strategy, incorrect poses are eliminated, and the priority of object grasping is determined, with the topmost object being given a high priority, to complete the initial matching. S605: Use the ICP fine registration algorithm to optimize the pose estimation of the initial matching result to complete the pose matching.

9. The out-of-order target recognition and positioning method based on binocular laser three-dimensional scanning imaging according to claim 8, characterized in that: Preprocessing of fused point cloud data includes: noise reduction, downsampling, and background separation.

Images