A precise grabbing method of a whole vehicle assembling robot based on binocular stereo vision

Abstract

The present application belongs to the field of robot technology, and particularly relates to a whole vehicle assembling robot precision grabbing method based on binocular stereo vision, which comprises the following steps: collecting a part image to be grabbed to generate a depth map, using an adaptive window to carry out distance weighted interpolation repair on invalid pixels to obtain a point cloud set; calculating a surface smoothness index of a candidate point; obtaining effective candidate points through edge distance screening and cylindrical virtual space collision detection; combining effective point cloud density, surface smoothness index and edge distance to calculate a grabbing suitability; selecting an optimal grabbing point according to the grabbing suitability and controlling a robot to perform grabbing. The present application effectively solves the data loss and noise interference problem caused by high light reflection surface, and improves the stability and safety of grabbing operation under complex lighting environment.

Description

Technical Field

[0001] This invention relates to the field of robotics. More specifically, this invention relates to a precise grasping method for a vehicle assembly robot based on binocular stereo vision. Background Technology

[0002] Industrial robots are widely used in scenarios such as vehicle assembly and parts handling. In these applications, 3D vision-based guided grasping technology is a core component for achieving automated operations. Typically, a binocular structured light camera is used to project coded patterns and acquire reflected images from the object's surface. A stereo matching algorithm is then used to calculate parallax, thereby generating a depth map and a 3D point cloud. This allows the robot's end effector to identify the object's position and orientation, guiding the robot's grasping action.

[0003] However, in actual industrial scenarios such as vehicle assembly, the objects to be captured are often highly reflective components such as body panels and automotive glass. These components have extremely smooth surfaces with significant specular reflection characteristics. When a structured light pattern is projected onto such surfaces, the light is prone to specular reflection, deviating from the camera's field of view, or directly reflecting into the lens, causing pixel overexposure. This optical characteristic makes it difficult for binocular stereo matching algorithms to extract effective feature points, resulting in large areas of missing data or the formation of false outliers in the generated original depth map and point cloud data.

[0004] To address the point cloud quality issues in highly reflective areas, existing processing methods often employ fixed-window interpolation algorithms or simple mean filtering. However, fixed windows struggle to adapt to reflective holes of varying sizes. Furthermore, in selecting gripping points, current gripping strategies typically focus primarily on surface geometric flatness. In sparse data areas caused by high reflectivity, even if a seemingly flat plane is fitted through interpolation, the data accuracy remains low. If the robot forcibly grips this area, problems such as suction cup leakage and slippage are highly likely. Additionally, if the gripping point is too close to the edge of the component or surrounded by obstructions, it may cause the workpiece to fall or collide during high-speed robotic arm movement. Summary of the Invention

[0005] To address the technical problems of existing technologies, such as insufficient point cloud data and severe noise interference in highly reflective scenes, the difficulty of balancing hole filling and feature preservation in traditional fixed window repair, and the lack of comprehensive consideration of data confidence and physical safety in grasping strategies, which lead to easy slippage or collisions during grasping, this invention provides a precise grasping method for a vehicle assembly robot based on binocular stereo vision. The method includes: controlling a binocular camera to acquire images of the parts to be grasped to generate an original depth map; cropping the region of interest; using an adaptive window to perform distance-weighted interpolation to fill the original depth map, obtaining a repaired depth map; obtaining a point cloud set from the repaired depth map; downsampling the point cloud set to obtain candidate points; and constructing the nearest neighbors of the candidate points. The covariance matrix of the point coordinates is decomposed to obtain three eigenvalues. Based on the three eigenvalues ​​and the angle between the candidate point and the normal vector of the nearest neighbor point, the surface stability index of the candidate point is obtained. The edge distance between the candidate point and the edge of the component is calculated, and candidate points with edge distances less than the safety threshold are eliminated. A cylindrical virtual space along the normal vector direction is established for collision detection, and the candidate points that pass the detection are taken as valid candidate points. The effective point cloud density in the neighborhood of the valid candidate point is calculated. The grasping suitability is obtained based on the surface stability index, edge distance, and effective point cloud density. The valid candidate point with the highest grasping suitability is selected as the best grasping point, and its coordinates are mapped to the robot coordinate system. The robot end effector is controlled to approach the best grasping point and grasp the component.

[0006] This invention effectively solves the problem of depth map voids caused by specular reflection on body sheet metal or glass surfaces through adaptive window distance weighted interpolation technology. While filling in data gaps, it avoids feature blurring caused by global smoothing, providing complete and high-quality point cloud data that retains local details for subsequent calculations. Secondly, this invention constructs a multi-level screening mechanism that includes surface stability, edge safety distance, and collision detection. This not only ensures the flatness of the gripping position to prevent air leakage from the suction cup during robot gripping, but also eliminates the risk of edge slippage and robot arm motion interference from a physical perspective. This invention also constructs gripping suitability based on effective point cloud density, avoiding the robot's attempt to grasp artifact regions that, although fitted to a plane, have sparse and unreliable actual data. Finally, this invention ensures perfect fit between the end effector posture and the workpiece surface by mapping visual coordinates to the robot coordinate system and performing precise approximation along the normal direction, achieving highly robust automated gripping under complex lighting conditions.

[0007] Preferably, the step of using an adaptive window to perform distance-weighted interpolation filling on the original depth map includes: designating pixels with depth values ​​of 0 and infinity within the region of interest as invalid pixels, and pixels outside the invalid pixels as valid pixels; designating any invalid pixel as the target pixel; establishing a search window of an initial size centered on the target pixel; and counting the number of valid pixels within the search window; if the number of valid pixels within the search window is lower than a preset threshold, then increasing the size of the search window by a fixed step size until the number of valid pixels within the window exceeds the threshold or reaches a preset maximum window limit; based on the finally determined search window, using the depth values ​​of all valid pixels within the search window, performing distance-weighted interpolation filling according to their Euclidean distance to the target pixel.

[0008] This invention achieves intelligent repair of reflective holes of different scales through an adaptive window based on the number of effective pixels. When dealing with small-area defects, it maintains a small search window to preserve the local curvature features and edge details of the component surface to the maximum extent. When dealing with large-area reflective areas, it automatically expands the window size to obtain enough reference pixels for interpolation, effectively balancing the contradiction between hole filling and detail preservation. This prevents image edge blurring caused by fixed large window interpolation or the inability to repair large holes due to fixed small window, thus improving the accuracy and naturalness of depth map repair.

[0009] Preferably, after obtaining the point cloud set, the point cloud set is processed using a statistical outlier filtering algorithm.

[0010] Preferably, the nearest neighbors of a candidate point are the 30 candidate points that are closest to the candidate point.

[0011] Preferably, the surface stability index satisfies the following relationship: In the formula, For the first The surface stability index of each candidate point For the first The normal vector of each candidate point For the first The first candidate point The normal vector of the nearest neighbor, The number of nearest neighbors. , and The first The three feature values ​​corresponding to each candidate point, among which , It is an inverse cosine function. The sign for the vector modulo operation. It is an exponential function with the natural constant as the base.

[0012] This invention can identify spherical protrusions or isotropic noise regions in point clouds by using the ratio relationship of eigenvalues; combined with the statistics of the angle between the neighborhood normal vectors, it can further distinguish regions with rough surfaces or drastic changes in normals, ensuring that the selected candidate points are located on highly flat and smooth surfaces, avoiding the risk of insufficient vacuum suction force or air leakage failure caused by uneven surfaces.

[0013] Preferably, the calculation of the edge distance between the candidate point and the edge of the component includes: identifying the set of edge points of the component using the maximum angle criterion, and for each candidate point, using the Euclidean distance between it and the nearest edge point as the edge distance.

[0014] Preferably, the step of establishing a cylindrical virtual space along the normal vector direction for collision detection includes: after removing candidate points whose edge distance is less than a safety threshold, establishing a cylindrical virtual space along the normal vector direction with any remaining candidate point as the center; if there is a non-target point cloud in the space, the candidate point fails the detection and is removed; if there is no non-target point cloud in the space, the candidate point passes the detection.

[0015] Preferably, the non-target point cloud is point cloud data whose distance from the central axis of the cylindrical virtual space is less than a preset safety radius.

[0016] This invention simulates the physical envelope of a robot's end effector suction cup or gripper by constructing a cylindrical virtual space extending along the normal vector direction of candidate points. Before path planning, it can identify obstacles or other non-target point clouds that may exist on the grasping path in advance, thereby directly eliminating candidate points that, although flat in surface, have obstructed spatial paths. This reduces the probability of collisions between the robotic arm and surrounding fixtures or adjacent parts in actual operations, protecting expensive equipment and workpieces.

[0017] Preferably, the grasp suitability satisfies the following relationship: In the formula, For the first The suitability of capturing each valid candidate point For the first The surface stability index of the effective candidate points For the first The edge distance of each valid candidate point For the first Effective point cloud density of each valid candidate point As a safety threshold, To achieve the maximum effective point cloud density, It is the hyperbolic tangent function. This is the sensitivity factor.

[0018] This invention integrates surface stability, edge distance, and point cloud density, and utilizes the saturation characteristics of the hyperbolic tangent function. When the grasping point is far enough from the edge and the data is dense enough, the grasping suitability tends to stabilize and no longer increases indefinitely, thus shifting the evaluation focus back to geometric stability. When any indicator is too low, the grasping suitability decays rapidly, which can forcibly filter out artifact regions caused by specular reflection, ensuring that the final selected optimal grasping point achieves an optimal balance in the three dimensions of geometry, physics, and data confidence.

[0019] Preferably, the process of controlling the robot end effector to approach and grasp the component at the optimal grasping point includes: reading a pre-calibrated hand-eye calibration matrix, converting the optimal grasping point position and normal vector in the visual coordinate system into the target position and target approach vector in the robot coordinate system; subsequently generating robot control commands to control the robot end effector's attitude axis to align with the target approach vector and quickly move to a position above the target position; finally, controlling the robot to approach the target position from above the optimal grasping point along the normal direction at low speed using linear interpolation until reaching the target position and activating the grasping mechanism to complete the precise grasping of the component.

[0020] The beneficial effects of this invention are as follows: Firstly, by employing distance-weighted interpolation based on adaptive window expansion, this invention achieves adaptive filling of the repair range by dynamically adjusting the depth data loss at different scales caused by specular reflection on the body sheet metal or glass surface. This effectively fills large-area data gaps while avoiding edge contour blurring caused by traditional global filtering or fixed window interpolation. This provides complete and highly accurate three-dimensional spatial information for robot visual guidance, establishing a data foundation for precise grasping. Secondly, by analyzing the consistency of local normal vectors and the characteristics of the covariance matrix, this invention identifies flat areas suitable for vacuum adsorption, avoiding suction cup leakage caused by excessive surface curvature. Simultaneously, by combining interference checks and edge distance screening in a cylindrical virtual space, it directly eliminates potential collision risks and edge slippage hazards during the robotic arm's movement from a physical perspective, ensuring the safety of equipment and workpieces on the vehicle assembly line. Furthermore, this invention addresses the problem that sparse data regions are easily misfitted as planes. By obtaining the appropriate grasping suitability through effective point cloud density, it ensures that the final selected optimal grasping point achieves the best balance in terms of geometric stability, physical security, and data reliability, thereby improving the robustness and production efficiency of industrial robot operations under complex lighting conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a precise grasping method for a vehicle assembly robot based on binocular stereo vision according to the present invention.

[0022] Figure 2This is a schematic oblique side view illustrating the point cloud data of automotive parts in this invention;

[0023] Figure 3 This is a schematic top view illustrating the point cloud data of automotive parts in this invention;

[0024] Figure 4 This is a schematic top-view diagram illustrating the distribution of the grasping suitability in this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention discloses a precise grasping method for a vehicle assembly robot based on binocular stereo vision, referring to... Figure 1 This includes steps S1 to S5:

[0028] S1. Image acquisition and point cloud restoration of highly reflective areas to obtain a restored point cloud set.

[0029] It's important to note that in the vehicle assembly scenario, the smooth surfaces of body panels and glass components are highly susceptible to specular reflection, causing disparity calculations by the binocular camera to fail and resulting in holes or flying point noise in the point cloud. To ensure the reliability of subsequent feature extraction, high-quality point cloud data must first be acquired. This step utilizes the assumption of depth continuity of neighboring pixels to interpolate and repair missing regions caused by reflections and remove background interference, providing a clean data foundation for subsequent accurate calculations.

[0030] Specifically, the binocular structured light camera mounted on the robot's end effector is moved to a position directly above the part to be grasped (such as a body panel), capturing images from both the left and right eyes and projecting a structured light coded pattern. A stereo matching algorithm is then used to generate an original depth map. Based on a pre-defined part template projection, a region of interest (ROI) is extracted from the original depth map. Pixels with depth values ​​of 0 or infinity within the ROI are designated as invalid pixels, while pixels outside the invalid pixels are considered valid. An adaptive window is used for distance-weighted interpolation to fill the invalid pixels, resulting in a repaired depth map. This repaired depth map is then converted into point cloud data, and statistical outlier filtering is used to remove outliers, yielding a repaired point cloud set.

[0031] In one embodiment, the adaptive window for distance-weighted interpolation filling includes: taking any invalid pixel in the region of interest caused by reflection or occlusion as the target pixel, establishing a search window of initial size centered on the target pixel, and counting the number of valid pixels in the search window; if the number of valid pixels in the search window is lower than a preset threshold, expanding the size of the search window in a fixed step increment until the number of valid pixels in the window is greater than the threshold or reaches a preset maximum window limit; based on the finally determined search window, using the depth values ​​of all valid pixels in the search window, performing distance-weighted interpolation filling according to their Euclidean distance to the target pixel, thereby filling large-area reflective holes while preserving the local curvature features and edge details of the component surface to the maximum extent.

[0032] For example, the initial size is 3×3, the maximum window size is limited to 15×15, the fixed step size is 2, and the number threshold is 10. For instance, if the number of effective pixels in the 3×3 search window is lower than the preset number threshold, then a 5×5 search window is used; if the number of effective pixels in the 5×5 search window is lower than the preset number threshold, then a 7×7 search window is used, and so on.

[0033] S2. Obtain the surface stability index of the candidate point based on the difference in normal vector between the candidate point and its nearest neighbor and the eigenvalue corresponding to the candidate point.

[0034] It should be noted that the gripping stability of a robot suction cup or gripper is highly dependent on the geometric characteristics of the contact surface. Therefore, this invention obtains the surface stability index of a candidate point based on the difference in the normal vector between the candidate point and its nearest neighbor and the eigenvalue corresponding to the candidate point.

[0035] Specifically, the repaired point cloud set is voxelized and downsampled to obtain a candidate point set. For each candidate point in the candidate point set, its nearest neighbor is searched within its neighborhood. A neighborhood covariance matrix is ​​constructed using the X, Y, and Z coordinates of the nearest neighbor of the candidate point, and eigenvalue decomposition is performed to obtain three eigenvalues. Based on the eigenvalues ​​corresponding to the candidate point and the average deviation of the candidate point from its nearest neighbor in the normal vector, the surface stability index of the candidate point is obtained.

[0036] In one embodiment, the neighborhood range of the candidate point is a range with a radius of 5mm centered on the candidate point, and the 30 nearest candidate points within the neighborhood range of the candidate point are the nearest neighbors.

[0037] Specifically, the surface stability index satisfies the following relationship:

[0038] ;

[0039] In the formula, For the first The surface stability index of each candidate point For the first The normal vector of each candidate point For the first The first candidate point The normal vector of the nearest neighbor, The number of nearest neighbors. , and The first The three feature values ​​corresponding to each candidate point, among which , It is an inverse cosine function. The sign for the vector modulo operation. It is an exponential function with the natural constant as the base.

[0040] in, The surface flatness factor represents the candidate point. When the candidate point corresponds to a spherical region or noise, its three eigenvalues ​​will be very close, making the sum of the three eigenvalues ​​close to 0. To avoid spherical areas or noise being identified as grab points, by Construct a surface stability index when the sum of the three eigenvalues ​​is close to When the surface flatness factor of a candidate point is smaller, the surface stability index of the candidate point is smaller; when the sum of the three eigenvalues ​​is greater than or equal to the surface flatness factor of the candidate point, the surface stability index of the candidate point is smaller. The greater the difference between the two, the larger the surface flatness factor of the candidate point, and the greater the surface stability index of the candidate point.

[0041] The normal consistency factor represents the candidate point, where Let be the angle between the normal vector of a candidate point and the normal vector of its nearest neighbor. When the candidate point and its nearest neighbor lie on a smoother plane, the normal vectors of the two points are closer to being parallel, and the smaller the angle between the two normal vectors is. As a result, the normal uniformity factor of the candidate point is larger, and the surface stability index of the candidate point is larger. When the surface where the candidate point and its nearest neighbor lie is more irregular, the angle between the normal vectors of the two points will be larger, and the normal uniformity factor of the candidate point will be smaller, and the surface stability index of the candidate point will be smaller.

[0042] S3. Perform candidate point screening to obtain valid candidate points.

[0043] It should be noted that during vehicle assembly, the gripping points must not only be flat but also safe. If the gripping point is too close to the edge of the component, the component may slip when the robotic arm accelerates; if there are obstacles around the gripping point, a collision may occur. Therefore, this invention uses a physical constraint filter to perform collision detection based on the distance from the candidate point to the point cloud boundary, eliminating points that, although flat, are located in the edge danger zone or pose a collision risk, thus ensuring the safety of the gripping operation.

[0044] Specifically, the edge point set of the component is identified using the maximum angle criterion. For each candidate point, the Euclidean distance between it and the nearest edge point is taken as the edge distance. If the edge distance is less than a preset safety threshold, the candidate point is discarded. Next, a cylindrical virtual space is established centered on the retained candidate points along their normal vector direction, and the presence of non-target point clouds within this space is detected. If they exist, the candidate points are discarded, and the remaining candidate points are considered valid candidate points. For example, if the safety threshold is 20mm, the radius of the cylindrical virtual space is 10mm, and the axial height is 50mm.

[0045] In one embodiment, the non-target point cloud is point cloud data whose distance from the central axis of the cylindrical virtual space is less than a preset safety radius, wherein the preset safety radius is half the radius of the cylindrical virtual space.

[0046] S4. Construct a crawling suitability based on point cloud density and edge distance metrics.

[0047] It should be noted that in the scenario of whole vehicle assembly, data sparsity or artifact noise is easily generated for highly reflective parts such as paint sheet metal parts or glass. Relying solely on stability screening may lead to the robot attempting to grasp unreliable areas. Therefore, this invention constructs grasping suitability based on point cloud density and edge distance indicators.

[0048] Specifically, the point cloud obtained through effective pixels is the effective point cloud, and the number of effective point clouds in the neighborhood of each effective candidate point is the effective point cloud density of the effective candidate point. The grasping suitability of the effective candidate point is obtained based on the surface stability index, edge distance, and effective point cloud density of the effective candidate point. The neighborhood of the effective candidate point is a spherical range with a radius of 10 mm centered on the effective candidate point.

[0049] Specifically, the crawl suitability satisfies the following relationship:

[0050] ;

[0051] In the formula, For the first The suitability of capturing each valid candidate point For the first The surface stability index of the effective candidate points For the first The edge distance of each valid candidate point For the first Effective point cloud density of each valid candidate point As a safety threshold, To achieve the maximum effective point cloud density, It is the hyperbolic tangent function. As a sensitivity factor, in this embodiment The value is 0.5, and the implementers can adjust it according to the actual situation. The value of .

[0052] in, The safety confidence factor represents the number of valid candidate points. The relative edge safety distance of valid candidate points; the farther the candidate point is from the edge of the component, the more likely it is to become safer. The larger the value, the lower the risk of the robotic arm's end effector slipping or colliding with the surrounding environment during the grasping process, thus resulting in a higher safety confidence factor. To determine the relative data confidence level, for highly reflective components in vehicle assembly, the effective point cloud density within the neighborhood of a candidate point is... A larger value indicates fewer holes or artifacts caused by reflected light interference in the area, resulting in more complete and reliable point cloud data, thus increasing the security confidence factor. The hyperbolic tangent function is then used. Utilizing its saturation characteristic, when a valid candidate point simultaneously satisfies the conditions of being far from the edge and having dense data, the safety confidence factor approaches 1 and no longer increases indefinitely; conversely, when a valid candidate point is close to the edge or in a sparse reflective region causing either index to be too low, the safety confidence factor drops rapidly. Furthermore, through the surface stability index... Multiplying the crawling suitability by a security confidence factor ensures that only regions that are geometrically stable, physically secure, and have reliable data can achieve a high crawling suitability, thereby enabling accurate crawling.

[0053] For example, Figure 2 and Figure 3 These are oblique side views and top views of the point cloud data of automotive parts in this invention. Figure 4 The top-down view of the gripping suitability distribution shows that the upper right corner of the part has a greater gripping suitability, while the other unsuitable gripping positions have a smaller gripping suitability, ensuring the stability of the robotic arm when performing gripping tasks.

[0054] S5, coordinate mapping and precise capture execution.

[0055] Specifically, the grasping suitability of all valid candidate points is sorted, and the point with the highest grasping suitability is selected as the optimal grasping point. A pre-calibrated hand-eye calibration matrix is ​​read, and the optimal grasping point position and normal vector in the visual coordinate system are converted into the target position and target approach vector in the robot coordinate system. Then, robot control commands are generated to strictly align the robot's end effector attitude axis with the target approach vector and quickly move it above the target position. Finally, the robot is controlled to approach the target position from above the optimal grasping point using low-speed linear interpolation along the normal direction until it reaches the target position and activates the grasping mechanism to complete the grasping of the part.

Claims

1. A precise grasping method for a vehicle assembly robot based on binocular stereo vision, characterized in that, include: The system controls a stereo camera to acquire images of the parts to be grasped, generating an original depth map. The region of interest is then cropped, and an adaptive window is used to perform distance-weighted interpolation to fill the original depth map, resulting in a repaired depth map. A point cloud set is then obtained from the repaired depth map. Candidate points are obtained by downsampling the point cloud set. The covariance matrix of the nearest neighbor coordinates of each candidate point is constructed and decomposed to obtain three eigenvalues. Based on the three eigenvalues ​​and the angle between the candidate point and the normal vector of its nearest neighbor, the surface stability index of the candidate point is obtained, satisfying the following relationship: , For the first The surface stability index of each candidate point For the first The normal vector of each candidate point For the first The first candidate point The normal vector of the nearest neighbor, The number of nearest neighbors. , and The first The three feature values ​​corresponding to each candidate point, among which , It is an inverse cosine function. The sign for the vector modulo operation. It is an exponential function with the natural constant as its base; Calculate the edge distance between candidate points and component edges, discard candidate points with edge distances less than a safety threshold, and establish a cylindrical virtual space along the normal vector direction for collision detection. Candidate points that pass the detection are considered valid candidate points. Calculate the effective point cloud density within the neighborhood of each valid candidate point. Based on the surface stability index, edge distance, and effective point cloud density, determine the grasping suitability, satisfying the following relationship: , For the first The suitability of capturing each valid candidate point For the first The surface stability index of the effective candidate points For the first The edge distance of each valid candidate point For the first Effective point cloud density of each valid candidate point As a safety threshold, To achieve the maximum effective point cloud density, It is the hyperbolic tangent function. Sensitivity factor; Select the effective candidate point with the highest grasping suitability as the optimal grasping point, map its coordinates to the robot coordinate system, and control the robot end effector to approach the optimal grasping point and grasp the part.

2. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, The step of using an adaptive window to perform distance-weighted interpolation filling on the original depth map includes: Pixels with depth values ​​of 0 and infinity within the region of interest are considered invalid pixels, while pixels outside the invalid pixels are considered valid pixels. Any invalid pixel is designated as the target pixel, and a search window of initial size is established centered on the target pixel. The number of valid pixels within the search window is counted. If the number of valid pixels within the search window is lower than a preset threshold, the search window size is increased incrementally with a fixed step size until the number of valid pixels within the window exceeds the threshold or reaches a preset maximum window limit. Based on the finally determined search window, distance-weighted interpolation is performed using the depth values ​​of all valid pixels within the search window, based on their Euclidean distance to the target pixel.

3. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, After obtaining the point cloud set, the point cloud set is processed using a statistical outlier filtering algorithm.

4. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, The nearest neighbors of a candidate point are the 30 candidate points that are closest to it.

5. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, The calculation of the edge distance between the candidate point and the edge of the component includes: identifying the set of edge points of the component using the maximum angle criterion, and for each candidate point, using the Euclidean distance between it and the nearest edge point as the edge distance.

6. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, The method of establishing a cylindrical virtual space along the normal vector direction for collision detection includes: after removing candidate points whose edge distance is less than the safety threshold, establishing a cylindrical virtual space along the normal vector direction with any remaining candidate point as the center; if there is a non-target point cloud in the space, the candidate point fails the detection and is removed; if there is no non-target point cloud in the space, the candidate point passes the detection.

7. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 6, characterized in that, The non-target point cloud refers to point cloud data whose distance from the central axis of the cylindrical virtual space is less than a preset safety radius.

8. The precise grasping method for a vehicle assembly robot based on binocular stereo vision according to claim 1, characterized in that, The process of controlling the robot's end effector to approach and grasp the component at the optimal grasping point includes: reading a pre-calibrated hand-eye calibration matrix, converting the optimal grasping point position and normal vector in the visual coordinate system into the target position and target approach vector in the robot coordinate system; subsequently generating robot control commands to align the robot's end effector attitude axis with the target approach vector and quickly move it to a position above the target position; finally, controlling the robot to approach the target position from above the optimal grasping point along the normal direction at low speed using linear interpolation until it reaches the target position and activates the grasping mechanism to complete the precise grasping of the component.

Images