Hole opening positioning method for automobile part machining

Abstract

The invention relates to the field of image processing, in particular to a trepanning positioning method for automobile part machining, and the method comprises the steps: obtaining an image of a to-be-trepanned region of a workpiece, and calibrating an ROI region; detecting key points and a main direction of the ROI image by adopting an ORB algorithm; the method comprises the following core steps: calculating the local illumination disturbance degree of each key point neighborhood image block, dynamically adjusting the normalization range according to the local illumination disturbance degree to carry out intensity remapping on the image blocks, and generating ICE-rBRIEF feature descriptors on the basis; matching the generated descriptor with a reference model; establishing a local non-uniform two-dimensional deformation field model by using the high-confidence matching points; and finally, the theoretical trepanning position is accurately compensated, and the trepanning mechanism is guided to work. According to the method, the local illumination disturbance characteristics are analyzed, the intensity normalization range of the image blocks is adjusted in a self-adaptive mode, ICE-rBRIEF feature descriptors are generated, and high-precision trepanning positioning is achieved in combination with deformation compensation.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and particularly to a hole opening positioning method for automotive part processing. Background Art

[0002] The automotive parts involved in the present invention, taking the battery tray of new energy vehicles as an example, are key structural components for carrying power battery packs. Such parts are usually made of lightweight and high-strength metals such as aluminum alloy, and have the characteristics of large size, thin wall, and relatively complex structure. During their production process, a large number of holes need to be precisely machined in specific planar areas for fixing battery modules, connecting the vehicle body, or installing other accessories. The accuracy of these hole positions has a direct impact on the assembly quality, safety performance, and endurance of the whole vehicle.

[0003] The traditional hole opening method based on special tooling fixtures can ensure accuracy to a certain extent, but it has limitations for large parts such as battery trays that are prone to non-uniform deformation due to processes such as welding and stamping. Therefore, using non-contact and highly adaptable vision-guided technology for hole opening positioning has become an important development direction.

[0004] In vision-based positioning methods, the ORB algorithm is often used for feature point extraction and description due to its good balance between speed and robustness. This algorithm first uses the FAST detector to identify key points, and then generates a binary string for each key point through the rBRIEF descriptor. However, when the ORB algorithm is applied to automotive parts with metal surfaces such as battery trays, the performance of its rBRIEF descriptor will be severely challenged. The metal surface is extremely prone to high-gloss reflection, and the lighting conditions in the industrial site are often difficult to be perfectly uniform. These factors cause the gray values of the point pairs on which rBRIEF depends for comparison to become very unstable in the image patches in the neighborhood of the key points. For the same physical feature point, when the lighting or viewing angle changes slightly, its high-gloss state may change drastically, resulting in a significant difference in its rBRIEF descriptor. This instability of the descriptor may directly lead to a decrease in the success rate of subsequent feature matching and an increase in incorrect matches, making the deformation compensation model established based on these unreliable matching results inaccurate, and ultimately unable to meet the requirements of high-precision hole opening positioning, restricting the application potential of vision technology in the field of precision machining. Summary of the Invention

[0005] Regarding the problem that the instability of the above descriptors may lead to subsequent feature matching errors, the present invention proposes an opening positioning method for automotive part processing, including: acquiring an image of the automotive part to be processed and calibrating the ROI region; detecting key points in the ROI region using the ORB algorithm and calculating their main directions; determining the rotated and aligned neighborhood image patches centered on each key point; counting the number of pixels in each neighborhood image patch whose gray values are greater than the set upper threshold and less than the set lower threshold, and the ratio of the number of pixels to the total number of pixels in the neighborhood image patch constitutes the local illumination perturbation degree; based on the local illumination perturbation degree, determining the adaptive effective normalization lower limit and upper limit for intensity normalization of the neighborhood image patch; applying the adaptive effective normalization lower limit and upper limit to the neighborhood image patch to perform intensity remapping to obtain a remapped image patch; generating a local intensity-normalized rBRIEF feature descriptor on the remapped image patch; matching the feature descriptor with the feature descriptor in the preset reference model to obtain a high-confidence matching feature point pair; using the high-confidence matching feature point pair to establish a deformation field model for characterizing the local non-uniform two-dimensional deformation between the ROI region and the preset reference model; accurately compensating the theoretical opening position according to the deformation field model to obtain the actual opening position, and guiding the opening actuator to complete the opening operation.

[0006] Compared with the prior art using the standard rBRIEF descriptor, the present invention preprocesses the neighborhood image patches of key points by adaptive intensity normalization based on the characteristics of local illumination perturbation, generates an ICE-rBRIEF feature descriptor that is more robust to highlights and uneven illumination on the metal surface, significantly improves the stability and accuracy of feature matching under harsh illumination conditions, and further improves the accuracy of the subsequent deformation compensation model and the accuracy of the final opening positioning. It is particularly suitable for automotive metal parts with complex surface reflection characteristics.

[0007] Further, the specific calculation method of the local illumination perturbation degree is as follows: ; where represents the local illumination perturbation degree of the neighborhood image patch corresponding to the key point ; represents the number of pixels in the image patch whose gray values are less than the set lower threshold; represents the number of pixels in the image patch whose gray values are greater than the set upper threshold; represents the total number of pixels in the neighborhood image patch.

[0008] Further, the specific calculation method of the adaptive effective normalization lower limit and upper limit is as follows: ; ; wherein represents the effective lower limit of adaptive normalization; represents the effective upper limit of adaptive normalization; , , are respectively the 10th, 50th, and 90th percentile gray values of the image block ; represents the local illumination disturbance degree of the neighborhood image block corresponding to the key point .

[0009] Compared with the normalization method using a fixed range or simple maximum and minimum values in the prior art, the present invention dynamically adjusts the effective range of normalization through the local illumination disturbance degree, making it contract towards the median when the illumination disturbance is severe, and can more effectively suppress the influence of extreme pixel values on the generation of subsequent rBRIEF descriptors, enhancing the robustness of the descriptors to illumination changes.

[0010] Furthermore, the calculation method of the remapped image block is specifically as follows: ; wherein represents the pixel gray value at the coordinate in the remapped image block ; represents the pixel gray value at the coordinate in the original rotation-aligned image block ; represents the effective lower limit of adaptive normalization; represents the effective upper limit of adaptive normalization; represents a very small positive number; The function represents a truncation function for restricting within the interval; represents the rounding function.

[0011] Compared with the prior art that directly generates descriptors on the original image block, the present invention effectively compresses the extreme gray values affected by highlights or shadows by linearly mapping the pixel gray values to the dynamic range determined by the adaptive effective normalization upper and lower limits, making the remapped image block have a more balanced contrast, providing a basis for generating a more stable local intensity-normalized rBRIEF feature descriptor.

[0012] Furthermore, the neighborhood image block is an image area with a set size determined centered on the key point.

[0013] Further, it also includes preprocessing the ROI region, specifically: grayscale processing the ROI region; applying adaptive histogram equalization processing to the grayscale ROI region; and applying a Gaussian smoothing filter or a median filter to the ROI region after the adaptive histogram equalization processing to denoise it.

[0014] Compared with the prior art that uses simple preprocessing such as standard Gaussian smoothing or median filtering, the present invention can more effectively process the complex lighting conditions and image noise in the industrial field through targeted background light correction and edge-preserving denoising algorithms, providing higher-quality image input for the stability and accuracy of subsequent ORB key point detection.

[0015] Further, after the ORB algorithm detects the key points, it also includes sorting and screening according to the Harris corner response values of the key points, and calculating the main direction of the key points based on the intensity centroid method.

[0016] Further, when matching the feature descriptors with the feature descriptors in the preset reference model to obtain high-confidence matching feature point pairs, it also includes: using nearest neighbor search based on the Hamming distance, and combining the nearest neighbor distance ratio test and the cross-validation strategy to screen the preliminary matching point pairs.

[0017] Further, the deformation field model is a two-dimensional thin plate spline interpolation model.

[0018] Further, guiding the hole-opening actuator to complete the hole-opening operation also includes: converting the actual hole-opening position into the three-dimensional physical coordinates in the working coordinate system of the hole-opening actuator through the pre-calibrated internal and external parameters of the camera and the hand-eye calibration relationship between the camera and the hole-opening actuator, and controlling the hole-opening actuator to operate according to the three-dimensional physical coordinates.

[0019] Compared with the prior art that only outputs image coordinates or relies on manual conversion, the present invention clarifies the complete conversion path from image coordinates to the physical coordinates of the actuator. Through precise camera calibration and hand-eye calibration, it ensures that the visual positioning result can be accurately transmitted and applied to the hole-opening operation in the physical space, which is the key link to achieve high-precision automated hole opening.

[0020] The technical effects of the present invention are: The present invention proposes a method for generating a light - robust adaptive local intensity - normalized rBRIEF feature descriptor and applies it to the hole - opening positioning of large automotive metal parts. Specifically, for the rotation - aligned neighborhood image patches of each ORB key point, the local light perturbation degree is calculated by analyzing the distribution of its gray - scale values. This index can effectively quantify the influence degree of common highlights or deep shadows on the metal surface on this local image patch. Then, based on the calculated local light perturbation degree, the effective upper and lower limits for subsequent intensity normalization are dynamically adjusted. This adjustment enables the normalization process to adaptively shrink its reference dynamic range according to the severity of local illumination, thereby suppressing the influence of extreme pixel values. Further, using the adaptively determined effective normalization range, intensity remapping is performed on the original neighborhood image patch to generate a new image patch that weakens the influence of uneven illumination such as highlights and shadows. The final rBRIEF binary descriptor is generated by performing point - pair gray - scale comparison on this specifically optimized and more light - robust image patch. This makes the descriptor more robust to the drastic changes in illumination caused by the reflection characteristics of the metal surface, significantly improving the stability and accuracy of feature matching and laying a solid foundation for subsequent high - precision deformation compensation and hole - opening positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] By reading the following detailed description with reference to the accompanying drawings, the above - mentioned and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understandable. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, where: Figure 1 FIG. is a flowchart of a hole - opening positioning method for automotive part processing according to an embodiment of the present invention; Figure 2 FIG. is a gray - scale schematic diagram of a battery tray of a new - energy vehicle according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0023] The following will describe the specific embodiments of the present invention in detail with reference to the drawings.

[0024] An embodiment of a hole - opening positioning method for automotive part processing: As Figure 1As shown in the figure, an opening positioning method for automobile part processing according to the present invention includes: S1. Acquisition of target area image and calibration of ROI area.

[0025] In the prior art, for the opening positioning of automobile parts, a special tooling fixture is often used in combination with a numerical control machine tool. However, for parts such as the lower shell of a new energy vehicle battery tray, which are large-sized thin-walled and formed by welding or complex extrusion, due to the welding deformation or forming tolerance that is difficult to fully predict and control, it is difficult to meet the high-precision requirements of all key hole positions by using the traditional rigid positioning method, and the tooling cost is high and the flexibility is poor.

[0026] In this embodiment, taking the battery tray of a new energy vehicle as an example, as Figure 2 shown, there are multiple planar areas for installing battery modules or connecting to the vehicle body on it, and the hole position accuracy requirements for these areas are relatively high. A lower shell of a battery tray that has just been welded or stamped and still needs to be drilled is fixed to the vision inspection station of the automated processing unit through a general tooling fixture. The vision system should include at least one industrial digital camera, which can be installed through a C-arm or a fixed bracket so that the optical axis of its lens is approximately perpendicular to the specific planar functional area on the workpiece where the hole is expected to be opened. Start the camera to collect the two-dimensional digital image of this planar area.

[0027] To focus computing resources and accurately define the scope of subsequent analysis, the operator can, through the human-machine interaction interface, on the workpiece image displayed in real time, with reference to the contour of the CAD two-dimensional engineering drawing of the corresponding area of this workpiece pre-loaded, interactively (for example, drag the mouse to select a box) calibrate the accurate region of interest (ROI). Or, the system can also automatically identify and frame the ROI through pre-rough positioning and template matching. Exemplary description: After the camera takes a picture of a battery tray, the operator selects a 200mm×300mm installation plane in the middle as the ROI. If the image resolution is 0.1mm / pixel, a possible situation of the ROI parameters may be: , where represents the center coordinates of the ROI area; represents the width and height of the ROI area.

[0028] S2. ROI image preprocessing and ORB feature point detection.

[0029] The acquired original ROI image data, especially the images from the industrial site, are often accompanied by interferences such as uneven illumination and random noise. To improve the quality of subsequent feature point detection, image preprocessing needs to be performed and the FAST part in the ORB algorithm is used to preliminarily detect key points.

[0030] First, convert the input color image into a single-channel grayscale image; subsequently, to reduce the impact of uneven illumination, adaptive histogram equalization can be applied to the grayscale image; then, to smooth random noise and protect edge details, a Gaussian smoothing filter or a median filter with a kernel size of, for example, 5x5 can be applied. The above preprocessing algorithms are well-known techniques and will not be elaborated here.

[0031] On the preprocessed ROI grayscale image, apply the key point detection process of the ORB algorithm standard. This process first uses the FAST (Features from Accelerated Segment Test) corner detection algorithm to quickly identify candidate corner points. Subsequently, to ensure the quality and stability of the corner points, the Harris corner response values of these candidate corner points are usually calculated, and they are sorted and filtered according to the response values. In one embodiment, the top corner points with the highest response values can be selected. In another embodiment, a response value threshold can be set, and only the corner points with response values greater than the threshold are selected. The ORB algorithm can also calculate the main direction of each selected key point based on the intensity centroid method, and this main direction is used for the rotational invariance of the subsequent feature descriptor.

[0032] S3. Extract the key point neighborhood image patches and rotate and align them; analyze the local illumination characteristics of the neighborhood image patches and calculate the normalization parameters; perform intensity remapping on the neighborhood image patches based on the adaptive effective normalization range.

[0033] After obtaining the ORB key points and their main directions in S2, it is also necessary to generate a feature descriptor that is robust to illumination changes (especially the specular reflection and shadow on the metal surface) for each key point. The standard rBRIEF descriptor directly performs point-to-point grayscale comparison on the rotated and aligned image patches around the key points. In the face of the drastic illumination changes commonly found on the metal surface, the descriptor becomes extremely unstable, seriously affecting the accuracy of subsequent matching.

[0034] In this embodiment, an illumination-robust contrast-enhanced rBRIEF (ICE-rBRIEF) descriptor is proposed. Its core idea is to perform an adaptive intensity normalization preprocessing based on the local illumination characteristics analysis on the image patches in the neighborhood of each key point before applying the binary string generation rule of the rBRIEF standard.

[0035] S3.1: Extract the key point neighborhood image patches and rotate and align them.

[0036] First, for each ORB key point output by S2, a neighborhood image patch of a fixed size is extracted from the preprocessed ROI grayscale image with the coordinates of the key point as the center. In this embodiment, the size of the image patch can be set to 31x31 pixels. Subsequently, using the principal direction information of the key point, this image patch is rotationally transformed so that its principal direction is aligned to a unified reference direction (e.g., the positive x-axis direction). This step is a standard part of the ORB algorithm and is used to ensure that the subsequent generated descriptors have rotational invariance. The neighborhood image patch that has been extracted and rotationally aligned for each key point 。

[0037] S3.2: Analysis of the local illumination characteristics of the neighborhood image patch and calculation of the normalization parameters.

[0038] Highlights or deep shadows on the metal surface can cause extreme polarization of the pixel grayscale distribution within the neighborhood image patch, making the rBRIEF descriptor based on a fixed comparison method unstable. Therefore, it is necessary to quantify the degree of this illumination perturbation. In this embodiment, for the rotationally aligned neighborhood image patch output in step S3.1 (assuming it contains pixels), calculate its local illumination perturbation degree 。

[0039] First, calculate the 10th percentile and the 90th percentile of all pixel grayscale values within the image patch . In this embodiment, can be used as the lower threshold, and can be used as the upper threshold. These two values can relatively robustly represent the main grayscale range within the image patch excluding extreme highlights and dark points. In addition to the 10th percentile and the 90th percentile, the implementer can also set them to the 15th percentile and the 85th percentile, etc. according to the actual situation or further experiments; then count the number of pixels within the image patch whose grayscale values are strictly less than , and the number of pixels whose grayscale values are strictly greater than . These pixels can be regarded as pixels more severely affected by the illumination perturbation; finally, calculate the proportion of these extreme pixels to obtain the local illumination perturbation degree , there is: ; where represents the local illumination perturbation degree of the neighborhood image patch corresponding to the key point ; represents the number of pixels within the image patch whose grayscale values are less than the lower threshold; represents the number of pixels within the image patch whose grayscale values are greater than the upper threshold; represents the total number of pixels within an image block, which is in this embodiment.

[0040] When the gray-scale distribution within the image block is relatively concentrated and there are few extremely bright or dark pixels (e.g., a diffuse reflection surface with uniform illumination), and are both small. If an image block has a large number of extremely dark or bright pixels (e.g., deep shadows and strong highlights coexist), then will have a high value, indicating that the image block is severely disturbed by illumination. That is, in the area of strong highlights or deep shadows on the surface of the battery tray, will be significantly higher; in an area with relatively uniform illumination, will be low.

[0041] According to the local illumination disturbance degree of the image block, an effective dynamic range for subsequent intensity normalization needs to be determined. If is very high, it means that the original dynamic range of the image block may have been damaged by extreme values, and directly using the original minimum and maximum values or fixed percentiles for normalization will not yield good results. The present invention adjusts the upper and lower limits for normalization dynamically through to make it contract towards the central gray-scale trend of the image block, thereby suppressing the influence of extreme values.

[0042] Let , (the median), be the 10th, 50th, and 90th percentile gray-scale values of the image block respectively, then there are: ; ; where represents the adaptive effective normalization lower limit; represents the adaptive effective normalization upper limit; represents the key point corresponding to the local illumination disturbance degree of the neighboring image block. The above formula linearly interpolates and contracts the original robust gray-scale range towards the median

[0043] When , that is, when the illumination disturbance is extremely small, and . At this time, normalization will use the original 10 - 90 percentile range; when , that is, when there is illumination disturbance, will be greater than , will be less than , both approaching , narrowing the normalized effective dynamic range. The larger it is, the greater the degree of contraction.

[0044] If reaches its theoretical maximum value (for example, assuming all pixels are below 10% or above 90%, representing extreme cases), and will both tend to , which will result in extremely low contrast of the normalized image patch (close to a constant image), making the subsequent rBRIEF point pair comparison almost random. In extreme cases, this may instead enhance the neglect of "structure" and rely more on a small number of pixels that are not fully compressed. However, a more practical effect is that by shrinking the dynamic range, pixels that were saturated due to highlights and pixels that were close to 0 due to shadows can be pulled into a more concentrated gray level interval after normalization, avoiding the generation of fixed 0 or 1 in the rBRIEF binary comparison, thus increasing the robustness of the descriptor to illumination changes.

[0045] In the highlight area of the battery tray, is high, resulting in a significant contraction of the range, suppressing highlights; in the uniformly illuminated area, is low, and this range is close to , maintaining the original contrast.

[0046] S3.3: Neighborhood Image Patch Intensity Remapping Based on Adaptive Effective Normalization Range.

[0047] Using the effective normalization lower limit and upper limit determined in S3.2, perform intensity remapping on the original neighborhood image patch obtained in S3.1 after rotation alignment to generate a new, more illumination-robust image patch . The subsequent rBRIEF descriptor will be calculated on this remapped image patch.

[0048] For each pixel in the original image patch , denote its gray value as , remap it to the new gray value , and form a new image patch .

[0049] ; where represents the coordinate in the remapped image patch The pixel gray value at Represents the original rotation-aligned image patch The coordinates in The pixel gray value at Represents the effective adaptive normalization lower limit; Represents the effective adaptive normalization upper limit; Represents a very small positive constant, which can be set to 1e-6 in this embodiment, and is used to prevent division-by-zero errors; The function represents a truncation function for Restricted to Interval; Represents the rounding function.

[0050] The above steps linearly transform the original pixel gray value From the dynamically adjusted input range Mapped to the standard output range , if the original gray value of a pixel Is lower than , then its remapped value Will be truncated to 0; if the original gray value of a pixel Is higher than , then its remapped value Will be truncated to 255, and the pixel values in between will be linearly stretched to Range.

[0051] By dynamically adjusting And (Based on ), this remapping process can effectively adjust the gray distribution of image patches severely affected by highlights or deep shadows to a more regular and contrast-controllable range, so that the subsequent rBRIEF point pair gray comparison is less susceptible to the original extreme lighting conditions.

[0052] For the image patches on the battery tray irradiated by highlights, its Is higher, resulting in And The range between them narrows and may shift upward as a whole. The normalization of the above formula will effectively suppress the brightness of the highlight area and enhance the contrast of the originally darker area. For areas with uniform lighting, Is lower, and the normalization range is close to the original 10 - 90 percentile range, and the transformation has little impact.

[0053] The standard rBRIEF descriptor selects predefined (e.g., 256) point pairs on the Gaussian-smoothed and rotation-aligned image patches , and compares their gray levels: if ， Then the bit of the descriptor is 1, otherwise it is 0. The ICE-rBRIEF descriptor of the present invention is an image block with stronger illumination robustness obtained after S3.3 intensity remapping. On this, the gray values of these 256 point pairs are and compared to generate the final 256-bit binary descriptor.

[0054] S4. Matching of ICE-rBRIEF feature descriptors.

[0055] After generating the ICE-rBRIEF descriptor with stronger illumination robustness, the next step is to match it with the descriptors in the pre-stored reference model to establish the corresponding relationship of feature points between the current workpiece ROI and the reference model.

[0056] First, it is necessary to process the ROI to be located of the ideal workpiece, and use the methods of S2 and S3 (i.e., ORB key point detection and ICE-rBRIEF descriptor generation) to extract the reference key points and their ICE-rBRIEF descriptors, and construct a reference descriptor library; then use a fast matching algorithm based on Hamming distance. For each ICE-rBRIEF descriptor in the current ROI, search for the descriptor with the smallest Hamming distance from it in the reference descriptor library as the best match; at the same time, to improve the reliability of the matching, cross-validation and / or nearest neighbor distance ratio test (Lowe's Ratio Test, that is, the Hamming distance of the best match is significantly smaller than that of the second-best match, for example, less than 0.75 times) can be used.

[0057] This step can finally output a set of geometrically verified high-confidence matching feature point pairs. Each pair contains the key point coordinates in the current ROI image and the corresponding key point coordinates in the reference model (in this embodiment, it is the CAD coordinate system) .

[0058] Exemplary illustration: Input 100 ICE-rBRIEF descriptors of the current ROI and 500 ICE-rBRIEF descriptors of the reference model. After matching and screening, 60 pairs of high-confidence matching point pairs may be obtained, such as [current point (100, 150)->reference point (50.5, 75.2)], [current point (320, 210)->reference point (160.1, 105.8)], etc.

[0059] S5. Modeling of local non-uniform two-dimensional deformation field within the ROI.

[0060] First, select a mathematical model that can fit non-rigid deformation, such as the two-dimensional thin plate spline interpolation (2D TPS) model. This model can find a smooth mapping function that minimizes the overall bending energy, and can thus accurately pass through all given pairs of matching control points. Then, by solving a system of linear equations related to the number of control points, the coefficients of the TPS model can be determined. After the model is established, for any point in the ROI in the reference coordinate system, its corresponding position in the current actual deformed workpiece image can be calculated through the TPS model. , and then the coefficients of the TPS model can be determined by solving a system of linear equations related to the number of control points. After the model is established, for any point in the ROI in the reference coordinate system, its corresponding position in the current actual deformed workpiece image can be calculated through the TPS model.

[0061] This step can output a parameterized two-dimensional non-rigid deformation field model, which is the coefficient matrix of the TPS model in this embodiment. Exemplary illustration: Based on the 60 pairs of matching points output by S4, a TPS model is calculated, and this model can be expressed as a function: ; where represents the key point coordinates in the reference model CAD coordinate system; represents the smooth mapping function.

[0062] S6. Precise compensation and guidance for the target hole opening position.

[0063] Based on the accurate local deformation field model established in S5, the theoretical hole opening position on the CAD design drawing can be accurately mapped to the current actual deformed workpiece ROI, thus guiding the hole opening actuator to perform high-precision operations.

[0064] First, record the list of theoretical two-dimensional coordinates of all target hole positions to be processed within the ROI in the workpiece CAD design coordinate system as ; then for each theoretical hole position , use the deformation field model output by S5 (such as ) for forward mapping to calculate its accurately compensated position in the current workpiece actual ROI image coordinate system, that is: ; further, convert the actual hole opening position in the calculated image coordinate system into the three-dimensional physical coordinates in the hole opening actuator (such as the end effector coordinate system of an industrial robot or the workbench coordinate system of a numerical control machine tool) through the pre-calibrated internal and external camera parameters and the hand-eye calibration relationship (if the hole opening tool is held by a robot). For hole openings perpendicular to the ROI plane, the Z-axis coordinate is usually determined by the workpiece surface height and the tool length. Finally, the control system drives the hole opening actuator (such as a drilling spindle, a laser head, or a punching module) to accurately move to the calculated three-dimensional physical coordinates and perform the hole opening operation.

[0065] Exemplary illustration: If there is a hole position at (100.0, 200.0) mm in CAD. After the TPS model transformation in S5 is performed, the position in the current image is obtained and then converted into the actual machining target point in the robot coordinate system, such as (X = 350.5, Y = 520.1, Z = -150.0) mm (Z is the machining depth). Then the robot moves to this position to complete the drilling.

Claims

1. An opening positioning method for automobile part processing, characterized in that, The method includes: obtaining an image of an automotive part to be processed and calibrating the ROI region; detecting key points in the ROI region using the ORB algorithm and calculating their main directions; determining neighborhood image patches after rotation alignment centered on each key point; counting the number of pixels in each neighborhood image patch whose gray values are greater than a set upper threshold and less than a set lower threshold, and the ratio of the number of pixels to the total number of pixels in the neighborhood image patch constitutes the local illumination disturbance degree; Based on the local illumination disturbance degree, determining an adaptive effective normalization lower limit and upper limit for intensity normalization of the neighborhood image patch; applying the adaptive effective normalization lower limit and upper limit to the neighborhood image patch to perform intensity remapping to obtain a remapped image patch; generating a local intensity-normalized rBRIEF feature descriptor on the remapped image patch; Matching the feature descriptor with the feature descriptor in a preset reference model to obtain a high-confidence matching feature point pair; using the high-confidence matching feature point pair to establish a deformation field model for characterizing the local non-uniform two-dimensional deformation between the ROI region and the preset reference model; accurately compensating the theoretical hole-opening position according to the deformation field model to obtain the actual hole-opening position, and guiding the hole-opening actuator to complete the hole-opening operation.

2. The hole opening positioning method for automotive part processing according to claim 1, characterized in that, The specific calculation method of the local illumination disturbance degree is: ； Among them represents the key point corresponding to the local illumination perturbation degree of the neighborhood image block; represents the number of pixels with gray values less than the set lower threshold within the image block; represents the number of pixels with gray values greater than the set upper threshold within the image block; represents the total number of pixels within the neighborhood image block.

3. The hole opening positioning method for automobile part processing according to claim 2, wherein, The specific calculation method of the adaptive effective normalization lower limit and upper limit is: ； ； wherein represents the effective normalization lower limit of self - adaptation; represents the effective normalization upper limit of self - adaptation; , , are respectively the 10th, 50th, and 90th percentile gray - scale values of the image block ; represents the local illumination disturbance degree of the neighborhood image block corresponding to the key point .

4. A hole opening positioning method for automotive part processing according to claim 3, characterized in that, The specific calculation method of the remapped image patch is: ； Among them represents the pixel gray value at the coordinate in the remapped image block ; represents the pixel gray value at the coordinate in the original rotation-aligned image block ; represents the adaptive effective normalization lower limit; represents the adaptive effective normalization upper limit; represents a very small positive number; The function represents a truncation function used to limit within the range of ; represents the rounding function.

5. A hole opening positioning method for automobile part processing according to claim 1, characterized in that, The neighborhood image patch is an image region of a set size determined centered on the key point.

6. A hole opening and positioning method for automobile part processing according to claim 1, characterized in that, It also includes preprocessing the ROI region, specifically: Performing grayscale processing on the ROI region; Applying adaptive histogram equalization processing to the grayscale ROI region; Applying a Gaussian smoothing filter or a median filter to the ROI region after adaptive histogram equalization processing for denoising.

7. A hole opening positioning method for automotive part processing according to claim 1, characterized in that After detecting key points using the ORB algorithm, it also includes sorting and screening according to the Harris corner response values of the key points, and calculating the main direction of the key points based on the intensity centroid method.

8. A hole opening positioning method for automotive part processing according to claim 1, characterized in that, Matching the feature descriptor with the feature descriptor in a preset reference model to obtain a high-confidence matching feature point pair includes: Adopting nearest neighbor search based on the Hamming distance, and combining the nearest neighbor distance ratio test and the cross-validation strategy to perform preliminary screening of the matching point pairs.

9. A hole opening positioning method for automobile part processing according to claim 1, characterized in that, The deformation field model is a two-dimensional thin plate spline interpolation model.

10. A hole opening positioning method for automotive part processing according to claim 1, characterized in that, Guiding the hole-opening actuator to complete the hole-opening operation includes: Converting the actual hole-opening position into the three-dimensional physical coordinates in the working coordinate system of the hole-opening actuator through the pre-calibrated internal and external parameters of the camera and the hand-eye calibration relationship between the camera and the hole-opening actuator, and controlling the hole-opening actuator to perform operations according to the three-dimensional physical coordinates.

Images