A method for object surface detection based on 3D reconstruction using a robotic arm hand-eye system

Abstract

This application discloses a method for surface detection of objects based on 3D reconstruction using a robotic arm hand-eye system. The method includes: a robotic arm, a 2D vision system, a 3D scanning system, and a system connector. The 2D vision system comprises a high-resolution camera, a liquid lens, and a ring-polarized light source, and is connected to the robotic arm via the system connector. The 2D vision system and the 3D scanning system are located on opposite sides of the system connector. The advantages of this invention are: by utilizing the hardware of the robotic arm hand-eye system and the special design of the vision system, it effectively reduces the problems of reflection and defocusing in automotive aluminum castings. By using 3D scanning technology to perform 3D reconstruction and optimal viewing angle prediction on automotive aluminum castings, flexible detection of complex surface objects of different sizes can be achieved. Through a software system equipped with an AI model and database, training can begin with a small number of normal samples, and the defect distribution results can be displayed to the user, accelerating the deployment time of the AI ​​model and simplifying the data labeling process.

Description

Technical Field

[0001] This invention belongs to the field of automotive component inspection and relates to a method for surface inspection of objects based on three-dimensional reconstruction using a robotic arm hand-eye system. Background Technology

[0002] With the continuous increase in labor costs and the rapid development of automated production lines, industrial automated visual defect inspection has been widely applied in various industries, such as mobile phone cover glass and LCD / OLED screen inspection. Industrial automated visual defect inspection boasts advantages such as high speed, high accuracy, good repeatability, easy expansion, and easy deployment, greatly improving production efficiency and reducing production costs. However, in the automotive parts manufacturing industry, the inspection of automotive aluminum castings still relies on manual inspection by workers. Automotive aluminum castings are produced by injecting smelted liquid aluminum into pre-prepared molds using methods such as pouring, injection, and suction. After cooling, they undergo subsequent processing to obtain parts with specific functional dimensions. Affected by factors such as temperature, filling speed, injection pressure, and mold wear, common defects in automotive aluminum castings include cold shuts, excess material, and film stretching.

[0003] Currently, the inspection of automotive aluminum castings relies on manual labor, primarily due to the following reasons: Automotive aluminum castings are diverse in type, size, and shape, exhibiting characteristics of low-volume, low-mix production. Current industrial automated vision defect inspection systems struggle to inspect such a wide variety of automotive parts. The complex structure of automotive aluminum castings makes data acquisition angle and position settings complex, typically requiring experienced professionals. Covering the entire surface of the aluminum casting is also challenging. This structural complexity leads to defocusing issues, making it difficult to simultaneously focus at different depths within the casting. Finally, the reflective properties of the surface material of automotive aluminum castings result in incomplete loss of defect information.

[0004] Patent document CN114418998A discloses a method, system, and electronic device for detecting defects in vehicle parts. The method includes: segmenting and recognizing a target vehicle body image to obtain segmented and recognized vehicle body images; identifying and extracting various parts of the vehicle body based on the segmented and recognized vehicle body images to obtain images of the extracted parts; detecting defects in the part images to obtain defect classification attributes; forming a multi-level attribute model of the vehicle body based on the defect classification attributes; and annotating component defect information in a 3D model of the target vehicle body based on the multi-level attribute model. This technology uses a mobile phone to collect data around the vehicle, resulting in low repeatability of the collection locations. It employs supervised deep learning for defect detection and vehicle body component recognition, requiring significant time for annotation. Annotating defects on a 3D model does not guarantee detection accuracy. Furthermore, it only obtains defect classification attributes, lacking a clear and complete understanding of defect distribution.

[0005] Patent document CN114769159A discloses a deep learning-based system and method for detecting surface defects in automotive parts. The system includes a control module, an image preprocessing module, an image recognition module, and a 3D stereo recognition module. The image recognition module consists of candidate area recognition and omnidirectional defect recognition. The image preprocessing module includes an image feature extraction module and an image feature recognition camera module. This invention, based on a deep model-based vehicle component detection strategy, maintains the reliability of automotive component detection. By simultaneously using image recognition and 3D stereo recognition, it effectively improves the system's effectiveness and reliability, ensuring effective detection of surface defects in automotive parts and guaranteeing detection results. While this technology is designed for external automotive components, this invention focuses on the detection of internal aluminum castings. The surface materials and component sizes differ. One method uses a conveyor belt mode with a fixed 3D scanner position. This technology uses a fixed coordinate data input method for 3D scanning, requiring experienced engineers to assist with input. This technology is based on 3D recognition and detection, performing deep learning by comparing the 3D scan results with the 3D input. Another method uses laser scanning for 3D scanning, estimating the total number of defects from the number of local defects, resulting in low detection accuracy and limited adaptability.

[0006] Patent document CN208536774U discloses a three-dimensional vision inspection camera system for automotive parts, including an injection molding machine, a six-axis robot, a three-dimensional vision camera, and a conveyor belt. The six-axis robot is positioned at the discharge port of the injection molding machine, and the three-dimensional vision camera is mounted on a column opposite the robot. The conveyor belt is inclinedly mounted on one side of the injection molding machine. An external light source is mounted on top of the three-dimensional vision camera, comprising a hollow rotating stage, LED lights, a power supply box, positive and negative electrode springs, a mounting ring, and contacts. The hollow rotating stage includes a fixed base, a second support, and a turntable. This invention controls the white, blue, and red LEDs on the external light source to be in the same position for three-dimensional vision inspection, preventing angular deviations from causing errors in the detected object and greatly improving the accuracy of the image capture. However, this technology, designed for automotive door handle inspection, uses a fixed light source and a three-dimensional camera. The object being inspected is fixed to the tail end of the six-axis robot, and the use of red, blue, and white LED light sources cannot effectively solve the defocusing problem, resulting in low inspection accuracy.

[0007] Furthermore, in terms of defect detection software, traditional image processing software in industrial production typically uses feature extraction and template matching methods. It selects a "golden sample" as a comparison object, and the sample being inspected is compared to this golden sample. Significant differences (exceeding a preset threshold) are identified as defects. However, automotive aluminum castings have complex surfaces and textured characteristics. Using traditional template matching methods greatly increases the over-detection rate, significantly impacting inspection efficiency. Recently, many visual defect detection systems have begun to adopt deep learning methods. The deep learning-based defect detection process can be summarized as image data acquisition, manual labeling, model training, and deployment and validation. These steps need to be repeated every time a product is changed. Image data acquisition and manual labeling are the most time-consuming. Since there are relatively few defective products on the production line, acquiring image data requires a significant amount of time to collect images of different types of defective products. Model training relies heavily on image labeling, which directly affects the accuracy of model training. To ensure consistent image labeling, training workers to use labeling tools and understand labeling criteria also consumes considerable time. While deep learning methods offer a high degree of automation, they require extensive preliminary work, and deploying them on production lines and in terms of manpower and resources is time-consuming and costly. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method for object surface detection based on three-dimensional reconstruction of a robotic arm hand-eye system, so as to solve the above-mentioned technical problems.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for detecting object surfaces based on three-dimensional reconstruction of a robotic arm hand-eye system, characterized in that it includes: providing a robotic arm, a 2D vision system, a 3D scanning system and a system connector, wherein the 2D vision system includes a high-resolution camera, a liquid lens and a ring-polarized light source, and is connected to the robotic arm through the system connector, and the 2D vision system and the 3D scanning system are located on the left and right sides of the system connector, respectively.

[0010] The robotic arm transports the car parts in the waiting area to the 3D camera area. The manufacturing data of the car parts is scanned by the 3D scanning system control module to obtain the 3D point cloud of the car parts.

[0011] Users control the robotic arm through the front-end web interface to move the 2D vision system to the optimal viewing angle, automatically or manually acquire sample images, complete the 3D reconstruction of car parts, and obtain the result of the optimal camera position.

[0012] The AI ​​model is continuously trained using the collected sample images. The defect distribution map of the surface defects of the car parts is obtained by calculation formula, the surface defects of the car parts in the waiting area are identified, and the defective car parts are identified and located.

[0013] Furthermore, the manufacturing data of the input automotive parts is scanned by the 3D scanning system control module to obtain a three-dimensional point cloud of the automotive parts, including the step of merging the three-dimensional point cloud.

[0014] Furthermore, it also includes: using the Voxel grid downsampling method to deduplicate and reduce noise in the point cloud, converting the point cloud into a mesh representation.

[0015] Furthermore, the three-dimensional reconstruction of the vehicle component includes: using a neural network with an autoencoder structure to encode the features of the vehicle component.

[0016] Furthermore, the user controls the robotic arm through the interface to move the 2D vision system to the optimal viewing angle, including: combining the pre-encoded position of the camera, using a decoder composed of MLPs to predict the probability that the camera belongs to the optimal viewing angle, and selecting the position with the highest probability as the optimal viewing angle.

[0017] Furthermore, it also includes: establishing a database to store the inspection data of all automotive parts during the inspection process.

[0018] Furthermore, the defects in the defect distribution map are divided into regions.

[0019] Furthermore, the AI ​​model is retrained using data from the database, and the new model will be updated into the software system after training is complete.

[0020] The beneficial effects of this invention are as follows:

[0021] First, this invention focuses on the inspection of aluminum castings inside automobiles. It is applicable to various sizes and shapes of aluminum castings for automobile interiors. This invention fixes a 3D scanner and a high-resolution 2D camera to the tail end of a six-axis robot via a connector, allowing the hand-eye system to be flexibly applied to aluminum castings of various sizes and shapes. It offers high repeatability of data acquisition positions, enabling automatic data collection and making it suitable for mass production. This invention uses a ring-shaped white polarized light source, effectively reducing glare on the surface of the aluminum castings. Furthermore, this technology is equipped with a liquid lens, effectively solving the defocusing problem for objects of different sizes.

[0022] Secondly, this invention uses structured light 3D scanning on the object being measured, resulting in high reconstruction accuracy. Based on the 3D scanning results, this invention automatically finds the optimal shooting angle, saving time and labor costs.

[0023] Third, this invention employs an unsupervised learning method, allowing training and display of defect distribution results with only a small number of unlabeled normal samples, reducing annotation time. Simultaneously, unsupervised learning can detect unknown defects. This invention uses an industrial-grade high-resolution 2D camera to acquire 2D image data for surface defect detection, achieving high accuracy. This invention acquires a defect distribution map, displaying the location of defects with pixel-level precision, effectively, intuitively, and completely displaying the defect distribution. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 This is a flowchart of a method for object surface detection based on three-dimensional reconstruction using a robotic arm hand-eye system, according to the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0028] In the accompanying drawings of this invention, the same or similar reference numerals correspond to the same or similar components. In the description of this invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" appear to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0029] In the description of this patent, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0030] Please refer to Figure 1 The present invention provides a method for object surface detection based on three-dimensional reconstruction using a robotic arm hand-eye system, comprising:

[0031] Step S101: Provide a robotic arm, a 2D vision system, a 3D scanning system and a system connector. The 2D vision system includes a high-resolution camera, a liquid lens and a ring-polarized light source. It is connected to the robotic arm through the system connector. The 2D vision system and the 3D scanning system are located on the left and right sides of the system connector, respectively.

[0032] In step S102, the robotic arm transports the car parts in the waiting area to the 3D camera area. The car parts are then transported to the 3D camera area, where the manufacturing data of the car parts is scanned by the 3D scanning system control module to obtain the 3D point cloud of the car parts.

[0033] In step S103, the user controls the robotic arm through the front-end web interface to move the 2D vision system to the optimal viewing angle, automatically or manually acquire sample images, complete the 3D reconstruction of the car parts, and obtain the result of the optimal camera position.

[0034] Step S104: The AI ​​model is continuously trained using the collected sample images. The defect distribution map of the surface defects of the car parts is obtained through the calculation formula. The surface defects of the car parts in the waiting area are identified, and the defective car parts are identified and located.

[0035] By implementing the above steps, utilizing the hardware of the robotic arm's hand-eye system and the special design of the vision system, the problems of glare and defocusing in automotive aluminum castings are effectively reduced. By employing 3D scanning technology to perform three-dimensional reconstruction and optimal viewing angle prediction of automotive aluminum castings, flexible detection of complex surface objects of different sizes is achieved. Through a software system equipped with AI models and databases, training can begin with a small number of normal samples, displaying defect distribution results to users, accelerating AI model deployment time and simplifying the data labeling process. This invention not only brings automation and intelligent solutions to the automotive parts industry, increasing production capacity, but it can also be used for the detection of complex objects in other industries.

[0036] The platform of this invention, based on a 3D scanning camera and reconstruction algorithm, can quickly reconstruct unknown and complex structural objects and predict the optimal camera position for detection. The platform has the following core functional modules:

[0037] In one embodiment, the robotic arm hand-eye system utilizes a six-axis robotic arm equipped with an industrial-grade 3D structured light depth camera and a 2D high-definition camera to form a hand-eye system, enabling clear imaging of the object to be inspected from all directions and multiple angles. Considering that different objects have different sizes, structures, and materials, and that reflections, distances that are too close or too far can cause camera defocusing, the system's camera lens is equipped with a ring-polarized light source and a liquid lens to achieve dynamic focusing capabilities and suppress the influence of surface reflections such as those from metals, effectively reducing data interference and allowing for flexible application to the inspection of objects of different sizes and variable working distances.

[0038] In one embodiment, the 3D reconstruction module: The platform uses a robotic arm equipped with a 3D scanning camera to move around the object and take pictures from multiple angles, merges the captured 3D point clouds, and uses the Voxel grid downsampling method to deduplicate and reduce noise in the point clouds. Finally, the point cloud is converted into a mesh representation, which ensures reconstruction accuracy while further reducing the file size of the reconstructed 3D object and improving the system running speed.

[0039] In one embodiment, the optimal viewpoint automatic prediction module utilizes an autoencoder neural network to encode features of 3D objects and predict camera position probabilities, aiming to achieve maximum surface coverage with as few camera positions as possible. The feature encoding network extracts implicit representations of visible object point clouds, then combines these with pre-encoded camera positions, and uses a decoder composed of MLPs to predict the probability that the camera belongs to the optimal viewpoint, selecting the position with the highest probability as the optimal viewpoint.

[0040] In one embodiment, the system software includes a front-end web interface through which users can control the trajectory and target position of the robotic arm, automatically or manually reconstructing the 3D of the object and obtaining the optimal camera position. Simultaneously, the system uses an SQLite database to efficiently store data from all inspected objects during the production process. The system also provides a defect detection AI algorithm; training can begin with a small number of unlabeled normal samples, displaying defect distribution results to the user. Users can continuously acquire images during practical applications and use the new images to continuously train the AI ​​model, enabling online continuous image acquisition, online updates to the AI ​​model, and online detection.

[0041] As a specific embodiment, the present invention provides a method for surface detection of objects based on 3D reconstruction using a robotic arm hand-eye system. It includes: a hardware device, namely the robotic arm hand-eye system, comprising a robotic arm, a 2D vision system, a 3D scanning system, and a system connector. The 2D vision system includes a high-resolution camera and lens, a liquid lens, and a ring-polarized light source. The vision system is connected to the robotic arm via the system connector, with the 2D vision system and 3D scanning system located on the left and right sides of the connector, respectively. The software system includes a user interface, a robotic arm control module, a 2D vision system control module, a 3D scanning system control module, a 3D reconstruction module, an optimal viewing angle automatic prediction module, an AI algorithm module, and a database system. This invention utilizes a robotic arm hand-eye system, which offers high repeatability of data acquisition positions and can automatically collect data, making it suitable for mass production. This invention employs an unsupervised learning method, allowing training and display of defect distribution results with only a small number of unlabeled normal samples, reducing annotation time. Simultaneously, unsupervised learning can detect unknown defects. This invention uses an industrial-grade high-resolution 2D camera to acquire 2D image data for surface defect detection, achieving high detection accuracy. This invention obtains a defect distribution map, displaying the location of defects with pixel-level precision, effectively, intuitively, and completely showing the defect distribution.

[0042] First, the robotic arm moves to the starting position and uses a 3D scanning system to perform a 3D scan on the object under test, obtaining a 3D point cloud of the object. The captured 3D point cloud is then merged, and the point cloud is deduplicated and denoised using the Voxel grid downsampling method, converting the point cloud into a mesh.

[0043] According to one aspect of the present invention, specifically, the Voxel grid down sample first requires setting the size of the voxel. Typically, the voxel is a cube; in this method, the side length of the voxel is set to 3mm. After obtaining the object's 3D point cloud, the maximum and minimum values ​​of the 3D point cloud on the X, Y, and Z axes are found and marked as (X...). min X max Y min Y max Z min Z max Based on these values, the bounding box of the point cloud is found. Then, according to the set voxel cube length, the bounding box is discretized into a combination of multiple voxel blocks at equal intervals. The object point cloud will fall into different voxel blocks. Within each voxel block, the point cloud closest to the center point of the voxel block is found and retained, while all other point clouds are deleted. After traversing all voxel blocks within the bounding box, the remaining point cloud is the point cloud after Voxel grid downsampling.

[0044] Secondly, a defect distribution map is obtained through an AI algorithm module. In a specific embodiment, a small amount of 2D image data of normal samples is first collected, and an initial AI model is trained through feature extraction and unsupervised learning. The AI ​​model is a set of feature distributions of normal samples. When detecting new objects, new 2D image data is input, and through feature extraction and comparison with the initial AI model, a defect distribution map is output.

[0045] Then, by combining the already scanned locations and the current scanned location with the 3D feature encoding vector and the 2D effective area, the next optimal viewpoint is optimized and predicted through an MLP network.

[0046] According to one aspect of an embodiment of the present invention, the inputs to the scanned positions and the current scanned position are the coordinates of a six-axis robotic arm, a 3D point cloud feature encoding vector, and a 2D image of the corresponding position. The loss optimization function is the coverage of the measured object and the two-dimensional effective area. An MLP network is trained, and the output is the coordinates of the six-axis robotic arm and the probability of the optimal viewing angle. The coordinates with the highest probability are selected as the optimal viewing angle. The coverage of the measured object refers to the ratio of the 3D point cloud obtained at a specific scanned position to the complete 3D point cloud of the measured object. The two-dimensional effective area refers to the ratio of the number of pixels without reflective parts in the 2D image data collected by the 2D high-resolution camera to the total number of pixels in the original 2D image data.

[0047] If the inspection is not completed, the robotic arm moves to the next predicted optimal viewpoint and repeats the above steps.

[0048] Once the inspection is complete, the original image and defect distribution map collected by the hardware device will be displayed on the user interface. Defects will be divided into regions and displayed in the format of "image name_defect region number". Users only need to confirm whether the defects displayed by the AI ​​algorithm are real defects. If the defect is misjudged by the AI ​​system and is not a real defect, the user can select "no".

[0049] According to one aspect of the present invention, supervised deep learning methods require the annotation of a large number of images before model training can begin, resulting in a long implementation cycle. An unsupervised initial AI model is trained using a small number of normal samples. Then, for detecting new objects, the 2D image of the new object is input into the initial AI model to obtain a defect distribution. Defects are categorized and numbered by region and displayed on the user interface. If a defect is a genuine defect, no further action is required from the user; if the AI ​​model misidentifies a defect, the user simply selects "No" on the user interface. This simplifies the data annotation process. The image data and defect results confirmed by the user are stored in a database, along with historical images and defect data, serving as input images for continuous AI model training, reducing the false positive rate and improving accuracy.

[0050] Finally, after user confirmation, the image data will be stored in SQLite, completing a single detection process.

[0051] According to one aspect of the present invention, training can begin with a small number of unlabeled normal samples, and the defect distribution results can be displayed to the user without spending time collecting labeled defective data. After each 50-100 detection cycles, the AI ​​model will be retrained using data from the database, and the new model will be updated to the software system after training. Users can continuously collect images during practical applications to improve accuracy. Users only need to select "no" for data that the AI ​​misclassifies to complete the data labeling process, simplifying the data labeling process and efficiently collecting data.

[0052] The beneficial effects of this invention are as follows:

[0053] First, this invention focuses on the inspection of aluminum castings inside automobiles. It is applicable to various sizes and shapes of aluminum castings for automobile interiors. This invention fixes a 3D scanner and a high-resolution 2D camera to the tail end of a six-axis robot via a connector, allowing the hand-eye system to be flexibly applied to aluminum castings of various sizes and shapes. It offers high repeatability of data acquisition positions, enabling automatic data collection and making it suitable for mass production. This invention uses a ring-shaped white polarized light source, effectively reducing glare on the surface of the aluminum castings. Furthermore, this technology is equipped with a liquid lens, effectively solving the defocusing problem for objects of different sizes.

[0054] Secondly, this invention uses structured light 3D scanning on the object being measured, resulting in high reconstruction accuracy. Based on the 3D scanning results, this invention automatically finds the optimal shooting angle, saving time and labor costs.

[0055] Third, this invention employs an unsupervised learning method, allowing training and display of defect distribution results with only a small number of unlabeled normal samples, reducing annotation time. Simultaneously, unsupervised learning can detect unknown defects. This invention uses an industrial-grade high-resolution 2D camera to acquire 2D image data for surface defect detection, achieving high accuracy. This invention acquires a defect distribution map, displaying the location of defects with pixel-level precision, effectively, intuitively, and completely displaying the defect distribution.

[0056] The foregoing description illustrates and describes several preferred embodiments of this application. However, as previously stated, it should be understood that this application is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the application concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of this application should be within the protection scope of the appended claims.

Claims

1. A method for object surface detection based on a mechanical arm hand-eye system three-dimensional reconstruction, characterized in that, Comprising: Providing a mechanical arm, a 2D vision system, a 3D scanning system and a system connector, the 2D vision system includes a high-resolution camera, a liquid lens and a ring-shaped polarized light source, connected to the mechanical arm through the system connector, and the 2D vision system and the 3D scanning system are respectively located on the left and right sides of the system connector; The mechanical arm transports the automobile parts in the inspection area to the three-dimensional camera area, and the manufacturing data of the input automobile parts are scanned by the 3D scanning system control module to obtain the three-dimensional point cloud of the automobile parts; The neural network with Auto encoder structure is used to encode the features of the three-dimensional point cloud of the automobile parts, combined with the pre-encoding position of the camera, the decoder composed of MLP is used to predict the probability of the best viewing angle of the camera, and the position with the highest probability is selected as the best viewing angle; The user controls the mechanical arm to move the 2D vision system to the best viewing angle through the front-end web interface, and automatically or manually collects sample images; The AI model is continuously trained using the collected sample images, and the defect distribution map of the surface defects of the automobile parts is obtained through the AI model.

2. The object surface detection method based on the three-dimensional reconstruction of the hand-eye system of the mechanical arm according to claim 1, characterized in that, The 3D scanning system control module scans the manufacturing data of the input automobile parts to obtain the three-dimensional point cloud of the automobile parts, including the step of merging and processing the three-dimensional point cloud.

3. The object surface detection method based on the three-dimensional reconstruction of the hand-eye system of the mechanical arm according to claim 2, characterized in that, Further comprising: using Voxel grid down sample method to remove and denoise the point cloud, and converting the point cloud into mesh representation.

4. The object surface detection method based on the three-dimensional reconstruction of the hand-eye system of the mechanical arm according to claim 1, characterized in that, Further comprising: Establishing a database to store the detection data of all automobile parts in the detection process.

5. The object surface detection method based on the three-dimensional reconstruction of the hand-eye system of the mechanical arm according to claim 1, characterized in that, The defects in the defect distribution map are divided by area.

6. The object surface detection method based on the three-dimensional reconstruction of the hand-eye system of the mechanical arm according to claim 4, characterized in that, The AI model is retrained combined with the data in the database, and the new model will be updated to the software system after training. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 5.

Images