Workpiece recognition method based on workpiece shape feature coding and storage medium

Abstract

The application discloses a workpiece recognition method based on workpiece shape feature coding and a storage medium, wherein the feature coding is a string coding generated by extracting three-dimensional shape features from a workpiece point cloud model, three-dimensional shape features are restored through decoding, and similarity comparison of shape features is carried out with workpiece coding in a database, so as to be used for workpiece recognition and retrieval. The method comprises the following steps: step 1, acquiring and inputting workpiece point cloud data; step 2, generating shape feature coding; step 3, decoding; and step 4, feature similarity comparison. The coding and recognition method can avoid the situation that the same part has multiple codings in different batches, can avoid waste caused by recognition coding errors, and can maximize the one-to-one correspondence between spare part coding and the actual object.

Description

Technical Field

[0001] This invention relates to the field of logistics and warehousing management technology for mechanical spare parts in industrial enterprises, specifically to a workpiece identification method and storage medium based on workpiece shape feature encoding, which is used for the logistics management mode of spare parts in tobacco industrial enterprises with complex types, many non-standard parts, and a small number of single types. Background Technology

[0002] In manufacturing, the management of equipment parts, including registration, procurement, storage, maintenance, and replacement, is a complex yet crucial task. The variety and quantity of parts make manual management not only costly but also prone to information clutter and omissions. With the continuous improvement of equipment informatization and intelligence, applying artificial intelligence technology to parts management is a future trend.

[0003] The most tedious task in workpiece management is workpiece identification and retrieval. Traditionally, this is done based on the workpiece's factory code, but the lack of a unified standard for coding rules easily leads to coding confusion and difficulty in interpreting the coded information. The trend in applying artificial intelligence technology is to use computer vision technology for automatic identification based on the workpiece's shape.

[0004] Designing a visual feature encoding and decoding algorithm based on workpiece shape for automatic identification and retrieval is an essential requirement for improving the intelligence level of workpiece management. This will also bring unprecedented changes to workpiece management, reducing the drawbacks of cumbersome, chaotic, and time-consuming entry and exit processes, greatly improving the work efficiency of staff and the utilization efficiency of workpieces, and reducing management costs. Summary of the Invention

[0005] The purpose of this invention is to provide a decoding and recognition method and storage medium based on workpiece 3D shape feature encoding. The feature encoding is a string encoding generated by extracting 3D shape features from a workpiece point cloud model. The 3D shape features are then decoded to reconstruct the 3D shape features, and a similarity comparison is performed with workpiece codes in a database for workpiece identification and retrieval. The method specifically includes the following steps:

[0006] Step 1, Input the workpiece point cloud model

[0007] Point clouds represent 3D models using a set of points with 3D coordinates. Point cloud data is acquired by scanning the workpiece using a depth camera. This invention employs non-contact structured light projection 3D shape measurement technology. A structured light grating is projected onto the workpiece to be scanned by a projector, and a high-definition camera captures multi-angle grating-modulated images. Phase measurement profilometry is used to calculate the pixel depth information of the images. Model files can be imported in STL, PLY, or PCD formats.

[0008] Step 2, generate shape feature encoding, including:

[0009] Step 2.1: Extract spin map features from the point cloud for workpiece shape encoding.

[0010] A spin graph is a local surface descriptor defined on each vertex of a 3D model, describing the relative positional relationships between the vertex and its neighboring data points. It possesses rigid geometric transformation invariance such as translation and rotation. First, a 3D mesh model is established on the point cloud model through downsampling and local surface normal estimation. A local coordinate system is established with each vertex in the mesh and its normal vector as the directed point, allowing surrounding points to be projected onto this coordinate system.

[0011] Let vertex p correspond to a directed point O, unit normal vector n correspond to p, tangent plane P be p, and line L be the line containing unit normal vector n. The distance α from other points x on the mesh to line L and the distance β from x to plane P are taken as the two spin coordinates of that point. For a given directed point O, the spin coordinate projection establishes a mapping x→(α,β).

[0012]

[0013] When all vertices within a certain range on the mesh surface are projected onto the coordinate system defined by the directed point O using formula (1), the spin graph of point O is obtained. To obtain the spin graph for each vertex in the 3D mesh model, the 3D model is converted into a set of 2D images, and then 3D shape features are extracted and quantized from the 2D images. The parameters of the spin graph include the resolution, size, and normal vector angle threshold. The resolution of the spin graph refers to the actual spatial distance corresponding to one pixel on the spin graph. With a fixed resolution, the size of the spin graph refers to the size of the projection area around the vertex. As the spin graph size increases, the features described by the spin graph gradually change from local features to global features. During the creation of the spin graph, to limit the self-occlusion of the target, the angle between the normal vectors of the directed point x and O cannot exceed the normal vector angle threshold. In this invention, the angle threshold is set to π / 3.

[0014] Step 2.2, Extract key points

[0015] Since spin graph features are defined on each vertex of a 3D model, the number of spin graphs is relatively large, making feature quantization and encoding inconvenient. Furthermore, features with closely spaced vertices exhibit significant information redundancy. Therefore, this invention generates feature descriptors by extracting a small number of key points from the point cloud model. This invention uses the points with the maximum curvature distributed across regions as key points, selecting the top n points with the maximum curvature in the point cloud model as key vertices, and setting a threshold for the minimum distance between vertices. Finally, it determines the n points with the maximum curvature distributed across regions on the surface of the 3D model.

[0016] Step 2.3, Feature Quantization

[0017] Spin map pixels are coordinate points defined in the α-β coordinate system. Since α-β are continuous real numbers in the interval [0,1]-[-1,1], each coordinate point in the parameter space α-β is interpolated into a coordinate grid of bin = (1 / Δ) × (2 / Δ) using bilinear interpolation with Δ as the discrete step size. The number of coordinate points in each grid is counted to form a discrete histogram feature. Furthermore, the values ​​of each bin of the histogram feature are binarized with a fixed threshold, and the binary codes of all key points are concatenated to form the binary digital code of the entire 3D model.

[0018] Step 2.4, Output Encoding

[0019] Twenty key points are extracted from the model point cloud. The spin map features of each key point are quantized into 153-bit binary codes, for a total of 3060 bits for the workpiece model shape. A 16-bit identification code is set at the beginning of the code to identify the valid code. The end is a reserved encoding position, padded with 0s, for future code expansion. The output code is a 3600-bit binary number string, which can be converted into a 900-character string and stored as a QR code.

[0020] Step 3, Decoding

[0021] The string is encoded as a hexadecimal number string. During decoding, each character in the string needs to be converted into a four-bit binary number according to the correspondence between hexadecimal and binary numbers, resulting in a binary number string with a total length of 3600 bits. The first 16 bits are the identification code, and bits 17 to 3076 are the shape code, which is divided into 20 key point shape feature codes with each feature code consisting of 153 bits.

[0022] Step 4, Feature Similarity Comparison

[0023] The input workpiece model generates a code, which is then compared one by one with codes in the database. The code is decoded to obtain 153-bit shape feature codes for all key points. The shape feature similarity metric is the binary distance between feature codes, with a minimum value of 0 and a maximum value of 153. A fixed threshold of 20 is used as a parameter; values ​​less than the threshold are considered a match. The code similarity metric is calculated between workpiece model codes using a cyclic matching method, such as... Figure 4 As shown in the experiment, if there are more than 3 matching points, the codes can be considered consistent, that is, the recognition is successful.

[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the steps of the workpiece recognition method based on workpiece shape feature encoding described in the present invention.

[0025] The beneficial effects of this invention are:

[0026] (1) It has a fast processing speed. The binary encoding can quickly calculate the encoding distance, which is suitable for large-scale retrieval.

[0027] (2) Matching robustness, it can still be identified even when the point cloud obtained by scanning the workpiece has missing parts or noise;

[0028] (3) Low data requirements, no need to pre-train model data;

[0029] (4) It has good scalability. When the input code cannot retrieve recognizable data, it can be expanded to a new code. Attached Figure Description

[0030] Figure 1 The flowchart of the decoding and recognition method based on the three-dimensional shape feature encoding of the workpiece of the present invention.

[0031] Figure 2 This is a schematic diagram of a point cloud model.

[0032] Figure 3 This is a diagram illustrating the encoding rules.

[0033] Figure 4 This is a diagram illustrating the encoding process. Detailed Implementation

[0034] This invention proposes a decoding and recognition method based on the three-dimensional shape feature encoding of a workpiece. The feature encoding is a string encoding generated by extracting three-dimensional shape features from the workpiece point cloud model. The three-dimensional shape features are then restored through decoding, and the similarity of the shape features is compared with the workpiece codes in the database for workpiece recognition and retrieval.

[0035] like Figure 1 As shown, a workpiece recognition method based on workpiece shape feature encoding includes:

[0036] Step 1, Input the workpiece point cloud model

[0037] like Figure 2 As shown, a point cloud represents a 3D model using a set of points with 3D coordinates. The point cloud data is acquired by scanning the workpiece using a depth camera. This invention employs a non-contact structured light projection 3D shape measurement technology. A structured light grating is projected onto the workpiece to be scanned by a projector, and a high-definition camera captures multi-angle grating-modulated images. The pixel depth information of the images is calculated using phase measurement profilometry. The model file can be imported in STL, PLY, or PCD formats.

[0038] Step 2, extract point cloud shape features

[0039] Step 2.1: Extract spin map features from the point cloud for workpiece shape encoding.

[0040] A spin graph is a local surface descriptor defined on each vertex of a 3D model, describing the relative positional relationships between vertices and neighboring data points. It possesses rigid geometric transformation invariance to translation and rotation. First, a 3D mesh model is established from the point cloud model through downsampling and local surface normal estimation. A local coordinate system is then established with each vertex in the mesh and its normal vector as the directed point, allowing surrounding points to be projected onto this coordinate system. Figure 3 As shown.

[0041] Let vertex p correspond to point O, unit normal vector n correspond to point p, tangent plane P be the plane of point p, and line L be the line containing unit normal vector n. The distance α from other points x on the mesh to line L and the distance β from x to plane P are taken as the two spin coordinates of that point. For a given directed point O, the spin coordinate projection establishes a mapping x→(α,β):

[0042]

[0043] When all vertices within a certain range on the mesh surface are projected onto the coordinate system defined by the directed point O using formula (1), the spin graph of point O is obtained. To obtain the spin graph for each vertex in the 3D mesh model, the 3D model is converted into a set of 2D images, and then 3D shape features are extracted and quantized from the 2D images. The parameters of the spin graph include the resolution, size, and normal vector angle threshold. The resolution of the spin graph refers to the actual spatial distance corresponding to one pixel on the spin graph. With a fixed resolution, the size of the spin graph refers to the size of the projection area around the vertex. As the spin graph size increases, the features described by the spin graph gradually change from local features to global features. During the creation of the spin graph, to limit the self-occlusion of the target, the angle between the normal vectors of the directed point x and O cannot exceed the normal vector angle threshold. In this invention, the angle threshold is set to π / 3.

[0044] Step 2.2, Extract key points

[0045] Since spin graph features are defined on each vertex of a 3D model, the number of spin graphs is relatively large, making feature quantization and encoding inconvenient. Furthermore, features with closely spaced vertices exhibit significant information redundancy. Therefore, this invention generates feature descriptors by extracting a small number of key points from the point cloud model. This invention uses the points with the maximum curvature distributed across regions as key points, selecting the top n points with the maximum curvature in the point cloud model as key vertices, and setting a threshold for the minimum distance between vertices. Finally, it determines the n points with the maximum curvature distributed across regions on the surface of the 3D model.

[0046] Step 2.3, Feature Quantization

[0047] Spin map pixels are coordinate points defined in the α-β coordinate system. Since α-β are continuous real numbers in the interval [0,1]-[-1,1], each coordinate point in the parameter space α-β is interpolated into a coordinate grid of bin = (1 / Δ) × (2 / Δ) using bilinear interpolation with Δ as the discrete step size. The number of coordinate points in each grid is counted to form a discrete histogram feature. Furthermore, the values ​​of each bin of the histogram feature are binarized with a fixed threshold, and the binary codes of all key points are concatenated to form the binary digital code of the entire 3D model.

[0048] Step 2.4, Output Encoding

[0049] Twenty key points were extracted from the model point cloud. The spin map features of each key point were quantized into 153-bit binary codes, resulting in a total of 3060-bit workpiece model shape codes. For example... Figure 4 As shown, the encoding header contains a 16-bit identification code to identify valid encodings, while the tail is a reserved encoding position, padded with 0s for future encoding expansion. The output encoding is a 3600-bit binary string, which can be converted into a 900-character string and stored as a QR code.

[0050] Step 3, Decoding

[0051] The string is encoded as a hexadecimal number string. During decoding, each character in the string needs to be converted into a four-bit binary number according to the correspondence between hexadecimal and binary numbers, resulting in a binary number string with a total length of 3600 bits. The first 16 bits are the identification code, and bits 17 to 3076 are the shape code, which is divided into 20 key point shape feature codes with each feature code consisting of 153 bits.

[0052] Step 4, Feature Similarity Comparison

[0053] The input workpiece model generates a code, which is then compared one by one with codes in the database. The code is decoded to obtain 153-bit shape feature codes for all key points. The shape feature similarity metric is the binary distance between feature codes, with a minimum value of 0 and a maximum value of 153. A fixed threshold of 20 is used as a parameter; values ​​less than the threshold are considered a match. The code similarity metric is calculated between workpiece model codes using a cyclic matching method, such as... Figure 4 As shown in the experiment, if there are more than 3 matching points, the codes can be considered consistent, that is, the recognition is successful.

[0054] In the above method, other three-dimensional shape features can be used for feature encoding.

Claims

1. A workpiece recognition method based on workpiece shape feature encoding, characterized in that, Includes the following steps: Step 1: Obtain and input the workpiece point cloud data. The point cloud model is represented by a set of points with three-dimensional coordinates. Step 2, generate shape feature encoding, including: Step 2.1: Extract spin map features from the point cloud for workpiece shape encoding. The spin map is defined on each vertex of the 3D model and describes the relative positional relationship between the vertex and its neighboring data points; including: Step 2.1.1: Establish a 3D mesh model from the point cloud model through downsampling and local surface normal estimation. Establish a local coordinate system for each vertex in the grid and its normal vector, such that the surrounding points are projected onto this coordinate system; For a given directed point O, spin coordinate projection establishes a mapping x→(α,β). （1） Where: p is the vertex corresponding to the directed point O, n is the unit normal vector corresponding to point p, and (α,β) are the two spin coordinates of the spin graph pixel x; When all vertices within a certain range of the mesh surface are projected onto the coordinate system defined by the directed point O using formula (1), the spin diagram of point O is obtained. Step 2.1.2: Obtain the spin graph for each vertex in the 3D mesh model, which converts the 3D model into a set of 2D images. Then, extract and quantize the 3D shape features from the 2D images. Step 2.2, extract key points: generate feature descriptors by extracting a small number of key points from the point cloud model, and use the point with the maximum curvature in the regional distribution as the key point; Step 2.3, Feature Quantization: Using Δ as the discrete step size, each coordinate point in the parameter space α-β is calculated into a coordinate grid of bin=(1 / Δ)×(2 / Δ) using bilinear interpolation. The number of coordinate points in each grid is counted to form a discrete histogram feature. The values ​​of each bin of the histogram feature are binarized with a fixed threshold. The binary codes of all key points are connected to form the binary digital code of the entire 3D model. Step 2.4, Output Encoding: Extract 20 key points from the model point cloud. The spin map feature of each key point is quantized into a 153-bit binary code, for a total of 3060 bits for the workpiece model shape encoding. A 16-bit identification code is set at the beginning of the encoding to identify the valid encoding. The end of the encoding is a reserved encoding position, padded with 0s for future encoding expansion. The output encoding is a 3600-bit binary number string that can be converted into a 900-character string, and the encoding information is stored in a QR code. Step 3, Decoding: Convert each character of the string into a four-bit binary number according to the correspondence between hexadecimal and binary numbers, resulting in a binary number string with a total length of 3600 bits; where: the first 16 bits are the identification code, and the 17th to 3076th bits are the shape code, which is divided into 20 key point shape feature codes according to the feature code of every 153 bits. Step 4, Feature Similarity Comparison: The input workpiece model generates an encoding, which is then compared one by one with the encodings in the database. The encoding is decoded to obtain the 153-bit shape feature encodings of all key points. The shape feature similarity measure is the binary distance between the feature encodings, with a minimum value of 0 and a maximum value of 153. A fixed threshold is used as a parameter, set to 20. If the value is less than the threshold, it is considered a match. The encoding similarity measure between the workpiece model encodings is calculated by cyclic matching. If a certain similarity measure is met, the recognition is successful.

2. The method according to claim 1, characterized in that, In step 2: The parameters of the spin graph include the spin graph resolution, spin graph size, and normal vector angle threshold.

3. The method according to claim 2, characterized in that: The resolution of the spin graph refers to the actual spatial distance corresponding to one pixel on the spin graph. When the resolution is constant, the size of the spin graph refers to the size of the projection area around the vertex. As the size of the spin graph increases, the features described by the spin graph gradually change from local features to global features.

4. The method according to claim 2, characterized in that: During the creation of a spin graph, the angle between the normal vectors of directed points x and O does not exceed the threshold of the angle between normal vectors.

5. The method according to claim 4, characterized in that: The included angle threshold is set to π / 3.

6. The method according to claim 1, characterized in that, In step 2: The key points are the n points with the largest curvature in the point cloud model, which are taken as key vertices. A threshold for the minimum distance between vertices is set, and finally the n points with the largest curvature distributed on the surface of the 3D model according to the region are determined.

7. The method according to claim 1, characterized in that, In step 4: The code similarity measure between workpiece model codes is calculated by cyclic matching. If more than 3 points are matched, the codes are considered to be consistent.

8. The method according to any one of claims 1-7, characterized in that: The input point cloud model is obtained by capturing multi-angle grating modulation images with a high-definition camera. The output represents the three-dimensional shape features of the workpiece using a QR code.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program can be executed by a processor to implement the steps of the workpiece recognition method based on workpiece shape feature encoding as described in any one of claims 1-8.

Images