Tire character recognition method based on three-dimensional point cloud

Abstract

The invention relates to a tire character recognition method, in particular to a tire character recognition method based on three-dimensional point cloud, and the method comprises the steps: calibrating an industrial 3D camera, managing and controlling the connection and communication of the camera through a tire point cloud data acquisition module, and visualizing and storing an acquisition result; performing combined filtering on the tire point cloud data by using voxel downsampling and a statistical filtering method, and then making a tire data set; putting the manufactured data set into a PointNet + + network added with a self-attention mechanism for training, and generating a training weight; and visualizing a tire identification result and storing a prediction result file. The invention relates to the technical field of three-dimensional recognition, geometric characteristics of the surface of a tire are more accurately captured by using point cloud, the influence of factors such as light and tread stains is reduced, hierarchical feature extraction of PointNet + + is combined with a Transform self-attention mechanism, detail features are more sufficiently extracted by using the relationship between different hierarchical features, and the recognition accuracy is improved. And the robustness of tire character recognition in a complex environment is improved.

Description

Technical Field

[0001] The invention relates to a tire character recognition method, in particular to a tire character recognition method based on three-dimensional point cloud. Background Art

[0002] In the context of industrial development, compared with traditional workshops, intelligent factories reduce the participation of manual labor and use more objective judgments to quantitatively measure the information of objects. The rapid development of computer vision technology has made the production process more efficient and further exerted the practical use of industrial intelligence. In the transportation industry containing tires, the identification and matching of tread information is spread throughout all aspects related to tires. If the identification information does not match the actual situation, it may lead to serious consequences. Therefore, accurate tire character recognition is even more important.

[0003] 3D point clouds can fully present the geometric information of the surface of an object, including shape, curvature, and depth, while 2D images can only provide surface projection information. For complex or non-planar objects, such as tires, 3D point clouds can more accurately capture their overall form and details, and are less affected by lighting, shadows, and reflections. They can maintain stable recognition performance under more environmental conditions, while 2D images may be deformed or have missing information, and are more affected by lighting and background interference. Point cloud data can be collected from multiple angles, so that data from multiple perspectives can be considered when recognizing characters, providing richer spatial information and more suitable for tire recognition application scenarios with higher precision requirements.

[0004] Compared with common flat characters, the background contrast before and after the embossed characters is smaller. Therefore, the recognition of convex and concave characters mainly depends on the intensity of light, and the environment is an important factor. Currently, most of the commonly used convex and concave character recognition technologies are based on two-dimensional vision, graying and binarizing the acquired image data, and then extracting the characteristic information of the tire. However, two-dimensional image sensors mainly rely on the contrast of the object being measured, and have a certain impact on the recognition of tire characters under non-specific lighting conditions. For example, if the tire surface becomes dim due to dirt, or produces reflections when illuminated, traditional two-dimensional vision systems will find it difficult to capture these details. Therefore, the application of two-dimensional vision in tire character detection is subject to certain limitations, and its tire feature recognition under non-ideal environments needs to be improved. Summary of the invention

[0005] The present invention provides a tire character recognition method based on three-dimensional point cloud. The method uses a 3D camera to collect tire information and obtains different position information of characters in point cloud data, thereby directly extracting the character bumps of the tire itself and avoiding the influence of environmental factors on the data.

[0006] The present invention is achieved through the following technical solutions:

[0007] A tire character recognition method based on three-dimensional point cloud includes camera calibration, point cloud data acquisition, tire point cloud preprocessing, tire recognition model training, and recognition result visualization.

[0008] S1. Industrial 3D camera calibration, by shooting the calibration plate image, converting the world coordinate system to the pixel coordinate system, and obtaining the camera intrinsic parameters, extrinsic parameters and distortion coefficients.

[0009] S2. Point cloud data collection part, the tire information is collected through the calibrated industrial 3D camera, the collection is triggered after the connection is successful, the collected data is visualized, and the data can be saved in various formats such as txt, pcd, ply, obj, depth map, etc.

[0010] S3. Preprocess the collected tire point cloud images, use voxel downsampling and statistical filtering to reduce the point cloud density and trim redundant points, perform data enhancement and annotation on the processed tire point cloud, each character corresponds to its own label information, and divide the training set, test set, and validation set into a tire dataset through a json file.

[0011] S4. Put the dataset produced in S3 into PointNet++ with self-attention mechanism for training and generate training weights.

[0012] S5. Use the training weights generated in S4 to realize the visualization of the tire character recognition results, including two-dimensional visualization and three-dimensional visualization, and save the prediction result file.

[0013] Furthermore, the S1 camera calibration measures and calculates the internal and external parameters of the camera to obtain the parameters of the camera imaging geometry model, and corresponds the camera's pixel coordinate system with the actual world coordinate system, thereby achieving accurate image acquisition and recognition. The camera calibration part adopts the Zhang Zhengyou calibration method.

[0014] Use a black and white grid calibration plate to capture images at different angles, detect the feature points of the images, and solve for the camera's internal and external parameters and distortion coefficients.

[0015] Furthermore, for the point cloud data acquisition module S2:

[0016] Tire point cloud data is collected through the Photoneo3D camera. After confirming that the camera is wired correctly, click the "Start Connecting" button in the window. The status bar on the right will display the device connection status, connected device name and other parameters. After the connection is successful, the scan can be triggered.

[0017] To visualize the point cloud object, put the points in the point cloud pointer into points one by one, and create attribute data for each point. Pass the color attribute of the point data into polydata, and use the scalar attribute data of the unsigned char type as the color value.

[0018] The scanned results are saved in a specified folder. The saved point cloud files can be opened through a file viewer to realize point cloud visualization and convert different file types.

[0019] Furthermore, for S3, the camera acquisition process will inevitably contain some redundant points. The use of voxel downsampling and statistical filtering combined filtering can simplify the tire point cloud data and remove outliers. The coordinates of the points are normalized to the grid to calculate the voxel index. The voxel grid is set to an appropriate size according to the point cloud data. The centroid of all points in each voxel is calculated as the representative point of downsampling and input into the subsequent SOR filtering. Then, the noise points that may affect the results are eliminated through statistical filtering, and the k-neighborhood of each point is statistically analyzed to calculate the distance distribution between the input point and the neighboring points. According to the degree of deviation between the distance from the point to the neighboring point and the global average distance, the points that exceed the threshold range are screened as outliers.

[0020] The filtered tire point cloud is enhanced and annotated using the point cloud processing software CloudCompare. The same characters are annotated with the same color. The generated txt annotation file information includes x, y, z, label, dx, dy, dz. (x, y, z) represents the point cloud position coordinates, and (dx, dy, dz) represents the point cloud normal vector information. The annotated tire point cloud dataset is adjusted to the format required by the network, and the training set, test set, and validation set are divided into json files.

[0021] For a feature enhancement method based on PointNet++ and Transformer provided by S4, tire character recognition training can be further subdivided as follows:

[0022] S4-1. Based on the hierarchical sampling and feature extraction method of PointNet++, a starting point is selected from the point cloud, and the Euclidean distance is used as the distance metric between points. Then, the point farthest from the selected point set is iteratively sampled. This method can ensure that the sampling points are relatively evenly distributed in space, thereby preserving the point cloud geometric structure to the greatest extent.

[0023] Use multi-resolution grouping to divide different levels of point cloud data grouping and feature extraction. Divide the input point cloud data into multiple local blocks according to its distribution characteristics, extract feature vectors from each block, and then connect the extracted feature vectors in series to represent the characteristics of the local area.

[0024] The features extracted by S4-2.PointNet++ are input into the Transformer module, and the relationship between different points is captured through the self-attention mechanism. The enhanced features are fused with the original features to form the final features. The inverse distance weight method is adopted, and the acquired point cloud features are transferred to the target point using interpolation. The decoded features are mapped to the classification space, and each point obtains the category probability distribution. The tire point cloud dataset prepared in S3 is put into the optimized network for training, and the corresponding training weights are generated after the training is completed.

[0025] Specifically, for S5 recognition result visualization, it includes two-dimensional and three-dimensional result visualization.

[0026] S5-1. Put all the data that need to be verified into the dictionary. A tire contains multiple characters, and each character has a unique label. Randomly sample the data in each category and return the index, and continue sampling according to the index. Receive and read multiple parameters including the path of point cloud data, the path to generate results, the number of categories, the test data loader, the dictionary of the model and its weight path, and set an optional color map to generate the color of the corresponding character recognition display, and recognize the same character as the same color. Process the image data and generate the corresponding prediction results. The generated file includes the point cloud location information and the predicted recognition information.

[0027] S5-2. Use Open3D to draw 3D images. First, load the prediction file containing point cloud data and color the points according to the category labels. Since the tire character information features are relatively fine, it is necessary to set appropriate sampling points for it in order to obtain better recognition results.

[0028] The advantages of the present invention are:

[0029] The 3D point cloud can accurately capture the geometric characteristics of the tire surface. It does not require a special light source to ensure lighting conditions, and will not be affected by objects such as stains and marks on the tire surface, which improves the anti-interference ability of recognition. The hierarchical feature extraction of PointNet++ is combined with the self-attention mechanism of Transformer to make full use of the relationship between features at different levels to realize the recognition of 3D tire point cloud characters. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of a tire character recognition method based on three-dimensional point cloud according to an embodiment of the present invention;

[0031] Figure 2 It is a partial schematic diagram of the filtering effect of the tire character recognition method based on three-dimensional point cloud according to an embodiment of the present invention;

[0032] Figure 3It is a schematic diagram of tire recognition effect of the tire character recognition method based on three-dimensional point cloud according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The technical solutions of the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] The present invention provides a tire character recognition method based on three-dimensional point cloud, such as Figure 1 ,include:

[0035] S1. During the camera calibration process, a checkerboard is printed as a calibration plate. Based on the camera's pinhole imaging model, the calculation process is simplified by using a flat calibration plate.

[0036] Assuming that the calibration plate is located on the Z=0 plane in the world coordinate system, the simplification of three-dimensional coordinates to two-dimensional coordinates is realized, and the world coordinate system is transformed into the camera coordinate system through rigid body transformation:

[0037]

[0038] The three-dimensional coordinates in the real world are represented by Xw, Yw, and Zw; the coordinates of the coordinate system with the optical center of the camera as the origin are represented by Xc, Yc, and Zc; R is the rotation matrix and T is the translation matrix.

[0039] It is also necessary to convert the camera coordinate system to the image coordinate system, and the image coordinate system to the pixel coordinate system. Finally, the conversion process is combined to obtain:

[0040]

[0041] Where (u, v) is the coordinate of the point in the pixel coordinate system; (u 0 ,v 0 ) is the origin of the image coordinate system; the first matrix on the right side of the equal sign is the camera intrinsic parameter, f x This is equivalent to using the number of pixels in the x direction to quantify the physical focal length f, f y It is equivalent to using the number of pixels in the y direction to quantify the physical focal length f; the second item on the right side of the equal sign is the camera extrinsic parameter; Xw, Yw, and Zw are the three-dimensional coordinates in the real world.

[0042] After solving the initial intrinsic and extrinsic parameters, the maximum likelihood estimation is used to fine-tune and optimize the camera's intrinsic and extrinsic parameters and distortion coefficients to improve the calibration accuracy.

[0043] S2. Data collection part creates an instance responsible for managing and controlling the connection and communication of the camera, checks whether the camera is running, and if false is returned, it means that it cannot connect to the device. At this time, a message is added to LogText and the execution ends. If the camera is running, the version information is obtained and displayed. Loop through the device list array and use assignment to build and display the detailed information of the device, including device name, hardware identification, type, and firmware version.

[0044] After connecting the Photoneo3D camera, you can start collecting tire point cloud data and visualize and store the collected results.

[0045] S3. The collected point cloud data consists of millions of points, which are often mixed with noise interference. If the point cloud data is not properly preprocessed, it will have a certain impact on subsequent operations. Specifically, in S3 of this embodiment, voxel downsampling and statistical filtering are used for joint filtering.

[0046] Normalize the coordinates of the points to the grid and calculate the voxel index. Set a suitable voxel grid size according to the point cloud data. As a preferred embodiment of this embodiment, the voxel grid is set to 0.01. Calculate each voxel V n All points in {P 1 ,P 2 ,…,P i} as the centroid of the representative point:

[0047]

[0048] The representative points obtained by downsampling are then filtered using statistical analysis techniques. First, the entire point cloud data set needs to be traversed to calculate the average distance between each point and its k nearest neighbors, assuming that there are n points in total. The global average distance μ and standard deviation σ of all points are calculated based on these average distances as follows.

[0049]

[0050] The threshold is obtained by multiplying the custom standard deviation coefficient α by the standard deviation σ, and then adding or subtracting it from the average distance μ:

[0051]

[0052] According to the set threshold, points that do not meet the range are regarded as outliers and removed from the point cloud. In this embodiment, after practical testing, the standard deviation coefficient is set to 1.5, and the number of points considered in the neighborhood k is set to 90. Figure 2 ,In the local schematic diagram of the filtering effect, the red part is the outlier point, and the green part is the point cloud retained after the joint filtering of voxel downsampling and statistical filtering.

[0053] Tire point cloud data is enhanced to improve the generalization ability of model training. The training dataset is augmented by adding noise, scaling, and translation methods.

[0054] Randomly translate the point cloud by adding and subtracting the point cloud coordinates at the same time; add noise to the point cloud data, perform small-scale disturbance, set the weight coefficient and the upper and lower limits of the random noise, and control the amplitude of the random noise; randomly scale the point cloud, set the appropriate range of multiplication factors to multiply the point cloud data.

[0055] Use CloudCompare to annotate the tire point cloud. The same characters are annotated with the same color. The generated txt annotation file information includes x, y, z, label, dx, dy, dz, where (x, y, z) represents the point cloud position coordinates, and (dx, dy, dz) represents the point cloud normal vector information. When importing point cloud information, you can choose whether to use the point cloud normal vector. If the input channel is 3, it will not be used; if the input channel is 6, it will use the normal vector information. Adjust the annotated tire point cloud dataset to the format required by the network, and use the json file to batch divide the training set, test set, and validation set in a ratio of 7:2:1.

[0056] S4. The tire recognition training process is as follows:

[0057] S4-1. Sampling uses the iterative farthest point sampling method FPS. Select a starting point from the point cloud, use the Euclidean distance as the distance measure between points, and then iteratively collect the point farthest from the selected point set. Compared with random sampling, it can sample the global tire point cloud data more evenly and completely.

[0058] Input and output: tensor(B,N,C)->tensor(B,N1,C), where B is the batch size, N and N1 are the number of points before and after sampling, and C is the number of features of each point.

[0059] The input of the grouping layer includes point cloud data and a set of center points N1. These center points are selected from the original point cloud through the farthest point sampling of the sampling layer to represent different local areas.

[0060] The multi-resolution grouping MRG method groups point cloud data and extracts features by dividing them into different levels. According to the distribution characteristics of the point cloud data, it is divided into regions, and feature vectors are extracted from each block. The extracted feature vectors are then connected in series to represent the characteristics of the local area.

[0061] For each sampling point q obtained by FPS j , the neighborhood point set uses radius r k The spherical query of gives:

[0062] N(q j ,rk )={p i ∈P|||q j -p i ||≤r k}

[0063] Where p i The input point cloud includes coordinates and initial features.

[0064] For each neighborhood point set, use shared MLP to extract features:

[0065] f k =MLP k (Relative(p i ,q j ))

[0066] Different radius r 1 ,r 2 ,…,r k Extracted feature concatenation:

[0067] f MRG (q j )=Concat(f 1 ,f 2 ,…,f k )

[0068] In this embodiment, MLP is used to transform f MRG Project to target dimension 128:

[0069]

[0070] S4-2. The extracted local area features are converted into inputs and enhanced through Transformer Encoder. The relationship between features is learned through the self-attention mechanism, and enhanced feature representation is generated.

[0071] The spatial position information of the key points is embedded into the input features. Use MLP to map the position information to a vector of the same dimension as the feature:

[0072] f pos (p i )=MLP pos (p i )

[0073] The embedded spatial information is fused with the MRG features to obtain the input features:

[0074]

[0075] The Transformer input dimension is (N, B, d), and the features need to be converted from (B, N, d) to capture the relationship between points through the attention head. For each point, the self-attention weight is calculated to obtain the global feature relationship:

[0076]

[0077] Where Q, K, and V are the query, key, and value obtained by linear transformation of the features; d k Represents the vector dimension, used for scaling and stable training.

[0078] Feed-forward Network consists of two linear transformations and an activation function to further process the features after the self-attention mechanism. When the input value is less than 0, the ReLU function will lose the associated features of small targets. For small features, mutations may occur due to their non-smooth characteristics. The smooth Swish function has better convergence and faster convergence speed. The calculation formula is as follows:

[0079]

[0080] The first-order differential of this function is non-zero and continuous, which makes the gradient back-propagation more stable and can effectively prevent the loss or mutation of features. The Swish function is suitable for processing targets with fine features. The expression of the feedforward network is:

[0081] FFN(x)=W 2 (Swish(W 1 x+b 1 ))+b 2

[0082] Where x is the eigenvector of the position; W 1 is the weight matrix of the first layer; b 1 is the bias vector of the first layer; W 2 is the weight matrix of the second layer; b 2 is the bias vector of the second layer.

[0083] The first layer of the two-layer linear transformation increases the feature dimension, and the second layer restores the feature dimension to facilitate residual connection. The activation function introduces nonlinearity, allowing the network to learn complex mapping relationships.

[0084] Add a residual connection after the feedforward network and normalize it:

[0085] f output =Norm(x+FFN(x))

[0086] The inverse distance weighted method is adopted to transfer the acquired point cloud features to the target point using interpolation, and the decoded features are mapped to the classification space, and the category probability distribution is obtained for each point.

[0087] Specifically, the tire dataset prepared in S3 is put into the optimized network for training for 300 rounds, and the corresponding training weights are generated after the training is completed.

[0088] S5. Visualize the recognition results in two dimensions and three dimensions. Put all the data to be verified into a dictionary, where the dictionary key represents the category to which the data belongs. Each tire contains multiple characters, and each character has its own label. Set the maximum capacity of the cache. Once the cache reaches this capacity limit, determine which data should be replaced or removed to make room for new data.

[0089] Performs random sampling of data of different categories and returns the sampling index, which is then used for sampling. Receives a dictionary containing the storage path of the point cloud data, the save path of the generated results, the total number of categories, the test data loader, the model and its weight path, and a color map, and is responsible for processing the image data and generating the corresponding prediction result files. These generated files contain both point cloud information and the identification information obtained through prediction.

[0090] Use the Open3D library to render 3D images. Load the prediction file containing point cloud data and color the point cloud according to the category label. In this example, the sampling points are set to 20,000 to obtain better results.

[0091] Figure 3 This is a schematic diagram of the tire recognition effect of the embodiment, where the same number is recognized as the same color.

Claims

1. A tire character recognition method based on three-dimensional point cloud, characterized in that: The method comprises the following steps: S1. Industrial 3D camera calibration, capture the chessboard calibration plate image, convert the world coordinate system to the pixel coordinate system, and obtain the camera intrinsic parameters, extrinsic parameters and distortion coefficients; S2. Point cloud data collection part, collects tire point cloud information through the industrial 3D camera calibrated by S1, triggers collection after confirming successful connection, visualizes the collected data, and can choose to save the data in various formats such as txt, pcd, ply, obj, depth map, etc. S3. Preprocess the point cloud image collected by S2, use voxel downsampling and statistical filtering to reduce the point cloud density and trim those points that do not meet the standards, perform data enhancement and labeling on the processed tire point cloud, each character corresponds to its own label information, and divide it into training set, test set, and validation set through json files to make a tire data set; S4. Put the tire dataset produced in S3 into the PointNet++ network with self-attention mechanism for training and generate training weights; S5. Use the training weights generated in S4 to realize the visualization of the tire character recognition results, including two-dimensional visualization and three-dimensional visualization, and save the prediction result file.

2. A tire character recognition method based on three-dimensional point cloud as claimed in claim 1, characterized in that: In S2, collecting tire point cloud data specifically includes: Create an instance to manage and control the camera's connection and communication. Check if the camera is running. If false is returned, it means that it cannot connect to the device. At this time, a message will be added to LogText and the execution will end. If the camera is running, the version information will be obtained and displayed. Loop through the device list array and use assignment to build and display the detailed information of the device, including device name, hardware identification, type, and firmware version. Visualize the point cloud object, put the points in the point cloud pointer into points one by one, create attribute data for each point, pass the color attribute of the point data into polydata, and use the scalar attribute data of unsigned char type as the color value; The scanned results are saved in a designated folder. You can choose to save the data in various formats such as txt, pcd, ply, obj, depth map, etc., which enhances the flexibility of the data. The saved tire point cloud file can be opened through a file viewer and supports conversion of different file types.

3. A tire character recognition method based on three-dimensional point cloud as claimed in claim 1, characterized in that: In S3, the tire data is preprocessed and a tire data set is generated. The specific method includes: Use voxel downsampling and statistical filtering to combine filtering, normalize the coordinates of the points to the grid to calculate the voxel index, set the appropriate voxel grid size according to the point cloud data, and calculate each voxel V n All points in {P1,P2,…,P i } as the centroid of the representative point: The representative points obtained by downsampling are filtered using statistical analysis technology. The entire point cloud data set is traversed to calculate the average distance between n points and their nearest k points. Based on these average distances, the global average distance μ and standard deviation σ of all points are calculated. According to the degree of deviation between the distance from the point to the neighboring point and the global average distance, the points that exceed the threshold range are screened as outliers. The preprocessed tire point cloud data is enhanced, and the tire point cloud is annotated using the point cloud processing software CloudCompare. The same characters are marked with the same color. The annotated tire point cloud dataset is adjusted to the format required by the network, and the training set, test set, and validation set are divided into batches in a ratio of 7:2:1 using a json file.

4. The tire character recognition method based on three-dimensional point cloud according to claim 1, characterized in that: In S4, the tire character recognition model is trained, and the specific method includes: Combine PointNet++’s hierarchical feature extraction with Transformer’s self-attention mechanism to fully capture the relationship between features at different levels; When sampling, first randomly select a point, then select the point farthest from this point as the starting point and continue iterating until the required number of sampling points is reached or all points are covered. In each iteration, the Euclidean distance from all remaining points to each point in the currently selected point set is calculated; Input and output: tensor(B,N,C)->tensor(B,N1,C), where B is the batch number, N and N1 are the number of points before and after sampling, and C is the number of features of each point; The input of the grouping layer includes point cloud data and a set of center points N1. These center points are selected from the original point cloud through the farthest point sampling of the sampling layer to represent different local areas. The multi-resolution grouping MRG method is used to group and extract features from the point cloud data at different levels. According to the distribution characteristics of the points, it is divided into multiple local areas, and feature vectors are extracted for each local area. Finally, the feature vectors are connected in series to represent the features of the local area. For each sampling point q obtained by FPS j , using radius r k The spherical query obtains the neighborhood point set, and for each neighborhood point set, the shared MLP is used to extract features, and the features extracted from different radii are concatenated and projected to the target dimension 128; The embedded spatial information is fused with the MRG features to obtain the input features: Where p i The input point cloud includes coordinates and initial features; is the MRG feature; f pos It is the spatial information after being mapped to 128 dimensions; After the Transformer Encoder performs feature enhancement, the relationship between features is learned through the self-attention mechanism, and an enhanced feature representation is generated. The relationship between points is captured by the attention head, the attention weights between points are calculated, and the features are weighted summed according to these weights. The feed-forward network consists of two linear transformations and an activation function. The Swish function is more suitable for tire character targets with fine features. The expression of the feed-forward network is: FFN(x)=W2(Swish(W1x+b1))+b2 Where x is the feature vector of the position; W1 is the weight matrix of the first layer; b1 is the bias vector of the first layer; W2 is the weight matrix of the second layer; b2 is the bias vector of the second layer; Add a residual connection after the feedforward network and normalize it: f output =Norm(x+FFN(x)) The inverse distance weighted method is adopted to transfer the acquired point cloud features to the target point using interpolation, and the decoded features are mapped to the classification space; The prepared tire point cloud dataset is put into the model for training, and the corresponding training weights are generated after the training is completed.

5. The tire character recognition method based on three-dimensional point cloud according to claim 1, characterized in that: In S5, the tire character recognition result is visualized, and the specific method includes: Put all the data that needs to be verified into a dictionary. The key of the dictionary is the category to which the data belongs. A tire has multiple characters, and each character corresponds to a unique label. Set the maximum number of data items that can be stored in the cache. When the cache reaches this size, decide to replace or delete the data to make room for new data. Perform random sampling for each tire character data and return the sampling index, and then sample based on these indexes, receive the dictionary covering the storage path of the point cloud data, the save path of the generated results, the total number of categories, the test data loader, the model and its weight path, and a color map, responsible for processing the image data and generating the corresponding prediction result files, which contain both the point cloud information and the recognition information obtained through the prediction; Use the Open3D library to render 3D images, load the prediction file containing point cloud data, color the point cloud according to the category label, and set appropriate sampling points according to the features of the identified object.

Images