An automatic target detection pre-labeling method based on point cloud data

Abstract

The present invention discloses an automatic target detection and pre-labeling method based on point cloud data, which relates to the field of computer vision and geometric processing technology, including S1, data acquisition and preprocessing; S2, feature extraction; S3, target detection; S4, pre-labeling; S5, quality inspection; S6, result output and visualization. The automatic target detection and pre-labeling method based on point cloud data makes full use of the three-dimensional spatial information of point cloud data, and can more accurately describe the shape, size and position of the target, thereby facilitating the accurate recognition and labeling of the target. Moreover, since the point cloud data is obtained by laser scanning and other methods, even if the target is partially blocked, it can be detected and labeled through the remaining point cloud data, which improves the robustness and accuracy of the method. In addition, the present invention not only realizes the automatic recognition and labeling of the target, but also can achieve faster speed and higher accuracy, and is particularly suitable for real-time application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and geometric processing, and in particular to an automatic target detection pre-labeling method based on point cloud data. Background Art

[0002] In the field of computer vision and geometric processing, target detection pre-labeling refers to the process of identifying and labeling targets in images or three-dimensional scenes. These labels usually include information such as the target's category, location, size, etc. Traditional technologies generally use manual methods to perform target detection pre-labeling, but manual labeling requires professional labelers to identify and label targets in images or scenes one by one. This process is time-consuming and costly, especially for large-scale data sets. Manual labeling is almost impossible. At the same time, due to the subjectivity and experience differences of labelers, the quality of manual labeling is often unstable.

[0003] With the rapid development of computer vision technology, the amount of data that needs to be processed is also increasing. Manual labeling cannot meet the labeling needs of large-scale data sets, especially in real-time or online application scenarios. Therefore, it is gradually replaced by automatic target detection pre-labeling. Automatic target detection pre-labeling automatically identifies and locates specific targets in images or scenes through algorithms and models, and adds corresponding labeling information to them. Compared with traditional manual labeling, it significantly improves the efficiency and accuracy of labeling.

[0004] For example, the patent document with application number 202210314507.9 discloses a pre-labeling method and storage medium based on target detection. The invention uses convolutional neural networks in combination with active learning to form a method for automatic labeling of target detection; using small-scale original image data, through active learning and automatic labeling of existing data, the possibility of model overfitting is reduced, taking into account the distribution characteristics of the entire sample, and iteratively training the model. New samples are added to the data set for iterative training, thereby training a target detection and recognition model with ideal accuracy, and improving the accuracy of the model after migration when the category changes and the rapid adaptability of the algorithm; compared with other methods, it realizes automatic expansion of labeled samples for training, which facilitates the acquisition of large-scale labeled data sets, can greatly reduce the time and economic cost of manual labeling, and ensure the effectiveness and efficiency of labeling.

[0005] However, the existing automatic target detection pre-labeling methods similar to the above-mentioned applications still have the following shortcomings:

[0006] Existing automatic target detection pre-labeling methods are usually based on deep learning models, and therefore require a large amount of labeled data for training. The preparation process of labeled data is often time-consuming and expensive, which limits the generalization ability of the model and the scope of practical application. In addition, in some applications, such as autonomous driving and robot navigation, high real-time requirements are placed on target detection. However, existing automatic target detection pre-labeling methods are inefficient when processing large-scale data sets and are difficult to meet application scenarios with high real-time requirements.

[0007] Therefore, it is urgent to improve this shortcoming. The present invention studies and improves the existing technology and its shortcomings, and provides an automatic target detection pre-labeling method based on point cloud data. Summary of the invention

[0008] The purpose of the present invention is to provide an automatic target detection pre-labeling method based on point cloud data to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: an automatic target detection pre-labeling method based on point cloud data, comprising the following steps:

[0010] S1. Data acquisition and preprocessing: The point cloud data describing the surface shape, posture, texture and other information of the object is acquired through the 3D scanning device, and the point cloud data is preprocessed;

[0011] S2, Feature extraction: Select deep learning algorithms according to the application scenario to extract useful features from point cloud data;

[0012] S3. Target detection: Based on feature extraction, deep learning models are used to detect targets in point cloud data to identify targets in images or scenes and determine their approximate locations.

[0013] S4, pre-labeling: automatically generate target labeling information based on the target detection results;

[0014] S5. Quality check: Perform quality check on the automatically generated annotation data to ensure the accuracy and consistency of the annotation, including checking whether the position of the annotation points is accurate and whether the annotation lines are smooth;

[0015] S6. Result output and visualization: Save and output the annotation results in a specified format (such as file, database, etc.), and select visualization tools such as VTuber and VTuber3D as needed to visualize the annotation results so that users can view and understand the annotation results.

[0016] Furthermore, the point cloud data is a three-dimensional data set represented in the form of points in a three-dimensional coordinate system, and each point contains three-dimensional coordinates (X, Y, Z), and some points contain color information or reflection intensity information; and the point cloud data is obtained as follows:

[0017] LiDAR: LiDAR measures the distance between an object and the sensor by emitting laser pulses and receiving reflected signals. LiDAR provides high-precision and high-density point cloud data, which is suitable for large-scale terrain mapping and object modeling.

[0018] Structured light scanning: Projects a known pattern of light onto the surface of an object, captures the changes in the reflected light pattern through a camera, and calculates the three-dimensional shape of the object. It is suitable for accurate modeling of small and medium-sized objects.

[0019] Time-of-flight sensor: calculates distance by emitting light pulses and measuring the time it takes for the light to travel from the emission to the return, generating point cloud data suitable for 3D perception in mobile devices;

[0020] Stereo vision: Use two or more cameras to capture images of the same scene from different angles, calculate the depth information of the object by comparing the parallax in the images, and generate a three-dimensional point cloud, which is suitable for navigation and modeling in indoor environments;

[0021] Multi-view reconstruction: Take multiple photos of an object from different angles, use computer vision algorithms to synthesize these photos into a complete 3D model, and generate point cloud data, which is suitable for 3D modeling and reconstruction of large-scale scenes.

[0022] Furthermore, the preprocessing specifically includes the following operations:

[0023] Noise removal: used to remove noise points contained in point cloud data due to sensor errors, environmental interference, and signal attenuation. Methods include: statistical denoising (based on statistical principles, data is analyzed and modeled to identify and eliminate outliers or noise points), radius filtering (set a search radius and a minimum number of neighboring points within the search radius. If the number of neighboring points of a point is less than the set minimum number, the point is considered to be a noise point and is removed; it is especially suitable for noise points in point cloud data caused by scanning unwanted objects or surface reflections, etc.), Gaussian filtering (de-noising is indirectly achieved by converting point cloud data into depth images or surface meshes and then applying Gaussian filtering);

[0024] Adjusting data: used to adjust and optimize point cloud data to meet the needs of subsequent tasks. Methods include: downsampling (using random sampling or voxel gridding to reduce the number of points to simplify point cloud data), registration (using feature matching or ICP algorithm to align point cloud data from different sensors or scanning devices to the same coordinate system), and smoothing (using moving least squares or Laplace smoothing to remove small fluctuations or unevenness in point cloud data caused by scanning equipment or environmental factors).

[0025] Furthermore, in step S2, the deep learning algorithm includes PointNet and its extension, convolution-based method, and graph neural network. The extraction logic of each method is as follows:

[0026] PointNet and its extensions (such as PointNet++, DGCNN, etc.): Directly extract features from point cloud data by sharing a multi-layer perceptron (MLP) and maximum pooling operations. The specific operation is: extract the local features of each point through a multi-layer perceptron, and finally obtain the global features through a maximum pooling operation;

[0027] Convolution-based methods: Apply convolution operations by converting point cloud data into a regular grid (such as voxelization) or an image (such as projection to a two-dimensional plane), and use convolutional neural networks to extract features of point cloud data;

[0028] Graph Neural Network (GNN): For point cloud data, each point is regarded as a node in the graph, and the graph structure is formed by building connection relationships (such as adjacency relationships) between points, and then the graph neural network is applied for feature extraction.

[0029] Furthermore, in step S2, the feature extraction method is specifically implemented as follows:

[0030] 1) Surface feature extraction

[0031] Normal estimation: Calculate the normal vector of each point in the point cloud to describe the shape and geometry of the surface;

[0032] Curvature estimation: Describe the curvature of a surface by calculating the local curvature of a point;

[0033] 2) Local feature extraction

[0034] SHOT (Signature of Histograms of OrienTations): Describes the distribution and geometric structure of the surrounding points by calculating the feature vector of each point in the point cloud;

[0035] FPFH (Fast Point Feature Histograms): A method for quickly calculating point feature histograms to describe the local shape information of points;

[0036] 3) Global feature extraction

[0037] VFH (Viewpoint Feature Histogram): Describes the overall shape and structure by projecting the point cloud into a two-dimensional histogram and counting the distribution of points in each bin.

[0038] Furthermore, the specific steps of the target detection are as follows:

[0039] Use voxel-based methods (such as VoxelNet) or point-based methods (such as PointRCNN) to generate candidate boxes, i.e. potential target areas;

[0040] After generating the candidate box, a convolutional neural network (CNN) or a recurrent neural network (RNN) is used to classify and regress the candidate box. The classification operation is used to determine whether the candidate box contains the target and give the target category. The regression operation is used to adjust the size and position of the candidate box to make it fit the actual target more accurately.

[0041] Use non-maximum suppression (NMS) and confidence threshold screening to remove redundant candidate boxes and retain the most likely target detection results.

[0042] Furthermore, the annotation information specifically includes:

[0043] Category label: the category name to which the target object belongs;

[0044] Bounding box: used to indicate the location and size of the target object in the image or point cloud data. It is represented by a rectangular box or a three-dimensional cube box and contains the location and size information of the object.

[0045] Confidence: Indicates the confidence of the model in the test results. The value range is [0, 1]. The higher the value, the higher the confidence of the model in the results.

[0046] Center point or center coordinates: mark the center point or center coordinates of the target object;

[0047] Rotation angle: For 3D target detection, it is used to indicate the rotation of the target object relative to a fixed axis or reference frame, especially when the target object changes its orientation;

[0048] Other attributes: Mark other attribute information according to specific application scenarios and task requirements.

[0049] Furthermore, the bounding box includes the following two formats:

[0050] Two-dimensional bounding box: For image data, it is represented in the format of (x, y, width, height), where (x, y) is the coordinate of the upper left corner of the rectangle and (width, height) is the width and height of the rectangle;

[0051] 3D bounding box: For point cloud data, it is represented using the format of (x, y, z, width, height, depth) or (x_min, y_min, z_min, x_max, y_max, z_max). These parameters define the position and size of the target object in 3D space.

[0052] Furthermore, the quality inspection specifically includes the following contents:

[0053] 1) Data integrity

[0054] Completeness check: ensure that all data samples are correctly labeled and there are no omissions or missing data;

[0055] Field completeness: Verify that the labeled data contains all necessary field information, such as category labels, bounding box coordinates, center point locations, etc.

[0056] 2) Labeling accuracy

[0057] Category accuracy: Check whether the labeled category matches the actual category of the target object to ensure that there is no category confusion or incorrect labeling;

[0058] Position accuracy: In addition to the position of the point, check whether the bounding box accurately selects the target object and whether there is any obvious offset or deviation;

[0059] 3) Logical consistency

[0060] Internal consistency: Check whether the annotation information of the same target object under different viewing angles is consistent, such as position, size, etc.;

[0061] External consistency: ensure that the annotation information matches the actual attributes and characteristics of the target object without logical conflicts or contradictions;

[0062] 4) Format standardization

[0063] File format: Check whether the file format of the annotation data meets the requirements;

[0064] Field naming and format: Verify whether the field naming is standardized and clear, and whether the data format is correct, such as integer, floating point number, string, etc.;

[0065] 5) Repeatability

[0066] Data reproducibility: Check whether there are repeated or redundant samples in the labeled data to avoid introducing bias during the training process;

[0067] Labeling repeatability: For the same target object, check whether there are multiple or repeated labels;

[0068] 6) Contextual relevance

[0069] Environmental relevance: Verify whether the labeled data takes into account the environment or contextual information of the target object, such as background, other objects, etc.

[0070] Logical relevance: Check whether the annotation information is consistent with the actual position and relationship of the target object in the scene, such as relative position, occlusion relationship, etc.

[0071] Furthermore, the operational process of the quality inspection is as follows: using automated tools to preliminarily screen and inspect the labeled data, identifying problematic data points or straight line segments, and recording and providing feedback on the problems found, and correcting and improving the labeled data.

[0072] The present invention provides an automatic target detection pre-marking method based on point cloud data, which has the following beneficial effects:

[0073] The pre-labeling method provided by the present invention makes full use of the three-dimensional spatial information of the point cloud data, and can more accurately describe the shape, size and position of the target, thereby helping to achieve accurate recognition and labeling of the target. Moreover, since the point cloud data is acquired by laser scanning and the like, even if the target is partially obscured, it can be detected and labeled through the remaining point cloud data, which improves the robustness and accuracy of the method. In addition, the present invention uses deep learning and other technologies to perform feature extraction, target detection, and pre-labeling on point cloud data to generate labeled data, and corrects and improves the labeled data through quality inspection, which not only realizes automatic recognition and labeling of the target, but also can achieve faster speed and higher accuracy, and can be well adapted to real-time application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of the operation flow of an automatic target detection and pre-labeling method based on point cloud data according to the present invention;

[0075] Figure 2 A schematic diagram of a data acquisition and preprocessing process of an automatic target detection and pre-labeling method based on point cloud data according to the present invention;

[0076] Figure 3 The present invention is a schematic diagram of a feature extraction process of an automatic target detection pre-labeling method based on point cloud data. DETAILED DESCRIPTION

[0077] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0078] like Figure 1-Figure 3 As shown, a method for automatic target detection and pre-labeling based on point cloud data includes the following steps:

[0079] S1. Data acquisition and preprocessing: The point cloud data describing the surface shape, posture, texture and other information of the object is obtained through the 3D scanning equipment. The point cloud data is a 3D data set represented in the form of points in a 3D coordinate system, and each point contains 3D coordinates (X, Y, Z). Some points contain color information or reflection intensity information. The acquisition methods are as follows:

[0080] LiDAR: LiDAR measures the distance between an object and the sensor by emitting laser pulses and receiving reflected signals. LiDAR provides high-precision and high-density point cloud data, which is suitable for large-scale terrain mapping and object modeling.

[0081] Structured light scanning: Projects a known pattern of light onto the surface of an object, captures the changes in the reflected light pattern through a camera, and calculates the three-dimensional shape of the object. It is suitable for accurate modeling of small and medium-sized objects.

[0082] Time-of-flight sensor: calculates distance by emitting light pulses and measuring the time it takes for the light to travel from the emission to the return, generating point cloud data suitable for 3D perception in mobile devices;

[0083] Stereo vision: Use two or more cameras to capture images of the same scene from different angles, calculate the depth information of the object by comparing the parallax in the images, and generate a three-dimensional point cloud, which is suitable for navigation and modeling in indoor environments;

[0084] Multi-view reconstruction: Take multiple photos of an object from different angles, use computer vision algorithms to combine these photos into a complete 3D model, and generate point cloud data, which is suitable for 3D modeling and reconstruction of large-scale scenes;

[0085] And preprocess the point cloud data, including the following operations:

[0086] Noise removal: used to remove noise points contained in point cloud data due to sensor errors, environmental interference, and signal attenuation. The methods include: statistical denoising (based on statistical principles, data is analyzed and modeled to identify and eliminate outliers or noise points. In this case, the local density or distance of each point in the point cloud is calculated, and a threshold is set to identify and remove isolated noise points), radius filtering (a search radius and a minimum number of neighboring points within the search radius are set. If the number of neighboring points of a point is less than the set minimum number, the point is considered to be a noise point and is removed; it is especially suitable for noise points in point cloud data caused by scanning unwanted objects or surface reflections; when using the radius filter in the PCL (Point Cloud Library) library, the search radius is set to 0.1, and the minimum number of neighboring points is set according to the point cloud's point spacing in the x, y, and z directions), Gaussian filtering (denoising is indirectly achieved by converting point cloud data into a depth image or surface mesh and then applying Gaussian filtering);

[0087] Adjusting data: used to adjust and optimize point cloud data to meet the needs of subsequent tasks. Methods include: downsampling (using random sampling or voxel gridding to reduce the number of points to simplify point cloud data), registration (using feature matching or ICP algorithm to align point cloud data from different sensors or scanning devices to the same coordinate system), smoothing (using moving least squares or Laplace smoothing to remove small fluctuations or unevenness in point cloud data caused by scanning equipment or environmental factors)

[0088] S2. Feature extraction: Select one of PointNet and its extensions, convolution-based methods, and graph neural networks according to the application scenario to extract useful features from point cloud data; the extraction logic of each method is as follows:

[0089] PointNet and its extensions (such as PointNet++, DGCNN, etc.): Directly extract features from point cloud data by sharing a multi-layer perceptron (MLP) and maximum pooling operations. The specific operation is: extract the local features of each point through a multi-layer perceptron, and finally obtain the global features through a maximum pooling operation;

[0090] Convolution-based methods: Apply convolution operations by converting point cloud data into a regular grid (such as voxelization) or an image (such as projection to a two-dimensional plane), and use convolutional neural networks to extract features of point cloud data;

[0091] Graph Neural Network (GNN): For point cloud data, each point is regarded as a node in the graph, and the graph structure is formed by building the connection relationship between the points (such as adjacency relationship). Then, the graph neural network is applied for feature extraction. The graph neural network can extract features by using the topological relationship in the point cloud data, which is particularly effective for processing point cloud data with complex structures.

[0092] The specific implementation of the feature extraction method is as follows:

[0093] 1) Surface feature extraction

[0094] Normal estimation: Calculate the normal vector of each point in the point cloud to describe the shape and geometry of the surface;

[0095] Curvature estimation: Describe the curvature of a surface by calculating the local curvature of a point;

[0096] 2) Local feature extraction

[0097] SHOT (Signature of Histograms of OrienTations): Describes the distribution and geometric structure of the surrounding points by calculating the feature vector of each point in the point cloud;

[0098] FPFH (Fast Point Feature Histograms): A method for quickly calculating point feature histograms to describe the local shape information of points;

[0099] 3) Global feature extraction

[0100] VFH (Viewpoint Feature Histogram): describes the overall shape and structure by projecting the point cloud into a two-dimensional histogram and counting the distribution of points in each bin;

[0101] S3. Target detection: Based on feature extraction, a deep learning model is used to detect targets in point cloud data to identify targets in images or scenes and determine their approximate locations. The specific steps are as follows:

[0102] Use VoxelNet or PointRCNN to generate candidate boxes, i.e. potential target areas;

[0103] After generating the candidate box, a convolutional neural network (CNN) or a recurrent neural network (RNN) is used to classify and regress the candidate box. The classification operation is used to determine whether the candidate box contains the target and give the target category. The regression operation is used to adjust the size and position of the candidate box to make it fit the actual target more accurately.

[0104] Use confidence threshold filtering to remove redundant candidate boxes and retain the most likely target detection results;

[0105] S4, pre-labeling: automatically generate labeling information of the target based on the target detection results; the labeling information specifically includes:

[0106] Category label: The category name to which the target object belongs. For example, in a car detection scenario, possible category labels include "car", "pedestrian", "bicycle", etc.

[0107] Bounding box: used to indicate the location and size of the target object in the image or point cloud data. It is represented by a rectangular box or a three-dimensional cube box and contains the location and size information of the object.

[0108] Confidence: Indicates the confidence of the model in the test results. The value range is [0, 1]. The higher the value, the higher the confidence of the model in the results.

[0109] Center point or center coordinates: mark the center point or center coordinates of the target object;

[0110] Rotation angle: For 3D target detection, it is used to indicate the rotation of the target object relative to a fixed axis or reference frame, especially when the target object changes its orientation;

[0111] Other attributes: According to the specific application scenarios and task requirements, other attribute information is marked, such as the color, material, speed, acceleration, etc. of the target object;

[0112] The bounding box includes the following two formats:

[0113] Two-dimensional bounding box: For image data, it is represented in the format of (x, y, width, height), where (x, y) is the coordinate of the upper left corner of the rectangle and (width, height) is the width and height of the rectangle;

[0114] 3D bounding box: For point cloud data, it is expressed in the format of (x, y, z, width, height, depth) or (x_min, y_min, z_min, x_max, y_max, z_max). These parameters define the position and size of the target object in 3D space.

[0115] S5. Quality check: Perform quality check on automatically generated annotation data. Use automated tools to preliminarily screen and check the annotation data, identify problematic data points or straight line segments, record and provide feedback on the problems found, and correct and improve the annotation data to ensure the accuracy and consistency of the annotation, including checking whether the position of the annotation points is accurate (superimpose the annotation data on the original image or point cloud data for visualization, and visually check whether the annotation points are accurately placed on the target object), whether the annotation lines are smooth (for the annotated straight line segments, the least squares method can be used to fit these points, and the error between the fitted straight line and the original annotation points can be calculated to evaluate the smoothness of the line; or the annotated straight line segments can be superimposed on the original image or point cloud data for visualization, and visually check whether the line is smooth and whether there are too many corners or fluctuations), etc.

[0116] The quality inspection specifically includes the following contents:

[0117] 1) Data integrity

[0118] Completeness check: ensure that all data samples are correctly labeled and there are no omissions or missing data;

[0119] Field completeness: Verify that the labeled data contains all necessary field information, such as category labels, bounding box coordinates, center point locations, etc.

[0120] 2) Labeling accuracy

[0121] Category accuracy: Check whether the labeled category matches the actual category of the target object to ensure that there is no category confusion or incorrect labeling;

[0122] Position accuracy: In addition to the position of the point, check whether the bounding box accurately selects the target object and whether there is any obvious offset or deviation;

[0123] 3) Logical consistency

[0124] Internal consistency: Check whether the annotation information of the same target object under different viewing angles is consistent, such as position, size, etc.;

[0125] External consistency: ensure that the annotation information matches the actual attributes and characteristics of the target object without logical conflicts or contradictions;

[0126] 4) Format standardization

[0127] File format: Check whether the file format of the annotation data meets the requirements;

[0128] Field naming and format: Verify whether the field naming is standardized and clear, and whether the data format is correct, such as integer, floating point number, string, etc.;

[0129] 5) Repeatability

[0130] Data reproducibility: Check whether there are repeated or redundant samples in the labeled data to avoid introducing bias during the training process;

[0131] Labeling repeatability: For the same target object, check whether there are multiple or repeated labels;

[0132] 6) Contextual relevance

[0133] Environmental relevance: Verify whether the labeled data takes into account the environment or contextual information of the target object, such as background, other objects, etc.

[0134] Logical relevance: Check whether the annotation information is consistent with the actual position and relationship of the target object in the scene, such as relative position, occlusion relationship, etc.

[0135] S6. Result output and visualization: Save and output the annotation results in a specified format (such as file, database, etc.), and select visualization tools such as VTuber and VTuber3D as needed to visualize the annotation results so that users can view and understand the annotation results.

[0136] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. An automatic target detection pre-labeling method based on point cloud data, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: acquiring point cloud data describing object information through a 3D scanning device and preprocessing the point cloud data; S2, Feature extraction: Select deep learning algorithms according to the application scenario to extract useful features from point cloud data; S3. Target detection: Based on feature extraction, deep learning models are used to detect targets in point cloud data to identify targets in images or scenes and determine their approximate locations. S4, pre-labeling: automatically generate target labeling information based on the target detection results; S5. Quality check: Perform quality check on the automatically generated annotation data to ensure the accuracy and consistency of the annotation; S6. Result output and visualization: Save and output the annotation results in a specified format, and select visualization tools as needed to visualize the annotation results; In step S2, the deep learning algorithms include PointNet and its extensions, convolution-based methods, and graph neural networks. The extraction logic of each method is as follows: PointNet and its extensions: Directly extract features from point cloud data by sharing multi-layer perceptrons and maximum pooling operations; the specific operation is: extract local features of each point through multi-layer perceptrons, and finally obtain global features through maximum pooling operations; Convolution-based methods: Apply convolution operations by converting point cloud data into a regular grid or image, and use convolutional neural networks to extract features of point cloud data; Graph neural network: For point cloud data, each point is regarded as a node in the graph, and the graph structure is formed by building the connection relationship between the points. Then, the graph neural network is applied for feature extraction. In step S2, the specific implementation of the feature extraction method is as follows: 1) Surface feature extraction Normal estimation: Calculate the normal vector of each point in the point cloud to describe the shape and geometry of the surface; Curvature estimation: Describe the curvature of a surface by calculating the local curvature of a point; 2) Local feature extraction SHOT: Describes the distribution and geometric structure of the surrounding points by calculating the feature vector of each point in the point cloud; FPFH: A method for quickly computing point feature histograms to describe the local shape information of points; 3) Global feature extraction VFH: Describes the overall shape and structure by projecting the point cloud into a two-dimensional histogram and counting the distribution of points in each bin.

2. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The point cloud data acquisition method is as follows: LiDAR: It measures the distance between an object and the sensor by emitting laser pulses and receiving reflected signals. It is suitable for large-scale terrain mapping and object modeling. Structured light scanning: Projects a known pattern of light onto the surface of an object, captures the changes in the reflected light pattern through a camera, and calculates the three-dimensional shape of the object. It is suitable for accurate modeling of small and medium-sized objects. Time-of-flight sensor: calculates distance by emitting light pulses and measuring the time it takes for the light to travel from the emission to the return, generating point cloud data suitable for 3D perception in mobile devices; Stereo vision: Use two or more cameras to capture images of the same scene from different angles, calculate the depth information of the object by comparing the parallax in the images, and generate a three-dimensional point cloud, which is suitable for navigation and modeling in indoor environments; Multi-view reconstruction: Take multiple photos of an object from different angles, use computer vision algorithms to synthesize these photos into a complete 3D model, and generate point cloud data, which is suitable for 3D modeling and reconstruction of large-scale scenes.

3. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The preprocessing specifically includes the following operations: Noise removal: used to remove noise points in point cloud data caused by sensor errors, environmental interference, and signal attenuation. Methods include: statistical denoising, radius filtering, and Gaussian filtering; Adjust data: used to adjust and optimize point cloud data to meet the needs of subsequent tasks. Methods include downsampling, registration, and smoothing.

4. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The specific steps of target detection are as follows: Generate candidate boxes, i.e. potential target regions, using voxel-based or point-based methods; After generating the candidate box, a convolutional neural network or a recurrent neural network is used to classify and regress the candidate box. The classification operation is used to determine whether the candidate box contains the target and give the target category. The regression operation is used to adjust the size and position of the candidate box to make it fit the actual target more accurately. Use non-maximum suppression and confidence threshold screening to remove redundant candidate boxes and retain the most likely target detection results.

5. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The annotation information specifically includes: Category label: the category name to which the target object belongs; Bounding box: used to indicate the location and size of the target object in the image or point cloud data. It is represented by a rectangular box or a three-dimensional cube box and contains the location and size information of the object. Confidence: Indicates the confidence of the model in the test results. The value range is [0, 1]. The higher the value, the higher the confidence of the model in the results. Center point or center coordinates: mark the center point or center coordinates of the target object; Rotation angle: For 3D target detection, it is used to indicate the rotation of the target object relative to a fixed axis or reference frame. It is applicable when the target object has a change in orientation.

6. The method for automatic target detection and pre-marking based on point cloud data according to claim 5, characterized in that: The bounding box includes the following two formats: Two-dimensional bounding box: For image data, it is represented in the format of (x, y, width, height), where (x, y) is the coordinate of the upper left corner of the rectangle and (width, height) is the width and height of the rectangle; 3D bounding box: For point cloud data, it is expressed in the format of (x, y, z, width, height, depth) or (x_min, y_min, z_min, x_max, y_max, z_max). These parameters define the position and size of the target object in 3D space.

7. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The quality inspection specifically includes the following contents: 1) Data integrity Completeness check: ensure that all data samples are correctly labeled and there are no omissions or missing data; Field completeness: Verify that the labeled data contains all necessary field information; 2) Labeling accuracy Category accuracy: Check whether the labeled category matches the actual category of the target object to ensure that there is no category confusion or incorrect labeling; Position accuracy: In addition to the position of the point, check whether the bounding box accurately selects the target object and whether there is any obvious offset or deviation; 3) Logical consistency Internal consistency: Check whether the annotation information of the same target object under different viewing angles is consistent; External consistency: ensure that the annotation information matches the actual attributes and characteristics of the target object without logical conflicts or contradictions; 4) Format standardization File format: Check whether the file format of the annotation data meets the requirements; Field naming and format: Verify whether the field naming is standardized and clear, and whether the data format is correct; 5) Repeatability Data reproducibility: Check whether there are repeated or redundant samples in the labeled data to avoid introducing bias during the training process; Labeling repeatability: For the same target object, check whether there are multiple or repeated labels; 6) Contextual relevance Environmental relevance: Verify whether the labeled data takes into account the environment or contextual information of the target object; Logical relevance: Check whether the annotation information is consistent with the actual position and relationship of the target object in the scene.

8. The method for automatic target detection and pre-marking based on point cloud data according to claim 1, characterized in that: The operational process of the quality inspection is as follows: use automated tools to preliminarily screen and inspect the labeled data, identify problematic data points or straight line segments, record and provide feedback on the problems found, and correct and improve the labeled data.

Images