3D point cloud segmentation target detection system

Abstract

The invention, which relates to the technical field of 3D point cloud segmentation, discloses a 3D point cloud segmentation target detection system comprising a data preprocessing module, a feature extraction module, a segmentation module, a target detection module, a post-processing module, a result verification module and a result output module. According to the 3D point cloud segmentation target detection system, filtering, down-sampling and enhancement processing are carried out on 3D point cloud data through the data preprocessing module, so that the influence of sparsity of the point cloud data is weakened, enough feature information can be provided even in a long-distance and complex environment, segmentation and detection of 3D point cloud are facilitated, and meanwhile, the detection efficiency is improved. The feature information of the target object can be rapidly extracted through local feature extraction and global feature fusion, and the detection precision and efficiency can be well balanced in cooperation with a targeted segmentation algorithm and a target detection technology.

Description

Technical Field

[0001] The present invention relates to the technical field of 3D point cloud segmentation, and in particular to a 3D point cloud segmentation target detection system. Background Art

[0002] 3D point cloud is a data set consisting of a series of points sampled in three-dimensional space. Each point is defined by three coordinate values: x, y, and z, representing a specific position in space. The advantage of point cloud is that it can provide rich geometric information for modeling and analyzing three-dimensional scenes; the disadvantage is that point cloud data is usually very large, and the processing and storage costs are high. Therefore, it is necessary to segment the point cloud data into multiple parts or regions according to certain rules or standards through 3D point cloud segmentation. Each part or region represents a meaningful object or structure, which helps to extract the target object in the point cloud and provides a basis for subsequent analysis and processing.

[0003] On the basis of segmentation, 3D point cloud target detection aims to identify the target objects in the point cloud and detect the position, size, posture and other information of these objects. Traditional 3D point cloud segmentation target detection methods usually rely on technologies such as geometric shape analysis, edge detection and model fitting. Although these methods can achieve segmentation and detection tasks to a certain extent, they have slow processing speeds and limited effects on complex scenes and irregularly shaped objects. In recent years, deep learning technology has made significant progress in the field of 3D point cloud segmentation target detection. Deep learning-based methods extract features from point clouds by training neural networks and use these features for segmentation and detection. These methods excel in processing speed, accuracy and robustness and are currently the mainstream technology.

[0004] However, in actual operations, since 3D point cloud data is usually sparse, especially in long distances or complex environments, effective point cloud data is even scarcer, and sparse data may not provide sufficient feature information, which increases the difficulty of segmentation and detection. In addition, due to the disorder and irregularity of point cloud data, existing deep learning algorithms often face challenges in computational efficiency when processing large-scale point cloud data. Although some methods such as sparse convolution and voxelization attempt to improve processing efficiency, they still cannot strike a good balance between the accuracy and efficiency of target detection.

[0005] Therefore, it is urgent to improve this shortcoming. The present invention studies and improves the existing technology and its shortcomings, and provides a 3D point cloud segmentation target detection system. Summary of the invention

[0006] The purpose of the present invention is to provide a 3D point cloud segmentation target detection system to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a 3D point cloud segmentation target detection system, comprising:

[0008] A data preprocessing module, which is responsible for preprocessing the original 3D point cloud data to improve the effect and efficiency of subsequent processing;

[0009] A feature extraction module, which is responsible for extracting useful feature information from the preprocessed point cloud data to provide a basis for subsequent segmentation and detection tasks;

[0010] A segmentation module, which is responsible for segmenting the point cloud data into a number of parts or regions, each of which represents an object or structure with clear semantics;

[0011] A target detection module, which is responsible for identifying and locating the target object in the segmented point cloud data;

[0012] A post-processing module, which is responsible for post-processing the target detection results to improve the accuracy and robustness of the detection;

[0013] A result verification module, which is responsible for detecting and verifying the 3D point cloud segmentation target detection results to ensure that the detected target position, size, category and other information are accurate;

[0014] The result output module is responsible for presenting the final results of the verified 3D point cloud segmentation target detection to the user in an appropriate manner.

[0015] Furthermore, the data preprocessing module specifically includes the following submodules:

[0016] Point cloud filtering submodule: used to filter the original point cloud data, remove noise points and useless information, retain key features, and improve the quality of point cloud data;

[0017] Point cloud downsampling submodule: In view of the sparsity of point cloud data, distance-based farthest point sampling (D-FPS) or semantic feature-based farthest point sampling (F-FPS) is used to reduce the amount of data and improve processing efficiency. By fusing D-FPS and F-FPS, it ensures that there are enough foreground points for regression and enough background points for classification.

[0018] Point cloud enhancement submodule: used to perform enhancement processing such as rotation, translation, and scaling on point cloud data to improve the generalization ability of the model.

[0019] Furthermore, the processing logic of the point cloud downsampling submodule includes:

[0020] Distance-based Farthest Point Sampling (D-FPS): Select an initial point from the point cloud, and then gradually select the points with the farthest distance from the selected point set as new sampling points until the required number of sampling points is reached; when processing sparse point cloud data, D-FPS can ensure that the sampling points can be evenly distributed throughout the point cloud, reducing data redundancy while retaining the key features of the point cloud;

[0021] Farthest Point Sampling (F-FPS) based on semantic features: Different from D-FPS, F-FPS considers not only the distance between points but also the semantic features of points during the sampling process. Specifically, F-FPS calculates the distance between points based on the feature vectors of points (such as color, texture, shape, etc.), thereby selecting points with different semantic features as sampling points. It is particularly suitable for processing point cloud data with rich semantic information, and can ensure that the sampling points contain objects or structures of different categories, thereby improving the accuracy of classification tasks. In sparse point cloud data, F-FPS can retain more foreground points, which usually have more important semantic information and are very helpful for regression tasks.

[0022] Fusion of D-FPS and F-FPS: During the sampling process, both the distance between points and the semantic features of the points are considered, so that points that are both representative and contain rich semantic information are selected as sampling points; this fusion method can make full use of the advantages of D-FPS and F-FPS, improve the diversity and representativeness of the sampling points, and thus improve the performance of the entire 3D point cloud segmentation object detection system.

[0023] Furthermore, the processing flow of the feature extraction module is as follows:

[0024] Local feature extraction: using convolutional neural network (CNN) or point cloud neural network (PointNet), graph convolutional network (GCN) and other deep learning models to extract local features of point cloud data, the local features include point cloud coordinates (three-dimensional coordinates of points), normals (reflecting the surface shape and direction of point cloud data), low-level features such as color, and high-level semantic features learned by deep learning models; in this embodiment, convolutional neural network (CNN) automatically learns the feature representation of point cloud data through structures such as convolutional layer, pooling layer, and fully connected layer; when processing point cloud data, CNN can convert point cloud data into a regular three-dimensional voxel grid or project it into a two-dimensional image, and then apply standard convolution operations to extract features; in addition to CNN, other deep learning models such as point cloud neural network (PointNet) and graph convolutional network (GCN) can also be used for local feature extraction of point cloud data, and these models can be customized and optimized according to the characteristics of point cloud data to extract more accurate and rich feature information;

[0025] Global feature fusion: Use direct concatenation (directly concatenate the extracted local features to form a large feature vector), weighted averaging (weighted averaging the extracted local features to obtain a global feature vector) or attention mechanism (use the attention mechanism to weight local features so that the model can pay more attention to important local features) to fuse local features to form a global feature representation to capture the overall structure and semantic information of point cloud data, which helps the model better capture the overall structure and semantic information of point cloud data.

[0026] Furthermore, the segmentation module uses the feature information obtained from the feature extraction module and adopts a feature-based segmentation algorithm to perform fine segmentation processing on the point cloud data.

[0027] Furthermore, the feature-based segmentation algorithm performs segmentation based on the geometric shape, color, texture and other features of the point cloud, including edge-based segmentation and region-based segmentation;

[0028] The edge-based segmentation algorithm relies on the detection of edge features in the point cloud. The segmentation area is obtained by detecting the boundaries of certain areas of the point cloud. The principle is to locate the intensity changes of edge points. The specific operations are: fitting line segments by calculating gradients, detecting changes in the direction of unit normal vectors on the surface, or performing fast segmentation based on grouping of scan lines;

[0029] The region-based segmentation algorithm classifies nearby points with similar attributes by using neighborhood information to obtain segmented regions and distinguish the differences between different regions.

[0030] Furthermore, the region-based methods are divided into two categories: seeded region (bottom-up) methods and non-seed region (top-down) methods;

[0031] Seed region method: start segmentation by selecting multiple seed points, and gradually form point cloud regions by adding neighboring points of the seeds starting from these seed points;

[0032] Non-seed region method: After all points are grouped into a region, the subdivision process is started to divide it into smaller regions.

[0033] Furthermore, the target detection module specifically includes the following submodules:

[0034] Anchor point generation submodule: Generates a series of candidate regions (i.e., anchor points) based on the segmentation results. These regions may contain target objects. Specifically, it analyzes the distribution characteristics of point cloud data, determines the potential object locations, and generates anchor points with different scales and aspect ratios based on the scale and shape information of the object.

[0035] Bounding box regression submodule: responsible for fine-tuning the anchor points to more accurately predict the position and boundary of the target object. Specifically, by comparing the position difference (such as offset) between the anchor points and the actual target object, a convolutional neural network is used to learn how to map this position difference to the adjustment parameters of the bounding box. By adjusting the position, scale and aspect ratio of the anchor points, a more accurate bounding box is obtained.

[0036] Classification evaluation submodule: classifies the candidate areas, determines whether they belong to the target object, and gives corresponding confidence scores, which helps to screen out the real target objects and reduce the false detection rate.

[0037] Furthermore, the post-processing module specifically includes the following post-processing operations:

[0038] Remove duplicate detection: Identify and remove duplicate detection results by comparing the location, shape, size and other information of different detection results, thereby avoiding redundancy and errors;

[0039] Non-maximum suppression (NMS): According to the confidence and overlap of the detection results, the optimal detection result is selected and other overlapping detection results are suppressed, thereby reducing false detection and missed detection;

[0040] Bounding box adjustment: fine-tune the bounding box, including the position, scale, and orientation of the bounding box, so that it more accurately surrounds the target object;

[0041] Category re-evaluation: Re-evaluate the classification results using additional information (such as context information, shape information, etc.) to improve the accuracy of classification;

[0042] Point cloud segmentation refinement: Through further point cloud segmentation refinement operations, the target object is more accurately separated from the background to improve the accuracy of detection;

[0043] Target tracking and association: In the point cloud data of videos or continuous frames, the target tracking algorithm is used to associate the target objects in different frames, which helps to maintain the consistency of the target objects in continuous frames and improve the stability of detection.

[0044] Furthermore, the result output module specifically includes the following submodules:

[0045] Result formatting submodule: converts the target information detected by the system into a format that is easy for users to understand and use. The target information includes the location, size, category, confidence level, etc. of the target;

[0046] Result visualization submodule: Use visualization techniques, such as 3D rendering, charts or reports, to present the targets detected by the system in an intuitive way;

[0047] Storage and export submodule: stores the detected target information in a database or file for subsequent analysis and processing, and provides an export function that allows users to export the results in a specific file format (such as OBJ, PLY or STL) to external devices or software to achieve information sharing and integration;

[0048] Report generation submodule: Automatically generates detailed reports based on the detection results, including the location, size, category, confidence level and other information of the target, and provides report templates and customization options, enabling users to generate reports in different formats as needed.

[0049] The present invention provides a 3D point cloud segmentation target detection system, which has the following beneficial effects:

[0050] The present invention filters, downsamples and enhances 3D point cloud data through a data preprocessing module, thereby weakening the impact of the sparsity of point cloud data, so that it can provide sufficient feature information even in long distances and complex environments, which is beneficial to the segmentation and detection of 3D point clouds. At the same time, through local feature extraction and global feature fusion, the feature information of the target object can be quickly extracted, and in conjunction with targeted segmentation algorithms and target detection technology, the detection accuracy and efficiency can be well balanced. In addition, a post-processing module is also provided to further process the target detection results, and cooperate with the result verification module for detection and verification, which is beneficial to further improve the accuracy of the target detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A logic block diagram of a 3D point cloud segmentation target detection system of the present invention;

[0052] Figure 2 It is a schematic diagram of the operation flow of a data preprocessing module-feature extraction module of a 3D point cloud segmentation target detection system of the present invention;

[0053] Figure 3 A schematic diagram of a segmentation algorithm of a 3D point cloud segmentation target detection system of the present invention;

[0054] Figure 4 Schematic diagram of a post-processing module of a 3D point cloud segmentation target detection system of the present invention. DETAILED DESCRIPTION

[0055] 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.

[0056] like Figure 1-Figure 4 As shown, a 3D point cloud segmentation target detection system includes:

[0057] Data preprocessing module: responsible for preprocessing the original 3D point cloud data to improve the effect and efficiency of subsequent processing, and the data preprocessing module specifically includes the following submodules:

[0058] Point cloud filtering submodule: used to filter the original point cloud data, remove noise points and useless information, retain key features, and improve the quality of point cloud data;

[0059] Point cloud downsampling submodule: In view of the sparsity of point cloud data, the farthest point sampling (D-FPS) based on distance or the farthest point sampling (F-FPS) based on semantic features is used to reduce the amount of data and improve processing efficiency. By fusing D-FPS and F-FPS, it is ensured that there are enough foreground points for regression and enough background points for classification. The processing logic of the point cloud downsampling submodule includes:

[0060] Distance-based Farthest Point Sampling (D-FPS): Select an initial point from the point cloud, and then gradually select the points with the farthest distance from the selected point set as new sampling points until the required number of sampling points is reached; when processing sparse point cloud data, D-FPS can ensure that the sampling points can be evenly distributed throughout the point cloud, reducing data redundancy while retaining the key features of the point cloud;

[0061] Farthest Point Sampling (F-FPS) based on semantic features: Different from D-FPS, F-FPS considers not only the distance between points but also the semantic features of points during the sampling process. Specifically, F-FPS calculates the distance between points based on the feature vectors of points (such as color, texture, shape, etc.), thereby selecting points with different semantic features as sampling points. It is particularly suitable for processing point cloud data with rich semantic information, and can ensure that the sampling points contain objects or structures of different categories, thereby improving the accuracy of classification tasks. In sparse point cloud data, F-FPS can retain more foreground points, which usually have more important semantic information and are very helpful for regression tasks.

[0062] Fusion of D-FPS and F-FPS: In the sampling process, the distance between points and the semantic features of the points are considered at the same time, so as to select points that are both representative and contain rich semantic information as sampling points; this fusion method can make full use of the advantages of D-FPS and F-FPS, improve the diversity and representativeness of the sampling points, and thus improve the performance of the entire 3D point cloud segmentation object detection system;

[0063] In this embodiment, the distribution of sampling points and the richness of semantic features are controlled by adjusting the weights of D-FPS and F-FPS, that is, the weight of F-FPS is increased to retain more foreground points, or the weight of D-FPS is increased to ensure uniform distribution of sampling points; in addition, the sampling process can be further optimized by introducing a dynamic stop mechanism and an adaptive sampling threshold;

[0064] Point cloud enhancement submodule: used to perform enhancement processing such as rotation, translation, and scaling on point cloud data to improve the generalization ability of the model;

[0065] Feature extraction module: responsible for extracting useful feature information from the preprocessed point cloud data, providing a basis for subsequent segmentation and detection tasks, and the processing flow of the feature extraction module is as follows:

[0066] Local feature extraction: using convolutional neural network (CNN) or other deep learning models such as point cloud neural network (PointNet) and graph convolutional network (GCN) to extract local features of point cloud data, including low-level features such as point cloud coordinates (three-dimensional coordinates of points), normals (reflecting the surface shape and direction of point cloud data), and colors, as well as high-level semantic features learned by deep learning models; in this embodiment, convolutional neural network (CNN) automatically learns feature representation of point cloud data through structures such as convolutional layers, pooling layers, and fully connected layers; when processing point cloud data, CNN can convert point cloud data into a regular three-dimensional voxel grid or project it into a two-dimensional image, and then apply standard convolution operations to extract features; in addition to CNN, other deep learning models such as point cloud neural network (PointNet) and graph convolutional network (GCN) can also be used for local feature extraction of point cloud data, and these models can be customized and optimized according to the characteristics of point cloud data to extract more accurate and rich feature information;

[0067] Global feature fusion: Use direct concatenation (directly concatenate the extracted local features to form a large feature vector), weighted averaging (weighted averaging the extracted local features to obtain a global feature vector) or attention mechanism (using the attention mechanism to weight local features so that the model can pay more attention to important local features) to fuse local features to form a global feature representation to capture the overall structure and semantic information of the point cloud data, which helps the model better capture the overall structure and semantic information of the point cloud data;

[0068] Segmentation module: responsible for segmenting the point cloud data into several parts or regions, each of which represents an object or structure with clear semantics;

[0069] In this embodiment, the segmentation module uses the feature information obtained from the feature extraction module and adopts a feature-based segmentation algorithm to perform fine segmentation processing on the point cloud data. Specifically, the feature-based segmentation algorithm performs segmentation according to the geometric shape, color, texture and other features of the point cloud, including edge-based segmentation and region-based segmentation;

[0070] 1) Edge-based segmentation algorithms rely on the detection of edge features in point clouds. They obtain segmentation regions by detecting the boundaries of certain areas of the point cloud. The principle is to locate the intensity changes of edge points. The specific operations are: fitting line segments by calculating gradients, detecting changes in the direction of unit normal vectors on the surface, or performing fast segmentation based on grouping of scan lines;

[0071] 2) Region-based segmentation algorithms use neighborhood information to classify nearby points with similar attributes to obtain segmented regions and distinguish the differences between different regions. Region-based methods are divided into two categories: seed region (bottom-up) methods and non-seed region (top-down) methods.

[0072] Seed region method: start segmentation by selecting multiple seed points, and gradually form point cloud regions by adding neighboring points of the seeds starting from these seed points;

[0073] Non-seed region method: After all points are grouped into a region, the subdivision process is started to divide it into smaller regions;

[0074] The target detection module is responsible for identifying and locating the target object in the segmented point cloud data, and the target detection module specifically includes the following submodules:

[0075] Anchor point generation submodule: Generates a series of candidate regions (i.e., anchor points) based on the segmentation results. These regions may contain target objects. Specifically, it analyzes the distribution characteristics of point cloud data, determines the potential object locations, and generates anchor points with different scales and aspect ratios based on the scale and shape information of the object.

[0076] Bounding box regression submodule: responsible for fine-tuning the anchor points to more accurately predict the position and boundary of the target object. Specifically, by comparing the position difference (such as offset) between the anchor points and the actual target object, a convolutional neural network is used to learn how to map this position difference to the adjustment parameters of the bounding box. By adjusting the position, scale and aspect ratio of the anchor points, a more accurate bounding box is obtained.

[0077] Classification and evaluation submodule: classifies candidate regions to determine whether they belong to target objects and gives corresponding confidence scores, which helps to screen out real target objects and reduce false detection rates;

[0078] The post-processing module is responsible for post-processing the target detection results. In this embodiment, the post-processing operations include:

[0079] Remove duplicate detection: Identify and remove duplicate detection results by comparing the location, shape, size and other information of different detection results, thereby avoiding redundancy and errors;

[0080] Non-maximum suppression (NMS): According to the confidence and overlap of the detection results, the optimal detection result is selected and other overlapping detection results are suppressed, thereby reducing false detection and missed detection;

[0081] Bounding box adjustment: fine-tune the bounding box, including the position, scale, and orientation of the bounding box, so that it more accurately surrounds the target object;

[0082] Category re-evaluation: Re-evaluate the classification results using additional information (such as context information, shape information, etc.) to improve the accuracy of classification;

[0083] Point cloud segmentation refinement: Through further point cloud segmentation refinement operations, the target object is more accurately separated from the background to improve the accuracy of detection;

[0084] Target tracking and association: In the point cloud data of videos or consecutive frames, target tracking algorithms are used to associate target objects in different frames, which helps to maintain the consistency of target objects in consecutive frames and improve the stability of detection;

[0085] Result verification module, which is responsible for detecting and verifying the 3D point cloud segmentation target detection results to ensure that the detected target position, size, category and other information are accurate;

[0086] In this embodiment, the specific operation of target detection verification is: using known ground truth data (GroundTruth) or test data sets to verify the targets detected by the system, and by comparing the differences between the position, size and category of the targets detected by the system and the real data, the accuracy of the system is evaluated;

[0087] The result output module is responsible for presenting the final results of the verified 3D point cloud segmentation target detection to the user in an appropriate manner, and the result output module specifically includes the following submodules:

[0088] Result formatting submodule: converts the target information detected by the system into a format that is easy for users to understand and use. The target information includes the location, size, category, confidence level, etc. of the target;

[0089] Result visualization submodule: Use visualization techniques, such as 3D rendering, charts or reports, to present the targets detected by the system in an intuitive way;

[0090] Storage and export submodule: stores the detected target information in a database or file for subsequent analysis and processing, and provides an export function that allows users to export the results in a specific file format (such as OBJ, PLY or STL) to external devices or software to achieve information sharing and integration;

[0091] Report generation submodule: Automatically generates detailed reports based on the detection results, including the location, size, category, confidence level and other information of the target, and provides report templates and customization options, enabling users to generate reports in different formats as needed.

[0092] 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. A 3D point cloud segmentation target detection system, characterized in that: include: Data preprocessing module, feature extraction module, segmentation module, target detection module, post-processing module, result verification module and result output module; The data preprocessing module is responsible for preprocessing the original 3D point cloud data; The feature extraction module is responsible for extracting useful feature information from the preprocessed point cloud data; The segmentation module is responsible for segmenting the point cloud data into several parts or regions, each part or region represents an object or structure with clear semantics; The target detection module is responsible for identifying and locating the target object in the segmented point cloud data; The post-processing module is responsible for post-processing the target detection results to improve the accuracy and robustness of the detection; The result verification module is responsible for detecting and verifying the 3D point cloud segmentation target detection results to ensure that the detected target information is accurate; The result output module is responsible for presenting the verified final result of 3D point cloud segmentation target detection to the user in an appropriate manner.

2. A 3D point cloud segmentation target detection system according to claim 1, characterized in that: The data preprocessing module specifically includes the following submodules: Point cloud filtering submodule: used to filter the original point cloud data, remove noise points and useless information, retain key features, and improve the quality of point cloud data; Point cloud downsampling submodule: In view of the sparsity of point cloud data, the farthest point sampling based on distance or the farthest point sampling based on semantic features is adopted to reduce the amount of data and improve processing efficiency. By fusing D-FPS and F-FPS, it is ensured that there are enough foreground points for regression and enough background points for classification. Point cloud enhancement submodule: used to rotate, translate, and scale point cloud data to improve the generalization ability of the model.

3. A 3D point cloud segmentation target detection system according to claim 2, characterized in that: The processing logic of the point cloud downsampling submodule includes: Distance-based farthest point sampling: select an initial point from the point cloud, and then gradually select the points farthest from the selected point set as new sampling points until the required number of sampling points is reached; Farthest point sampling based on semantic features: The distance between points is calculated based on the feature vectors of the points, so that points with different semantic features are selected as sampling points. This method is especially suitable for processing point cloud data with rich semantic information. Fusion of D-FPS and F-FPS: In the sampling process, both the distance between points and the semantic features of the points are considered, so that points that are both representative and contain rich semantic information are selected as sampling points.

4. The 3D point cloud segmentation target detection system according to claim 1, characterized in that: The processing flow of the feature extraction module is as follows: Local feature extraction: using convolutional neural networks, point cloud neural networks, or graph convolutional networks to extract local features of point cloud data, including low-level features and high-level semantic features; Global feature fusion: Local features are fused by direct concatenation, weighted averaging or attention mechanism to form a global feature representation to capture the overall structure and semantic information of point cloud data.

5. The 3D point cloud segmentation target detection system according to claim 1, characterized in that: The segmentation module uses the feature information obtained from the feature extraction module and adopts a feature-based segmentation algorithm to perform fine segmentation processing on the point cloud data.

6. A 3D point cloud segmentation target detection system according to claim 5, characterized in that: The feature-based segmentation algorithm performs segmentation according to the features of the point cloud, including edge-based segmentation and region-based segmentation; The edge-based segmentation algorithm relies on the detection of edge features in the point cloud, and obtains the segmentation area by detecting the boundaries of certain areas of the point cloud. The specific operations are: fitting line segments by calculating gradients, detecting changes in the direction of unit normal vectors on the surface, or performing fast segmentation based on grouping of scan lines; The region-based segmentation algorithm classifies nearby points with similar attributes by using neighborhood information to obtain segmented regions and distinguish the differences between different regions.

7. A 3D point cloud segmentation target detection system according to claim 6, characterized in that: The region-based methods are divided into two categories: seeded region methods and non-seed region methods; Seed region method: start segmentation by selecting multiple seed points, and gradually form point cloud regions by adding neighboring points of the seeds starting from these seed points; Non-seed region method: After all points are grouped into a region, the subdivision process is started to divide it into smaller regions.

8. The 3D point cloud segmentation target detection system according to claim 1, characterized in that: The target detection module specifically includes the following submodules: Anchor point generation submodule: Generates a series of candidate regions, namely anchor points, based on the segmentation results. Specifically, it analyzes the distribution characteristics of point cloud data, determines the potential location of objects, and generates anchor points with different scales and aspect ratios based on the scale and shape information of the objects. Bounding box regression submodule: responsible for fine-tuning the anchor points to more accurately predict the position and boundary of the target object. Specifically, by comparing the position difference between the anchor points and the actual target object, a convolutional neural network is used to learn how to map this position difference to the adjustment parameters of the bounding box. By adjusting the position, scale and aspect ratio of the anchor points, a more accurate bounding box is obtained. Classification evaluation submodule: classifies the candidate area, determines whether it belongs to the target object, and gives a corresponding confidence score.

9. The 3D point cloud segmentation target detection system according to claim 1, characterized in that: The post-processing module specifically includes the following post-processing operations: Remove duplicate detection: Identify and remove duplicate detection results by comparing the location, shape and size information of different detection results; Non-maximum suppression: According to the confidence and overlap of the detection results, the optimal detection result is selected and other overlapping detection results are suppressed; Bounding box adjustment: fine-tune the bounding box to more accurately surround the target object; Category re-evaluation: re-evaluating the classification results using additional information; Point cloud segmentation refinement: Through further point cloud segmentation refinement operations, the target object is separated from the background more accurately; Target tracking and association: In point cloud data of videos or consecutive frames, target tracking algorithms are used to associate target objects in different frames.

10. The 3D point cloud segmentation target detection system according to claim 1, characterized in that: The result output module specifically includes the following submodules: Result formatting submodule: converts the target information detected by the system into a format that is easy for users to understand and use. The target information includes the location, size, category, and confidence level of the target. Result visualization submodule: Use visualization techniques, such as 3D rendering, charts, or reports, to present the targets detected by the system in an intuitive way; Storage and export submodule: stores the detected target information in a database or file for subsequent analysis and processing, and provides an export function that allows users to export the results to external devices or software in a specific file format; Report generation submodule: Automatically generates detailed reports based on the detection results, and provides report templates and customization options.

Images