Traffic scene-oriented multi-modal point cloud semantic segmentation method and device

Abstract

The invention relates to a computer vision technology, and discloses a traffic scene-oriented multi-modal point cloud semantic segmentation method, which comprises the following steps: acquiring three-dimensional point cloud data and RGB image data, and querying a corresponding relationship between pixel points and point cloud to obtain a pixel-point cloud mapping relationship; converting the three-dimensional point cloud data into a voxel grid view and a range plane view, and extracting point cloud features of the three-dimensional point cloud data, extracting image features of the image data, extracting range plane view features and extracting voxel features; and summarizing the point cloud features, the range plane view features and the voxel features to obtain multi-view fusion features, performing feature fusion on the multi-view fusion features and the image features to obtain multi-modal fusion features, and identifying objects in the traffic scene according to the multi-modal fusion features. The invention further provides a traffic scene-oriented multi-modal point cloud semantic segmentation device, electronic equipment and a storage medium. According to the invention, effective identification and detection of objects in a traffic scene can be ensured.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a multimodal point cloud semantic segmentation method and device for traffic scenes. Background Art

[0002] As autonomous driving technology places increasingly higher demands on environmental perception, accurate traffic scene perception can significantly improve the safety and reliability of autonomous driving systems. Implementing point-by-point semantic segmentation in 3D point cloud data is one of the core tasks in this field, and it is also a fundamental and challenging problem that has long existed in computer vision and computer graphics. The main goal of 3D semantic segmentation is to develop computational methods that can predict the semantic labels of each point in a 3D scene, which is widely used in autonomous driving, mobile robots, industrial control, augmented reality, and medical image analysis.

[0003] In recent years, single-input point cloud segmentation methods (such as deep learning-based techniques) have made significant progress in semantic segmentation at the point-by-point level, especially in processing three-dimensional point cloud data. However, with the increasing demand for higher accuracy and robustness in complex scenes and dynamic environments, multimodal point cloud segmentation techniques have emerged. Multimodal methods enhance the understanding of the scene by fusing data from different sensors (such as lidar and cameras), providing richer geometry, color, and texture information, and overcoming the limitations of single-modality methods in dealing with sparse point clouds and occlusions. Compared with single-input methods, multimodal point cloud segmentation utilizes the complementary information of multiple modalities to improve segmentation accuracy and robustness. It is suitable for complex application scenarios such as autonomous driving and avoids the problem of incomplete point cloud information.

[0004] Although multimodal point cloud segmentation has shown good potential in road scenes, the following problems still exist:

[0005] The feature types provided by different sensors (such as lidar and cameras) vary greatly, and how to effectively extract and fuse information from these modalities remains a challenge. When processing these multimodal data, existing methods often fail to fully utilize the information or interfere with each other, affecting the final segmentation effect. In addition, multimodal point cloud segmentation relies on the intersection of views from different sensors, but due to the limited viewing angle coverage of each sensor, it cannot always be guaranteed that all areas are covered by multimodal data. This perspective inconsistency is particularly evident at long distances or in occluded conditions, affecting perception accuracy. In addition, existing multimodal methods perform poorly in identifying small objects and distant objects in areas where point cloud data is relatively sparse. This defect will result in small obstacles or distant objects in traffic scenes that cannot be effectively detected and segmented, affecting the safety of the system. Summary of the invention

[0006] The present invention provides a multimodal point cloud semantic segmentation method and device for traffic scenes, which can ensure effective recognition and detection of objects in traffic scenes.

[0007] To achieve the above object, the present invention provides a multimodal point cloud semantic segmentation method for traffic scenes, comprising:

[0008] Acquire 3D point cloud data and RGB image data in a traffic scene, use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the 3D point cloud data, and obtain a pixel-point cloud mapping relationship;

[0009] The three-dimensional point cloud data is converted into a voxel grid view and a range plane view respectively, and geometric features of the three-dimensional point cloud data are extracted to obtain point cloud features, image features of the RGB image data are extracted, image features of the range plane view are extracted to obtain range plane view features, and geometric features of the voxel grid view are extracted to obtain voxel features;

[0010] The point cloud features, the range plane view features and the voxel features are summarized to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, the multi-view fusion features are fused with the image features using a multimodal feature fusion algorithm to obtain multimodal fusion features, and objects in the traffic scene are identified according to the multimodal fusion features.

[0011] Optionally, before using the pre-constructed multi-view index to query the correspondence between the pixel points in the RGB image data and the point cloud in the three-dimensional point cloud data, it also includes constructing the multi-view index.

[0012] Optionally, the constructing a multi-view index includes:

[0013] Projecting the point cloud in the three-dimensional point cloud data to a target view according to a preset projection function to obtain a target view projection mapping, wherein the target view includes a voxel grid view and a range plane view;

[0014] Establishing a mapping between the target view and the three-dimensional point cloud data according to a preset hash mapping to obtain a hash view mapping;

[0015] The multi-view index is obtained according to the target view projection mapping and the hash view mapping.

[0016] Optionally, obtaining the multi-view index according to the target view projection mapping and the hash view mapping further includes multi-view feature propagation correction.

[0017] Optionally, the multi-view feature propagation correction includes:

[0018] The point cloud features of the three-dimensional point cloud data are projected to the target view using a weighted average algorithm, and the features of the target view are mapped and calculated using a linear interpolation algorithm. Among them, the features of the voxel grid view returned to the three-dimensional point cloud data are mapped and calculated using a trilinear interpolation algorithm, and the features of the range plane view returned to the three-dimensional point cloud data are mapped and calculated using a bilinear interpolation method.

[0019] Optionally, the step of fusing the multi-view fusion feature with the image feature using a multi-modal feature fusion algorithm according to the pixel-point cloud mapping relationship to obtain the multi-modal fusion feature comprises:

[0020] According to the geometric relationship between the laser radar that collects the 3D point cloud data and the multi-camera that collects the RGB image data, the points in the 3D point cloud data are aligned with the pixels in the RGB image data;

[0021] After the alignment process is completed, the multi-view fusion feature and the image feature are mapped to the same feature space, and the multi-view fusion feature and the image feature are spliced ​​to obtain a multimodal fusion feature.

[0022] Optionally, when splicing the multi-view fusion features with the image features, it also includes using a pre-built gated fusion algorithm to evaluate the importance of the multi-view fusion features and the image features to obtain an importance evaluation result; and performing weighted fusion of the multi-view fusion features and the image features according to the importance evaluation result.

[0023] In order to solve the above problems, the present invention also provides a multimodal point cloud semantic segmentation device for traffic scenes, the device comprising:

[0024] A data processing module is used to obtain three-dimensional point cloud data and RGB image data in a traffic scene, and use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the three-dimensional point cloud data to obtain a pixel-point cloud mapping relationship;

[0025] a feature extraction module, configured to convert the three-dimensional point cloud data into a voxel grid view and a range plane view, respectively, and extract geometric features of the three-dimensional point cloud data to obtain point cloud features, extract image features of the RGB image data, extract image features of the range plane view to obtain range plane view features, and extract geometric features of the voxel grid view to obtain voxel features;

[0026] The object recognition module is used to aggregate the point cloud features, the range plane view features and the voxel features to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, use a multimodal feature fusion algorithm to fuse the multi-view fusion features with the image features to obtain multimodal fusion features, and identify the objects in the traffic scene according to the multimodal fusion features.

[0027] In order to solve the above problem, the present invention further provides an electronic device, the electronic device comprising:

[0028] at least one processor; and,

[0029] a memory communicatively connected to the at least one processor; wherein,

[0030] The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned multimodal point cloud semantic segmentation method for traffic scenes.

[0031] In order to solve the above problems, the present invention also provides a computer-readable storage medium, in which at least one computer program is stored. The at least one computer program is executed by a processor in an electronic device to implement the above-mentioned multimodal point cloud semantic segmentation method for traffic scenarios.

[0032] The present invention uses a pre-constructed multi-view index to query the correspondence between pixel points in RGB image data and point clouds in 3D point cloud data, so as to establish a one-to-one mapping from 3D point cloud data to target view and then back to 3D point cloud data, share point cloud location information, and promote feature interaction of multiple views. In addition, according to the pixel-point cloud mapping relationship, a multi-modal feature fusion algorithm is used to fuse multi-view fusion features with image features to obtain multi-modal fusion features, which can realize the effective combination of multi-view fusion features and image features, ensure the feature representation capability of the acquired multi-modal information and retain the effective information in the multi-modal information, and further extract relatively rich information in areas where the point cloud data is relatively sparse, so as to ensure the effective recognition and detection of objects in traffic scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of a flow chart of a multimodal point cloud semantic segmentation method for traffic scenes provided by an embodiment of the present invention;

[0034] Figure 2 An algorithm structure flow chart of a multimodal point cloud semantic segmentation method for traffic scenes provided by an embodiment of the present invention;

[0035] Figure 3 A functional module diagram of a multimodal point cloud semantic segmentation device for traffic scenes provided by an embodiment of the present invention;

[0036] Figure 4 A schematic diagram of the structure of an electronic device for implementing the multimodal point cloud semantic segmentation method for traffic scenarios provided in one embodiment of the present invention.

[0037] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0038] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0039] The embodiment of the present application provides a multimodal point cloud semantic segmentation method for traffic scenes. The execution subject of the multimodal point cloud semantic segmentation method for traffic scenes includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the multimodal point cloud semantic segmentation method for traffic scenes can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms.

[0040] Reference Figure 1 FIG. 1 is a flow chart of a multimodal point cloud semantic segmentation method for traffic scenes provided by an embodiment of the present invention. In this embodiment, the multimodal point cloud semantic segmentation method for traffic scenes includes:

[0041] S1. Acquire three-dimensional point cloud data and RGB image data in a traffic scene, use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the three-dimensional point cloud data, and obtain a pixel-point cloud mapping relationship.

[0042] In an embodiment of the present invention, the three-dimensional point cloud data is point cloud data acquired by a laser radar device.

[0043] In the embodiment of the present invention, the RGB image data is an image acquired by using multiple cameras.

[0044] As an embodiment of the present invention, before using the pre-constructed multi-view index to query the correspondence between the pixel points in the RGB image data and the point cloud in the three-dimensional point cloud data, the method further includes constructing a multi-view index.

[0045] Furthermore, the constructing of the multi-view index includes:

[0046] Projecting the point cloud in the three-dimensional point cloud data to a target view according to a preset projection function to obtain a target view projection mapping, wherein the target view includes a voxel grid view and a range plane view;

[0047] Establishing a mapping between the target view and the three-dimensional point cloud data according to a preset hash mapping to obtain a hash view mapping;

[0048] The multi-view index is obtained according to the target view projection mapping and the hash view mapping.

[0049] Exemplarily, the projecting of the point cloud in the three-dimensional point cloud data to the target view according to a preset projection function to obtain the target view projection mapping may be implemented by the following steps:

[0050] The point cloud P in the three-dimensional point cloud data is represented as Where N is the number of points, 3 is the coordinate dimension, and C is the feature channel;

[0051] Project the point cloud to the target view using the projection function Where M is the number of elements in the target view and D represents the position dimension.

[0052] Exemplarily, a mapping between a target view and the three-dimensional point cloud data is established according to a preset hash mapping to obtain a hash view mapping, and the following implementation steps may be adopted:

[0053] Through hash function A mapping between the target view and the three-dimensional point cloud data is established to obtain a hash view mapping, where N is the number of points, M is the number of elements in the target view, and D represents the position dimension.

[0054] Furthermore, obtaining the multi-view index according to the target view projection mapping and the hash view mapping also includes multi-view feature propagation correction.

[0055] Furthermore, the multi-view feature propagation correction includes:

[0056] The point cloud features of the three-dimensional point cloud data are projected to the target view using a weighted average algorithm, and the features of the target view are mapped and calculated using a linear interpolation algorithm. Among them, the features of the voxel grid view returned to the three-dimensional point cloud data are mapped and calculated using a trilinear interpolation algorithm, and the features of the range plane view returned to the three-dimensional point cloud data are mapped and calculated using a bilinear interpolation method.

[0057] Exemplarily, obtaining the multi-view index according to the target view projection mapping and the hash view mapping also includes multi-view feature propagation correction, which may be implemented in the following steps:

[0058] In the feature propagation process from 3D point cloud data to the target view, each point P in the point cloud affects the corresponding element j in the target view (such as voxel or pixel in range plane view). The propagation process is carried out by weighted averaging and other methods, and the formula is as follows:

[0059]

[0060] Among them, F X (j) represents the feature of the corresponding element j in the target view X, F P Represents the features of point cloud P in 3D point cloud data, represents multiple key point clouds where the corresponding element j in the target view X is affected by the point cloud P, To influence the number of key point clouds of the corresponding element j in the target view X, Ω is a weighting function, usually using weighted averaging. Multiple points in the point cloud can transfer features to the corresponding voxel grid view or pixels of the range plane view through weighted averaging.

[0061] Similarly, the features in the target view can also be propagated back to the point cloud. The features propagated back from the voxel grid view to the point cloud can be calculated by the trilinear interpolation method; the features propagated back to the point cloud from the range plane view can be calculated by the bilinear interpolation method, and its formula and partial derivatives are as follows:

[0062]

[0063]

[0064] Where i is the i-th point cloud in the 3D point cloud data, φ(·) is the weight used to calculate the bilinear interpolation, δ(j) represents the neighboring grid of point cloud j, point cloud u is selected from the neighboring grid of point cloud j, and F R Refers to the range plan view feature, Represents the integer part of the coordinate difference between point cloud j and point cloud u in the x direction, Represents the integer part of the coordinate difference between point cloud j and point cloud u in the y direction.

[0065] The embodiment of the present invention measures the difference between the semantic label and the true label of each point cloud or voxel through the cross entropy loss function. The formula is:

[0066]

[0067] Where: y i represents the true semantic label, represents the predicted probability distribution, and N is the total number of point clouds.

[0068] The embodiment of the present invention uses a pre-built multi-view index to query the correspondence between pixel points in RGB image data and point clouds in the three-dimensional point cloud data, so as to establish a one-to-one mapping from three-dimensional point cloud data to the target view and then back to the three-dimensional point cloud data, share the location information of the point cloud, and promote feature interaction of multiple views.

[0069] S2. Convert the three-dimensional point cloud data into a voxel grid view and a range plane view respectively, and extract the geometric features of the three-dimensional point cloud data to obtain point cloud features, extract the image features of the RGB image data, extract the image features of the range plane view to obtain range plane view features, and extract the geometric features of the voxel grid view to obtain voxel features.

[0070] In an embodiment of the present invention, the voxel grid view refers to a method for representing three-dimensional spatial data, which can divide the space into a regular cubic grid. Each cubic unit is called a voxel. The voxel grid view represents three-dimensional objects through voxels, where each voxel can store information about the spatial volume it represents, such as color, density, and material properties.

[0071] In an embodiment of the present invention, the range plane view refers to a method for representing three-dimensional spatial data. In the lidar data processing, it directly reflects the sampling process of the lidar, can map points in the three-dimensional space to a two-dimensional plane, and can convert the three-dimensional point cloud data into an image obtained when observing the three-dimensional scene from a certain perspective.

[0072] The embodiment of the present invention is to Figure 1 The consistency loss ensures that the features between point clouds and other views (voxels, range images) are consistent. Figure 1 Consistency loss can reduce the differences between different modalities, thereby improving segmentation accuracy and Figure 1 The consistency loss is usually implemented by measuring the similarity between features of different modalities, such as through distance metrics or similarity metrics of feature vectors.

[0073] In an embodiment of the present invention, three views of the point cloud are extracted with geometric features through a specific network (three-dimensional point cloud data through MLP, voxel grid view through SPVCNN network), and RGB image data is extracted with rich texture and color information through HRNet or other image backbone networks.

[0074] S3. Summarizing the point cloud features, the range plane view features and the voxel features to obtain a multi-view fusion feature, and based on the pixel-point cloud mapping relationship, using a multimodal feature fusion algorithm to fuse the multi-view fusion feature with the image feature to obtain a multimodal fusion feature, and identifying objects in the traffic scene based on the multimodal fusion feature.

[0075] As an embodiment of the present invention, the multi-view fusion feature is fused with the image feature using a multi-modal feature fusion algorithm according to the pixel-point cloud mapping relationship to obtain the multi-modal fusion feature, which includes:

[0076] According to the geometric relationship between the laser radar that collects the 3D point cloud data and the multi-camera that collects the RGB image data, the points in the 3D point cloud data are aligned with the pixels in the RGB image data;

[0077] After the alignment process is completed, the multi-view fusion feature and the image feature are mapped to the same feature space, and the multi-view fusion feature and the image feature are spliced ​​to obtain a multimodal fusion feature.

[0078] Furthermore, when the multi-view fusion features and the image features are spliced, the importance of the multi-view fusion features and the image features is evaluated using a pre-built gated fusion algorithm to obtain an importance evaluation result; and weighted fusion of the multi-view fusion features and the image features is performed according to the importance evaluation result.

[0079] Exemplarily, the weighted fusion of the multi-view fusion feature and the image feature according to the importance evaluation result may be implemented by the following steps:

[0080]

[0081] Where: L represents the number of input feature vectors (i.e., the number of features from different views), represents the feature vector of the μth view branch, N μ is the number of points in the branch feature, C μ is the number of channels, G μ is the gating vector corresponding to each view, representing the importance of each channel, G μ =sigmoid(w μ*X μ ) indicates that the gate vector is generated using the sigmoid function, w μ is the weight vector of the feature vector of the μth view branch, · represents the element-wise multiplication operation, and the softmax function normalizes the gated vector into a probability weight.

[0082] The embodiment of the present invention aims to ensure the consistency of features of different modalities through cross-modal consistency loss, especially points outside the camera field of view (points outside FOV). By using pseudo camera features for completion and comparing these pseudo features with actual features, the model optimizes the consistency between cross-modalities. Its loss function can be expressed as

[0083]

[0084] Among them: F P,i represents the features of the i-th point cloud, Pseudo camera features representing points outside the camera's field of view.

[0085] The embodiment of the present invention can evaluate the effectiveness and accuracy of the semantic segmentation method using different evaluation indicators, and the following implementation steps can be adopted:

[0086] Assume that there are N point clouds in a point cloud set. These point clouds can be divided into M semantic labels in total. α represents the αth semantic category, and the value range of α is {0,…,M}; c is an M×M confusion matrix, c αβ The first subscript α represents the true label category of the sample, and the second subscript β represents the predicted label category of the sample, so c αβ Indicates the number of samples whose true label is the αth class and whose predicted label is the βth class. The intersection over union ratio of each class is called IoU α The specific calculation method of each indicator is as follows:

[0087] Overall accuracy OAcc is a simple metric that calculates the ratio between the number of true classified samples and the total number of samples:

[0088]

[0089] Average precision is the average of OAcc with respect to classes, calculated by calculating OAcc in each class and then averaging over the total number of classes:

[0090]

[0091] The average intersection-over-union (IoU) is a standard metric for semantic segmentation. It calculates the intersection ratio between the true value and the predicted value, and then averages it over the total number of classes:

[0092]

[0093] The present invention uses a pre-constructed multi-view index to query the correspondence between pixel points in RGB image data and point clouds in 3D point cloud data, so as to establish a one-to-one mapping from 3D point cloud data to target view and then back to 3D point cloud data, share point cloud location information, and promote feature interaction of multiple views. In addition, according to the pixel-point cloud mapping relationship, a multi-modal feature fusion algorithm is used to fuse multi-view fusion features with image features to obtain multi-modal fusion features, which can realize the effective combination of multi-view fusion features and image features, ensure the feature representation capability of the acquired multi-modal information and retain the effective information in the multi-modal information, and further extract relatively rich information in areas where the point cloud data is relatively sparse, so as to ensure the effective recognition and detection of objects in traffic scenes.

[0094] Reference Figure 2 As shown, it is an algorithm structure flow chart of a multimodal point cloud semantic segmentation method for traffic scenes provided by one embodiment of the present invention.

[0095] like Figure 3 , is a functional module diagram of a multimodal point cloud semantic segmentation device for traffic scenes provided by one embodiment of the present invention.

[0096] The multimodal point cloud semantic segmentation device 100 for traffic scenes of the present invention can be installed in an electronic device. According to the functions implemented, the multimodal point cloud semantic segmentation device 100 for traffic scenes can include a low data processing module 101, a feature extraction module 102 and an object recognition module 103.

[0097] The module described in the present invention may also be referred to as a unit, which refers to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, and is stored in a memory of the electronic device.

[0098] In this embodiment, the functions of each module / unit are as follows:

[0099] The data processing module 101 is used to obtain three-dimensional point cloud data and RGB image data in a traffic scene, and use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the three-dimensional point cloud data to obtain a pixel-point cloud mapping relationship.

[0100] In an embodiment of the present invention, the three-dimensional point cloud data is point cloud data acquired by a laser radar device.

[0101] In the embodiment of the present invention, the RGB image data is an image acquired by using multiple cameras.

[0102] As an embodiment of the present invention, before using the pre-constructed multi-view index to query the correspondence between the pixel points in the RGB image data and the point cloud in the three-dimensional point cloud data, the method further includes constructing a multi-view index.

[0103] Furthermore, the constructing of the multi-view index includes:

[0104] Projecting the point cloud in the three-dimensional point cloud data to a target view according to a preset projection function to obtain a target view projection mapping, wherein the target view includes a voxel grid view and a range plane view;

[0105] Establishing a mapping between the target view and the three-dimensional point cloud data according to a preset hash mapping to obtain a hash view mapping;

[0106] The multi-view index is obtained according to the target view projection mapping and the hash view mapping.

[0107] Exemplarily, the projecting of the point cloud in the three-dimensional point cloud data to the target view according to a preset projection function to obtain the target view projection mapping may be implemented by the following steps:

[0108] The point cloud P in the three-dimensional point cloud data is represented as Where N is the number of points, 3 is the coordinate dimension, and C is the feature channel;

[0109] Project the point cloud to the target view using the projection function Where M is the number of elements in the target view and D represents the position dimension.

[0110] Exemplarily, a mapping between a target view and the three-dimensional point cloud data is established according to a preset hash mapping to obtain a hash view mapping, and the following implementation steps may be adopted:

[0111] Through hash function A mapping between the target view and the three-dimensional point cloud data is established to obtain a hash view mapping, where N is the number of points, M is the number of elements in the target view, and D represents the position dimension.

[0112] Furthermore, obtaining the multi-view index according to the target view projection mapping and the hash view mapping also includes multi-view feature propagation correction.

[0113] Furthermore, the multi-view feature propagation correction includes:

[0114] The point cloud features of the three-dimensional point cloud data are projected to the target view using a weighted average algorithm, and the features of the target view are mapped and calculated using a linear interpolation algorithm. Among them, the features of the voxel grid view returned to the three-dimensional point cloud data are mapped and calculated using a trilinear interpolation algorithm, and the features of the range plane view returned to the three-dimensional point cloud data are mapped and calculated using a bilinear interpolation method.

[0115] Exemplarily, obtaining the multi-view index according to the target view projection mapping and the hash view mapping also includes multi-view feature propagation correction, which may be implemented in the following steps:

[0116] In the feature propagation process from 3D point cloud data to the target view, each point P in the point cloud affects the corresponding element j in the target view (such as voxel or pixel in range plane view). The propagation process is carried out by weighted averaging and other methods, and the formula is as follows:

[0117]

[0118] Among them, F X (j) represents the feature of the corresponding element j in the target view X, F P Represents the features of point cloud P in 3D point cloud data, represents multiple key point clouds where the corresponding element j in the target view X is affected by the point cloud P, To influence the number of key point clouds of the corresponding element j in the target view X, Ω is a weighting function, usually using weighted averaging. Multiple points in the point cloud can transfer features to the corresponding voxel grid view or pixels of the range plane view through weighted averaging.

[0119] Similarly, the features in the target view can also be propagated back to the point cloud. The features propagated back from the voxel grid view to the point cloud can be calculated by the trilinear interpolation method; the features propagated back to the point cloud from the range plane view can be calculated by the bilinear interpolation method, and its formula and partial derivatives are as follows:

[0120]

[0121]

[0122]

[0123] Where i is the i-th point cloud in the 3D point cloud data, φ(·) is the weight used to calculate the bilinear interpolation, δ(j) represents the neighboring grid of point cloud j, point cloud u is selected from the neighboring grid of point cloud j, and F R Refers to the range plan view feature, Represents the integer part of the coordinate difference between point cloud j and point cloud u in the x direction, Represents the integer part of the coordinate difference between point cloud j and point cloud u in the y direction.

[0124] The embodiment of the present invention measures the difference between the semantic label and the true label of each point cloud or voxel through the cross entropy loss function. The formula is:

[0125]

[0126] Where: y i represents the true semantic label, represents the predicted probability distribution, and N is the total number of point clouds.

[0127] The embodiment of the present invention uses a pre-built multi-view index to query the correspondence between pixel points in RGB image data and point clouds in the three-dimensional point cloud data, so as to establish a one-to-one mapping from three-dimensional point cloud data to the target view and then back to the three-dimensional point cloud data, share the location information of the point cloud, and promote feature interaction of multiple views.

[0128] The feature extraction module 102 is used to convert the three-dimensional point cloud data into a voxel grid view and a range plane view respectively, and extract the geometric features of the three-dimensional point cloud data to obtain point cloud features, extract the image features of the RGB image data, extract the image features of the range plane view to obtain range plane view features, and extract the geometric features of the voxel grid view to obtain voxel features.

[0129] In an embodiment of the present invention, the voxel grid view refers to a method for representing three-dimensional spatial data, which can divide the space into a regular cubic grid. Each cubic unit is called a voxel. The voxel grid view represents three-dimensional objects through voxels, where each voxel can store information about the spatial volume it represents, such as color, density, and material properties.

[0130] In an embodiment of the present invention, the range plane view refers to a method for representing three-dimensional spatial data. In the lidar data processing, it directly reflects the sampling process of the lidar, can map points in the three-dimensional space to a two-dimensional plane, and can convert the three-dimensional point cloud data into an image obtained when observing the three-dimensional scene from a certain perspective.

[0131] The embodiment of the present invention is to Figure 1 The consistency loss ensures that the features between point clouds and other views (voxels, range images) are consistent. Figure 1 Consistency loss can reduce the differences between different modalities, thereby improving segmentation accuracy and Figure 1 The consistency loss is usually implemented by measuring the similarity between features of different modalities, such as through distance metrics or similarity metrics of feature vectors.

[0132] In an embodiment of the present invention, three views of the point cloud are extracted with geometric features through a specific network (three-dimensional point cloud data through MLP, voxel grid view through SPVCNN network), and RGB image data is extracted with rich texture and color information through HRNet or other image backbone networks.

[0133] The object recognition module 103 is used to aggregate the point cloud features, the range plane view features and the voxel features to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, use a multimodal feature fusion algorithm to fuse the multi-view fusion features with the image features to obtain multimodal fusion features, and identify objects in the traffic scene according to the multimodal fusion features.

[0134] As an embodiment of the present invention, the multi-view fusion feature is fused with the image feature using a multi-modal feature fusion algorithm according to the pixel-point cloud mapping relationship to obtain the multi-modal fusion feature, which includes:

[0135] According to the geometric relationship between the laser radar that collects the 3D point cloud data and the multi-camera that collects the RGB image data, the points in the 3D point cloud data are aligned with the pixels in the RGB image data;

[0136] After the alignment process is completed, the multi-view fusion feature and the image feature are mapped to the same feature space, and the multi-view fusion feature and the image feature are spliced ​​to obtain a multimodal fusion feature.

[0137] Furthermore, when the multi-view fusion features and the image features are spliced, the importance of the multi-view fusion features and the image features is evaluated using a pre-built gated fusion algorithm to obtain an importance evaluation result; and weighted fusion of the multi-view fusion features and the image features is performed according to the importance evaluation result.

[0138] Exemplarily, the weighted fusion of the multi-view fusion feature and the image feature according to the importance evaluation result may be implemented by the following steps:

[0139]

[0140] Where: L represents the number of input feature vectors (i.e., the number of features from different views), represents the feature vector of the μth view branch, N μ is the number of points in the branch feature, C μ is the number of channels, G μ is the gating vector corresponding to each view, representing the importance of each channel, G μ=sigmoid(w μ *X μ ) indicates that the gate vector is generated using the sigmoid function, w μ is the weight vector of the feature vector of the μth view branch, · represents the element-wise multiplication operation, and the softmax function normalizes the gated vector into a probability weight.

[0141] The embodiment of the present invention aims to ensure the consistency of features of different modalities through cross-modal consistency loss, especially points outside the camera field of view (points outside FOV). By using pseudo camera features for completion and comparing these pseudo features with actual features, the model optimizes the consistency between cross-modalities. Its loss function can be expressed as

[0142]

[0143] Among them: F P,i represents the features of the i-th point cloud, Pseudo camera features representing points outside the camera's field of view.

[0144] The embodiment of the present invention can evaluate the effectiveness and accuracy of the semantic segmentation method using different evaluation indicators, and the following implementation steps can be adopted:

[0145] Assume that there are N point clouds in a point cloud set. These point clouds can be divided into M semantic labels in total. α represents the αth semantic category, and the value range of α is {0,…,M}; c is an M×M confusion matrix, c αβ The first subscript α represents the true label category of the sample, and the second subscript β represents the predicted label category of the sample, so c αβ Indicates the number of samples whose true label is the αth class and whose predicted label is the βth class. The intersection over union ratio of each class is called IoU α The specific calculation method of each indicator is as follows:

[0146] Overall accuracy OAcc is a simple metric that calculates the ratio between the number of true classified samples and the total number of samples:

[0147]

[0148] Average precision is the average of OAcc with respect to classes, calculated by calculating OAcc in each class and then averaging over the total number of classes:

[0149]

[0150] The average intersection-over-union (IoU) is a standard metric for semantic segmentation. It calculates the intersection ratio between the true value and the predicted value, and then averages it over the total number of classes:

[0151]

[0152] like Figure 4 Detailed description is given below, which is a schematic diagram of the structure of an electronic device for implementing a multimodal point cloud semantic segmentation method for traffic scenarios provided by one embodiment of the present invention.

[0153] The electronic device may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program stored in the memory 11 and executable on the processor 10, such as a multimodal point cloud semantic segmentation method program for traffic scenes.

[0154] Among them, the processor 10 can be composed of an integrated circuit in some embodiments, for example, it can be composed of a single packaged integrated circuit, or it can be composed of multiple integrated circuits with the same function or different functions, including one or more central processing units (CPU), microprocessors, digital processing chips, graphics processors and various control chips. The processor 10 is the control core (ControlUnit) of the electronic device, which uses various interfaces and lines to connect the various components of the entire electronic device, and executes or executes the program or module stored in the memory 11 (for example, executes a multimodal point cloud semantic segmentation method program for traffic scenes, etc.), and calls the data stored in the memory 11 to execute various functions of the electronic device and process data.

[0155] The memory 11 includes at least one type of readable storage medium, and the readable storage medium includes a flash memory, a mobile hard disk, a multimedia card, a card-type memory (for example, SD or DX memory, etc.), a magnetic memory, a disk, an optical disk, etc. In some embodiments, the memory 11 may be an internal storage unit of an electronic device, such as a mobile hard disk of the electronic device. In other embodiments, the memory 11 may also be an external storage device of an electronic device, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device. Further, the memory 11 may also include both an internal storage unit of the electronic device and an external storage device. The memory 11 can not only be used to store application software and various types of data installed in the electronic device, such as a code of a multimodal point cloud semantic segmentation method program for traffic scenes, but can also be used to temporarily store data that has been output or is to be output.

[0156] The communication bus 12 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The bus is configured to realize connection and communication between the memory 11 and at least one processor 10, etc.

[0157] The communication interface 13 is used for communication between the above-mentioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is generally used to establish a communication connection between the electronic device and other electronic devices. The user interface may be a display (Display), an input unit (such as a keyboard (Keyboard)), and optionally, the user interface may also be a standard wired interface, a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, etc. Among them, the display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device and to display a visual user interface.

[0158] Figure 4 Only an electronic device with components is shown, and those skilled in the art will understand that Figure 4 The structure shown does not constitute a limitation on the electronic device, and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

[0159] For example, although not shown, the electronic device may also include a power source (such as a battery) for supplying power to each component. Preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that the power management device can realize functions such as charging management, discharging management, and power consumption management. The power source may also include one or more DC or AC power sources, recharging devices, power failure detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include a variety of sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be repeated here.

[0160] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0161] The program of a multimodal point cloud semantic segmentation method for traffic scenes stored in the memory 11 of the electronic device is a combination of multiple instructions. When running in the processor 10, it can achieve:

[0162] Acquire 3D point cloud data and RGB image data in a traffic scene, use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the 3D point cloud data, and obtain a pixel-point cloud mapping relationship;

[0163] The three-dimensional point cloud data is converted into a voxel grid view and a range plane view respectively, and geometric features of the three-dimensional point cloud data are extracted to obtain point cloud features, image features of the RGB image data are extracted, image features of the range plane view are extracted to obtain range plane view features, and geometric features of the voxel grid view are extracted to obtain voxel features;

[0164] The point cloud features, the range plane view features and the voxel features are summarized to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, the multi-view fusion features are fused with the image features using a multimodal feature fusion algorithm to obtain multimodal fusion features, and objects in the traffic scene are identified according to the multimodal fusion features.

[0165] Specifically, the specific implementation method of the processor 10 for the above instructions can refer to the description of the relevant steps in the corresponding embodiment of the accompanying drawings, which will not be repeated here.

[0166] Furthermore, if the module / unit integrated in the electronic device 1 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0167] The present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor of an electronic device, the computer program can implement:

[0168] Acquire 3D point cloud data and RGB image data in a traffic scene, use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the 3D point cloud data, and obtain a pixel-point cloud mapping relationship;

[0169] The three-dimensional point cloud data is converted into a voxel grid view and a range plane view respectively, and geometric features of the three-dimensional point cloud data are extracted to obtain point cloud features, image features of the RGB image data are extracted, image features of the range plane view are extracted to obtain range plane view features, and geometric features of the voxel grid view are extracted to obtain voxel features;

[0170] The point cloud features, the range plane view features and the voxel features are summarized to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, the multi-view fusion features are fused with the image features using a multimodal feature fusion algorithm to obtain multimodal fusion features, and objects in the traffic scene are identified according to the multimodal fusion features.

[0171] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0172] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0173] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0174] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0175] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

[0176] The blockchain referred to in this invention is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0177] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0178] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system claim can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A multimodal point cloud semantic segmentation method for traffic scenes, characterized in that: The method comprises: Acquire 3D point cloud data and RGB image data in a traffic scene, use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the 3D point cloud data, and obtain a pixel-point cloud mapping relationship; The three-dimensional point cloud data is converted into a voxel grid view and a range plane view respectively, and geometric features of the three-dimensional point cloud data are extracted to obtain point cloud features, image features of the RGB image data are extracted, image features of the range plane view are extracted to obtain range plane view features, and geometric features of the voxel grid view are extracted to obtain voxel features; The point cloud features, the range plane view features and the voxel features are summarized to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, the multi-view fusion features are fused with the image features using a multimodal feature fusion algorithm to obtain multimodal fusion features, and objects in the traffic scene are identified according to the multimodal fusion features.

2. The multimodal point cloud semantic segmentation method for traffic scenes according to claim 1, characterized in that: Before using the pre-constructed multi-view index to query the correspondence between the pixel points in the RGB image data and the point cloud in the three-dimensional point cloud data, the method further includes constructing the multi-view index.

3. The multimodal point cloud semantic segmentation method for traffic scenes according to claim 1 or 2, characterized in that: The step of constructing a multi-view index includes: Projecting the point cloud in the three-dimensional point cloud data to a target view according to a preset projection function to obtain a target view projection mapping, wherein the target view includes a voxel grid view and a range plane view; Establishing a mapping between the target view and the three-dimensional point cloud data according to a preset hash mapping to obtain a hash view mapping; The multi-view index is obtained according to the target view projection mapping and the hash view mapping.

4. The multimodal point cloud semantic segmentation method for traffic scenes according to claim 3, characterized in that: The obtaining the multi-view index according to the target view projection mapping and the hash view mapping also includes multi-view feature propagation correction.

5. The multimodal point cloud semantic segmentation method for traffic scenes as claimed in claim 4, characterized in that: The multi-view feature propagation correction includes: The point cloud features of the three-dimensional point cloud data are projected to the target view using a weighted average algorithm, and the features of the target view are mapped and calculated using a linear interpolation algorithm. Among them, the features of the voxel grid view returned to the three-dimensional point cloud data are mapped and calculated using a trilinear interpolation algorithm, and the features of the range plane view returned to the three-dimensional point cloud data are mapped and calculated using a bilinear interpolation method.

6. The multimodal point cloud semantic segmentation method for traffic scenes according to claim 1, characterized in that: According to the pixel-point cloud mapping relationship, the multi-view fusion feature is fused with the image feature by using a multi-modal feature fusion algorithm to obtain the multi-modal fusion feature, which includes: According to the geometric relationship between the laser radar that collects the 3D point cloud data and the multi-camera that collects the RGB image data, the points in the 3D point cloud data are aligned with the pixels in the RGB image data; After the alignment process is completed, the multi-view fusion feature and the image feature are mapped to the same feature space, and the multi-view fusion feature and the image feature are spliced ​​to obtain a multimodal fusion feature.

7. The multimodal point cloud semantic segmentation method for traffic scenes according to claim 6, characterized in that: When the multi-view fusion feature is spliced ​​with the image feature, it also includes: Using a pre-built gated fusion algorithm to evaluate the importance of the multi-view fusion feature and the image feature, and obtaining an importance evaluation result; The multi-view fusion feature and the image feature are weightedly fused according to the importance evaluation result.

8. A multimodal point cloud semantic segmentation device for traffic scenes, characterized in that: The device is used to implement the multimodal point cloud semantic segmentation method for traffic scenes according to any one of claims 1 to 7, and the device includes: A data processing module is used to obtain three-dimensional point cloud data and RGB image data in a traffic scene, and use a pre-built multi-view index to query the correspondence between pixel points in the RGB image data and point clouds in the three-dimensional point cloud data to obtain a pixel-point cloud mapping relationship; a feature extraction module, configured to convert the three-dimensional point cloud data into a voxel grid view and a range plane view, respectively, and extract geometric features of the three-dimensional point cloud data to obtain point cloud features, extract image features of the RGB image data, extract image features of the range plane view to obtain range plane view features, and extract geometric features of the voxel grid view to obtain voxel features; The object recognition module is used to aggregate the point cloud features, the range plane view features and the voxel features to obtain multi-view fusion features, and according to the pixel-point cloud mapping relationship, use a multimodal feature fusion algorithm to fuse the multi-view fusion features with the image features to obtain multimodal fusion features, and identify the objects in the traffic scene according to the multimodal fusion features.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the multimodal point cloud semantic segmentation method for traffic scenes as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the multimodal point cloud semantic segmentation method for traffic scenes as described in any one of claims 1 to 7 is implemented.

Images