A method for extracting point cloud features, a method for identifying objects and application products

Abstract

The embodiment of the present application provides a method for extracting point cloud features, a method for identifying objects, and an application product, the method comprising: obtaining an object to be encoded, wherein the object to be encoded includes a pillar to be encoded or a voxel to be encoded; extracting and recording the distribution properties of the point cloud of the object to be encoded within different height ranges; obtaining a feature vector corresponding to the object to be encoded at least based on the distribution properties of the point cloud. The embodiment of the present application can effectively process the distribution features of the point cloud of the object to be encoded at different height levels, improve the characterization effect of the feature vector obtained based on these features on the object to be encoded, and the encoding method of the embodiment of the present application does not rely on maximum pooling or similar pooling operations, so that the encoding result can retain the relevant details of the point cloud distribution in the original column, thereby improving the accuracy of identifying the object based on the encoding.

Description

Technical Field

[0001] The present application relates to the field of image processing. Specifically, embodiments of the present application relate to a method for extracting point cloud features, a method for identifying objects, and application products. Background Art

[0002] The relevant technology obtains the point cloud data of the target three-dimensional space through point cloud data acquisition sensors, and then obtains the feature vectors of the corresponding three-dimensional space based on the point cloud data. These feature vectors can be further used in scenarios such as three-dimensional modeling, scene reconstruction, robot navigation, object recognition, virtual reality and augmented reality.

[0003] Taking object recognition as an example, the LiDAR-based 3D object detection provided by the related technology plays a key role in a variety of applications, especially in the fields of autonomous driving and robotics. It is very important to quickly and accurately detect objects in the surrounding environment, which places high demands on performance and latency. The grid-based point cloud method provided by the related technology grids the points into regular grids (voxels / pillars), encodes each grid and uses 2D or 3D convolution for feature learning, and then can complete the object recognition of the scene based on the collected point cloud in the corresponding scene (for example, point cloud data collected by LiDAR (Light Detection and Ranging) technology).

[0004] The grid coding methods of the related art include pillar coding methods and voxel coding methods. The pillar coding method (i.e., the method of encoding pillars) is efficient, but it lacks the ability to encode highly perceptual information. Figure 2 The pillar codes and Figure 3 The maximum pooling and other operations involved in the voxel encoding method will reduce the representation effect of the feature vector on the distribution characteristics of the original point cloud data, thereby reducing the accuracy of object recognition based on these feature vectors.

[0005] Therefore, how to obtain an efficient encoding method for point clouds has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the embodiments of the present application is to provide a method for extracting point cloud features, a method for identifying objects and an application product. The use of the embodiments of the present application can improve the expression effect of the encoded feature vector on height information and thereby improve the accuracy of applications such as identifying objects based on the feature vector.

[0007] In a first aspect, an embodiment of the present application provides a method for extracting point cloud features, the method comprising: obtaining an object to be encoded, wherein the object to be encoded comprises a pillar to be encoded or a voxel to be encoded; extracting and recording the distribution properties of the point cloud of the object to be encoded within different height ranges; and obtaining a feature vector corresponding to the object to be encoded at least based on the distribution properties of the point cloud.

[0008] Some embodiments of the present application obtain and record the height distribution attributes of the point cloud in the pillar to be encoded, and use this information for encoding to obtain the corresponding feature vector, so that the feature vector includes the distribution attribute information of the point cloud in each pillar at different heights, thereby improving the accuracy of using these feature vectors for applications such as object recognition, that is, the pillar encoding method or point column encoding method of the embodiment of the present application can characterize the height information with less calculation; some embodiments of the present application obtain the point cloud distribution attributes of the voxel to be encoded in the height direction so that the obtained feature vector can reflect the distribution attributes of the point cloud of the corresponding voxel at different heights in a finer granularity, thereby improving the characterization effect of the obtained feature vector on the distribution attributes of the point cloud data within different height ranges in the voxel, and ultimately improving the accuracy of applications such as object recognition based on feature vectors.

[0009] In some embodiments, the extracting and recording the distribution properties of the point cloud of the object to be encoded within different height ranges includes: dividing the object to be encoded along the height direction of the object to be encoded to obtain a plurality of boxes arranged in sequence from bottom to top; performing feature extraction on the point cloud in each box to obtain the distribution properties of the point cloud.

[0010] In order to obtain the distribution characteristics of the point cloud of the object to be encoded within different height ranges, some embodiments of the present application can first further divide the object to be encoded in the height direction to obtain multiple boxes, and use each box as the smallest unit for feature extraction to extract the distribution characteristics of the point cloud of the object to be encoded within different height ranges, thereby improving the perception and representation effect of the feature vector on the height feature.

[0011] In some embodiments, extracting and recording the distribution properties of the point cloud of the object to be encoded within different height ranges further includes: using at least one histogram to characterize the distribution properties of the point cloud.

[0012] Some embodiments of the present application use a histogram to record the distribution pattern of the point cloud of each box in a plurality of boxes arranged sequentially from bottom to top, which can improve the characterization effect of the distribution pattern of the point cloud data along the height direction of the cylinder, and thus improve the encoding effect.

[0013] In some embodiments, the distribution attributes of the point cloud include at least one of the following information: the total number of box point clouds, box reflection intensity statistics, and box point cloud acquisition timestamp information.

[0014] Some embodiments of the present application use a box as the smallest unit for feature extraction, and count the total number of point clouds of each smallest unit to obtain the total number of box point clouds, count the positions of all point clouds in the corresponding box to obtain box position statistics, count the reflection intensity data of all point clouds in the corresponding box to obtain box reflection intensity statistics, or count the acquisition time information of all point clouds in the box to obtain box point cloud data timestamp information. The embodiments of the present application obtain the distribution characteristics of point clouds of as many types as possible in different height ranges, thereby improving the characterization effect of the obtained feature vector on the original point cloud data.

[0015] In some embodiments, the feature extraction of the point cloud in each box includes: counting the number of points in each box to obtain the total number of box point clouds corresponding to each box; using at least one histogram to characterize the distribution properties of the point cloud includes: using a first histogram to characterize the total number of box point clouds corresponding to each box in the multiple boxes.

[0016] Some embodiments of the present application use a histogram to characterize the number of point clouds included in each box, so as to obtain the distribution density of the point cloud of the object to be encoded in the height direction, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0017] In some embodiments, the feature extraction of the point cloud in each box includes: obtaining the reflection intensity statistics of all points in each box to obtain the box reflection intensity statistics corresponding to each box; and using at least one histogram to characterize the distribution properties of the point cloud includes: using a second histogram to characterize the box reflection intensity statistics corresponding to each box in the multiple boxes.

[0018] Some embodiments of the present application use a histogram to characterize the statistical values ​​of the reflection intensity of the point cloud included in each box, so as to obtain the distribution law of the reflection intensity of the point cloud of the object to be encoded in the height direction, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0019] In some embodiments, obtaining the reflection intensity statistics of all points in each box includes: calculating the sum of the reflection intensities of all points in the i-th box, wherein the i-th box is any box; calculating the quotient of the sum and the total number of points in the i-th box to obtain the reflection intensity mean.

[0020] Some embodiments of the present application use the average reflection intensity of all point clouds in each box as the reflection intensity value of the box, so as to improve the accuracy and objectivity of the obtained reflection intensity statistical value.

[0021] In some embodiments, the feature extraction of the point cloud in each box includes: obtaining the acquisition time information of all points in each box to obtain the box point cloud acquisition timestamp information; the use of at least one histogram to characterize the distribution properties of the point cloud includes: using a third histogram to characterize the box point cloud acquisition timestamp information corresponding to each box in the multiple boxes.

[0022] Some embodiments of the present application use a histogram to characterize the acquisition time characteristics of the point cloud included in each box, so as to obtain the attributes of the acquisition time of the point cloud of the object to be encoded in the height direction, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0023] In some embodiments, obtaining a feature vector corresponding to the object to be encoded at least based on the distribution properties of the point cloud includes: acquiring position information of the object to be encoded, wherein the position information of the object to be encoded is used to characterize the global position of each object to be encoded; and obtaining the feature vector based on the position information of the object to be encoded and the distribution properties of the point cloud.

[0024] Some embodiments of the present application fuse the position information of the object to be encoded with the distribution attribute information of the point cloud of the object to be encoded in the height direction, so that the obtained feature vector can better reflect the distribution pattern of the original point cloud at different heights in the object to be encoded.

[0025] In some embodiments, the object to be encoded is the pillar to be encoded, wherein the position information of the object to be encoded is the two-dimensional plane coordinates of the center position of the pillar to be encoded.

[0026] Some embodiments of the present application use the two-dimensional coordinates of the center position of the pillar to be encoded to represent the position of the object to be encoded, thereby improving the uniformity of the position information.

[0027] In some embodiments, obtaining the feature vector based on the position information of the object to be encoded and the distribution properties of the point cloud includes: connecting the data information of each histogram in the at least one histogram and the position information of the object to be encoded to obtain combined information; and using a linear layer to project the combined information to obtain the feature vector.

[0028] Some embodiments of the present application can improve the way in which the obtained feature vector represents the distribution law of the original point cloud data by connecting the distribution characteristics of the corresponding categories of point cloud data statistically obtained in each histogram and the position information of the object to be encoded.

[0029] In a second aspect, some embodiments of the present application provide a method for identifying an object, the method comprising: obtaining a feature vector obtained by using a method provided by any one of the embodiments included in the first aspect; and identifying an object in a three-dimensional space based on the feature vector.

[0030] In a third aspect, some embodiments of the present application provide a device for identifying an object, the device comprising: a feature vector acquisition module, configured to acquire a feature vector; and a recognition module, configured to identify an object in a three-dimensional space based on the feature vector.

[0031] In a fourth aspect, some embodiments of the present application provide a device for extracting point cloud features, the device comprising: a point cloud data acquisition module, configured to acquire an object to be encoded, wherein the object to be encoded comprises a pillar to be encoded or a voxel to be encoded; a height feature extraction module, configured to extract and record the distribution properties of the point cloud of the object to be encoded within different height ranges; and a coding module, configured to obtain a feature vector corresponding to the column to be encoded at least based on the distribution properties of the point cloud.

[0032] In a fifth aspect, some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any one of the embodiments included in the first aspect or the second aspect.

[0033] In a sixth aspect, some embodiments of the present application provide an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of the embodiments included in the first or second aspect above can be implemented.

[0034] In the seventh aspect, some embodiments of the present application provide a computer program product, wherein the computer program product includes a computer program, wherein when the computer program is executed by a processor, it can implement the method described in any one of the embodiments included in the first aspect or the second aspect above. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 One of the network architecture diagrams of the pillar coding based detector provided for related art;

[0037] Figure 2 A schematic diagram of a point column feature encoding process (or pillar encoding process) provided for related technologies;

[0038] Figure 3 Schematic diagram of the voxel encoding process provided for related technologies;

[0039] Figure 4 One of the flow charts of the method for extracting point cloud features provided in an embodiment of the present application;

[0040] Figure 5 Flow chart 2 of the method for extracting point cloud features provided in an embodiment of the present application;

[0041] Figure 6 A flowchart of a method for identifying an object provided in an embodiment of the present application;

[0042] Figure 7 A block diagram of the device for extracting point cloud features provided in an embodiment of the present application;

[0043] Figure 8 A block diagram of a device for identifying an object provided in an embodiment of the present application;

[0044] Fig. 9 A schematic diagram of the composition of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0046] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0047] In application areas such as autonomous driving, point cloud-based object detection is a critical task. When it comes to end-side applications, accurate and effective pillar feature encoding can effectively benefit from pillar-based point cloud detection methods, which often have the advantages of friendly deployment, compact structure, and high performance. At present, the existing pillar encoding methods not only ignore the height information, but also cause information loss in the encoding process, which directly and seriously affects the subsequent point cloud detection performance. The existing voxel encoding method also reduces the representation of the feature vector to the original point cloud distribution characteristics due to the use of operations such as maximum pooling. Therefore, as described in the background technology section, how to explore an efficient grid encoding method to enhance feature expression while avoiding information loss is an important issue.

[0048] The detector structure based on pillar coding provided by the related art is as follows Figure 1 As shown, the pillar-coded detector includes: a point-pillar feature coding network, a backbone network and a detection head, wherein, in practical applications, the collected point cloud of the scene to be identified is input into the point-pillar feature coding network, and the prediction result (i.e., the three-dimensional object recognition result for the scene to be identified) is output through the backbone network and the detection head.

[0049] exist Figure 1 In the data processing process of the point-pillar feature coding network, the process includes: dividing the point cloud on a two-dimensional plane to obtain multiple pillars (that is, the pillars to be encoded in the embodiment of the present application are also divided in this way), stacking the pillars, extracting and encoding features of each pillar and constructing a pseudo image based on the feature vector obtained by encoding, and finally performing three-dimensional object recognition based on the pseudo image. Figure 1 The backbone network architecture is also provided as an example, which includes multiple convolution processing and splicing operations. Figure 1 The specific implementation of the backbone network and the detection head can be found in existing literature, so these two modules will not be described in detail.

[0050] The following are combined Figure 2 and Figure 3 The pillar encoding process and voxel encoding process provided by the related art and the defects of each process are explained by way of example.

[0051] like Figure 2 As shown, the pillar encoding method provided in this figure can be applied to Figure 1 The point pillar feature encoding network includes: (1) For a pillar containing N*3 point clouds, first use a multi-layer perceptron to map it to a higher dimension to obtain a point-by-point feature N*C. (2) Apply a maximum pooling operation to the feature N*C along the point cloud point number dimension to obtain a feature 1*C representing the current pillar. The whole process can be expressed as: 1*C=MaxPool(N*C). (3) Repeat the above process to extract features from all pillars in the entire scene to obtain the corresponding pillar features. (4) Map the features of all pillars obtained into a H*W BEV feature map according to the pixel coordinates of each pillar for input into the backbone network.

[0052] The inventors of this application found in their research that Figure 2 The provided pillar encoding method process has at least the following technical defects: a maximum pooling operation is adopted, which directly selects the eigenvalue with the largest value along the point dimension to form the feature vector of each pillar. Therefore, this operation does not take into account the position information of point clouds at different heights and positions within the pillar, and does not consider the distribution of point clouds at different heights. It is a biased feature encoding method.

[0053] That is to say, the inventors of the present application found in their research that Figure 2 The pillar encoding method provided by the related art shown has the following disadvantages: (1) information loss significantly affects performance; (2) the encoding scheme causes huge changes in activation values, making quantization challenging. Among them, the cause of problem (1) can be attributed to at least two factors: first, there is no detailed height-aware feature encoding for each pillar, which will lead to incomplete representation of the 3D scene, and then make it difficult to distinguish height differences when performing 3D object recognition in a scene under different circumstances, for example, distinguishing pedestrians in a scene from people standing on a car in the scene; second, the use of maximum pooling operation in the pillar encoding process leads to the loss of height information. This is because the use of maximum pooling operation in each pillar to obtain pillar features will result in the loss of fine-grained information in the height direction, which limits the ability of the encoded features to distinguish different objects. The cause of the above problem (2) can at least be attributed to the numerical difference in the input data when encoding the pillar features. The existing method will input the original point cloud coordinates and point cloud density, as well as the encoded pillar relative coordinates. The huge difference between these values ​​makes it difficult to effectively quantize the activation.

[0054] like Figure 3 As shown, the voxel encoding method provided by the related technology includes the following examples: (1) For a point cloud scene, which contains M points, represented as M*3, in order to be able to use regular convolution operations on irregular point clouds, VoxelNet encodes the point cloud of the entire scene into voxels of length H*W*D space, and each voxel is represented as a cube (which can be used as the voxel to be encoded in the embodiment of the present application). When a voxel contains N points, these points are represented as N*3. Similar to the Pillar encoding in PointPillars, VoxelNet first uses a multi-layer perceptron to map it to a higher dimension to obtain point-by-point features N*C. (2) The feature N*C is subjected to a maximum pooling operation along the point cloud point number dimension to obtain a feature 1*C representing the current pillar. The whole process can be expressed as: 1*C=MaxPool(N*C). (3) Repeat the above process to extract features from all voxels in the entire scene to obtain the corresponding pillar features. (4) The features of all voxels obtained are mapped into a 4D feature map of C*D*H*W according to the spatial position coordinates of each voxel for input into the backbone network. C represents the feature dimension of each voxel, D represents the number of voxels divided along the height direction, and H and W represent the number of voxels divided along the X-axis and Y-axis directions, respectively. Therefore, the voxel encoding of the related technology can encode all M point clouds in a fixed range of space into a 4D feature map of C*D*H*W for subsequent convolutional neural network processing.

[0055] It is understandable that due to Figure 2 and Figure 3 For providing a prior art grid coding method, thus Figure 2 and Figure 3 The clarity does not affect the disclosure of the technical solution of this application.

[0056] In summary, it is not difficult to find that the above voxel encoding process provided by the related art has similar technical problems as the pillar encoding method. For example, both the voxel encoding and pillar encoding of the related art use the maximum pooling operation, which directly selects the eigenvalue with the largest value along the point dimension to form the corresponding eigenvector, so this operation will cause information loss. In addition, compared with pillar encoding, the voxel encoding process of the related art includes high-dimensional encoding, but it also makes the representation of the encoded eigenvector more complex, which will increase the computational complexity of the model.

[0057] Therefore, how to obtain more efficient feature expression in pillar coding and voxel coding and improve the balance between performance and speed has high application value in autonomous systems for actual deployment.

[0058] At least in order to achieve the technical goal and solve the above-mentioned problems, some embodiments of the present application provide a method for extracting point cloud features, which uses a histogram to encode the height perception information of each pillar or each voxel into multiple boxes, and counts the distribution attributes of the point cloud in each box, and then obtains a feature vector based on the distribution attributes of the point cloud of each box. For example, through some embodiments of the present application, a weighted intensity (for example, the intensity value is the reflection intensity of each LiDAR point) histogram can be obtained (that is, the distribution attributes of multiple types of point clouds corresponding to multiple boxes are used as weight values), thereby providing height perception information for scenes such as three-dimensional object recognition and detection. For example, in some embodiments of the present application, the obtained histogram is connected with the pillar center coordinates of the pillar to be encoded, and passed through a linear layer to obtain a feature vector representing the pillar to be encoded. In addition, the method for extracting point cloud features of the embodiment of the present application can be seamlessly integrated into the detector provided by the relevant technology as a plug-in module. For example, the method for extracting point cloud features provided in the embodiment of the present application is inserted as a plug-in Figure 1 The pillar detector, at this time Figure 1 The point-column feature encoding network adopts the method of extracting point cloud features in the embodiment of the present application to obtain the feature vector corresponding to each pillar.

[0059] It should be noted that before introducing the method for extracting point cloud features provided in some embodiments of the present application, the meanings of the point cloud, the pillars to be encoded and the voxels to be encoded involved in the method are first explained.

[0060] A point cloud is a set of 3D data points, each of which corresponds to a point, and each point is represented by a 3D coordinate in a Cartesian coordinate system or other coordinate systems. It is considered to be a set of unstructured 3D points that symbolize the geometry of a 3D object and is used in many computer vision tasks such as classification and segmentation, object recognition, reconstruction, etc. For example, a point cloud of a corresponding scene can be obtained through LiDAR (Light Detection And Ranging).

[0061] The pillars to be encoded and the voxels to be encoded in the embodiments of the present application correspond to the pillars and voxels in the related art respectively. A pillar (called a point pillar in some literature) is a data structure obtained by uniformly discretizing the three-dimensional point cloud data collected in the target space on the xy plane. Each pillar can be regarded as an area with a fixed size in the x and y directions, which is used to accommodate all the point cloud data in the area. For example, Figure 1 A pillar is shown, or Figure 2 In the voxel encoding of the related art, a voxel (such as Figure 3 The four stacked columns shown are examples of four voxels) are used as the basic unit for representing point cloud data in three-dimensional space, and the point cloud data can be converted into voxel representation through the voxelization process. Voxelization is the process of uniformly discretizing point cloud data in three-dimensional space, and each voxel can contain a certain number of points, and the features of these points can be extracted and represented by voxel encoding. Through the embodiments of the present application, the distribution characteristics of point cloud data at different height levels in the column (the column can be a pillar in the pillar encoding or each voxel in the voxel encoding) can be effectively processed, so that the obtained feature vector can better reflect the distribution characteristics of point cloud data at different heights in the object to be encoded, and improve the characterization effect of the feature vector on the original point cloud. In addition, the embodiments of the present application do not rely on maximum pooling or similar pooling operations in the process of encoding pillars or voxels, so that the encoding result can retain the relevant details of the point cloud distribution in the original column, thereby improving the accuracy of tasks such as object recognition based on feature encoding (or feature information).

[0062] Combine the following Figure 4 The method for extracting point cloud features provided by some embodiments of the present application is exemplified.

[0063] Please see Figure 4 , Figure 4 A method for extracting point cloud features provided in some embodiments of the present application includes:

[0064] S110, obtaining an object to be encoded.

[0065] It should be noted that the object to be encoded is a voxel or a pillar of the related technology, wherein the process of dividing the spatial point cloud data to obtain voxels and pillars or the definitions of pillars and voxels can be referred to above, and no further details will be given here to avoid repetition.

[0066] For example, in some embodiments of the present application, the object to be encoded is a pillar to be encoded. For example, the column to be encoded is a column obtained by dividing the collected point cloud data of the three-dimensional space on a two-dimensional plane. In other embodiments of the present application, the object to be encoded is a voxel to be encoded.

[0067] It should be noted that the voxels to be encoded are the voxels in the voxel encoding provided by the related art. The method for obtaining the pillars corresponding to the pillars to be encoded or the voxels corresponding to the voxels to be encoded can refer to the existing methods or the methods described above, for example, Figure 2 The pillar shown can be used as a pillar to be encoded in the embodiment of the present application (in some documents, the pillar is called a point pillar). Figure 3 The four provided voxels can all be used as voxels to be encoded.

[0068] It is not difficult to understand that the encoding method for three-dimensional object recognition provided in the embodiment of the present application can be used to improve the existing voxel encoding process as well as the existing pillar encoding process.

[0069] S120: extract and record distribution properties of the point cloud of the object to be encoded within different height ranges.

[0070] In some embodiments of the present application, the distribution attribute information of the point cloud is used to characterize the distribution law of the point cloud data in the corresponding box. For example, in some embodiments of the present application, the distribution attribute of the point cloud includes at least one of the following information: the total number of box point clouds (i.e., the total number of all points included in the corresponding box), the box reflection intensity statistics (i.e., the average value of the reflection intensity of all points included in the box), and the box point cloud acquisition timestamp information (i.e., the acquisition time information of all points in the box). For example, in some embodiments of the present application, obtaining the distribution attributes of point clouds in different height ranges means obtaining the total number of point clouds of each box obtained along the height direction, that is, obtaining the total number of box point clouds; in some embodiments of the present application, obtaining the distribution attributes of point clouds in different height ranges means obtaining the statistical value of the reflection intensity of the point cloud data of each box obtained along the height direction, that is, obtaining the statistical value of the box reflection intensity; in some embodiments of the present application, obtaining the distribution attributes of point clouds in different height ranges means listing the position information of the point cloud data of each box, that is, obtaining the box position statistics; in some embodiments of the present application, obtaining the distribution attributes of point clouds in different height ranges means listing the timestamp information corresponding to all point clouds in each box, that is, obtaining the box point cloud acquisition timestamp information. It can be understood that those skilled in the art can also use two or three of the three types of information, namely, the total number of box point clouds, the statistical value of box reflection intensity, and the box point cloud acquisition timestamp information, as the distribution attributes of the point cloud of the corresponding box (i.e., as an example of the pillars to be encoded in different height ranges).

[0071] The following is an illustrative description of the distribution properties of the point cloud obtained in some embodiments of the present application. In some embodiments of the present application, obtaining the distribution properties of the point cloud exemplarily includes obtaining the number of points in all boxes; in some embodiments of the present application, obtaining the distribution properties of the point cloud exemplarily includes obtaining the reflection intensity corresponding to the point cloud in all boxes; in some embodiments of the present application, obtaining the distribution properties of the point cloud exemplarily includes obtaining the acquisition time of all points in the box; in some embodiments of the present application, obtaining the distribution properties of the point cloud exemplarily includes obtaining the total number of points in each box and the average value of the reflection intensity of the point cloud; in some embodiments of the present application, obtaining the distribution properties of the point cloud is to obtain the number of points in each box, the reflection intensity of the point cloud and the point acquisition time. Due to limited space, the distribution properties of various point clouds are not listed one by one. It can be understood that those skilled in the art can select the corresponding distribution property type of the point cloud according to actual needs, and the present application does not limit the selected distribution property type of the point cloud.

[0072] It is not difficult to understand that some embodiments of the present application use the box as the smallest unit for point cloud feature extraction and count the distribution characteristics of the point cloud of each smallest unit, which can improve the characterization effect of the encoding vector obtained based on these data on the height feature.

[0073] Unlike the related technology that does not perform height perception on the point cloud within the pillars and voxels, the embodiments of the present application need to obtain and record the distribution properties of the point cloud of the pillars to be encoded or the voxels to be encoded in different height ranges, so that the feature vector can reflect the height distribution properties of the point cloud within the pillars or voxels, thereby improving the accuracy of three-dimensional object recognition.

[0074] S130, obtaining a feature vector corresponding to the object to be encoded at least according to a distribution attribute of the point cloud.

[0075] For example, in some embodiments of the present application, a feature vector is obtained directly according to the distribution properties of the point cloud, that is, the distribution properties of the point cloud are mapped through a linear layer to obtain the feature vector.

[0076] For example, in some embodiments of the present application, the distribution attributes of the point cloud are fused with the position information of the object to be encoded, and the fused information is mapped through a linear layer to obtain a feature vector.

[0077] It is understandable that the feature vector obtained in S130 can be used in applications such as three-dimensional modeling, scene reconstruction, robot navigation, virtual reality and augmented reality, and some embodiments of the present application do not limit the application scenario of the feature vector.

[0078] It is not difficult to understand that some embodiments of the present application obtain and record the height distribution properties of the point cloud in the pillar to be encoded, and use this information for encoding to obtain the corresponding feature vector, so that the feature vector includes the perceived height information of each pillar, thereby improving the accuracy of applications such as object recognition using these feature vectors; some embodiments of the present application further obtain the point cloud distribution properties of the voxels to be encoded in the height direction so that the obtained feature vector can reflect the distribution properties of the point cloud in the corresponding voxel at different heights in a more fine-grained manner, thereby improving the characterization effect of the obtained feature vector on the distribution properties of the point cloud within different height ranges within the voxel, and ultimately improving the accuracy of applications such as object recognition based on feature vectors.

[0079] For example, the method for extracting point cloud features provided in some embodiments of the present application can be applied in the scene of 3D object detection based on LiDAR (Light Detection and Ranging). For example, the method for extracting point cloud features in some embodiments of the present application can be applied in the field of autonomous driving and robotics, thereby quickly and accurately detecting objects in the surrounding environment, thereby improving processing performance and processing speed.

[0080] The implementation process of S110 is described below by way of example.

[0081] In some embodiments of the present application, obtaining the object to be encoded described in S110 is to read from a storage unit a pillar obtained by evenly dividing the point cloud of a certain scene collected in a two-dimensional plane; in other embodiments of the present application, obtaining the object to be encoded described in S110 is to read from a storage unit a voxel obtained by evenly dividing the point cloud of a certain scene collected in a three-dimensional space.

[0082] In some embodiments of the present application, the acquisition of the object to be encoded described in S110 exemplarily includes: reading the point cloud of the first space (referring to any three-dimensional space) acquired by the acquisition unit; averaging the point cloud in the two-dimensional plane to obtain multiple pillars, and each pillar is used as a pillar to be encoded. In some other embodiments of the present application, the acquisition of the object to be encoded described in S110 exemplarily includes: reading the point cloud of the first space acquired by the data acquisition unit; averaging the point cloud data in the three-dimensional space to obtain multiple voxels, and each voxel is used as a pillar to be encoded.

[0083] The implementation process of S120 is described below by way of example.

[0084] In some embodiments of the present application, the process of extracting and recording the distribution attributes of the point cloud of the object to be encoded in different height ranges in S120 exemplarily includes:

[0085] In the first step, the object to be encoded is divided along the height direction of the object to be encoded to obtain a plurality of boxes arranged in sequence from bottom to top.

[0086] In the second step, feature extraction is performed on the point cloud in each box to obtain the distribution properties of the point cloud.

[0087] It should be noted that histogram technology is also introduced in some embodiments of the present application. For example, in some embodiments of the present application, S120 further includes, after the above-mentioned second step: a third step, using at least one histogram to characterize the distribution properties of the point cloud. In some embodiments of the present application, a histogram corresponds to the distribution properties of a type of point cloud in the box, and the distribution properties of each type of point cloud exemplarily include the density of the point cloud, the statistical value of the reflection intensity corresponding to the point cloud, the position corresponding to the point cloud or the acquisition time information, etc.

[0088] In order to obtain the distribution characteristics of the point cloud data of the object to be encoded within different height ranges, some embodiments of the present application can further divide the object to be encoded in the height direction to obtain multiple boxes, and use each box as the minimum unit for feature extraction, thereby improving the effect of the feature vector on the perception and characterization of height features. Some embodiments of the present application use a histogram to record the distribution law of the point cloud data of each box in a plurality of boxes arranged in sequence from bottom to top, which can improve the characterization effect of the distribution law of the point cloud along the height direction of the cylinder, thereby improving the encoding effect.

[0089] It is not difficult to understand that some embodiments of the present application can effectively retain point cloud information in different height directions by encoding the point cloud in the pillar or voxel in the form of a histogram along the height direction, while avoiding various pooling operations (such as Figure 2 or Figure 3 The voxel encoding and pillar encoding of the related technologies provided both use pooling operations) to reduce information loss. In addition to height information, some embodiments of the present application will also add information such as the reflection intensity of the point cloud and the center coordinates of the pillars to enhance the geometric semantic expression of the pillars by the feature vector, and can convert the computational complexity of the original multi-layer perceptron from point-by-point calculation to pillar-by-pillar calculation, effectively reducing computational overhead, improving computational efficiency, and at the same time improving the accuracy of point cloud detection.

[0090] The implementation process of the second and third steps is further explained below through multiple examples.

[0091] Example 1

[0092] For example, in some embodiments of the present application, the process of extracting features from the point cloud in each box in the second step exemplarily includes: counting the number of points in each box to obtain the total number of box point clouds corresponding to each box. Correspondingly, the process of using at least one histogram to characterize the distribution properties of the point cloud in the third step exemplarily includes: using the first histogram to characterize the total number of box point clouds corresponding to each box in the multiple boxes. It is not difficult to understand that some embodiments of the present application use a histogram to characterize the total number of points included in each box, so as to obtain the point cloud distribution density of the object to be encoded in the height direction, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0093] Example 2

[0094] For example, in some embodiments of the present application, the process of extracting features from the point cloud in each box in the second step exemplarily includes: obtaining the reflection intensity statistics of all points in each box, and obtaining the box reflection intensity statistics corresponding to each box. Correspondingly, the process of using at least one histogram to characterize the distribution properties of the point cloud in the third step exemplarily includes: using a second histogram to characterize the box reflection intensity statistics corresponding to each box in the multiple boxes. For example, the box reflection intensity statistics are obtained by calculating the average value of the reflection intensity of all points in the corresponding box. For example, in some embodiments of the present application, obtaining the reflection intensity statistics of all points in each box includes: calculating the sum of the reflection intensities of all points in the i-th box, wherein the i-th box is any box; calculating the quotient of the sum and the total number of points in the i-th box to obtain the reflection intensity mean. It is not difficult to understand that some embodiments of the present application use a histogram to represent the statistical values ​​of the reflection intensity of the points included in each box, so as to obtain the distribution law of the reflection intensity of the point cloud data of the object to be encoded in the height direction, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0095] Example 3

[0096] For example, in some embodiments of the present application, the process of extracting features from the point cloud in each box in the second step exemplarily includes: obtaining the acquisition time information of all points in each box to obtain the timestamp information of the box point cloud data. The corresponding third step of using at least one histogram to characterize the distribution properties of the point cloud exemplarily includes: using a third histogram to characterize the timestamp information of the box point cloud data corresponding to each box in the multiple boxes. It is not difficult to understand that some embodiments of the present application use a histogram to characterize the acquisition time characteristics of the point cloud included in each box, and then the attributes of the acquisition time of the point cloud of the object to be encoded in the height direction can be obtained, thereby improving the technical effect of the obtained feature vector on the perception of height information characterization.

[0097] It should be noted that those skilled in the art can freely select an embodiment according to actual needs when executing the second and third steps, or freely combine any two embodiments from the above examples 1 to 3, or choose to execute the above examples 1 to 3 at the same time. It is not difficult to understand that the more combined examples there are, the more fully the features of each box can be extracted, thereby improving the characterization effect of the feature vector on the height information of the pillar to be encoded or the voxel to be encoded, and ultimately improving the accuracy of three-dimensional object recognition.

[0098] The implementation process of S130 is described below by way of example.

[0099] In some embodiments of the present application, the process of obtaining the feature vector corresponding to the object to be encoded at least according to the distribution attribute of the point cloud in S130 exemplarily includes: obtaining the position information of the object to be encoded, wherein the position information of the object to be encoded is used to characterize the global position of each object to be encoded; obtaining the feature code according to the position information of the object to be encoded and the distribution attribute of the point cloud. For example, in some embodiments of the present application, the object to be encoded is the pillar to be encoded, wherein the position information is the two-dimensional plane coordinates of the center position of the pillar to be encoded. For example, the process of obtaining the feature vector according to the position information of the object to be encoded and the distribution attribute of the point cloud includes: connecting the data information of each histogram in the at least one histogram and the position information of the object to be encoded to obtain the combined information; projecting the combined information using a linear layer to obtain the feature vector. Some embodiments of the present application can improve the representation method of the distribution law of the original point cloud data by connecting the distribution characteristics of the corresponding categories of point cloud data statistically obtained by each histogram and the position information of the object to be encoded. For example, in some embodiments of the present application, at least one histogram includes a first histogram and a second histogram, and the two histograms are respectively used to characterize the distribution attributes of the point cloud of five boxes from bottom to top, then the data information of connecting each histogram in the at least one histogram exemplarily includes: feature splicing of the distribution attributes of the point cloud used to characterize the same box on the first histogram and the second histogram (for example, splicing the features of each box corresponding to the second histogram behind the features of the box corresponding to the first histogram to obtain a histogram feature with a longer length).

[0100] It is not difficult to understand that some embodiments of the present application fuse the position information of the object to be encoded with the distribution attribute information of the point cloud of the object to be encoded in the height direction, so that the obtained feature vector can better reflect the distribution law of the original point cloud. Some embodiments of the present application use the two-dimensional coordinates of the center position of the pillar to be encoded to represent the position of the object to be encoded, thereby improving the uniformity of the position information.

[0101] The following takes the object to be coded as the pillar to be coded and combines Figure 5 The method for extracting point cloud features provided by some embodiments of the present application is exemplified.

[0102] Some embodiments of the present application propose a method for extracting point cloud features, which is a new histogram-based quantization-aware pillar feature encoder, in which the feature vector obtained is intended to meet at least the following two conditions: 1. Eliminate the need for sampling operations while still maintaining the ability to capture the height perception of each point within the pillar, which means that the encoding method should effectively handle the point density changes at different height levels within the pillar. 2. Preserve the semantic and geometric information within the pillar without relying on maximum pooling or similar pooling operations to ensure that the encoded representation retains the relevant details and features of the original pillar.

[0103] The method for extracting point cloud features provided by some embodiments of the present application uses a histogram distribution technology to record the distribution properties of the point cloud in the box, which can effectively capture the point distribution at different height levels in the column. This enables the embodiments of the present application to maintain the height perception of the point without the need for sampling operations, and the implementation of the histogram operation is relatively simple. For example, some embodiments of the present application divide the height range of the pillar to be encoded into multiple boxes, each box corresponding to a specific height range. For example, if the height range of the pillar to be encoded is from 0 to 10 meters, the pillar to be encoded can be divided into multiple boxes in units of 1 meter, for example, 0-1 meters as the first box, 1-2 meters as the second box, 2-3 meters as the third box, and so on. Next, by counting the number of point clouds in each box in a given pillar, this counting process generates a count value for each box, indicating the number of point clouds located in the specific height range (i.e., the total number of box point clouds). For example, if more points are found in boxes with a lower height range, it means that the point density at the bottom of the pillar is higher. Conversely, if there is a higher concentration of points in bins at higher height ranges, this indicates a higher density of points at the tops of the pillars. This approach provides the model with semantic and geometric information derived from the height dimension, thus enhancing its understanding and representation of point cloud data.

[0104] The code corresponding to the algorithm of the method for extracting point cloud features provided in some embodiments of the present application is as follows:

[0105] 1: Divide the height H of the pillar into B bins, each bin having height h, where

[0106] 2: where H p represents the histogram of points in each bin

[0107] 3: where H irepresents the weighted histogram ofintensity in each bin

[0108] 4:for p i ={p i |i=1,2,...,N v }do

[0109] 5:Compute the corresponding bin index b i based on the z coordinate z i of point p i :

[0110] 6:H p [b i ]←H p [b i ]+1

[0111] 7:H i [b i ]←H i [b i ]+r i {r i is the LiDAR reflection intensity of p i }

[0112] 8:end for

[0113] 9:H i [b i ]:=H i [b i ] / H p [b i ]where H i [b i ]!=0

[0114] 10:Compute the pillar center coordinates x center and y cenler

[0115] 11:

[0116] 12:Linear projection:Y=f θ ([H p [H i |xcenler |y center ])

[0117] The above algorithm flow is exemplarily described below. It can be understood that the feature vector described in S130 can be obtained through this algorithm (corresponding to the operations from S110 to S130).

[0118] The method for extracting point cloud features provided by some embodiments of the present application includes:

[0119] 1. Divide the pillar to be coded into boxes along the height H to obtain B boxes. That is, divide the pillar to be coded with a height of H into B boxes, where B is an integer greater than 1.

[0120] 2. Initialize the height histogram and reflection intensity histogram to 0.

[0121] 3. Calculate the histogram of each pillar to be encoded, denoted as H P ∈R B , where B is the number of bins.

[0122] 4. Combine the LiDAR reflection intensity r at each point i The average reflection intensity of all points in each box is calculated as the average reflection intensity of the box. The weighted intensity histogram is obtained according to the average reflection intensity of all boxes in the height direction, denoted as H. i ∈R B The weighted histogram represents the average LiDAR reflection intensity for each bin, providing valuable height-aware information for object detection.

[0123] 5. Provide the global position of each pillar. For example, use the x and y coordinates of the center of the pillar (i.e. x center and center ) as the global position. Since the coordinates of the cylinder are fixed, no additional calculations are required in this step.

[0124] 6. Connect the calculated histogram and the center coordinates of the cylinder to get [H p |H i |x center |y center ]∈R 2B+2 , that is, the combined information is obtained.

[0125] 7. In order to represent the characteristics of the pillar, a linear layer is used to project the combined information into the feature vector to obtain the feature vector used to characterize the pillar to be encoded.

[0126] It should be noted that the x centerand center The relative spatial relationship between different points in the point cloud is provided. The spatial position information of these pillars reflects the relative position of the target in the scene, helping the model to better understand the environmental context. For example, if there are two pillars with the same point cloud distribution, located at [0,10] and [10,0] in space respectively, and there are no pillar coordinates to indicate the spatial relationship, the two pillars will not be effectively distinguished. This will pose a major challenge to the object perception ability of the detector. Therefore, combining the pillar center x_{center} and y_{center} with other features such as height and intensity can further enhance the perception ability of the model. This is necessary to compensate for the missing spatial position information in the original point cloud data. In some embodiments of the present application, by combining the pillar center coordinates, the model can better capture the geometric structure and spatial relationship between different parts of the scene. This can improve the effect of accurately detecting objects from point cloud data.

[0127] like Figure 5 As shown, the method for extracting pillar features shown in the figure exemplarily includes: dividing the pillar to be encoded into 5 boxes along the height direction, using the first histogram to characterize the number of point cloud data extracted from each box (for example, Figure 5 There are five boxes in total, from the first box to the fifth box, arranged from bottom to top. Among them, if the total number of box point clouds in the first box is 0, the corresponding value in the first histogram is 0, the total number of box point clouds in the second box is 1, the corresponding value in the first histogram is 1, the total number of box point clouds in the third box is 2, the corresponding value in the first histogram is 2, the total number of box point clouds in the fourth box is 1, the corresponding value in the first histogram is 1, and the total number of box point clouds in the fifth box is 3, the corresponding value in the first histogram is 3). The second histogram is used to represent the reflection intensity statistics of the point cloud data in each box (the histogram and the first histogram correspond to the average value of the reflection intensity of all point clouds in each box in the five boxes to obtain the corresponding box reflection intensity statistics). The two-dimensional center of the bottom surface of the pillar to be encoded is used as the global position (that is, the position information of the pillar to be encoded), and then this information is spliced ​​and passed through the multi-layer perceptron MLP (Multi-Layer The feature vector of the pillar to be encoded is obtained by processing at the Perceptron layer, and then all the feature vectors are concatenated to obtain a set of feature vectors corresponding to multiple pillars, and finally object recognition is performed based on the set of feature vectors.

[0128] Understandably, Figure 5 The example only illustrates the three-dimensional object recognition method of the embodiment of the present application by using the first histogram and the second histogram. It can be understood that those skilled in the art can select more histograms to count the distribution characteristics of the point clouds of each box in multiple types of pillars to be encoded according to actual needs. The embodiment of the present application does not limit the number of histograms.

[0129] Some embodiments of the present application provide Figure 5 The histogram-based pillar feature encoding method shown can alleviate the loss of height information and reduce the computational cost of the pillar encoding process. It is worth noting that the pillar encoding method of some embodiments of the present application is a plug-in module that can be easily inserted into existing mainstream pillar-based methods while seamlessly improving their performance. At the same time, the feature vectors obtained in some embodiments of the present application can effectively eliminate the problem of severe performance degradation after quantization caused by existing pillar encoding methods, thereby easily bringing performance improvement.

[0130] It should be noted that the method of extracting point cloud features in some embodiments of the present application can be applied to the field of autonomous driving or other solutions that require the use of point cloud data for target detection. For example, the three-dimensional object recognition method of the embodiment of the present application can express the pillar features by encoding the point cloud in the pillar in the form of a histogram according to the height direction to obtain a feature vector, and then perform three-dimensional object recognition based on the feature vector. The purpose of the three-dimensional object recognition method provided in the embodiment of the present application is to ensure the improvement of detection performance while saving computing overhead, thereby improving the perception ability of the autonomous driving system to the external environment.

[0131] like Figure 6 As shown, some embodiments of the present application provide a method for identifying an object, the method comprising:

[0132] S210, obtaining a feature vector, that is, obtaining a feature vector obtained by using the above-mentioned method for extracting point cloud features. S220, identifying an object in a three-dimensional space according to the feature vector.

[0133] It should be noted that the implementation process of S210 may be specifically referred to the above description, and will not be described in detail here to avoid repetition.

[0134] The process of identifying an object in a three-dimensional space according to the feature vector described in S210 above exemplarily includes: obtaining a pseudo image according to the feature vector of each pillar to be encoded and then identifying the three-dimensional object according to the pseudo image. The specific process can refer to the implementation details of the relevant technology, and no further details will be given here to avoid repetition.

[0135] The 3D object recognition method of some embodiments of the present application captures the distribution of point clouds in different height intervals within the pillar when encoding. This enables some embodiments of the present application to maintain the semantic and geometric information within the pillar. In addition, some embodiments of the present application only operate at the pillar encoding stage without interacting with other modules, so that it can be seamlessly inserted into the existing pillar encoding-based detector method (for example, replace the encoding method of obtaining feature vectors in the embodiment of the present application with ... Figure 1The encoding algorithm used by the point-pillar feature encoding network is used as a plug-in module to improve the performance of the existing pillar-based point cloud detector. At the same time, it effectively converts the point-by-point operation of the multi-layer perceptron in the original pillar encoding process into a pillar-by-pillar operation, effectively improving the computational efficiency.

[0136] Please refer to Figure 7 , Figure 7 The device for extracting point cloud features provided in the embodiment of the present application is shown. It should be understood that the device is similar to the above-mentioned Figure 4 The method embodiment corresponds to the method embodiment and can execute each step involved in the above method embodiment. The specific functions of the device can be referred to in the above description. To avoid repetition, the detailed description is appropriately omitted here. The device includes at least one software function module that can be stored in the memory in the form of software or firmware or solidified in the operating system of the device. The device for extracting point cloud features includes: a point cloud data acquisition module 510, a height feature extraction module 520 and an encoding module 530.

[0137] The point cloud data acquisition module 510 is configured to acquire an object to be encoded, wherein the object to be encoded includes a pillar to be encoded or a voxel to be encoded.

[0138] The height feature extraction module 520 is configured to extract and record the distribution properties of the point cloud of the object to be encoded within different height ranges.

[0139] The encoding module 530 is configured to obtain a feature vector corresponding to the cylinder to be encoded at least according to the distribution attribute of the point cloud.

[0140] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method for extracting point cloud features, and will not be elaborated here.

[0141] Please refer to Figure 8 , Figure 8 The device for identifying an object provided in an embodiment of the present application is shown. It should be understood that the device is similar to the above-mentioned Figure 6 The method embodiment corresponds to the method embodiment, and can execute each step involved in the above method embodiment. The specific functions of the device can be referred to in the above description. To avoid repetition, the detailed description is appropriately omitted here. The device includes at least one software function module that can be stored in the memory in the form of software or firmware or fixed in the operating system of the device. The device for identifying objects includes: a feature vector acquisition module 610 (the algorithm executed by the module is as follows Figure 4 As shown) and identification module 620.

[0142] The feature vector acquisition module is configured to acquire the feature vector.

[0143] The recognition module is configured to recognize objects in the three-dimensional space according to the feature vector.

[0144] Technicians in the relevant field can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method for identifying three-dimensional objects or the method for extracting point cloud features, and will not be elaborated here.

[0145] Some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method described in any one of the embodiments of the method for identifying a three-dimensional object can be implemented.

[0146] Some embodiments of the present application provide a computer program product, which includes a computer program. When the computer program is executed by a processor, it can implement the method described in any one of the embodiments of the method for identifying a three-dimensional object.

[0147] like Fig. 9 As shown, some embodiments of the present application provide an electronic device 700, including a memory 710, a processor 720, and a computer program stored in the memory 710 and executable on the processor, wherein the processor 720 reads the program from the memory 710 through a bus 730 and executes the program to implement a method described in any one of the embodiments of the method for identifying a three-dimensional object as described above.

[0148] Processor 720 can process digital signals and can include various computing structures, such as complex instruction set computer structure, reduced instruction set computer structure, or a structure that implements a combination of multiple instruction sets. In some examples, processor 720 can be a microprocessor.

[0149] The memory 710 may be used to store instructions executed by the processor 720 or data related to the execution of instructions. These instructions and / or data may include code to implement some or all functions of one or more modules described in the embodiments of the present application. The processor 720 of the present embodiment may be used to execute the instructions in the memory 710 to implement Figure 4 or Figure 6 The memory 710 includes a dynamic random access memory, a static random access memory, a flash memory, an optical memory or other memory known to those skilled in the art.

[0150] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0151] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0152] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0153] The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0154] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0155] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. A method for extracting point cloud features, characterized in that: The method comprises: Acquire an object to be encoded, wherein the object to be encoded includes a pillar to be encoded or a voxel to be encoded; Extracting and recording the distribution attributes of the point cloud of the object to be encoded within different height ranges; A feature vector corresponding to the object to be encoded is obtained at least according to a distribution attribute of the point cloud.

2. The method according to claim 1, characterized in that The extracting and recording the distribution attributes of the point cloud of the object to be encoded within different height ranges includes: Segmenting the object to be encoded along the height direction of the object to be encoded to obtain a plurality of boxes arranged in sequence from bottom to top; Feature extraction is performed on the point cloud in each box to obtain the distribution properties of the point cloud.

3. The method according to claim 2, characterized in that The step of extracting and recording the distribution attributes of the point cloud of the object to be encoded within different height ranges also includes: At least one histogram is used to characterize the distribution properties of the point cloud.

4. The method according to claim 1, characterized in that The distribution attributes of the point cloud include at least one of the following information: the total number of box point clouds, box reflection intensity statistics, and box point cloud acquisition timestamp information.

5. The method according to claim 3, characterized in that The feature extraction of the point cloud in each box includes: Counting the number of points in each box to obtain the total number of box point clouds corresponding to each box; The adopting at least one histogram to characterize the distribution attribute of the point cloud comprises: The first histogram is used to represent the total number of box point clouds corresponding to each box in the multiple boxes.

6. The method according to any one of claims 1 to 5, characterized in that The feature extraction of the point cloud in each box includes: Obtain the reflection intensity statistics of all points in each box to obtain the box reflection intensity statistics corresponding to each box; The adopting at least one histogram to characterize the distribution attribute of the point cloud comprises: A second histogram is used to represent the box reflection intensity statistics corresponding to each box in the plurality of boxes.

7. The method according to claim 6, characterized in that The obtaining of the reflection intensity statistics of all points in each box includes: Calculate the sum of the reflection intensities of all points in the i-th box, where the i-th box is any box; The reflection intensity mean is obtained by calculating the quotient of the sum and the total number of midpoints of the i-th box.

8. The method according to claim 3, characterized in that The feature extraction of the point cloud in each box includes: Get the acquisition time information of all points in each box, and get the box point cloud acquisition timestamp information; The adopting at least one histogram to characterize the distribution attribute of the point cloud comprises: A third histogram is used to represent the box point cloud acquisition timestamp information corresponding to each box in the multiple boxes.

9. The method according to claim 3, characterized in that The obtaining a feature vector corresponding to the object to be encoded at least according to the distribution attribute of the point cloud includes: Acquire the position information of the object to be encoded, wherein the position information of the object to be encoded is used to represent the global position of each object to be encoded; The feature vector is obtained according to the position information of the object to be encoded and the distribution attribute of the point cloud.

10. The method according to claim 9, characterized in that The object to be encoded is the pillar to be encoded, wherein the position information of the object to be encoded is the two-dimensional plane coordinates of the center position of the pillar to be encoded.

11. The method according to claim 9, characterized in that The obtaining the feature vector according to the position information of the object to be encoded and the distribution attribute of the point cloud includes: Connecting data information of each histogram in the at least one histogram and position information of the object to be encoded to obtain combined information; The combined information is projected using a linear layer to obtain the feature vector.

12. A method for identifying an object, characterized in that: The method comprises: Obtaining a feature vector obtained by the method according to any one of claims 1 to 11; Objects in the three-dimensional space are identified according to the feature vector.

13. A device for identifying an object, characterized in that: The device comprises: A feature vector acquisition module is configured to acquire a feature vector; The recognition module is configured to recognize objects in the three-dimensional space according to the feature vector.

14. A device for extracting point cloud features, characterized in that: The device comprises: A point cloud data acquisition module is configured to acquire an object to be encoded, wherein the object to be encoded includes a pillar to be encoded or a voxel to be encoded; A height feature extraction module is configured to extract and record distribution properties of the point cloud of the object to be encoded within different height ranges; The encoding module is configured to obtain a feature vector corresponding to the cylinder to be encoded at least according to the distribution attribute of the point cloud.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 12 can be implemented.

16. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 12 can be implemented.

17. A computer program product, characterized in that The computer program product comprises a computer program, wherein the computer program can implement the method according to any one of claims 1 to 12 when executed by a processor.

Images