Lidar point cloud data pseudo-image encoding method for three-dimensional target detection

Abstract

The application discloses a laser radar point cloud data pseudo-image coding method for three-dimensional target detection, solves the problems of complex coding process, high information loss and low coding efficiency in the prior art, realizes efficient coding of laser radar point cloud data, and has high precision and real-time performance in target detection, and the method comprises the following steps: acquiring laser radar point cloud data, calculating the grid space coordinates corresponding to each point of the laser radar point cloud data and the reflection intensity corresponding to the grid space coordinates according to the detection range and the grid space division resolution parameter of the laser radar point cloud data; coding the grid space coordinates corresponding to each point of the laser radar point cloud data to obtain position distribution information of a height perception space as a first pseudo-image; coding the reflection intensity to obtain maximum intensity distribution information as a second pseudo-image; and splicing the first pseudo-image and the second pseudo-image to obtain a coding feature pseudo-image.

Description

Pseudo-image encoding method for LiDAR point cloud data for 3D target detection Technical Field

[0001] This invention relates to the field of lidar point cloud data processing and 3D target detection technology, and in particular to a pseudo-image encoding method for lidar point cloud data for 3D target detection. Background Technology

[0002] With the rapid development of technologies such as autonomous driving and intelligent sensing, LiDAR, as a sensor capable of acquiring high-precision 3D environmental information, has seen its collected point cloud data widely applied in target detection tasks. Compared with traditional image data, LiDAR point cloud data exhibits significant sparsity and disorder. When directly using deep learning methods for target detection, it needs to be encoded first. Encoding can solve the problems of disorder and sparsity in the original point cloud by generating features independent of point order through voxelization, attention mechanisms, or pooling, compensating for the uneven spatial distribution, adapting to the ordered input requirements of models such as convolutional neural networks, and ensuring high accuracy and real-time performance in 3D target detection.

[0003] Currently, mainstream LiDAR point cloud encoding methods are mainly divided into three categories: point-based methods, voxel-based methods, and column-based methods. Point-based methods directly process the raw point cloud, preserving its original features, but require complex networks to handle irregular data, resulting in high computational cost, demanding hardware, and difficulty in achieving real-time detection. Voxel-based methods mesh the point cloud and aggregate features; the voxelized 3D features require 3D sparse convolution, leading to high network complexity and demanding hardware. Column-based methods are widely used in real-time target detection. They divide the point cloud data into vertical columnar units, aggregate features within these units to form encoded data, and encode the raw point cloud data into a 2D pseudo-image, allowing direct feature extraction using 2D neural networks. PointPillars and DynamicPFE methods involve fully connected layers and batch normalization during encoding, causing redundant computational resource consumption and failing to fully utilize height information, resulting in the loss of some height features and affecting the final target detection accuracy.

[0004] In conclusion, it is particularly important to find a lidar point cloud data encoding method with low computational complexity and high accuracy. Summary of the Invention

[0005] This invention provides a pseudo-image encoding method for LiDAR point cloud data for 3D target detection, which solves the problems of complex encoding process, high information loss and low encoding efficiency in the prior art, and realizes efficient encoding of LiDAR point cloud data with high accuracy and real-time performance in target detection.

[0006] This invention provides a pseudo-image encoding method for lidar point cloud data for 3D target detection, the method comprising:

[0007] Acquire lidar point cloud data, and calculate the grid space coordinates and reflection intensity corresponding to each point in the lidar point cloud data based on the detection range and grid space division resolution parameters of the lidar point cloud data.

[0008] Encode the grid spatial coordinates corresponding to each point in the lidar point cloud data to obtain the positional distribution information of the height perception space as the first pseudo image;

[0009] The reflection intensity is encoded to obtain the distribution information of the maximum intensity value as a second pseudo image;

[0010] The first pseudo-image and the second pseudo-image are concatenated to obtain the coded feature pseudo-image.

[0011] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0012] This invention employs the acquisition of LiDAR point cloud data. Based on the detection range and grid space division resolution parameters of the LiDAR point cloud data, the grid space coordinates and reflection intensity corresponding to each point in the LiDAR point cloud data are calculated. This step converts the original unordered point cloud data into a structured grid representation, thereby simplifying the data processing flow and improving computational efficiency, while retaining key spatial locations and reflection intensities, laying the foundation for subsequent encoding. The grid space coordinates corresponding to each point in the LiDAR point cloud data are encoded to obtain the positional distribution information of the height perception space as the first pseudo-image. The three-dimensional point cloud data is converted into a two-dimensional image format, making the height information visible and facilitating the application of mature image processing technologies. The algorithm simultaneously enhances the perception of changes in object height within a scene, aiding in terrain analysis and object recognition. It encodes reflection intensity to obtain the distribution information of maximum intensity values ​​as a second pseudo-image. This step utilizes the reflection intensity data to encode an image, highlighting differences in the physical properties of object surfaces, thus providing supplementary information and enhancing the distinguishability of different objects, especially in low-light or complex environments. The first and second pseudo-images are then concatenated to obtain an encoded feature pseudo-image. This concatenation operation integrates spatial location and intensity information to generate a multi-channel feature image, providing a more comprehensive and richer feature representation, which helps improve the performance of subsequent machine learning or computer vision tasks, such as object detection, classification, and scene understanding. Attached Figure Description

[0013] Figure 1 is a flowchart of the pseudo-image encoding method for lidar point cloud data for three-dimensional target detection provided in an embodiment of the present invention;

[0014] Figure 2 is a schematic diagram of the encoding process of the pseudo-image encoding method for lidar point cloud data for three-dimensional target detection provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] This invention provides a pseudo-image encoding method for lidar point cloud data for three-dimensional target detection. As shown in Figure 1, the method includes the following steps S101 to S104.

[0017] S101, acquire lidar point cloud data, and calculate the grid space coordinates and reflection intensity corresponding to each point in the lidar point cloud data according to the detection range and grid space division resolution parameters of the lidar point cloud data.

[0018] Specifically, in step S101, based on the detection range and grid space division resolution parameters of the lidar point cloud data, the grid space coordinates corresponding to each point of the lidar point cloud data and the reflection intensity corresponding to the grid space coordinates are calculated, including the following steps S1011 to S1013.

[0019] S1011, determine the coordinate information and reflection intensity of each point in the lidar point cloud data, and obtain the grid coordinates and reflection intensity corresponding to each point based on the coordinate information and reflection intensity;

[0020] Here, the grid coordinates corresponding to each point are represented as follows:

[0021] (1.1);

[0022] in, Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; This indicates the coordinate system of the current point in the lidar point cloud data. Axis coordinates; Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; Indicates the current point in the lidar point cloud data The intensity of reflection; Indicates the number of points in the lidar point cloud data; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis.

[0023] S1012, calculate the grid space range in integer form based on the detection range and the resolution parameters of the grid space division;

[0024] S1013, based on the detection range, grid coordinates, and grid space division resolution parameters, calculates the grid space coordinates and reflection intensity corresponding to each point in the lidar point cloud data.

[0025] Here, the grid space coordinates are represented as:

[0026] (1.2);

[0027] (1.3);

[0028] (1.4);

[0029] in, Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; This indicates the coordinate system of the current point in the lidar point cloud data. Axis coordinates; Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; This indicates that the numerical value will be converted into an integer form.

[0030] For example, the number of points in the lidar point cloud data is Each point has , , Coordinate information and reflection intensity , The axis represents the forward direction of the radar. The axis is the leftward direction of the radar. The axis represents the vertical height direction of the radar, and the grid coordinates corresponding to each point are expressed by formula (1.1).

[0031] Based on the detection range of LiDAR point cloud data and grid space partitioning resolution parameters Calculate the grid space range in integer form. , , The specific calculation method is as follows:

[0032] (1.5);

[0033] (1.6);

[0034] (1.7);

[0035] in, This indicates that the numerical value is converted into an integer form. Represents the real spatial region along The total number of grid cells divided along the axial direction. Represents the real spatial region along The total number of grid cells divided along the axial direction. Represents the real spatial region along The total number of grids divided along the axis.

[0036] Specifically, assuming the preset lidar point cloud detection range is: The grid space partitioning resolution parameter is Then the grid space range , , .

[0037] Calculate the grid space coordinates corresponding to the lidar point cloud data , , For specific calculation methods, please refer to formulas (1.2), (1.3) and (1.4).

[0038] S102, Encode the grid spatial coordinates corresponding to each point in the lidar point cloud data to obtain the location distribution information of the height perception space as the first pseudo image;

[0039] Specifically, in step S102, the grid spatial coordinates corresponding to each point in the lidar point cloud data are encoded to obtain the position distribution information of the height perception space as the first pseudo image, including the following steps S1021 to S1023.

[0040] S1021, Create shape as The first all-zero tensor, and the initial first pseudo-image is obtained based on the first all-zero tensor; where, Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grid cells divided along the axis;

[0041] S1022, Traverse each point in the lidar point cloud data and use the grid space coordinates corresponding to the current input point as the first index;

[0042] S1023, Modify the element label state at the corresponding position in the initial first pseudo-image according to the first index to obtain the first pseudo-image (spatial feature).

[0043] For example, create a shape as The first all-zero tensor is used to store the first pseudo-image.

[0044] Traverse all points in the lidar point cloud data, based on the grid spatial coordinates of the input points. , , As the first index, the value of the corresponding position in the first pseudo-image is set to 1, marking the existence of a point cloud within that grid.

[0045] The specific formula is as follows:

[0046] (1.8);

[0047] Based on the grid spatial coordinates of each point in the lidar point cloud data ( , , ), marking the presence of location-related features in the first pseudo-image.

[0048] Specifically, the encoded data indicates whether there are points within a grid space block divided by coordinates; the coordinates of the block implicitly contain their values. , , The first pseudo-image describes the spatial distribution of point clouds within the grid space, specifically the distance and contour of the target within the grid space. Furthermore, the data format of the first pseudo-image is consistent with the input data (C, H, W) of a standard 2D neural network. Dividing the pseudo-image along the xy-space coordinates represents the width and height dimensions of the neural network input data, while dividing it along the height represents the channel dimensions of the neural network input data. Therefore, we can use operations such as convolutional neural networks to efficiently process the encoded data.

[0049] S103, Encode the reflection intensity to obtain the maximum intensity distribution information as the second pseudo-image;

[0050] Specifically, in step S103, the reflection intensity is encoded to obtain the intensity maximum value distribution information as the second pseudo image, including the following steps S1031 to S1033.

[0051] S1031, Create shape as The second all-zero tensor is used to initialize the second pseudo-image, resulting in an initial second all-zero tensor; where, Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grid cells divided along the axis;

[0052] S1032, iterate through each point in the LiDAR point cloud data, and input the grid space coordinates corresponding to the current point. Axis coordinates and The axis coordinates are used as the second index;

[0053] S1033, using the second index, update the reflection intensity at the corresponding position in the initial second all-zero tensor according to the reflection intensity corresponding to the current point, and obtain the second all-zero tensor.

[0054] Here, using the second index, the reflection intensity at the corresponding position in the initial second all-zero tensor is updated according to the reflection intensity corresponding to the current point, resulting in the second all-zero tensor, which includes:

[0055] Get the first reflection intensity of the current point corresponding to the second index at the corresponding position in the initial second all-zero tensor;

[0056] If the reflection intensity at the current point is greater than the first reflection intensity, then update the first reflection intensity to the reflection intensity at the current point;

[0057] If the reflection intensity at the current point is less than or equal to the first reflection intensity, then the first reflection intensity remains unchanged.

[0058] For example, create a shape as The second all-zero tensor is used to store the second pseudo-image (intensity feature).

[0059] Traverse all points in the lidar point cloud data, based on the grid space coordinates of the input points. , As a second index, it determines the current input. If the reflection intensity is greater than the value of the corresponding position in the second pseudo-image, then change the reflection intensity value of the corresponding position in the second pseudo-image to the current input. .

[0060] (1.9);

[0061] To better utilize the reflection intensity in point cloud data, the maximum intensity of all grids along the xy-space coordinates was statistically analyzed, yielding reflection intensity data. This data represents the intensity characteristics of each grid within the xy-space, i.e., from the point cloud bird's-eye view, and can help distinguish targets of different materials and distances. Furthermore, the second pseudo-image data format is consistent with the input data (C, H, W) format of a standard 2D neural network. The pseudo-image is divided along the xy-space coordinates into the width and height dimensions of the neural network input data, and one information dimension of the intensity feature information is the channel dimension of the neural network input data.

[0062] S104, the first pseudo-image and the second pseudo-image are concatenated to obtain the coded feature pseudo-image.

[0063] Specifically, in step S104, the first pseudo-image and the second pseudo-image are concatenated to obtain the coded feature pseudo-image, including:

[0064] The first pseudo-image and the second pseudo-image are concatenated along the channel dimension to obtain the encoded feature pseudo-image, wherein the shape of the first pseudo-image is as follows: The shape of the second pseudo-image is The shape of the encoded feature pseudo-image is .

[0065] For example, a concatenation operation is used to concatenate the first pseudo-image (spatial feature) and the second pseudo-image (intensity feature) along the i-th channel dimension to obtain an encoded feature pseudo-image. The shape of the encoded feature pseudo-image is... .

[0066] By stitching together information from the first and second pseudo-images, spatial distribution features with high-sensitivity point cloud data and point cloud intensity features from a bird's-eye view are generated. This dual-feature fusion enhances feature representation capabilities. Furthermore, the encoding process involves only simple operations such as grid coordinate mapping, feature labeling, and intensity statistics, directly manipulating the original point cloud without requiring voxelization or complex operations like full connection and batch normalization. This reduces operational complexity, enables efficient parallel processing, significantly lowers encoding latency, and meets real-time detection requirements.

[0067] Furthermore, after the above data encoding, the lidar point cloud data forms a shape of... The pseudo-image. Specifically, assuming a preset LiDAR point cloud detection range. and grid space partitioning resolution parameters Then the coded feature pseudo-image is the shape. The pseudo-image has a length and width of 496 and 432, and the number of feature channels is 33.

[0068] Referring to Figure 2, firstly, LiDAR point cloud data is acquired, with each point including grid spatial coordinates and reflection intensity. The real space containing the LiDAR point cloud data is divided into a three-dimensional grid space, and the number of grids is calculated based on preset detection range and resolution parameters.

[0069] Then, the grid spatial coordinates corresponding to each point in the lidar point cloud data are encoded to obtain the location distribution information of the height perception space as the first pseudo-image.

[0070] Next, the reflection intensity of the point cloud data is encoded: each point is traversed, and the maximum reflection intensity of each point in the two-dimensional grid is counted to generate a second pseudo-image.

[0071] Finally, the first and second pseudo-images are concatenated along the channel dimension to obtain the encoded feature pseudo-image. This pseudo-image can be directly input into a two-dimensional convolutional neural network for three-dimensional object detection, simplifying the processing flow and improving efficiency.

[0072] The effectiveness of this invention can be further illustrated by the following experimental data.

[0073] This invention employs the method described in this paper, along with the PointPillars, DynamicPFE, and TinypillarNet methods. Experiments are conducted on the KITTI dataset validation set for comparative analysis. The KITTI dataset, jointly created by the Karlsruhe Institute of Technology in Germany and Toyota Research Institute of America, is one of the most commonly used international datasets for evaluating computer vision algorithms in autonomous driving scenarios.

[0074] Below, the average accuracy (3DmAP@R40) and coding latency are used to quantitatively compare and evaluate the performance of the proposed pseudo-image coding method for LiDAR point cloud data for 3D target detection and existing methods.

[0075] 3DmAP@R40 is the official evaluation metric provided by the KITTI dataset. It represents the average precision calculated based on 40 recall interpolation points at a 3DIoU threshold, reflecting the performance of the object detection model. Encoding latency refers to the time required to complete the operation. A lower encoding latency indicates that the method requires less time to implement.

[0076] To ensure fairness in the experiment, the comparative test used the feature extraction network and detection head proposed by PointPillars. The encoder was replaced by the encoder proposed in this invention, the DynamicPFE method, and the TinypillarNet method, respectively. Experiments were conducted on the KITTI dataset validation set to verify the effect of the encoder replacement. The detection range of the LiDAR point cloud data was [missing information]. The resolution of the coding grid is The grid space range is The encoding latency test was conducted on a CPU (specifically, an AMD EPYC7542).

[0077] The experimental results of the comparative test are shown in Table 1.

[0078] Table 1 Comparison of Quantitative Parameters in the Comparative Test Results

[0079]

[0080] As shown in Table 1, the method proposed in this invention was tested on the KITTI dataset validation set. The average accuracy (3DmAP@R40) of this method outperforms the PointPillars, DynamicPFE, and TinypillarNet methods, demonstrating that the proposed encoding method can improve the accuracy of 3D object detection compared to other encoding methods. In terms of encoding latency, the proposed method has lower encoding latency on CPU devices compared to the PointPillars, DynamicPFE, and TinypillarNet methods, indicating that the proposed method can effectively reduce encoding latency.

[0081] In summary, the pseudo-image encoding method for lidar point cloud data for 3D target detection proposed in this invention can effectively improve the average accuracy of target detection while reducing the latency of encoding operations.

[0082] This invention provides a pseudo-image encoding method for lidar point cloud data for 3D target detection. The encoding process involves only simple operations such as grid coordinate mapping, feature labeling, and intensity statistics. It does not require voxelization and directly processes the point cloud data. It eliminates redundant calculations such as fully connected layers and batch normalization, enabling efficient parallel processing, significantly reducing encoding latency, and meeting real-time detection requirements.

[0083] This invention provides a pseudo-image encoding method for LiDAR point cloud data for 3D target detection. The first pseudo-image fully preserves the 3D spatial distribution features of the point cloud, especially the height direction information, solving the problem of lost height features in traditional cylindrical encoding. The second pseudo-image aggregates the maximum reflection intensity feature, which can help distinguish targets of different materials and distances. The fusion of the two features improves the feature expression capability.

[0084] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A pseudo-image encoding method for lidar point cloud data for 3D target detection, characterized in that, include: Acquire lidar point cloud data, and calculate the grid space coordinates and reflection intensity corresponding to each point of the lidar point cloud data based on the detection range and grid space division resolution parameters of the lidar point cloud data; encode the grid space coordinates corresponding to each point of the lidar point cloud data to obtain the position distribution information of the height perception space as the first pseudo image. The step of encoding the grid spatial coordinates corresponding to each point in the lidar point cloud data to obtain the location distribution information of the height-sensing space as the first pseudo-image includes: creating a shape of The first all-zero tensor, and the initial first pseudo-image is obtained based on the first all-zero tensor; wherein, Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grids divided along the axis; traversing each point in the lidar point cloud data, using the grid space coordinates corresponding to the current input point as the first index; modifying the element label state at the corresponding position in the initial first pseudo-image according to the first index to obtain the first pseudo-image; encoding the reflection intensity to obtain the intensity maximum value distribution information as the second pseudo-image; the encoding of the reflection intensity to obtain the intensity maximum value distribution information as the second pseudo-image includes: creating a shape of The second all-zero tensor is used to initialize the second pseudo-image, thus obtaining the initial second all-zero tensor; where, Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grids divided along the axis; traversing each point in the lidar point cloud data, and inputting the grid space coordinates corresponding to the current point. Axis coordinates and The axis coordinates are used as the second index; using the second index, the reflection intensity at the corresponding position in the initial second all-zero tensor is updated according to the reflection intensity corresponding to the current point to obtain the second all-zero tensor, and the second pseudo-image is obtained according to the second all-zero tensor; the first pseudo-image and the second pseudo-image are concatenated to obtain the coded feature pseudo-image; the concatenation of the first pseudo-image and the second pseudo-image to obtain the coded feature pseudo-image includes: concatenating the first pseudo-image and the second pseudo-image along the channel dimension to obtain the coded feature pseudo-image, wherein the shape of the first pseudo-image is... The shape of the second pseudo-image is The shape of the coded feature pseudo-image is ,in, Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grid cells divided along the axis; Represents the real spatial region along The total number of grids divided along the axis.

2. The pseudo-image encoding method for lidar point cloud data for three-dimensional target detection according to claim 1, characterized in that, The step of calculating the grid space coordinates and reflection intensity corresponding to each point in the lidar point cloud data based on the detection range and grid space division resolution parameters of the lidar point cloud data includes: determining the coordinate information and reflection intensity of each point in the lidar point cloud data; obtaining the grid coordinates and reflection intensity corresponding to each point based on the coordinate information and reflection intensity; calculating the grid space range in integer form based on the detection range and the grid space division resolution parameters; and calculating the grid space coordinates and reflection intensity corresponding to each point in the lidar point cloud data based on the detection range, the grid coordinates, and the grid space division resolution parameters.

3. The pseudo-image encoding method for lidar point cloud data for three-dimensional target detection according to claim 2, characterized in that, The grid coordinates corresponding to each point are represented as follows: ;in, Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; This indicates the coordinate system of the current point in the lidar point cloud data. Axis coordinates; Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; Indicates the current point in the lidar point cloud data The intensity of reflection; Indicates the number of points in the lidar point cloud data; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Maximum value of the axis; This indicates the first [item] in the lidar point cloud data. One point.

4. The pseudo-image encoding method for lidar point cloud data for three-dimensional target detection according to claim 1, characterized in that, The grid space coordinates are represented as follows: ； ； ;in, Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; This indicates the coordinate system of the current point in the lidar point cloud data. Axis coordinates; Indicates the current point in the lidar point cloud data In the real coordinate system Axis coordinates; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Minimum value of axis; Represents lidar point cloud data Minimum value of axis; This indicates that the numerical value has been converted to an integer form; Indicates the current point In grid space coordinates Axis coordinates; Indicates the current point In grid space coordinates Axis coordinates; Indicates the current point In grid space coordinates Axis coordinates; This represents the resolution parameter for grid space partitioning.

5. The pseudo-image encoding method for lidar point cloud data for three-dimensional target detection according to claim 1, characterized in that, The step of using the second index to update the reflection intensity at the corresponding position in the initial second all-zero tensor according to the reflection intensity corresponding to the current point to obtain the second all-zero tensor includes: obtaining the first reflection intensity at the corresponding position of the current point corresponding to the second index in the initial second all-zero tensor; if the reflection intensity of the current point is greater than the first reflection intensity, then updating the first reflection intensity to the reflection intensity of the current point; if the reflection intensity of the current point is less than or equal to the first reflection intensity, then the first reflection intensity remains unchanged.

Images