A high-resolution processing method, device and computer storage medium

Abstract

The present application provides a high-resolution processing method, device and computer storage medium. The processing method includes: obtaining a low-resolution point cloud; projecting the low-resolution point cloud into a first distance map according to the angle parameters of the laser radar, including a distance map of a coordinate channel and a distance map of a distance channel; inputting the first distance map into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map; based on the addition of the attention distance map and the high-resolution distance map, a second distance map is obtained; and the second distance map is reversely projected to obtain a high-resolution point cloud. In the above manner, the high-resolution processing method of the present application can improve the resolution of the point cloud data collected by the laser radar, and can achieve better display effects with low-cost resources.

Description

Technical Field

[0001] The present application relates to the field of laser point cloud technology, and in particular to a high-resolution processing method, device, and computer storage medium. Background Art

[0002] The working principle of lidar is to emit a beam of laser to the surroundings, then compare the reflected signal from the surroundings with the emitted signal, and then process it to obtain information (for example, the distance and intensity to surrounding objects). LiDAR is often used for map construction, positioning, and perception of autonomous driving. In indoor environments, the scenes are relatively limited, and it is generally assumed that the robot moves on a fixed plane, so it is possible to use only a single-line laser and reduce costs. However, outdoors, the environment is more complex, and height information cannot be directly simplified. Therefore, single-line lidar is difficult to apply in outdoor scenes, so try to use high-resolution lidar or single-line lidar with the ability to rotate in additional degrees of freedom to improve the scanning data.

[0003] However, the ability to rotate in additional degrees of freedom with a single-beam LiDAR does not change the total number of points captured, thus failing to capture more spatial information in the same amount of time. In many cases, the point clouds scanned by low-beam LiDAR are sparse, making it difficult to distinguish road information when creating point cloud maps. Therefore, current technical solutions for acquiring high-resolution point cloud data are costly, both in terms of time and hardware, making them uneconomical. Summary of the Invention

[0004] The present application provides a high-resolution processing method, device, and computer storage medium.

[0005] The present application provides a high-resolution processing method, which includes:

[0006] Get low-resolution point cloud;

[0007] Projecting the low-resolution point cloud into a first distance map according to the angle parameter of the laser radar, wherein the first distance map includes a distance map of a coordinate channel and a distance map of a distance channel;

[0008] Inputting the first distance map into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, wherein the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map;

[0009] Obtaining a second distance map based on the addition of the attention distance map and the high-resolution distance map;

[0010] The second distance map is back-projected to obtain a high-resolution point cloud.

[0011] The projecting of the low-resolution point cloud into a first distance map according to the angle parameter of the laser radar includes:

[0012] Calculating a horizontal angle set of the low-resolution point cloud in the first distance map based on a horizontal angle parameter of the laser radar;

[0013] Based on the three-dimensional coordinates of each point cloud data in the low-resolution point cloud and the horizontal angle set, obtaining an index of each point cloud data in the horizontal direction;

[0014] Calculating a vertical angle set of the low-resolution point cloud in the first distance map based on a vertical angle parameter of the laser radar;

[0015] Based on the three-dimensional coordinates of each point cloud data in the low-resolution point cloud and the vertical angle set, obtaining an index of each point cloud data in the vertical direction;

[0016] The relevant point cloud information is projected according to the horizontal index and the vertical index of each point cloud data to form the first distance map.

[0017] Wherein, obtaining a low-resolution point cloud includes:

[0018] Obtaining the coordinates of each point cloud data in the low-resolution point cloud;

[0019] The distance of each point cloud data is calculated based on the coordinates of each point cloud data.

[0020] Before inputting the first distance map into a pre-trained neural network, the high-resolution processing method further includes:

[0021] Acquire a horizontal resolution of the first distance map;

[0022] Calculating a left padding amount based on the horizontal resolution;

[0023] Calculating a right padding number based on the left padding number and the horizontal resolution;

[0024] The first distance map is cyclically filled based on the left filling number and the right filling number.

[0025] The cyclically filling the first distance map based on the left filling number and the right filling number includes:

[0026] Filling the left side of the first distance map with a pixel set whose number of columns on the right side of the first distance map is equal to the number of left padding columns;

[0027] Filling the right side of the first distance map with a pixel set whose number of columns on the left side of the first distance map is equal to the number of right-side padding;

[0028] Repeat the following steps until the new left padding number is 0: calculate a new left padding number based on the filled first distance map, calculate a right padding number based on the left padding number and the horizontal resolution; and cyclically fill the first distance map based on the left padding number and the right padding number.

[0029] The resolution network includes an encoder network and a decoder network, and the encoder network has an opposite structure to the decoder network; the encoder network and / or the decoder network include several convolutional layers, batch normalization layers and activation functions, and the input of the encoder network also includes a transposed convolutional layer.

[0030] The channel attention network includes a convolutional layer and a channel attention module; the channel attention module is used to reshape the distance map of the distance channel into a first distance channel map and a second distance channel map, multiply the first distance channel map and the second distance channel map to obtain a channel attention map, perform matrix multiplication on the channel attention map and the first distance channel map, reshape it back to a reshaped image of the size of the distance map of the distance channel, and finally add it to the distance map of the distance channel to obtain the output of the channel attention module;

[0031] The first distance channel image and the second distance channel image are transposed images.

[0032] Wherein, the high-resolution processing method further includes:

[0033] Using a low-resolution laser radar to collect low-resolution point cloud data to obtain a first training point cloud;

[0034] Using a high-resolution laser radar to collect high-resolution point cloud data to obtain a second training point cloud;

[0035] The first training point cloud and the second training point cloud are input into the neural network to train the neural network.

[0036] The present application also provides a high-resolution processing device, the high-resolution processing device comprising a memory and a processor, wherein the memory is coupled to the processor;

[0037] The memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned high-resolution processing method.

[0038] The present application also provides a computer storage medium, which is used to store program data. When the program data is executed by a processor, it is used to implement the above-mentioned high-resolution processing method.

[0039] The beneficial effects of the present application are as follows: a high-resolution processing device obtains a low-resolution point cloud; the low-resolution point cloud is projected into a first distance map according to the angle parameters of the laser radar, including a distance map of a coordinate channel and a distance map of a distance channel; the first distance map is input into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map; a second distance map is obtained by adding the attention distance map and the high-resolution distance map; the second distance map is reversely projected to obtain a high-resolution point cloud. In the above manner, the high-resolution processing method of the present application can improve the resolution of the point cloud data collected by the laser radar, and can achieve better display effects with low-cost resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without inventive efforts. Among them:

[0041] Figure 1 This is a flow chart of an embodiment of a high-resolution processing method provided by the present application;

[0042] Figure 2 This is a schematic diagram of the structure of an embodiment of a neural network architecture provided by this application;

[0043] Figure 3 It is a flow chart of the overall framework of the high-resolution processing method provided in this application;

[0044] Figure 4 is a schematic structural diagram of the first distance map provided by this application;

[0045] Figure 5 yes Figure 1 The specific flow diagram of step S12 of the high-resolution processing method is shown;

[0046] Figure 6 This is a schematic diagram of the principle of edge enhancement provided by this application;

[0047] Figure 7 This is a schematic diagram of the edge enhancement effect provided by this application;

[0048] Figure 8 This is a schematic diagram of the structure of an embodiment of a channel attention module provided by the present application;

[0049] Figure 9 It is a structural schematic diagram of an embodiment of a high-resolution processing device provided by the present application;

[0050] Figure 10 It is a structural diagram of an embodiment of a computer storage medium provided by this application. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0052] LiDAR is indispensable in the field of autonomous driving. Although high-resolution (or multi-beam) LiDAR can generate more points, it is very expensive. Therefore, in order to effectively reduce costs, this application proposes a new network structure and a super-resolution processing method for laser point cloud data based on this network structure, which is used to generate high-resolution point clouds from low-resolution point clouds. In this way, better results can be achieved at a lower cost for the work to be completed using LiDAR.

[0053] Please refer to the following for details: Figures 1 to 3 , Figure 1 This is a flow chart of an embodiment of a high-resolution processing method provided by this application. Figure 2 This is a schematic diagram of an embodiment of a neural network architecture provided by this application. Figure 3 It is a flow chart of the overall framework of the high-resolution processing method provided in this application.

[0054] Specifically, if Figure 1 and Figure 3 As shown, the high-resolution processing method of the embodiment of the present application specifically includes the following steps:

[0055] Step S11: Obtain low-resolution point cloud.

[0056] In the embodiment of the present application, point cloud data is first collected. On the one hand, a low-resolution laser radar can be used to collect point cloud data as raw data for high-resolution processing methods; on the other hand, a high-resolution laser radar can be used to collect point cloud data as training images for the neural network. Specifically, during the training process, the neural network downsamples the vertical direction of the high-resolution point cloud as input for training, and uses the high-resolution data as the true value. Figure 3 As shown, compared with high-precision point clouds, low-precision point clouds lack a large amount of point cloud data, resulting in limited information carried by low-precision point clouds. The purpose of the high-resolution processing method provided in this application is to process low-precision point clouds into high-precision point clouds through neural networks to improve the resolution of point clouds.

[0057] Specifically, the point cloud collected by the processing device using the laser radar is N×4, where N represents the number of points and 4 represents the information of each point, including x, y, z, i (where i is intensity, indicating reflectivity). Generally, laser radars are divided into 16-line, 32-line, 64-line and 128-line. The embodiment of the present application takes the effect of obtaining 64 lines after super-resolution processing of 16 lines as an example. Of course, the super-resolution processing method of the present application can also be used to achieve 16 lines to 32 lines or 32 lines to 128 lines, etc., which are not listed here one by one.

[0058] Step S12: Projecting the low-resolution point cloud into a first distance map according to the angle parameters of the laser radar, wherein the first distance map includes a distance map of a coordinate channel and a distance map of a distance channel.

[0059] In an embodiment of the present application, the processing device projects the collected low-resolution point cloud into a first distance map according to the angle parameters of the laser radar.

[0060] For example, a low-resolution LiDAR uses 64 lines, and the horizontal angular resolution of the collected point cloud is 0.2°, and the vertical angular resolution is approximately 0.47°. The calculation is as follows: the vertical angle is from -15° to +15°, and the horizontal angle is from 0° to 360°. The number of vertical laser beams is 64, and after the LiDAR rotates, the horizontal beams in one frame are 1800. In this case, we can calculate:

[0061]

[0062]

[0063] Among them, V res Indicates the vertical angular resolution, H res Indicates the horizontal angular resolution.

[0064] Based on the above concepts, the horizontal angle and vertical angle corresponding to each data point in the point cloud collected by the lidar in the first distance map can be calculated, thereby obtaining the horizontal angle set H and vertical angle set V of the point cloud.

[0065] Furthermore, the processing device also needs to calculate the horizontal index and vertical index of each data point in the first distance map based on the horizontal angle set H and the vertical angle set V and the three-dimensional coordinates of each data point. The specific calculation process is as follows:

[0066] H=arctan 2 (-y, x)

[0067]

[0068] H index =argmin(|HH|)

[0069] V index =argmin(|VV|)

[0070] Among them, V index Indicates the vertical index, H index Indicates the horizontal index.

[0071] After calculating the horizontal and vertical indices of each data point, the processing device projects the information in the point cloud into the first distance map based on the index of each data point. Furthermore, since the neural network input requires a four-channel projection map, each channel corresponds to information x, y, z, and range, where range represents distance. Since the original information of the point cloud generally includes reflectivity, the processing device needs to calculate distance information based on the three-dimensional information of the point cloud. The calculation formula is as follows:

[0072]

[0073] Through the above processing, the processing device can obtain Figure 4 Schematic diagram of the first distance map shown. Figure 4 The image includes four channels of projection, which can be represented from top to bottom as the projected x-channel image, y-channel image, z-channel image, and range channel image. After the above processing, we can obtain data of size C × V × H, where C represents the number of channels, such as 4, V represents the vertical resolution, such as 64, and H represents the horizontal resolution, such as 1800.

[0074] After the processing device obtains the distance map that meets the input requirements of the neural network, it can output the distance map to the neural network for resolution improvement processing.

[0075] Before this, the processing device can perform loop filling and completion on the distance map before inputting it into the network, which can solve the edge enhancement problem of the distance map projected into the distance channel. Figures 5 to 7 ,in, Figure 5 yes Figure 1 The specific flow diagram of step S12 of the high-resolution processing method is shown in FIG. Figure 6 This is a schematic diagram of the principle of edge enhancement provided by this application. Figure 7 This is a schematic diagram of the edge enhancement effect provided by this application.

[0076] Specifically, if Figure 5 As shown, the high-resolution processing method of the embodiment of the present application specifically includes the following steps:

[0077] Step S121: Acquire the horizontal resolution of the first distance map.

[0078] In the embodiment of the present application, the processing device obtains the horizontal resolution H of the range map of the range channel.

[0079] Step S122: Calculate the left padding amount based on the horizontal resolution.

[0080] In the embodiment of the present application, the processing device calculates the left padding amount based on the horizontal resolution H. The specific calculation process and formula are as follows:

[0081] r=H%16

[0082]

[0083] Among them, r is the intermediate quantity, and lp is the number of left paddings of the distance map of the distance channel.

[0084] Step S123: Calculate the right padding number based on the left padding number and the horizontal resolution.

[0085] In the embodiment of the present application, the processing device calculates the right padding amount based on the left padding amount lp and the middle amount r. The specific calculation formula is as follows:

[0086] rp=16-r-lp

[0087] Where rp is the right padding number of the distance map of the distance channel.

[0088] Step S124: cyclically filling the first distance map based on the left filling number and the right filling number.

[0089] In the embodiment of the present application, the processing device cyclically fills the first distance map based on the left filling number and the right filling number respectively. The specific filling process is as follows:

[0090] See also Figure 6, a region with a left padding number lp is copied to the right side of the distance map of the distance channel. That is, the pixel set is spliced to the left side of the distance map of the distance channel. After splicing, a new distance map is obtained. The processing device recalculates the new left and right padding numbers based on the new distance map and pads the new distance map based on the new left and right padding numbers. The above steps are repeated until the left and right padding numbers calculated based on the latest distance map are zero, completing the cyclic padding and completion process before input into the neural network.

[0091] Step S13: Input the first distance map into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map.

[0092] In the embodiment of the present application, the processing device inputs the first distance map completed by cyclic filling and completion into the following Figure 2 The neural network shown is composed of Figure 2 The neural network improves the feature space resolution of the first distance map, thereby obtaining a high-resolution point cloud.

[0093] Specifically, the neural network of the embodiment of the present application is mainly divided into two parts, the U-net classic network, that is, the resolution network and the reconstruction block based on the channel attention mechanism, that is, the channel attention network. Among them, the reconstruction block based on the channel attention mechanism of the neural network processes the input x, y, z data, that is, the distance map of the coordinate channel, while the U-net classic network of the neural network processes the input range data, that is, the distance map of the distance channel. Finally, the data obtained after processing by the two branch networks of the neural network are added together, and the final result is obtained after processing through the convolution layer.

[0094] The input data first passes through two transposed convolutional layers: the distance map of the coordinate channel passes through the transposed convolutional layer in the reconstruction block of the channel attention mechanism, and the distance map of the distance channel passes through the transposed convolutional layer in the classic U-net network to process the distance map to the required size. Each convolutional block in the neural network consists of two convolutional layers, each with a convolution kernel of 3, followed by input to the batch normalization layer and the ReLU function (activation function).

[0095] Specifically, the encoder network of the classic U-net network consists of four convolutional blocks and an average pooling layer. Correspondingly, the decoder network has the opposite structure of the encoder network. The classic U-net network uses transposed convolutional layers to increase the feature space resolution. In each convolutional block, all convolution results are subjected to batch normalization layers and ReLU functions.

[0096] The channel attention reconstruction block is composed of a convolutional layer and a channel attention module. For the specific structure of the channel attention module, please refer to Figure 8 .

[0097] For example, the input image is data of size C×V×H, where C is 3, including x channel, y channel, and z channel. The working steps of the channel attention module are as follows: the channel attention module reshapes the input image into a C×N first distance channel map and an N×C second distance channel map, where N is V×H. The channel attention module multiplies the first distance channel map and the second distance channel map, and then performs softmax (normalization) processing to obtain the channel attention map (C×C). The channel attention module then performs matrix multiplication on the channel attention map and the reshaped first distance channel map, multiplies it by the scale coefficient, and then reshapes it to its original shape to obtain the reshaped image; finally, the reshaped image is added to the original input image to obtain the final output image.

[0098] It should be noted that the scale coefficient in the channel attention module is initially 0 and then increases through continuous iterative learning.

[0099] During the training process, the processing device collects low-resolution point cloud data using a low-resolution lidar to obtain a first training point cloud, and collects high-resolution point cloud data using a high-resolution lidar to obtain a second training point cloud. The processing device inputs the first and second training point clouds into the neural network to train the neural network so that the result obtained by the neural network after performing high-resolution processing on the first training point cloud approximates the second training point cloud.

[0100] Step S14: Obtain a second distance map based on the addition of the attention distance map and the high-resolution distance map.

[0101] In an embodiment of the present application, the neural network adds the high-resolution distance map output by the U-net classic network and the attention distance map output by the channel attention reconstruction block to obtain a second distance map as output.

[0102] Step S15: Perform inverse projection on the second distance map to obtain a high-resolution point cloud.

[0103] In an embodiment of the present application, the processing device reversely projects the second distance map and reversely projects a high-resolution point cloud based on the second distance map, thereby realizing the entire process of obtaining a high-resolution point cloud based on low-resolution point cloud processing.

[0104] In an embodiment of the present application, a high-resolution processing device obtains a low-resolution point cloud; projects the low-resolution point cloud into a first distance map according to the angle parameters of the laser radar, including a distance map of the coordinate channel and a distance map of the distance channel; inputs the first distance map into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map; a second distance map is obtained by adding the attention distance map and the high-resolution distance map; the second distance map is reverse-projected to obtain a high-resolution point cloud. In the above manner, the high-resolution processing method of the present application can improve the resolution of the point cloud data collected by the laser radar and can achieve better display effects with low-cost resources.

[0105] In order to implement the high-resolution processing method of the above embodiment, this application also proposes a high-resolution processing device, please refer to Figure 9 , Figure 9 It is a structural schematic diagram of an embodiment of a high-resolution processing device provided by this application.

[0106] The high-resolution processing device 500 of the embodiment of the present application includes a memory 51 and a processor 52 , wherein the memory 51 and the processor 52 are coupled.

[0107] The memory 51 is used to store program data, and the processor 52 is used to execute the program data to implement the high-resolution processing method described in the above embodiment.

[0108] In this embodiment, the processor 52 may also be referred to as a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip having signal processing capabilities. The processor 52 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor, or the processor 52 may be any conventional processor.

[0109] This application also provides a computer storage medium, such as Figure 10 As shown, the computer storage medium 600 is used to store program data 61. When the program data 61 is executed by the processor, it is used to implement the high-resolution processing method described in the above embodiment.

[0110] The present application also provides a computer program product, wherein the computer program product includes a computer program, and the computer program is operable to enable a computer to execute the high-resolution processing method described in the embodiment of the present application. The computer program product can be a software installation package.

[0111] The high-resolution processing method described in the above embodiments of the present application, when implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a device, such as a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, 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.

[0112] The above description is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A high-resolution processing method based on laser radar, characterized in that: The high-resolution processing method includes: Get low-resolution point cloud; The low-resolution point cloud is projected into a first distance map according to the angle parameter of the laser radar, wherein the first distance map includes a distance map of a coordinate channel and a distance map of a range channel, the distance map of the coordinate channel includes an x channel map, a y channel map and a z channel map, x, y and z represent three-dimensional information of the point cloud, the distance map of the range channel is a range channel map, range represents distance, and Inputting the first distance map into a pre-trained neural network, wherein the neural network includes a resolution network and a channel attention network, wherein the resolution network processes the distance map of the distance channel to obtain a high-resolution distance map, and the channel attention network processes the distance map of the coordinate channel to obtain an attention distance map; Obtaining a second distance map based on the addition of the attention distance map and the high-resolution distance map; Performing back-projection on the second distance map to obtain a high-resolution point cloud; Before inputting the first distance map into a pre-trained neural network, the high-resolution processing method further includes: Acquire a horizontal resolution of the first distance map; Calculating a left padding amount based on the horizontal resolution; Calculating a right padding number based on the left padding number and the horizontal resolution; The first distance map is cyclically filled based on the left filling number and the right filling number.

2. The high-resolution processing method according to claim 1, characterized in that: The projecting the low-resolution point cloud into a first distance map according to the angle parameter of the laser radar includes: Calculating a horizontal angle set of the low-resolution point cloud in the first distance map based on a horizontal angle parameter of the laser radar; Based on the three-dimensional coordinates of each point cloud data in the low-resolution point cloud and the horizontal angle set, obtaining an index of each point cloud data in the horizontal direction; Calculating a vertical angle set of the low-resolution point cloud in the first distance map based on a vertical angle parameter of the laser radar; Based on the three-dimensional coordinates of each point cloud data in the low-resolution point cloud and the vertical angle set, obtaining an index of each point cloud data in the vertical direction; The relevant point cloud information is projected according to the horizontal index and the vertical index of each point cloud data to form the first distance map.

3. The high-resolution processing method according to claim 1, characterized in that: The obtaining of the low-resolution point cloud comprises: Obtaining the coordinates of each point cloud data in the low-resolution point cloud; The distance of each point cloud data is calculated based on the coordinates of each point cloud data.

4. The high-resolution processing method according to claim 1, characterized in that: The cyclically filling the first distance map based on the left filling number and the right filling number includes: Filling the left side of the first distance map with a pixel set whose number of columns on the right side of the first distance map is equal to the number of left padding columns; Filling the right side of the first distance map with a pixel set whose number of columns on the left side of the first distance map is equal to the number of right-side padding; Repeat the following steps until the new left padding number is 0: calculate a new left padding number based on the filled first distance map, calculate a right padding number based on the left padding number and the horizontal resolution; and cyclically fill the first distance map based on the left padding number and the right padding number.

5. The high-resolution processing method according to claim 1, characterized in that: The resolution network includes an encoder network and a decoder network, and the encoder network has an opposite structure to the decoder network; the encoder network and / or the decoder network includes several convolutional layers, batch normalization layers and activation functions, wherein the input of the encoder network also includes a transposed convolutional layer.

6. The high-resolution processing method according to claim 5, characterized in that: The channel attention network includes a convolutional layer and a channel attention module; the channel attention module is used to reshape the distance map of the distance channel into a first distance channel map and a second distance channel map, multiply the first distance channel map and the second distance channel map to obtain a channel attention map, perform matrix multiplication on the channel attention map and the first distance channel map, reshape it back to a reshaped image of the size of the distance map of the distance channel, and finally add it to the distance map of the distance channel to obtain the output of the channel attention module; The first distance channel image and the second distance channel image are transposed images.

7. The high-resolution processing method according to claim 6, characterized in that: The high-resolution processing method further includes: Using a low-resolution laser radar to collect low-resolution point cloud data to obtain a first training point cloud; Using a high-resolution laser radar to collect high-resolution point cloud data to obtain a second training point cloud; The first training point cloud and the second training point cloud are input into the neural network to train the neural network.

8. A high-resolution processing device, characterized in that The high-resolution processing device includes a memory and a processor, wherein the memory is coupled to the processor; The memory is used to store program data, and the processor is used to execute the program data to implement the high-resolution processing method according to any one of claims 1 to 7.

9. A computer storage medium, characterized in that The computer storage medium is used to store program data, and when the program data is executed by the processor, it is used to implement the high-resolution processing method according to any one of claims 1 to 7.

Images