A reflection difference plane mark design and identification method of a semi-solid state laser radar

Abstract

The application discloses a kind of semi-solid laser radar's reflection difference plane mark design and identification method, wherein identification method involves semi-solid laser radar and reflection difference plane mark, reflection difference plane mark includes matte background and reflection information;Identification method obtains dense point cloud in scanning field by solid-state laser radar, capture the identification information with strong reflectivity, then filter and voxelize to point cloud, finally through the plane mark neural network detection algorithm special for semi-solid laser radar extraction, identification obtains the information of plane mark;The application can solve the problem that single laser radar cannot identify plane mark information, compared with camera plane mark recognition, accurate three-dimensional information can be obtained, and it has the robustness of detection under dark and weak light.The application can provide plane semantic detection support for semi-solid laser radar SLAM, AGV navigation, underground autonomous exploration label identification and other tasks that must use laser radar or camera failure.

Description

Technical Field

[0001] This invention belongs to the field of lidar detection, specifically relating to a method for designing and recognizing reflectance difference plane markers in semi-solid-state lidar. Background Technology

[0002] With the continuous development of autonomous driving and navigation technologies, semi-solid-state LiDAR, as a highly efficient 3D surface scanning device, is widely used in autonomous driving and testing equipment. Unlike traditional mechanical LiDAR, which can only collect sparse point clouds, semi-solid-state LiDAR, due to its structural advantages, can collect high-resolution dense point clouds. Therefore, it can obtain richer point cloud information within the scanning field for 3D detection. Currently, LiDAR detection typically uses the surface contour of an object for pattern recognition to achieve 3D detection. However, planar signs, such as text, symbols, and QR codes, cannot be detected due to the lack of contour variations. For devices that rely solely on LiDAR or require precise 3D detection, this problem limits the detection range. This invention can be applied in scenarios such as indoor positioning, mapping, and AGV navigation, as well as in situations where passive planar sign detection using LiDAR is necessary, such as in low-light conditions. Summary of the Invention

[0003] To address the aforementioned issues, this invention discloses a method for designing and recognizing planar markers with reflection differences using a semi-solid-state LiDAR. By utilizing a semi-solid-state LiDAR to scan, extract, and recognize the features of the planar markers designed in this invention, the detection information of the planar markers is ultimately obtained. This invention, to a certain extent, solves the problem that semi-solid-state LiDAR cannot recognize planar markers, and compared to camera recognition of planar markers, it can obtain accurate three-dimensional information. This invention can provide planar semantic detection support for tasks that must be completed using LiDAR, such as solid-state LiDAR SLAM, AGV navigation, and underground autonomous exploration tag recognition, or in situations where cameras are disabled.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A method for designing and recognizing reflection difference plane markers in a semi-solid-state lidar system, comprising:

[0006] Step 1: Start the system and acquire dense point cloud data of the reflection difference plane markers and the environment using a semi-solid-state lidar.

[0007] Step 2: The dense point cloud is filtered, sampled, and rasterized using a point cloud processing algorithm to obtain a rasterized, regular point cloud.

[0008] Step 3: The regular point cloud is extracted and identified using a neural network detection algorithm specifically designed for semi-solid-state lidar to obtain information about the planar markers.

[0009] Furthermore, the semi-solid-state lidar is an active surface reflection imaging device commonly used in autonomous driving and detection. It can obtain high-resolution lidar point clouds and the reflection intensity of the echo at each point. For ease of description, the range of a single frame imaging of the semi-solid-state lidar is referred to as the radar scanning field.

[0010] Furthermore, the reflective difference plane marker consists of a matte background and a reflective marker. During imaging with a semi-solid-state lidar, the matte portion generates stray reflections, while the reflective portion generates directional reflections. Since the semi-solid-state lidar can obtain the reflection intensity of the echo, the reflective portion acquires a strong reflective point cloud compared to the background and the radar scanning field environment, ultimately obtaining the marker's imaging information. Because the intensity of reflection between the matte and reflective portions is relative, and vice versa, the background of the reflective difference plane marker can also be considered as the reflective portion, and the marker as the matte portion.

[0011] Furthermore, the semi-solid-state LiDAR-specific planar marker neural network detection algorithm described in step three is a dedicated algorithm improved upon the current LiDAR neural network detection algorithm to address the insufficient accuracy of the task in this invention. This algorithm, based on the current anchorless LiDAR neural network detection algorithm CenterPoint, incorporates a voxel encoding algorithm for reflection intensity perception and a height feature refinement backbone algorithm, making it specifically designed for the detection task in this invention. The main process of the semi-solid-state LiDAR-specific planar marker neural network detection algorithm is as follows:

[0012] Step 1: Filter the collected point cloud into voxel features;

[0013] Step 2: The voxel features are encoded a second time using a voxel encoding algorithm that senses reflection intensity to obtain the voxel features of the reflection intensity gain.

[0014] Step 3: Extract features from voxel features and downsample them using the sparse feature extraction backbone;

[0015] Step 4: Further feature extraction is performed on the output of Step 3 using the highly refined backbone to obtain highly refined features;

[0016] Step 5: Use a point cloud no-anchor-frame detector to extract the final information from the highly refined features and obtain the final detection result.

[0017] The present invention has the following beneficial effects:

[0018] (1) Planar signs designed by reflection difference can obtain clear sign images with semi-solid-state lidar. Combined with the reflection difference sign recognition algorithm, the problem that a single lidar sensor cannot recognize planar signs is solved.

[0019] (2) Compared with the imaging and recognition of passive planar signs by cameras, the imaging and recognition of planar signs by semi-solid-state lidar can obtain accurate position and size information of the signs, and the active scanning can have detection robustness under low light conditions. Attached Figure Description

[0020] Figure 1 This is a structural diagram of the apparatus for implementing the present invention;

[0021] Figure 2 These are three design and imaging examples of the reflective differential plane marker of the present invention;

[0022] Figure 3 This is a schematic diagram of a reflection difference plane marker imaging scene of the implementation device of the present invention;

[0023] Figure 4 This is an example of a reflective differential plane marker imaging device of the present invention;

[0024] Figure 5 This is the basic flow of the reflectance difference planar mark detection method of the present invention;

[0025] Figure 6 This is the basic flow of the semi-solid-state lidar-specific planar marker neural network detection algorithm of the present invention;

[0026] Figure 7 This is the basic structure of the voxel encoding algorithm for reflection intensity sensing in this invention;

[0027] Figure 8 This is the basic structure of the high-feature refinement backbone algorithm of the present invention.

[0028] List of identifiers in attached diagrams:

[0029] 1. Semi-solid-state lidar; 2. Power supply; 3. Computing unit; 11. Dense point cloud imaging; 12. Reflection difference plane marker. Detailed Implementation

[0030] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0031] like Figure 1As shown, the main components of the implementation device described in this embodiment are a semi-solid-state lidar 1, used to acquire dense point cloud images within the radar scanning field; a power supply 2, used to power the device; and a computing unit 3, which contains a computing unit and a controller, used for communication, imaging, and algorithm processing. The power supply 2 is electrically connected to the semi-solid-state lidar 1 and the computing unit 3.

[0032] like Figure 2 As shown, this embodiment illustrates the design and imaging examples of three reflectivity difference plane signs: a "left turn symbol," the Arabic numeral "20," and a 4x4 ArUco QR code "0." The identification information of the reflectivity difference plane signs ("left turn symbol," "20," and the QR code itself) is made of reflective material, while the background information is made of matte material. During semi-solid-state lidar imaging, the matte portion produces stray reflections, while the reflective portion produces directional reflections. Because the semi-solid-state lidar can obtain the reflection intensity of the echo, the reflective portion acquires a strong reflective point cloud compared to the background and the radar scanning field environment, ultimately obtaining the imaging information of the sign. The imaging results obtained by the semi-solid-state lidar are as follows: Figure 2 The comparison below shows that, since the intensity of reflection of the matte part and the reflective part is relative, and vice versa, the background of the reflective plane mark can be used as the reflective part, and the mark as the matte part.

[0033] like Figure 3 As shown in the figure, this embodiment illustrates an imaging scenario for a reflective difference plane marker. In this embodiment, the reflective difference plane marker 12 is fixed to a building structure, and the reflective difference plane marker 12 is imaged by a semi-solid-state lidar 1, ultimately resulting in a dense point cloud image 11.

[0034] like Figure 4 As shown, this embodiment creates two reflective difference plane markings: a "left turn symbol" and a "speed limit 70". The "left turn symbol" uses a matte material as the background and a reflective material as the marking body; the "speed limit 70" uses a reflective material as the background and a matte material as the marking, resulting in the final semi-solid-state lidar imaging example.

[0035] like Figure 5 As shown, the embodiment utilizes a planar marker design and recognition method based on semi-solid-state LiDAR. The basic process is as follows:

[0036] Step 1: Start the system and acquire dense point cloud data of the reflection difference plane markers and the environment using a semi-solid-state lidar.

[0037] Step 2: The dense point cloud is filtered, sampled, and rasterized using a point cloud processing algorithm to obtain a rasterized, regular point cloud.

[0038] Step 3: The regular point cloud is extracted and identified using a neural network detection algorithm specifically designed for semi-solid-state lidar to obtain information about the planar markers.

[0039] like Figure 6 As shown, the semi-solid-state LiDAR-specific planar marker neural network detection algorithm is a dedicated algorithm improved from current LiDAR neural network detection algorithms to address the insufficient accuracy of the task in this invention. Based on the current anchorless LiDAR neural network detection algorithm CenterPoint, this algorithm incorporates a voxel encoding algorithm for reflection intensity perception and a height feature-refined backbone, making it specifically designed for the detection task in this invention. The main process of the semi-solid-state LiDAR-specific planar marker neural network detection algorithm is as follows:

[0040] Step 1: Filter the collected point cloud to obtain voxel features;

[0041] Step 2: The voxel features are encoded a second time using a voxel encoding algorithm that senses reflection intensity to obtain the voxel features of the reflection intensity gain.

[0042] Step 3: Extract features from voxel features and downsample them using the sparse feature extraction backbone;

[0043] Step 4: Further feature extraction is performed on the output of Step 3 using the highly refined backbone to obtain highly refined features;

[0044] Step 5: Use a point cloud no-anchor-frame detector to extract the final information from the highly refined features and obtain the final detection result.

[0045] To highlight the main content of this invention and for ease of description, steps 1, 3, and 5 of the semi-solid-state lidar-specific planar marker neural network detection algorithm, as well as the neural network training process, are omitted, as those skilled in the art can easily find them in relevant materials. The following only provides a detailed explanation of the voxel encoding algorithm for reflection intensity perception and the height feature refinement backbone designed to achieve point cloud planar target detection in this invention.

[0046] like Figure 7 As shown, the voxel encoding algorithm for reflection intensity perception mainly functions to discover the feature relationships among all points within a non-empty voxel. The algorithm primarily extracts reflection intensity enhancement features using a reflection feature extractor and distinguishes foreground-background features using a foreground feature filtering method.

[0047] like Figure 7 As shown, the reflection feature extractor is designed as follows. Assume any non-empty voxel is... Where, p ip represents the point cloud points in a non-empty voxel. i ={x i ,y i ,z i ,r i}, x i ,y i ,z i ,r i These represent the three-dimensional coordinates and reflection intensity of the point cloud points, respectively; N v C represents the maximum number of point cloud points in a voxel. p The number of data channels for the point cloud is 4 in this case. First, one-dimensional convolution is used on the input voxels to discover local feature relationships. The goal here is to use the one-dimensional convolution kernel to perform operations on local points within the voxels, with the point features used as the sequence length (number of input channels) for concatenation. Then, the voxel features with discovered local relationships are used with a linear layer and a sigmoid function to discover global relationships, finally yielding the refined reflection intensity features.

[0048] like Figure 7 As shown, the foreground feature filter is designed as follows. After obtaining F, the next step is to filter out the foreground and background features from the voxels. Since F has already undergone preliminary feature extraction, the filtering of points in the foreground feature filter can now be achieved by the reflection feature extractor learning their distribution in the original voxels. Here, 0.5 is used as the threshold for foreground and background separation for the points in F. Subsequently, foreground feature points and background are selected from the original voxels using indices. The final voxel encoding result is shown below:

[0049]

[0050] Among them, f fgc To represent foreground features, we take three-dimensional coordinates and reflection intensity, respectively. v(·) represents a channel feature taken at a point in voxel v. Represents logical operations. n fg For the number of foreground points, it is obvious that... f fgc This can be understood as obtaining the center position of the foreground point and the average reflection intensity of the foreground representation. Similarly, the background point features f can be obtained. bgc Since the voxel encoding algorithm for reflectance intensity perception aims to highlight the foreground rather than the background, only the average reflectance intensity of background points is calculated. In the formula... Indicates and Conversely, n bg The number of background points can be represented by n. bg =N v -n fg get.

[0051] like Figure 8 As shown, the height feature refinement backbone, positioned between the sparse convolutional backbone and the subsequent dense convolutional backbone, is used to enhance the learning ability of the height-direction representation of the solid-state LiDAR while maintaining the original point cloud target detection capability. The height feature refinement backbone includes a bird's-eye view feature branch and a height refinement branch. The output features of the bird's-eye view feature branch and the height refinement branch are multiplied together, i.e., integrated using spatial attention.

[0052] like Figure 8 As shown, the bird's-eye view feature branch uses the more abstract features in the height direction of the sparse convolutional network (SPConvNet) as input, and extracts the BEV features using a dilated convolution module. Assume the input features are... Projected onto the input sparse features through the index matrix Where C1 is the number of channels of the input feature, N1 is the number of voxels of the input feature, and D1, H, and W represent the three-dimensional dimensions of the grid containing the voxels, namely the height, depth, and width, respectively. First, a height compression module is needed to convert the features into BEV features. Then, the dilated convolution module is used to extract features to obtain the desired features.

[0053] like Figure 8 As shown, the height feature refinement branch uses features with richer contextual features in the height direction from the sparse convolutional network as input, and refines the features in the height direction using the height feature refinement module. The height feature refinement module utilizes the features of the sparse convolutional network... As input, C2 represents the number of channels in the sparse convolutional network feature, and D2 represents the height dimension of the sparse convolutional network feature. As described above, since the input feature S2 is a sparse voxel feature, the height feature module extracts sparse features in the height direction by specifying the shape of the sparse convolution kernel. Subsequently, the sparse features are converted into dense three-dimensional features. Then, it is connected using linear layers to generate the final highly refined features.

[0054] To highlight the main content of this invention and for ease of description, this embodiment omits, simplifies, and abstracts some details such as modeling, electrical connection, imaging results, and algorithm design and neural network model training, which are common-sense axioms. Those skilled in the art can easily supplement and implement these technical details using relevant knowledge.

Claims

1. A method for designing and recognizing reflectance difference planar markers in a semi-solid-state lidar system, characterized in that: Includes the following steps: Step 1: Start the system and acquire dense point cloud data of the reflection difference plane markers and the environment using a semi-solid-state lidar. Step 2: The dense point cloud is filtered, sampled, and rasterized using a point cloud processing algorithm to obtain a rasterized, regular point cloud. Step 3: Extract and identify the regularized point cloud using a dedicated planar marker neural network detection algorithm for semi-solid-state LiDAR to obtain the information of the planar marker; The aforementioned neural network detection algorithm for planar markers in semi-solid-state lidar includes: Step 1: Filter the collected point cloud to obtain voxel features; Step 2: The voxel features are encoded a second time using a voxel encoding algorithm that senses reflection intensity to obtain the voxel features of the reflection intensity gain. The aforementioned voxel encoding algorithm for perceiving reflection intensity includes a reflection feature extractor and a foreground feature filter, which are cascaded together to obtain output features. Step 21, the reflection feature extractor is designed as follows: assuming any non-empty voxel is... , ;in, Points in a non-empty voxel cloud. , These represent the three-dimensional coordinates and reflection intensity of the point cloud points, respectively. This represents the maximum number of point cloud points in the voxel. To determine the number of data channels for point cloud points, we first use one-dimensional convolution to discover local feature relationships among the input voxels. Then, we use a linear layer and a sigmoid function to discover global relationships among the voxel features with discovered local relationships. Finally, we obtain the refined reflection intensity features. ; Step 22, the foreground feature filter is designed as follows: after obtaining Next, we need to filter out the foreground and background features from the voxels; here, the corresponding pair The points in the dataset are assigned a threshold of 0.5 for foreground-background separation. Then, foreground and background feature points are selected from the original voxels using an index. The final voxel encoding result is shown below: （2） in, To represent foreground features, three-dimensional coordinates and reflection intensity are taken respectively. Indicates the collection of body elements A certain channel feature at the midpoint, Represents logical operations. , For the number of foreground points, it is obvious that... Similarly, there are background point features. ; in the formula Indicates and The opposite logic, The number of background dots is represented by get; Step 3: Extract features from voxel features and downsample them using the sparse feature extraction backbone; Step 4: Further feature extraction is performed on the output of Step 3 using the highly refined backbone to obtain highly refined features; Step 5: Use a point cloud no-anchor-frame detector to extract the final information from the highly refined features and obtain the final detection result.

2. The method for designing and recognizing reflection difference plane markers for semi-solid-state lidar according to claim 1, characterized in that: The components of the apparatus for implementing the method are: Semi-solid-state lidar is used to acquire dense point cloud images within a radar scanning field; Power supply, used to provide power to the equipment; The computing unit contains computing units and a controller, and is used for communication, imaging and algorithm processing. The power supply is electrically connected to the semi-solid-state lidar and the computing unit.

3. The method for designing and recognizing reflection difference plane markers for semi-solid-state lidar according to claim 1, characterized in that: The background of the reflective plane sign is made of matte material, while the sign itself is made of reflective material, and vice versa.

4. The method for designing and recognizing reflection difference plane markers for semi-solid-state lidar according to claim 1, characterized in that: By using semi-solid-state lidar, the intensity of echo reflection at each point in a dense point cloud can be obtained. Due to the differences in reflectivity between reflective and matte materials of planar markers and the environment, an image of the marker itself can be obtained in the point cloud.

5. The method for designing and recognizing reflection difference plane markers for semi-solid-state lidar according to claim 1, characterized in that: The high-level feature refinement backbone described in step 4 adopts a parallel structure of bird's-eye view feature branch and high-level refinement branch, and the output feature is obtained by multiplying the outputs of the two. 5.1 The aforementioned bird's-eye view feature branch utilizes more abstract features in the height direction from a sparse convolutional network as input, and extracts BEV features using a dilated convolution module; assuming the input features are... Projected through the index matrix to ;in, The number of channels for the input feature. The number of voxels for the input feature. , , These represent the three-dimensional dimensions of the mesh containing the voxel: height, depth, and width. First, the features need to be converted into BEV features using a height compression module. Then, the dilated convolution module is used to extract features to obtain the feature set. ; 5.2 The high-level feature refinement branch adopts a structure of cascading sparse convolutional networks and high-level feature refinement modules; the high-level feature refinement branch utilizes the features of sparse convolutional networks. As input, where The number of channels in the sparse convolutional network feature. The height dimension of the sparse convolutional network features; due to the input features As sparse voxel features, the height feature refinement module extracts sparse features in the height direction by specifying the kernel shape of the sparse convolution; then, the sparse features are converted into dense three-dimensional features. And by connecting them with linear layers, highly refined features are obtained. .

Images