A measurement method for Lidar odometer

Abstract

The present invention relates to a measurement method for a Lidar odometer, and belongs to the technical field of remote sensing data processing. Based on the HRegNet network, the present invention adds initial registration error detection and cumulative error detection. The initial registration error detection is to determine whether the registration of two consecutive frames of point clouds (paired registration) is successful based on the carrier speed and the Lidar scanning frequency in each frame registration, and apply a pre-set transformation matrix to replace the registration that fails the detection, with the purpose of reducing the cumulative impact of a certain frame containing a large error on subsequent frames; the cumulative error detection is to strictly detect the gross error in the intermediate registration process and the accumulation of control errors based on the conversion between the end point cloud frame and the first frame point cloud during the continuous registration process. The present invention was tested on the Kitti odometry dataset, and the test results show that compared with the comparison algorithm, the present invention can realize Lidar odometer measurement more accurately, quickly and robustly.

Description

Technical Field

[0001] The present invention relates to a measurement method of a Lidar odometer, and belongs to the technical field of remote sensing data processing. Background Art

[0002] As a non-contact sensing method, remote sensing technology can acquire the spectral characteristics of observed objects from satellite and airborne platforms. Subsequent data processing methods primarily utilize the spectral emission and reflectance properties of the objects to interpret the observed scene. With the explosive development of remote sensing data acquisition technology, multimodal remote sensing data often exists within the same observation scene, providing diverse and complementary information for understanding and perceiving the geospatial environment. Therefore, the joint processing of multimodal remote sensing data is a promising paradigm in the field of remote sensing image interpretation.

[0003] In unmanned driving systems, vehicle odometers play an important role. Their main task is to determine the position and direction of the vehicle. They do not require additional signal assistance (such as GNSS signals) and can only rely on their own local sensors to perceive the vehicle's location. They can provide more accurate and reliable vehicle posture information than GNSS measurements when the GNSS signal is weak or the noise is large. Vehicle odometers can be divided into two categories: visual odometers and Lidar odometers. Visual odometers are limited by lighting conditions during application, while Lidar odometers can avoid the influence of lighting changes by actively emitting laser beams, making Lidar odometers more suitable for vehicle odometer measurement tasks than visual odometers. Lidar odometer measurement algorithms can be further divided into algorithms that rely solely on Lidar and algorithms that use third-party (such as INS, GNSS) auxiliary data. For odometer measurement methods that rely solely on Lidar, these algorithms can be further divided into the following three categories based on the type of correspondence used in the point cloud registration step: (1) point correspondence-based algorithms, (2) distribution correspondence-based algorithms, and (3) network correspondence-based algorithms.

[0004] The algorithm based on point correspondence extracts the key point set of the Lidar frame point cloud, establishes conjugate key point pairs, and then calculates the transformation relationship between consecutive Lidar frame point clouds. This type of algorithm has become the most commonly used method due to its simple principle, direct calculation process, and good compatibility with existing key point matching algorithms. When the data quality is high and the overlap is large, this type of algorithm can achieve high-precision registration between Lidar frame point clouds, but it is sensitive to noise, easily falls into local optimality, and the calculation process is relatively time-consuming. The core of the algorithm based on distribution correspondence is to represent the original point cloud data with a cell-based NDT. The size of the grid directly affects the final calculation results, but how to set the optimal grid size based on the input data is a very challenging task.

[0005] Compared to the first two algorithms, network correspondence-based Lidar odometry measurement algorithms are relatively new. They embed a neural network with millions of parameters in two Lidar point clouds and train it using a large number of data samples. This allows the constructed neural network to learn the point cloud patterns in specific scenes, thereby accurately estimating the transformation relationship between the Lidar frame point clouds in the test dataset. This type of algorithm has been widely applied to paired point cloud registration, with representative algorithms such as 3DRegNet, D3Feat, and HRegNet. However, in Lidar odometry measurement tasks, in addition to considering high-precision registration between paired point clouds, registration failures and error accumulation must also be handled.

[0006] The main purpose of the HRegNet network is to S and target point cloud P T , using the constructed registration network, a coarse-to-fine approach is used to predict an optimized rotation matrix and an optimized translation vector In order to well integrate the source point cloud P S Convert to target point cloud P T The registration process is as follows: Figure 1 As shown in the figure, for a given set of point clouds, we first use three consecutive feature extraction modules to perform layered downsampling on the original point cloud to obtain multiple key point sets with small data volumes. Feature Description Set and significant uncertainty Where l = {1, 2, 3} represents the layer number, M l Represents the number of key points, C l Represents the number of channels (dimensions) of the feature descriptor; then in the coarse registration stage (the third layer), a learning-based correspondence network is used to identify globally consistent conjugate point pairs in the feature space, and then a coarse transformation relationship R3 and t3 is estimated; finally, in the fine registration stage (the first and second layers), based on the coarse registration, more accurate corresponding point pairs are further determined in the three-dimensional coordinate space to obtain the refined transformation matrix and

[0007] The HRegNet registration network is designed for paired point cloud registration, aiming to achieve high-precision fusion between two sets of point clouds. While the registration success rate is as high as 99%, registration failures still occur, and no processing is performed on these failed point clouds. However, in Lidar odometry tasks, which require the registration of long sequences of continuous frame point clouds, not only registration accuracy but also registration robustness and control of registration error accumulation are crucial. In other words, the HRegNet paired point cloud registration network pursues registration accuracy and success rate, while Lidar odometry measurement is more concerned with registration robustness, remedial measures for registration failures, and handling of registration error accumulation. Therefore, directly applying the HRegNet network to Lidar odometry measurement will not meet its requirements. Summary of the Invention

[0008] The purpose of the present invention is to provide a Lidar odometer measurement method to solve the problem of poor measurement accuracy caused by the lack of registration cumulative error in the current Lidar odometer measurement.

[0009] In order to solve the above technical problems, the present invention provides a Lidar odometer measurement method, which includes the following steps:

[0010] 1) Obtain the point cloud data measured by the Lidar odometer, the scanning rate of the Lidar odometer, and the speed of the carrier on which the Lidar odometer is located;

[0011] 2) Use the HregNet network to register two adjacent frames of point clouds, and determine whether the initial registration error requirements are met based on the scanning rate, carrier speed, and the registration results of the two adjacent frames of point clouds;

[0012] 3) Using the HregNet network, the coordinate transformation matrix from the first frame to the last frame in a point cloud sequence is calculated according to the registration method of adjacent frames, which is recorded as the first coordinate transformation matrix. The coordinate transformation matrix from the first frame to the last frame in a point cloud sequence is also directly calculated, which is recorded as the second coordinate transformation matrix. The overall error detection of the point cloud sequence is performed based on the obtained first and second coordinate transformation matrices.

[0013] 4) When the overall error requirement is not met, the HregNet network is used to calculate the coordinate transformation matrix from the first frame to the last frame in a point cloud sequence according to the interval frame registration method, which is recorded as the third coordinate transformation matrix. Based on the obtained first coordinate transformation matrix, second coordinate transformation matrix and third coordinate transformation matrix, the local recalculation error detection is performed on the point cloud sequence.

[0014] Based on the HRegNet network, the present invention adds initial registration error detection and cumulative error detection. The initial registration error detection is to determine whether the registration of two consecutive frames of point clouds (paired registration) is successful based on the carrier speed and Lidar scanning frequency in each frame registration, and apply a pre-set transformation matrix to replace the registration that fails the detection. The purpose is to reduce the cumulative impact of a certain frame containing large errors on subsequent frames; the cumulative error detection is to strictly detect the gross errors in the intermediate registration process and control the accumulation of errors based on the conversion between the end point cloud frame and the first frame point cloud during the continuous registration process. Through the method of the present invention, the image of the cumulative error is taken into account in the measurement process, which greatly improves the measurement accuracy.

[0015] Furthermore, the initial registration error requirement in step 2) is that the translation distance between two adjacent frames of point clouds is no greater than a set threshold. The translation distance is decomposed from the conversion matrix between the two adjacent frames of point clouds. The set threshold is determined by the scanning rate and the carrier speed. If the initial registration error requirement is not met, the conversion matrix between the two adjacent frames of point clouds is set to:

[0016]

[0017] in is the conversion matrix between two adjacent frames of point clouds, and Δd is the set threshold.

[0018] Furthermore, the calculation formula for setting the threshold is:

[0019]

[0020] Where Δd is the set threshold; v is the vehicle speed in m / s; h is the Lidar scanning frequency in Hz.

[0021] The present invention determines a set threshold value according to the scanning rate and the carrier speed, and realizes initial registration error detection according to the relationship between the translation distance between two adjacent frames of point clouds and the set threshold value, which can accurately realize initial registration error detection.

[0022] Furthermore, the overall error requirement is that both the relative rotation and the relative translation are less than corresponding thresholds, where:

[0023]

[0024]

[0025] n is the number of point cloud frames contained in each segment in the overall error detection; trace(ΔR) represents the trace of the matrix ΔR; Indicates direct application of F i+n Frame point cloud and F iThe frame point cloud calculates the transformation matrix between the two frame point clouds, which is the second transformation matrix; Indicates that the Fth frame is calculated by applying the HRegNet registration network according to the registration method of adjacent frames. i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is the first transformation matrix.

[0026] The present invention uses the first transformation matrix and the second transformation matrix to calculate the relative rotation and relative translation. When the relative rotation and relative translation are both relatively small, it is considered that the overall error requirement is met. In this way, whether the overall error requirement is met can be quickly and accurately screened.

[0027] Furthermore, when the overall error meets the requirements, the F i+n Frame point cloud to F i The transformation matrix of the frame point cloud coordinate system is set as and The average value of .

[0028] Furthermore, the calculation formula of the third coordinate transformation matrix during the local recalculation error detection is:

[0029]

[0030] n is the number of point cloud frames contained in each segment in the overall error detection, and n is an even number; Indicates F i+n-2 Frame point cloud to F i+n-4 The transformation matrix of the coordinate system of the frame point cloud.

[0031] The present invention adopts a registration method with one frame interval to calculate the conversion matrix between paired point cloud frames, and based on this, calculates the conversion matrix between the first frame and the last frame.

[0032] Furthermore, the local recalculation error requirement used in local recalculation error detection is and The relative rotation and translation between and The relative rotation and translation between them meet the corresponding thresholds.

[0033] Furthermore, if If both meet the corresponding threshold requirements, then the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of

[0034] like If both meet the corresponding threshold, then the F i+nFrame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of

[0035] They are and The relative rotation and translation between They are and The relative rotation and translation between them.

[0036] Furthermore, if and If the thresholds do not meet the corresponding requirements, the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is set as

[0037] The calculation process of this transformation matrix is ​​the most direct, involves the least number of related point cloud frames, and is not affected by the F i+1 Frame point cloud to F i+n-1 The impact of continuous registration errors on frame point clouds. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the flowchart of the current HRegNet registration network;

[0039] Figure 2 is a flow chart of the measurement method of the Lidar odometer of the present invention;

[0040] Figure 3a Schematic diagram of visualization of Lidar odometry measurement results along the x, y, and z coordinate axes of the present invention and existing algorithms on the first set of test data sets;

[0041] Figure 3b Schematic diagram of visualization of Lidar odometry measurement results along the x, y, and z coordinate axes of the present invention and existing algorithms on the second test dataset;

[0042] Figure 3c Schematic diagram of visualization of the Lidar odometry measurement results along the x, y, and z coordinate axes of the present invention and the existing algorithm on the third test dataset;

[0043] Figure 4a Bird's-eye view of Lidar odometry measurements of the present invention and existing algorithms on the first set of test data sets;

[0044] Figure 4b Bird's-eye view of Lidar odometry measurement results of the present invention and existing algorithms on the second set of test data sets;

[0045] Figure 4c Bird's-eye view of the Lidar odometry measurement results of the present invention and the existing algorithm on the third test dataset;

[0046] Figure 5a Schematic diagram of the relative error of the fixed distance between the present invention and the existing algorithm on the first set of test data sets;

[0047] Figure 5b Schematic diagram of the relative error of the fixed distance between the present invention and the existing algorithm on the second set of test data sets;

[0048] Figure 5c Schematic diagram of the relative error of fixed distance between the present invention and the existing algorithm on the third set of test data sets. DETAILED DESCRIPTION

[0049] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0050] The HRegNet registration network only considers the registration between two station clouds (pairwise registration) and does not involve the sequential registration problem between continuous Lidar frame point clouds. Based on the HRegNet network, this paper proposes the HRegNet-odo network. This improved network achieves high-precision measurement of Lidar odometry by adding initial registration error detection and cumulative error detection. Figure 2 As shown in the figure, the following takes the Lidar odometer installed on an unmanned vehicle as an example to explain its measurement process in detail.

[0051] 1. Get relevant data.

[0052] The relevant data obtained by the present invention include point cloud data, scanning rate and carrier speed, wherein the point cloud data refers to the point cloud data measured by the Lidar odometer, the scanning rate refers to the scanning rate of the Lidar odometer, and the carrier speed refers to the speed of the carrier on which the Lidar odometer is located, which in this embodiment refers to the vehicle speed.

[0053] 2. Perform initial registration error detection.

[0054] This step mainly determines whether the registration of two consecutive point cloud frames (paired registration) is successful based on certain rules, and applies a pre-set transformation matrix to replace the registration that failed the test. In other words, this step is gross error detection, the purpose of which is to reduce the cumulative impact of large errors in a certain frame on subsequent frames.

[0055] This paper proposes an initial registration error detection method based on the vehicle speed and Lidar scanning frequency, that is, a rule for judging whether the registration of two consecutive point clouds is successful or not. In the process of sequential registration using the HregNet network, the calculated transformation matrix between the two point clouds is recorded as If from If the translation distance between the two decomposed point clouds is greater than 2Δd, the current registration is considered to have failed. If the current registration is considered to have failed, assuming that the vehicle moves along the y-axis, the registration is re-registered. Set to:

[0056]

[0057] In the above process, the threshold Δd is calculated according to the following formula:

[0058]

[0059] Where v is the vehicle speed in m / s, and h is the LiDAR scanning frequency in Hz. The parameter 1 / 2 is set because the displacement calculated from the vehicle speed and the LiDAR scanning frequency is the sum of the displacement along the three coordinate axes, but the forward direction has a greater contribution.

[0060] 3. Perform cumulative error detection.

[0061] During the registration of continuous point cloud frames, the registration of each previous frame will cumulatively affect the registration of subsequent frames. In order to detect and control the error accumulation, the present invention proposes a segmented detection method. The specific process is divided into two steps. The first is the overall error detection. If it fails, the second step of local recalculation error detection is performed.

[0062] (1) Overall error detection

[0063] In a registration sequence containing n point cloud frames, the Fth point cloud frame calculated by the HRegNet registration network is used. i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is recorded as but:

[0064]

[0065] Where n is a parameter in the overall error detection, representing the number of point cloud frames contained in the overall detection segment. Due to the need in the local recalculation error detection step, this value must be an even number; Indicates F i+n-1 Frame point cloud to F i+n-2 The transformation matrix of the coordinate system of the frame point cloud.

[0066] In the overall error detection process, the first step is to directly apply the Fi+n Frame point cloud and F i The frame point cloud calculates the transformation matrix between the two frame point clouds, which is recorded as Then calculate and The relative rotation dR and relative translation dt between the two are calculated as shown in formula (4):

[0067]

[0068] Where trace(ΔR) represents the trace of the matrix ΔR; ΔR and Δt can be calculated using formula (5), which is as follows:

[0069]

[0070] If dR and dt meet the requirements shown in formula (6), then the segment of the registration sequence is judged to have passed the overall error detection. Formula (6) is as follows:

[0071]

[0072] At this time, the F i+n Frame point cloud to F i The transformation matrix of the frame point cloud coordinate system is reset to 1 T i i+n and T i i+n The average value of , that is:

[0073]

[0074] If the segment of the registration sequence fails to pass the global error detection, it will continue to perform the second step of local recalculation error detection.

[0075] (2) Local recalculation error detection

[0076] For the registration sequence containing n point cloud frames, this step still uses the HRegNet registration network to gradually calculate the F i+n Frame point cloud to F i The transformation relationship of the coordinate system of the frame point cloud is different from that of the surface in that: in this step, the transformation matrix between two consecutive frames of point cloud is not calculated, but the transformation matrix between paired point cloud frames is calculated every frame. In other words, the previous calculation of F i+1 Frame point cloud and F i The conversion relationship between frame point clouds, and this step calculates the F i+2 Frame point cloud and F i The conversion relationship between frame point clouds. Therefore, the F i+n Frame point cloud to F iThe transformation matrix of the coordinate system of the frame point cloud can be expressed as:

[0077]

[0078] Where n is the number of point cloud frames contained in each segment of the overall error detection. Since the transformation relationship is recalculated every frame, n must be an even number. Indicates F i+n-2 Frame point cloud to F i+n-4 The transformation matrix of the coordinate system of the frame point cloud. As another implementation, the interval may be two frames or more.

[0079] Judge separately and and The relative rotation and translation between and

[0080] like If the requirements shown in formula (6) are met, the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of

[0081] like If the requirements shown in formula (6) are met, the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of

[0082] like as well as If both meet the corresponding requirements, then the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud can be and The average value of and or the average of both.

[0083] like and If the requirements shown in formula (6) are not met, then simply i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is set as Because the calculation process of the transformation matrix is ​​the most direct, it involves the least number of related point cloud frames and is not affected by the F i+1 Frame point cloud to F i+n-1The impact of continuous registration errors on frame point clouds.

[0084] Experimental verification

[0085] To better illustrate the effectiveness of the present invention, the effectiveness of the Lidar odometry measurement method proposed in the present invention is tested on the Kitti odometry dataset and compared with the point-to-point ICP (ICP-P2Point), point-to-plane ICP (ICP-P2Plane), and HRegNet algorithms.

[0086] All algorithms were completed on a computer with Ubuntu 20.04 system, Intel(R) Xeon(R) E-2176G CPU 3.70GHz 64.0-GB RAM. Among them, the ICP-P2Point and ICP-P2Plane algorithms were implemented in Python based on the Open3D open source library; the HRegNet and HRegNet-odo algorithms were implemented in Python based on PyTorch, and Adam was used as the optimizer. The deep neural network was trained on an NVIDIA RTX 3090 GPU.

[0087] (1) Data used in the test

[0088] Kitti odometry is a public dataset commonly used to test LiDAR odometry measurement algorithms. It consists of 21 subsequences, of which the first 11 sequences (00-10) contain the ground truth transformation matrix. In this experiment, sequences 00-05 are used for training, sequences 06 and 07 are used for validation, and sequences 08-10 are used for testing. Table 1 describes the details of the dataset used in the experiment.

[0089] Table 1

[0090]

[0091]

[0092] (2) Evaluation criteria

[0093] The proposed relative accuracy evaluation method, which is widely used in visual / Lidar odometry measurement and SLAM technology accuracy evaluation, is used to evaluate the Lidar odometry measurement algorithm involved in this paper. The relative error is defined as the deviation between the relative pose change estimate and the ground truth value within a distance interval. It includes two parts: relative rotation error (RRE) and relative translation error (RTE). The calculation formula is as follows:

[0094]

[0095] Where, and Indicates F j Frame point cloud to F i The ground truth rotation matrix and translation vector of the frame point cloud; and Represents the rotation matrix and translation vector between two frames of point cloud estimated by the algorithm; trace(ΔR) represents the trace of the matrix ΔR.

[0096] (3) Parameter settings

[0097] The measurement method of the present invention mainly involves two important parameters: the parameter Δd in the initial registration error detection and the number n of point cloud frames contained in a whole in the cumulative error detection.

[0098] For the Kitti odometry dataset, the vehicle speed during acquisition was approximately 50 km / h, and the onboard Lidar scanning frequency was 10 Hz. Therefore, according to Equation (2), Δd = 0.69 m can be calculated. The parameter n was set to 10 in this experiment. In addition, for the HRegNet and HRegNet-odo networks, the initial learning rate was set to 0.001, decreasing by 50% every 10 cycles, and the training cycle was set to 50. For both the ICP-P2Point and ICP-P2Plane algorithms, their initial registration matrices were set to the unit matrix, the distance threshold was set to 1.0 m, and the maximum number of iterations was set to 30.

[0099] (4) Qualitative analysis of Lidar odometry measurement results

[0100] like Figure 3a 、 Figure 3b and Figure 3c The figure shows the visualization results of the Lidar odometry measurements along the x, y, and z coordinate axes for the proposed algorithm (HRegNet-odo) and three comparison algorithms (HRegNet, ICP-P2Point, and ICP-P2Plane) on three test datasets (sequences 08-10). From this, we can find the following two points:

[0101] ① Among the four Lidar odometry measurement methods, the calculation results based on deep network algorithms (HRegNet-odo and HRegNet) have higher accuracy than those of traditional algorithms (ICP-P2Point and ICP-P2Plane), and the measurement results of the HRegNet-odo algorithm designed in the present invention are closest to the ground truth results. This qualitatively shows that the HRegNet-odo network can effectively realize the measurement of Lidar odometry in the three test datasets.

[0102] ② The measurement accuracy of the four Lidar odometry measurement methods in the x- and z-axis directions is significantly better than that in the y-axis direction. This is because the ground truth transformation matrix provided in the Kitti odometry dataset is defined in the vehicle carrier left camera coordinate system, with its x-axis pointing to the right, y-axis pointing down, and z-axis pointing forward. The scanned point cloud has more constraints in the horizontal direction and fewer constraints in the vertical direction. Therefore, the calculation accuracy in the horizontal direction is better than that in the vertical direction.

[0103] In practical applications, the horizontal component of the Lidar odometry measurement result is more important than the vertical component. In order to more clearly observe the measurement results of the four algorithms in the horizontal direction, as shown in the following figure: Figure 4a 、 Figure 4b and Figure 4c The figure shows a bird's-eye view of the Lidar odometry measurement results of the four algorithms on three test datasets. The following two points can be seen:

[0104] I. The calculation results of the HRegNet-odo and HRegNet algorithms are closer to the ground truth results than those of the ICP-P2Point and ICP-P2Plane algorithms. This qualitatively illustrates that deep learning methods are more accurate and robust than traditional algorithms when processing long sequences of continuous Lidar frame point cloud registration without initialization.

[0105] II. The calculation results of the HRegNet-odo and HRegNet algorithms on sequences 09 and 10 are better than those on sequence 08. This is because the vehicle in sequence 08 makes multiple large turns, which makes it difficult to align the point cloud frames. This also shows that relying solely on LiDAR scanned point cloud data to achieve high-precision odometry measurement is still a huge challenge. In subsequent research, multi-system fusion odometry measurement should be considered, such as odometry measurement technology that integrates vision, inertial navigation, and LiDAR.

[0106] (5) Quantitative analysis of Lidar odometry measurement results

[0107] Figure 5a 、 Figure 5b and Figure 5c The fixed distance relative error of the four algorithms used in this experiment is shown on three test datasets. Table 2 records the measurement accuracy and time consumption of the four algorithms on the three test datasets. It also summarizes the average measurement accuracy and time consumption of each algorithm on the three test datasets, as well as the root mean square error (RMSE) of the measurement accuracy. Bold font in the table indicates the optimal value.

[0108] Table 2

[0109]

[0110]

[0111] according to Figure 5a 、 Figure 5b 、 Figure 5c As well as the information shown in Table 2, we can find the following two pieces of information:

[0112] (1) In terms of odometry calculation accuracy, the relative error of the deep neural network algorithm (HRegNet and HRegNet-odo) is smaller than that of the traditional algorithm (ICP-P2Point and ICP-P2Plane algorithm), which quantitatively illustrates that the deep learning-based method effectively improves the problems of low accuracy and poor stability of the traditional algorithm when dealing with long sequences and continuous Lidar point cloud frame registration without initial values; and the calculation accuracy of the HRegNet-odo algorithm is higher than that of the HRegNet algorithm, which proves that the improvement made by the present invention to the HRegNet network enables it to perfectly convert from paired point cloud registration to Lidar odometry measurement of continuous point cloud registration; the average relative rotation and translation errors of the HRegNet-odo algorithm per 100 meters are 1.0° and 3.01m, and the RMSE is 1.0° and 2.65m, which quantitatively illustrates that the HRegNet-odo network can achieve high-precision measurement of Lidar odometry in the test dataset.

[0113] (2) In terms of computational time, HRegNet-odo and HRegNet algorithms are better than ICP-P2Point and ICP-P2Plane algorithms. The reason may be that HRegNet-odo and HRegNet algorithms are implemented based on GPU, while ICP-P2Point and ICP-P2Plane algorithms are completed on CPU. Further comparison of HRegNet-odo and HRegNet algorithms shows that HRegNet-odo consumes about 0.11s more time than HRegNet in each registration, and the average registration time per frame is about 0.2s.

[0114] The present invention applies the improved HRegNet-odo network to LiDAR odometry measurement and obtains good test results on the Kitti odometry dataset. The relative rotation accuracy and relative translation accuracy per 100 meters are better than 1.2° and 4.2m, respectively. The calculation time per frame is about 0.2s, which can meet the measurement accuracy and efficiency requirements of LiDAR odometry to a certain extent.

Claims

1. A Lidar odometer measurement method, characterized in that: The measurement method includes the following steps: 1) Obtain the point cloud data measured by the Lidar odometer, the scanning rate of the Lidar odometer, and the speed of the carrier on which the Lidar odometer is located; 2) Use the HregNet network to register two adjacent frames of point clouds, and determine whether the initial registration error requirements are met based on the scanning rate, carrier speed, and the registration results of the two adjacent frames of point clouds; 3) Using the HregNet network, the coordinate transformation matrix from the first frame to the last frame in a point cloud sequence is calculated according to the registration method of adjacent frames, which is recorded as the first coordinate transformation matrix. The coordinate transformation matrix from the first frame to the last frame in a point cloud sequence is also directly calculated, which is recorded as the second coordinate transformation matrix. The overall error detection of the point cloud sequence is performed based on the obtained first and second coordinate transformation matrices. 4) When the overall error requirement is not met, the HregNet network is used to calculate the coordinate transformation matrix from the first frame to the last frame in a point cloud sequence according to the registration method with an interval of at least one frame, which is recorded as the third coordinate transformation matrix. Based on the obtained first coordinate transformation matrix, second coordinate transformation matrix and third coordinate transformation matrix, the local recalculation error detection is performed on the point cloud sequence.

2. The Lidar odometer measurement method according to claim 1, characterized in that: The initial registration error requirement in step 2) is that the translation distance between two adjacent frames of point clouds is no greater than a set threshold. The translation distance is decomposed from the conversion matrix between the two adjacent frames of point clouds. The set threshold is determined by the scanning rate and the carrier speed. If the initial registration error requirement is not met, the conversion matrix between the two adjacent frames of point clouds is set to: Where T i i+1 is the conversion matrix between two adjacent frames of point clouds, and Δd is the set threshold.

3. The Lidar odometer measurement method according to claim 2, characterized in that: The calculation formula for setting the threshold is: Where Δd is the set threshold; v is the vehicle speed in m / s; h is the Lidar scanning frequency in Hz.

4. The Lidar odometer measurement method according to claim 1, characterized in that: The overall error requirement is that both the relative rotation and the relative translation are less than the corresponding thresholds, where: n is the number of point cloud frames contained in each segment in the overall error detection; trace(ΔR) represents the trace of the matrix ΔR; Indicates direct application of F i+n Frame point cloud and F i The frame point cloud calculates the transformation matrix between the two frame point clouds, which is the second transformation matrix; Indicates that the Fth frame is calculated by applying the HRegNet registration network according to the registration method of adjacent frames. i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is the first transformation matrix.

5. The Lidar odometer measurement method according to claim 4, characterized in that: When the overall error meets the requirements, the F i+n Frame point cloud to F i The transformation matrix of the frame point cloud coordinate system is set as and The average value of .

6. The Lidar odometer measurement method according to claim 4, characterized in that: The calculation formula of the third coordinate transformation matrix during the local recalculation error detection is: n is the number of point cloud frames contained in each segment in the overall error detection, and n is an even number; Indicates F i+n-2 Frame point cloud to F i+n-4 The transformation matrix of the coordinate system of the frame point cloud.

7. The Lidar odometer measurement method according to claim 6, characterized in that: The local recalculation error requirement used in local recalculation error detection is and The relative rotation and translation between and The relative rotation and translation between them meet the corresponding thresholds.

8. The Lidar odometer measurement method according to claim 6, characterized in that: like If both meet the corresponding threshold requirements, then the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of like If both meet the corresponding threshold, then the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is reset to and The average value of They are and The relative rotation and translation between They are and The relative rotation and translation between them.

9. The Lidar odometer measurement method according to claim 8, characterized in that: like and If the thresholds do not meet the corresponding requirements, the F i+n Frame point cloud to F i The transformation matrix of the coordinate system of the frame point cloud is set as

Images