A vehicle pose estimation method based on point cloud features

Abstract

This invention relates to the technical field of parking AGVs, specifically a vehicle pose estimation method based on point cloud features. The method includes: preprocessing point cloud data of a first region of interest; indexing the point cloud data and then segmenting it; vertically projecting the point cloud data onto the XOY plane; fitting line segments to the two-dimensional point cloud data to obtain two tire contour segments; extracting two line segments whose angle with the X-axis is less than a set threshold based on the two tire contour segments, determining whether the extracted two line segments are parallel (if parallel, recognition is successful); determining whether the parking AGV can safely enter under the vehicle based on the distance between the two nearest points of the two line segments; fitting a new line segment based on the midpoint of the two line segments, using the midpoint of the new line segment as the vehicle position, and obtaining the vehicle's orientation angle relative to the AGV through the tilt angle of the new line segment relative to the X-axis. This invention avoids the visual errors associated with manual AGV centering and solves the problems of high difficulty and low efficiency associated with traditional manual operation.

Description

Technical Field

[0001] This invention relates to the technical field of parking AGVs, and more specifically, to a vehicle pose estimation method based on point cloud features. Background Technology

[0002] In practical AGV (Automated Guided Vehicle) transport scenarios, target vehicles are typically parked at fixed locations. The AGV must first move to the vicinity of the parking area, accurately identify the vehicle's position relative to the AGV, and then navigate to directly beneath the vehicle to complete the lifting and transport tasks. However, current technologies often rely on manual operation of the AGV for alignment. This method is characterized by high operational difficulty, large errors, and low efficiency. Summary of the Invention

[0003] The purpose of this invention is to overcome the problems of high operational difficulty, large error, and low efficiency in the existing technology of relying on manual operation of AGV for centering, and to provide a vehicle pose estimation method based on point cloud features.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A vehicle pose estimation method based on point cloud features is provided, the method comprising the following steps: S1: Preprocess the point cloud data of the first region of interest using voxel filtering; S2: Index the preprocessed point cloud data, and then segment the indexed point cloud data; S3: Use point cloud dimensionality reduction to vertically project the segmented point cloud data onto the XOY plane to obtain two-dimensional point cloud data; S4: Use a random sampling algorithm to fit line segments to the two-dimensional point cloud data to obtain the outline lines of the two tires; S5: Extract two line segments with an angle less than a set threshold between them and the X-axis based on the two tire contour line segments. Analyze and determine whether the extracted two line segments are parallel. If they are parallel, the recognition is successful. S6: Determine whether the parking AGV can safely enter under the vehicle based on the distance between the two nearest points of the two line segments, and finally determine whether the vehicle recognition is reasonable; S7: A new line segment is obtained by fitting the midpoint of the two line segments, the midpoint of the new line segment is taken as the vehicle position, and the orientation angle of the vehicle relative to the AGV is obtained by the inclination angle between the new line segment and the X-axis.

[0005] Preferably, the method further includes step S8, which specifically includes: S81: Preprocess the point cloud data of the second region of interest using voxel filtering, and index the preprocessed point cloud data; S82: Segment the indexed point cloud data and calculate the centroid positions of the point cloud clusters of the two wheels using the arithmetic mean center calculation formula of the point cloud clusters. S83: Calculate the tilt angle between the line segment between the centroids of the two wheel point cloud clusters relative to the X-axis based on the positions of the centroids of the two wheel point cloud clusters to obtain the orientation angle of the vehicle relative to the AGV. The midpoint of the line segment between the centroids of the two wheel point cloud clusters is the position of the vehicle center relative to the AGV.

[0006] Preferably, step S4 specifically includes the following steps: S41: The minimum number of data points required to uniquely determine the model parameters are randomly selected from the two-dimensional point cloud data using a random sampling consensus algorithm to form a minimum sample set; S42: Next, model fitting is performed; then, interior point identification is performed. The algorithm substitutes all points in the entire dataset into the current model to calculate the error, and classifies points with errors less than a preset threshold as "interior points" and those with errors greater than the threshold as "outer points". Then, it is iterated until the preset number of iterations is reached or a model with a sufficiently high proportion of interior points is found. Finally, all the identified interior points are used to refit the model using methods such as least squares, thereby outputting a final model, which extracts line segment features from the point cloud and identifies line segments with lengths greater than a set threshold.

[0007] Preferably, in step S42, a certain point cloud P Whether a line segment conforms to the fitted model is determined by the following formula:

[0008] in, It is a point on a line segment. It is a direction vector. It is the maximum distance between point clouds that satisfies the fitted line segment.

[0009] Preferably, in step S5, the angle between the two line segments is calculated using the cosine formula for the vector angle. According to the angle between the two line segments Analyze and determine whether the two extracted line segments are parallel.

[0010] Preferably, step S1 specifically includes the following steps: S11: Define a first region of interest for the point cloud. The first region of interest is a cuboid region. The length and width of the first region of interest are the length of the car directly in front of the AGV. The height of the first region of interest is 3 cm to 5 cm. S12: Perform voxel filtering downsampling on the point cloud within the first region of interest. First, calculate a cuboid bounding box that can just enclose the entire point cloud data to create a voxel mesh. Then, divide the bounding box into countless small, fixed-volume cubic units, i.e., voxels, according to the set voxel size. Next, aggregate the points within the voxels. Each point in the point cloud is classified into its specific voxel. At the same time, all the points contained within each voxel are "aggregated" and replaced with a representative point. Finally, a new point cloud is generated, and the representative points of all voxels together constitute the downsampled new point cloud.

[0011] Preferably, in step S2, the specific steps for indexing the preprocessed point cloud data are as follows: the preprocessed point cloud is indexed by constructing a KD-Tree. First, a split dimension is selected, and then the median of all data points in the current dimension is selected as the split point. Finally, subtrees are recursively constructed. After determining the split dimension and split point of the current node, the remaining data points are divided into left and right subtrees. In the current split dimension, points with values ​​less than the split point are assigned to the left subtree, and points with values ​​greater than or equal to the split point are assigned to the right subtree. Then, the above steps are recursively performed on the left and right subsets until the subset contains only one point.

[0012] Preferably, in step S2, the preprocessed point cloud has an approximately symmetrical "L" shape. The specific steps for segmenting the indexed point cloud data are as follows: S21: For the point cloud with this feature, the Euclidean clustering extraction algorithm is used. The algorithm traverses every point in the point cloud. If the point has not been processed, it is used as a seed point of a new cluster, and all its nearest neighbors within the specified Euclidean distance tolerance range are found. S22: Euclidean clustering extraction recursively marks these nearest neighbors as part of the current cluster and continues to grow regions using them as new seed points until the cluster can no longer be expanded; S23: The Euclidean clustering extraction algorithm will filter the cluster based on the preset minimum and maximum number of points, discarding invalid clusters with too many or too few points; S24: Save the clusters that meet the conditions as independent point cloud sets, and repeat the above process until all points in the point cloud have been processed, thereby dividing it into multiple independent point cloud clusters, each cluster representing a potential wheel.

[0013] Preferably, in step S3, the segmented point cloud data is in three-dimensional space, but the point cloud data has a relatively low height range along the z-axis. The three-dimensional point cloud data is then vertically projected onto the XOY plane.

[0014] Preferably, in step S81, the second region of interest is a cuboid region, the length and width of the second region of interest are the length of the car directly in front of the AGV, and the height of the second region of interest is greater than the height of the first region of interest.

[0015] Compared with the prior art, the beneficial effects of the present invention are: the present invention realizes the function of accurately identifying the vehicle's position relative to the AGV in the application scenario of AGV transport vehicles, and then navigating the AGV to directly under the vehicle body, which improves the success rate of AGV transport vehicle tasks, avoids the visual error of human operation of AGV centering, reduces labor costs, and effectively solves the problems of high difficulty and low efficiency of traditional manual operation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the second embodiment of the vehicle pose estimation method based on point cloud features of the present invention.

[0017] Figure 2 This is a schematic diagram of the point cloud cluster detection algorithm for vehicle pose estimation in the point cloud feature-based vehicle pose estimation method of the present invention.

[0018] Figure 3 This is a schematic diagram of a frontal scan. Figure 4 This is a simulated point cloud image after frontal preprocessing; Figure 5 Simulated image of successful frontal straight line detection and recognition; Figure 6 Simulated image of successful frontal point cloud cluster detection and recognition; Figure 7 This is a schematic diagram of a side scan. Figure 8 This is a simulated point cloud image after side preprocessing; Figure 9 Simulated image of successful side line detection and recognition; Figure 10 Simulated image of successful detection and recognition of side point cloud clusters; Figure 11 This is a schematic diagram of the process of Embodiment 1 of the vehicle pose estimation method based on point cloud features of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments.

[0020] Example 1 like Figure 1 As shown, a vehicle pose estimation method based on point cloud features is proposed. The method includes the following steps: S1: Preprocess the point cloud data of the first region of interest using voxel filtering; S2: Index the preprocessed point cloud data, and then segment the indexed point cloud data; S3: Use point cloud dimensionality reduction to vertically project the segmented point cloud data onto the XOY plane to obtain two-dimensional point cloud data; S4: Use a random sampling algorithm to fit line segments to the two-dimensional point cloud data to obtain the outline lines of the two tires; S5: Extract two line segments with an angle less than a set threshold between them and the X-axis based on the two tire contour lines. Analyze and determine whether the two extracted line segments are parallel. If they are parallel, the recognition is successful. S6: Determine whether the parking AGV can safely enter under the vehicle based on the distance between the two nearest points of the two line segments, and finally determine whether the vehicle recognition is reasonable; S7: Fit a new line segment based on the midpoint of the two line segments, take the midpoint of the new line segment as the vehicle position, and obtain the vehicle's orientation angle relative to the AGV through the tilt angle between the new line segment and the X-axis.

[0021] It should be noted that in step S5, the threshold angle can be set to 20°, or it can be adjusted as needed, so that two line segments can be extracted based on the two tire contour line segments.

[0022] like Figure 2 As shown, the method further includes step S8, which specifically includes: S81: Preprocess the point cloud data of the second region of interest using voxel filtering, and index the preprocessed point cloud data; S82: Segment the indexed point cloud data and calculate the centroid positions of the point cloud clusters of the two wheels using the arithmetic mean center calculation formula of the point cloud clusters. S83: Calculate the tilt angle between the line segment between the centroids of the two wheel cloud clusters relative to the X-axis based on the positions of the centroids of the two wheel cloud clusters to obtain the orientation angle of the vehicle relative to the AGV. The midpoint of the line segment between the centroids of the two wheel cloud clusters is the position of the vehicle center relative to the AGV.

[0023] In addition, step S4 specifically includes the following steps: S41: The minimum number of data points required to uniquely determine the model parameters are randomly selected from the two-dimensional point cloud data using a random sampling consensus algorithm to form a minimum sample set; S42: Next, model fitting is performed; then, interior point identification is performed. The algorithm substitutes all points in the entire dataset into the current model to calculate the error, and classifies points with errors less than a preset threshold as "interior points" and those with errors greater than the threshold as "outer points". Then, it is iterated until the preset number of iterations is reached or a model with a sufficiently high proportion of interior points is found. Finally, all the identified interior points are used to refit the model using methods such as least squares, thereby outputting a final model, which extracts line segment features from the point cloud and identifies line segments with lengths greater than a set threshold.

[0024] In step S42, a certain point cloud P Whether a line segment conforms to the fitted model is determined by the following formula:

[0025] in, It is a point on a line segment. It is a direction vector. It is the maximum distance between point clouds that satisfies the fitted line segment.

[0026] In addition, in step S5, the angle between the two line segments is calculated using the cosine formula of the vector angle. According to the angle between the two line segments Analyze and determine whether the two extracted line segments are parallel.

[0027] Step S1 specifically includes the following steps: S11: Define the first region of interest (ROI) for the point cloud. The first ROI is a cuboid region. The length and width of the first ROI are the length of the car directly in front of the AGV. The height of the first ROI is a suitable value, which can be 3cm to 5cm. S12: Perform voxel filtering downsampling on the point cloud within the first region of interest. First, calculate a cuboid bounding box that can just enclose the entire point cloud data to create a voxel mesh. Then, divide the bounding box into countless small, fixed-volume cubic units, i.e., voxels, according to the set voxel size. Next, aggregate the points within the voxels. Each point in the point cloud is classified into its specific voxel. At the same time, all the points contained within each voxel are "aggregated" and replaced with a representative point. Finally, a new point cloud is generated, and the representative points of all voxels together constitute the downsampled new point cloud.

[0028] In addition, the specific steps for indexing the preprocessed point cloud data in step S2 are as follows: The preprocessed point cloud is indexed by constructing a KD-Tree. First, a split dimension is selected, and then the median of all data points in the current dimension is selected as the split point. Finally, subtrees are recursively constructed. After determining the split dimension and split point of the current node, the remaining data points are divided into left and right subtrees. In the current split dimension, points with values ​​less than the split point are assigned to the left subtree, and points with values ​​greater than or equal to the split point are assigned to the right subtree. Then, the above steps are recursively performed on the left and right subsets until the subset contains only one point.

[0029] In step S2, the preprocessed point cloud has an approximately symmetrical "L" shape. The specific steps for segmenting the indexed point cloud data are as follows: S21: For the point cloud with this feature, the Euclidean clustering extraction algorithm is used. The algorithm traverses every point in the point cloud. If the point has not been processed, it is used as a seed point of a new cluster, and all its nearest neighbors within the specified Euclidean distance tolerance range are found. S22: Euclidean clustering extraction recursively marks these nearest neighbors as part of the current cluster and continues to grow regions using them as new seed points until the cluster can no longer be expanded; S23: The Euclidean clustering extraction algorithm will filter the cluster based on the preset minimum and maximum number of points, discarding invalid clusters with too many or too few points; S24: Save the clusters that meet the conditions as independent point cloud sets, and repeat the above process until all points in the point cloud have been processed, thereby dividing it into multiple independent point cloud clusters, each cluster representing a potential wheel.

[0030] In addition, in step S3, the segmented point cloud data is in three-dimensional space, but the point cloud data has a relatively low height range on the z-axis. The three-dimensional point cloud data is then vertically projected onto the XOY plane.

[0031] In step S81, the second region of interest is a cuboid region, the length and width of which are equal to the length of the car directly in front of the AGV, and the height of the second region of interest is greater than the height of the first region of interest.

[0032] Example 2 like Figure 1 As shown, a vehicle pose estimation method based on point cloud features is presented, specifically for AGV frontal pose recognition in this example. Figure 3 This includes the following steps: Step 1: Box Clipping and Voxel Filtering. After obtaining point cloud data from radar, in order to achieve efficient real-time vehicle pose estimation on an industrial control computer with low computing resources, the point cloud information of the vehicle first needs to be preprocessed. The preprocessing steps are as follows: First, define the region of interest (ROI) in the point cloud. The ROI is a cuboid region directly in front of the AGV with the length and width of the car and a relatively small height. Then, perform voxel filtering downsampling on the point cloud within the ROI to remove some noise points and outliers.

[0033] Step 2: Euclidean clustering and point cloud dimensionality reduction. The preprocessed point cloud simulation image is shown below. Figure 4 As shown, the shape is an approximately symmetrical "L" shape. For this feature of the point cloud, the Euclidean Cluster Extraction algorithm is used to segment the point cloud into multiple independent point cloud clusters, each cluster representing a potential wheel. To improve the segmentation efficiency of the algorithm, the input point cloud is indexed by constructing a KD-Tree. Although the point cloud is currently in three-dimensional space, its z-axis height range is relatively low. The three-dimensional point cloud can be vertically projected onto the XOY plane. By reducing the dimensionality of the point cloud, the efficiency of fitting line segments to the wheel contour is improved.

[0034] Step 3: RANSAC line feature extraction and stacking of line segments meeting the conditions. For the 2D point cloud obtained after the above processing, line segment features are extracted using the Random Sample Consensus (RANSAC) algorithm. Line segments with lengths greater than a set threshold are identified to describe the vehicle's wheel contours. The following formula is used to determine the line segment feature of a given point cloud. P Does it conform to the fitted line segment model?

[0035] in It is a point on a line segment. It is a direction vector. It is the maximum distance between point clouds that satisfies the fitting line segment. In the RANSAC algorithm, the number of iterations is a key parameter. Too few iterations may not find a good model, while too many iterations will waste computational resources and affect the efficiency of recognition. Therefore, "adaptive iteration count" is selected, and the required number of iterations is calculated using the adaptive RANSAC iteration count calculation formula.

[0036] Step four: Calculate the geometric relationships between line segments. For cases where the angle calculated using the cosine formula for the included angle is successfully identified, such as... Figure 5 Then, calculate the distance between the midpoints of the two line segments to estimate the width of the vehicle (i.e., the wheelbase); and calculate the distance between the nearest endpoints of the two line segments to determine whether the AGV can safely enter the bottom of the vehicle; the calculations of both confirm whether a reasonable vehicle identification is constituted.

[0037] Step 5: Calculate the midpoint of the line segment and the inclination angle of the line. Calculate the midpoint between the two line segments as the vehicle's position; simultaneously calculate the inclination angle of the new line segment fitted by the contour line segment to obtain the vehicle's orientation angle relative to the AGV.

[0038] Step six, point cloud cluster detection and pose estimation, such as Figure 2 After the first method, "straight-line detection," the AGV vehicle has roughly stopped in front of the vehicle being transported. However, considering the relatively small allowable error in the passage between the AGV and the vehicle, more precise identification is required. This embodiment still uses a point cloud preprocessing step similar to the first method. The difference is that the region of interest is a cuboid region directly in front of the AGV, with the length and width being the length of the car and the height being a small, suitable value. Then, the Euclidean clustering method is used to classify the point cloud, dividing the point clouds of the left and right wheels into two point clouds. Successful point cloud cluster identification is shown below. Figure 6 .

[0039] Step 7: Calculate the centroid position of each wheel point cloud cluster using the arithmetic mean center calculation formula. The vehicle's pose data is obtained by calculating the geometric relationship between the centroids of two point cloud clusters. The slope of the line segment between the centroids allows calculation of the vehicle's tilt angle relative to the AGV; the midpoint of the line segment is the position of the vehicle's center relative to the AGV.

[0040] Example 3 like Figure 7 The vehicle's pose can be identified by the two tires on the side of the vehicle. The steps are the same as in Example 2, requiring two main steps: line detection and point cloud cluster detection.

[0041] Step 1, Line Detection. Line detection requires preprocessing of the radar scan point cloud, specifically downsampling the point cloud within the region of interest to obtain the following: Figure 8 The simulated point cloud was then used, and Euclidean clustering was performed using KD-Tree indexing to segment the point cloud into multiple independent point cloud clusters, each cluster representing a potential wheel; then, the reduced point cloud was randomly sampled consistently for line segment extraction to describe the wheel outline, as shown in the figure. Figure 9 The simulated line segments are used to determine whether the recognition was successful and the vehicle's two positions based on the geometric relationship of the line segments.

[0042] Step two, point cloud cluster detection. Point cloud cluster detection differs from point cloud preprocessing in line detection in that the region of interest is different. The point cloud is segmented into clusters representing the front and rear wheels, such as... Figure 10 The vehicle's pose information is obtained by calculating the geometric relationship between the centroids of two cloud clusters.

[0043] In the specific implementation of the above embodiments, the technical features can be combined in any non-contradictory way. For the sake of brevity, not all possible combinations of the above technical features are described. However, as long as the combination of these technical features is not contradictory, it should be considered to be within the scope of this specification.

[0044] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A vehicle pose estimation method based on point cloud features, characterized in that, The method includes the following steps: S1: Preprocess the point cloud data of the first region of interest using voxel filtering; S2: Index the preprocessed point cloud data, and then segment the indexed point cloud data; S3: Use point cloud dimensionality reduction to vertically project the segmented point cloud data onto the XOY plane to obtain two-dimensional point cloud data; S4: Use a random sampling algorithm to fit line segments to the two-dimensional point cloud data to obtain the outline line segments of the two tires; S5: Extract two line segments with an angle less than a set threshold between them and the X-axis based on the two tire contour line segments. Analyze and determine whether the extracted two line segments are parallel. If they are parallel, the recognition is successful. S6: Determine whether the parking AGV can safely enter under the vehicle based on the distance between the two nearest points of the two line segments, and finally determine whether the vehicle recognition is reasonable; S7: A new line segment is obtained by fitting the midpoint of the two line segments, the midpoint of the new line segment is taken as the vehicle position, and the orientation angle of the vehicle relative to the AGV is obtained by the inclination angle between the new line segment and the X-axis.

2. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, The method further includes step S8, which specifically includes... S81: Preprocess the point cloud data of the second region of interest using voxel filtering, and index the preprocessed point cloud data; S82: Segment the indexed point cloud data and calculate the centroid positions of the point cloud clusters of the two wheels using the arithmetic mean center calculation formula of the point cloud clusters. S83: Calculate the tilt angle between the line segment between the centroids of the two wheel point cloud clusters relative to the X-axis based on the positions of the centroids of the two wheel point cloud clusters to obtain the orientation angle of the vehicle relative to the AGV. The midpoint of the line segment between the centroids of the two wheel point cloud clusters is regarded as the position of the vehicle relative to the AGV.

3. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, Step S4 specifically includes the following steps. S41: The minimum number of data points required to uniquely determine the model parameters are randomly selected from the two-dimensional point cloud data using a random sampling consensus algorithm to form a minimum sample set; S42: Next, model fitting is performed; then, interior point identification is performed. The algorithm substitutes all points in the entire dataset into the current model to calculate the error, and classifies points with errors less than a preset threshold as "interior points" and those with errors greater than the threshold as "outer points". Then, the process is iterated until the preset number of iterations is reached or a model with a sufficiently high proportion of interior points is found. Finally, all the identified interior points are used to refit the model using methods such as least squares, thereby outputting a final model. This allows for the extraction of line segment features from the point cloud and the identification of line segments with a length greater than a set threshold.

4. The vehicle pose estimation method based on point cloud features according to claim 3, characterized in that, In step S42, a certain point cloud P Whether a line segment conforms to the fitted model is determined by the following formula: in, It is a point on a line segment. It is a direction vector. It is the maximum distance between point clouds that satisfies the fitted line segment.

5. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, In step S5, the angle between the two line segments is calculated using the cosine formula for the vector angle. According to the angle between the two line segments Analyze and determine whether the two extracted line segments are parallel.

6. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Define the first region of interest in the point cloud. The first region of interest is a cuboid region. The length and width of the first region of interest are the length of the car directly in front of the AGV. The height of the first region of interest is 3cm-5cm. S12: Perform voxel filtering downsampling on the point cloud within the first region of interest. First, calculate a cuboid bounding box that can just enclose the entire point cloud data to create a voxel mesh. Then, divide the bounding box into countless small, fixed-volume cubic units, i.e., voxels, according to the set voxel size. Next, aggregate the points within the voxels. Each point in the point cloud will be classified into its specific voxel. At the same time, all the points contained within each voxel are "aggregated" and replaced with a representative point. Finally, a new point cloud is generated, and the representative points of all voxels together constitute the downsampled new point cloud.

7. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, In step S2, the specific steps for indexing the preprocessed point cloud data are as follows: The preprocessed point cloud is indexed by constructing a KD-Tree. First, a split dimension is selected, and then the median of all data points in the current dimension is selected as the split point. Finally, subtrees are recursively constructed. After determining the split dimension and split point of the current node, the remaining data points are divided into left and right subtrees. In the current split dimension, points with values ​​less than the split point are assigned to the left subtree, and points with values ​​greater than or equal to the split point are assigned to the right subtree. Then, the above steps are recursively performed on the left and right subsets until the subset contains only one point.

8. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, In step S2, the preprocessed point cloud has an approximately symmetrical "L" shape. The specific steps for segmenting the indexed point cloud data are as follows: S21: For the point cloud with this feature, the Euclidean clustering extraction algorithm is used. The algorithm traverses every point in the point cloud. If the point has not been processed, it is used as a seed point of a new cluster, and all its nearest neighbors within the specified Euclidean distance tolerance range are found. S22: Euclidean clustering extraction recursively marks these nearest neighbors as part of the current cluster and continues to grow regions using them as new seed points until the cluster can no longer be expanded; S23: The Euclidean clustering extraction algorithm will filter the cluster based on the preset minimum and maximum number of points, discarding invalid clusters with too many or too few points; S24: Save the clusters that meet the conditions as independent point cloud sets, and repeat the above process until all points in the point cloud have been processed, thereby dividing it into multiple independent point cloud clusters, each cluster representing a potential wheel.

9. The vehicle pose estimation method based on point cloud features according to claim 1, characterized in that, In step S3, the segmented point cloud data is in three-dimensional space, but the point cloud data has a relatively low height range along the z-axis. The three-dimensional point cloud data is then vertically projected onto the XOY plane.

10. The vehicle pose estimation method based on point cloud features according to claim 2, characterized in that, In step S81, the second region of interest is a cuboid region, the length and width of the second region of interest are the length of the car directly in front of the AGV, and the height of the second region of interest is greater than the height of the first region of interest.

Images