A road vehicle contour recognition method for 3D point cloud data

Abstract

The application provides a road vehicle contour identification method for 3D point cloud data, comprising the following steps: scanning a to-be-identified road vehicle scene by a scanning device to obtain non-uniformly distributed 3D point cloud data; and identifying the contour of the non-uniformly distributed 3D point cloud data based on an improved DBSCAN algorithm to determine the contour information of the to-be-identified road vehicle scene. The application solves the problem that the existing identification method has poor performance for non-uniformly distributed data by processing the non-uniformly distributed 3D point cloud obtained by scanning in a road scene based on the improved DBSCAN algorithm.

Description

Technical Field

[0001] This invention relates to the field of road vehicle contour recognition technology, and in particular to a road vehicle contour recognition method for 3D point cloud data. Background Technology

[0002] With the development of science and technology, research on autonomous unmanned intelligent vehicles has increasingly become a research hotspot for research institutions worldwide. Autonomous unmanned intelligent vehicles can effectively reduce traffic accident fatalities, complete operations in dangerous environments without human intervention, and significantly improve the level of intelligence in human life. Environmental perception technology is one of the core technologies of autonomous intelligent vehicles. LiDAR and cameras are currently the core environmental perception sensors in unmanned vehicle technology. Target recognition methods using 3D point cloud data are key technologies for improving the environmental perception, real-time localization and mapping (SLAM) technology, and autonomous navigation of autonomous intelligent vehicles.

[0003] However, existing recognition methods suffer from poor performance when processing 3D point cloud segmentation, particularly for data with uneven density, leading to a decline in recognition effectiveness. Therefore, there is an urgent need for a road vehicle contour recognition method for 3D point cloud data. This method would process the non-uniformly distributed 3D point cloud obtained from scanning in a road scene using an improved DBSCAN (including 3D local density calculation and 3D point cloud plane segmentation), thereby solving the problem of poor performance of existing recognition methods for data with uneven density. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a road vehicle contour recognition method for 3D point cloud data. By processing the non-uniformly distributed 3D point cloud obtained from scanning in a road scene using an improved DBSCAN (including 3D local density calculation and 3D point cloud plane segmentation), the method solves the problem of poor performance of existing recognition methods for data with non-uniform density.

[0005] A method for recognizing road vehicle contours based on 3D point cloud data includes: scanning a road scene to be recognized using a scanning device to obtain non-uniformly distributed 3D point cloud data; and performing contour recognition on the non-uniformly distributed 3D point cloud data based on an improved DBSCAN algorithm to determine the contour information of the road scene to be recognized.

[0006] As an embodiment of the present invention, the DBSCAN algorithm includes: 3D local density calculation and 3D point cloud plane segmentation.

[0007] As one embodiment of the present invention, contour recognition is performed on non-uniformly distributed 3D point cloud data based on an improved DBSCAN algorithm to determine the contour information of the road scene to be identified. This includes: querying the EPS neighborhood of n nodes in the non-uniformly distributed 3D point cloud data using a density detection method to determine the density variation coefficient (CV) of each node; segmenting the non-uniformly distributed 3D point cloud data based on the CV to obtain several segmented regions; setting parameters for different segmented regions according to predetermined parameters to obtain several parameter-set regions; clustering these parameter-set regions based on the improved DBSCAN algorithm to obtain several different planes; and performing a reasonable verification on each plane to determine the contour information of the road scene to be identified.

[0008] As an embodiment of the present invention, the density variation coefficient CV of each node is determined by querying the EPS neighborhood of n nodes in a non-uniformly distributed 3D point cloud data according to the density detection method, including the following steps: Step 1: Select any node i from the n nodes and calculate the density of node i according to the density formula; Step 2: Determine the spherical neighborhood of node i and calculate the density of other points in the spherical neighborhood according to the density calculation formula; Step 3: Calculate the average density of the spherical neighborhood based on the density of node i and the density of other points in the spherical neighborhood; Step 4: Calculate the density variance of the neighborhood of node i according to the variance formula based on the density of node i and the average density; Step 5: Calculate the density variation coefficient CV of node i according to the density of node i and the density variance of the neighborhood of node i according to the density variation coefficient formula.

[0009] As one embodiment of the present invention, region segmentation is performed on non-uniformly distributed 3D point cloud data based on CV to obtain several segmented regions, including: region segmentation is performed on non-uniformly distributed 3D point cloud data based on equal depth block method and CV to obtain several different 3D segmented regions.

[0010] As an embodiment of the present invention, parameters are set for different segmented regions according to predetermined parameters to obtain several parameter setting regions, including the following steps: Step A: Determine the value of MinPts based on the CV corresponding to each node; wherein the value of CV is inversely proportional to the value of MinPts; Step B: Calculate the number of nodes in the Eps neighborhood of each node based on the node number calculation formula; Step C: Calculate the radius of each node in each segmented region based on the Eps calculation formula according to the number of nodes, MinPts and the preset Eps of each segmented region; Step D: Calculate the expected value of the radius of each node in the segmented region to obtain the EPs of the segmented region; Step E: Set the corresponding EPs and MinPts parameters based on each segmented region to obtain several parameter setting regions.

[0011] As an embodiment of the present invention, a clustering process is performed on several parameter-defined regions based on the improved DBSCAN algorithm to obtain several different planes, including the following steps: Step 1.1: Obtain the core point and core neighboring points within any parameter-defined region, and select three non-collinear candidate points from the core point and core neighboring points to form a fitting plane; Step 1.2: Perform outlier detection and maximum distance detection on the fitting plane; Step 1.3: After the fitting plane in Step 1.2 has been tested, determine the coplanarity of all other nodes within the current parameter-defined region according to the coplanarity condition, and identify coplanar points; Step 1.4: Perform clustering processing on all coplanar points and the core point in Step 1.1 to obtain the plane corresponding to the current parameter-defined region.

[0012] As an embodiment of the present invention, the reasonableness test is performed on each plane to determine the contour information of the road scene to be identified, including the following steps: Step 2.1: Select any plane as a reference plane, and calculate the similarity Sim between other planes and the current plane in turn based on the similarity formula; Step 2.2: Classify all reference planes corresponding to Sim based on the preset Sim interval to obtain the classification results; wherein, the classification results include abnormal, reasonable and transition sets; Step 2.3: After the classification is completed, repeat steps 2.1 and 2.2 in each set where the classification result is a transition set, but do not create new sets, and only retain the two results of abnormal and reasonable; Step 2.4: After calculating all planes, select the reference plane with the highest proportion of reasonable values ​​among all reference planes as the fitting plane of the set; Step 2.5: Combine the fitting planes of each set pair to form the final clustering result as the contour information of the road scene to be identified.

[0013] The beneficial effects of this invention are as follows:

[0014] This invention provides a road vehicle contour recognition method for 3D point cloud data. By processing the non-uniformly distributed 3D point cloud obtained by scanning in a road scene according to the improved DBSCAN (including 3D local density calculation and 3D point cloud plane segmentation), the method solves the problem that the existing recognition methods perform poorly on data with non-uniform density.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of a road vehicle contour recognition method based on 3D point cloud data, as described in an embodiment of the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] Please see Figure 1 This invention provides a method for recognizing road vehicle contours based on 3D point cloud data, comprising: scanning a road scene to be recognized using a scanning device to obtain non-uniformly distributed 3D point cloud data; and performing contour recognition on the non-uniformly distributed 3D point cloud data based on an improved DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm to determine the contour information of the road scene to be recognized.

[0021] The beneficial effects of the above technical solution are as follows: By using the above technical solution, the non-uniformly distributed 3D point cloud obtained by scanning in the road scene is processed according to the improved DBSCAN (including 3D local density calculation and 3D point cloud plane segmentation), thereby solving the problem that the existing recognition methods have poor performance for data with non-uniform density.

[0022] In one embodiment, the DBSCAN algorithm includes: 3D local density calculation and 3D point cloud plane segmentation.

[0023] In one embodiment, contour recognition is performed on non-uniformly distributed 3D point cloud data using an improved DBSCAN algorithm to determine the contour information of the road scene to be identified. This includes: querying the EPS neighborhood of n nodes in the non-uniformly distributed 3D point cloud data using a density detection method to determine the density variation coefficient (CV) of each node; segmenting the non-uniformly distributed 3D point cloud data based on the CV to obtain several segmented regions; setting parameters for different segmented regions according to predetermined parameters to obtain several parameter-set regions; clustering these parameter-set regions using the improved DBSCAN algorithm to obtain several different planes; and performing a reasonable verification on each plane to determine the contour information of the road scene to be identified. Here, the node refers to a data point constituting the 3D point cloud.

[0024] In one embodiment, the density coefficient of variation (CV) of each node is determined by querying the EPS neighborhood of n nodes in a non-uniformly distributed 3D point cloud using a density detection method, including the following steps: Step 1: Select any node i from the n nodes and calculate the density of node i using the density formula; Step 2: Determine the spherical neighborhood of node i and calculate the density of other points in the spherical neighborhood using the density calculation formula; Step 3: Calculate the average density of the spherical neighborhood based on the density of node i and the density of other points in the spherical neighborhood; Step 4: Calculate the density variance of the neighborhood of node i using the variance formula based on the density of node i and the average density; Step 5: Calculate the density coefficient of variation (CV) of node i using the density coefficient of variation formula based on the density coefficient of variation of node i and the density variance of the neighborhood of node i.

[0025] In one embodiment, non-uniformly distributed 3D point cloud data is segmented into several segmented regions based on CV, including: segmenting non-uniformly distributed 3D point cloud data into several different 3D segmented regions based on equal depth block method and CV.

[0026] In one embodiment, parameters are set for different segmented regions according to predetermined parameters to obtain several parameter setting regions, including the following steps: Step A: Determine the value of MinPts based on the CV corresponding to each node; wherein the value of CV is inversely proportional to the value of MinPts; Step B: Calculate the number of nodes in the Eps neighborhood of each node based on the node number calculation formula; Step C: Calculate the radius of each node in each segmented region based on the Eps calculation formula according to the number of nodes, MinPts, and the preset Eps of each segmented region; Step D: Calculate the expected value of the radius of each node in the segmented region to obtain the EPs of the segmented region; Step E: Set the corresponding EPs and MinPts parameters based on each segmented region to obtain several parameter setting regions.

[0027] In one embodiment, a clustering process is performed on several parameter-defined regions based on an improved DBSCAN algorithm to obtain several different planes, including the following steps: Step 1.1: Obtain the core point and core neighboring points within any parameter-defined region, and select three non-collinear candidate points from the core point and core neighboring points to form a fitting plane; Step 1.2: Perform outlier detection and maximum distance detection on the fitting plane; Step 1.3: After the fitting plane in Step 1.2 has been tested, determine the coplanarity of all other nodes within the current parameter-defined region according to the coplanarity condition, and identify coplanar points; Step 1.4: Perform clustering processing on all coplanar points and the core point in Step 1.1 to obtain the plane corresponding to the current parameter-defined region.

[0028] In one embodiment, a reasonableness check is performed on each plane to determine the contour information of the road scene to be identified, including the following steps: Step 2.1: Select any plane as a reference plane, and calculate the similarity Sim between the current plane and other planes in turn based on the similarity formula; Step 2.2: Classify all reference planes corresponding to Sim based on the preset Sim interval to obtain classification results; wherein, the classification results include abnormal, reasonable, and transition sets; Step 2.3: After classification, repeat steps 2.1 and 2.2 in each set where the classification result is a transition set, but do not create new sets, and only retain the two results of abnormal and reasonable; Step 2.4: After calculating all planes, select the reference plane with the highest proportion of reasonable values ​​among all reference planes as the fitting plane of the set; Step 2.5: Combine the fitting planes of each set pair to form the final clustering result as the contour information of the road scene to be identified.

[0029] In one specific embodiment, contour recognition is performed on non-uniformly distributed 3D point cloud data based on the improved DBSCAN algorithm to determine the contour information of the road scene to be identified, including: Step 1, selecting a node i and calculating the corresponding density according to the density formula, as follows:

[0030]

[0031] Where Eps is the computational radius of node i, and its size can be determined by testing in actual application scenarios. Pts(i) is the set of points centered at i with a radius of Eps, and |Pts(i)| is the number of elements in the set of points.

[0032] Step 2, similarly, represents and calculates the density of other points in the spherical neighborhood of node i. Taking node k as an example, the specific details are as follows:

[0033] ρ k =get ρ (Pts k ),Pts k ∈Pts(i)

[0034] Among them, Pts k ρ is a point in the set Pts(i). k Pts k The density of the Eps neighborhood, since the calculation formula is the same as the calculation of ρ in the above formula. i The formula is the same, so get is used for simplification. ρ To make a substitution.

[0035] Step 3: Based on the results of Steps 1 and 2, calculate the average density, as follows:

[0036]

[0037] Step 4: Based on the results of Steps 1 and 3, calculate the density variance of the neighborhood of node i (representing the dispersion of the data and reflecting the degree of deviation of each point in the Eps neighborhood from the mean), as follows:

[0038]

[0039] Step 5: Based on the results of the above steps, calculate the density variation coefficient, which represents the density change of node i in its Eps neighborhood. The more points there are, the smaller the value (s is the variance; as the density increases, the dispersion decreases, thus the overall CV decreases), as detailed below:

[0040]

[0041] Step 6: Repeat the above steps to obtain the CV value corresponding to each node. Based on the equal-depth block method, the data is segmented to obtain multiple different 3D partitions, and Eps is set for each partition. α The MinPts value is as follows:

[0042] Step 6-1: First, determine the value of MinPts based on cv. A smaller cv indicates more nearby points, so MinPts can be set larger; a larger cv indicates more nearby points, so MinPts can be set smaller. The specific value needs to be set according to the actual application conditions such as the acquisition device and acquisition density.

[0043] Step 6-2: Calculate the number of nodes in the Eps neighborhood of each point, as follows, where ρi can be obtained from step 1:

[0044]

[0045] Step 6-3, calculate EPS for each point within the partition, as follows:

[0046]

[0047] Step 6-4: Calculate the expected value of Eps(i) to obtain the Eps of this partition. α The details are as follows:

[0048]

[0049] Step 7: For each partition, perform 3D point cloud plane segmentation, as follows:

[0050] Step 7-1: Select three non-collinear candidate points (with similar normal directions) from the core point and its neighboring points to form a fitting plane. Core point: A point is considered a core point if the density of points around it reaches a set threshold, meaning the number of points in its neighborhood is no less than minPts. Neighboring points: Points that are close to the selected core point and also meet the core point requirements.

[0051] Step 7-2: Perform outlier testing and maximum distance testing on the selected fitting plane. Outlier testing: Define the concept of the observed value D (the distance from the core point and its adjacent points to the fitting plane). Outlier values ​​greater than the mean of the observed values ​​are considered outliers. Points summing to the standard deviation σ are considered outliers. If more than 25% of the samples are outliers, the plane is discarded. Maximum distance test: Define the concept of maximum distance MD (the median of the average distances between the three nearest neighbors of each core point). If the mean of the observations is greater than the maximum distance, the plane is discarded, as follows:

[0052]

[0053]

[0054] Where D is the set of observations, and the elements d in this set are... i That is, the distance from the core point and its adjacent points to the fitted plane; since three points are used to construct the plane, the number of points available for calculation is n-3, meaning the set contains n-3 elements, and d i The calculation method is the formula described below; specifically, it is explained as follows: point (x i ,y i ,z i The first line represents the distance from the plane Ax + By + Cz + D = 0; the second line is the formula for calculating MD, where median refers to taking the median, and c... i This refers to the i-th core point (see the text description in step 7-1 for the definition of a core point), p ij It refers to c i The j-th nearest neighbor (i.e., j can only take the values ​​1, 2, and 3), ||c i -p ij || refers to the Euclidean distance between points ci and pij;

[0055] Step 7-3: After completing the verification, subsequent judgments can be made on all other points in the partition based on the coplanarity condition. When the distance from a point to the plane is less than the coplanarity threshold δ, the data point is considered coplanar with the fitted plane. Specifically:

[0056]

[0057]

[0058] Where δ is the coplanarity threshold, used to determine whether a point is coplanar with the fitted plane (i.e., whether the point can be clustered into the fitted plane). σ is the mean of the observed values ​​(see the text description in step 7-2 for the concept of observed values). Since the set has n-3 elements (see the text description in step 7-2 for the reason), the mean should be calculated according to the number of n-3. σ is the standard deviation, which is also calculated according to the number of n-3.

[0059] It should be noted that since no sampling method was used, the population standard deviation was used here;

[0060] Step 7-4: Cluster all coplanar points (for valid fitted planes) and density-reachable points (for SEED points) into one class; SEED points: c in step 7-2 i Density reachable point: The density reachable point is the point that the density can directly reach. Density reachability: If a point p is in the neighborhood of q, and q is the core point, then p is the density reachable point of q.

[0061] If a point belongs to the same cluster as the SEED point (i.e., can be clustered with the SEED point), then two conditions must be met: (1) the point is a density reachable by the SEED point; (2) the distance between the point and the fitted plane (determined by the test group containing the SEED point) is less than the coplanar threshold.

[0062] Step 8: After clustering is complete, a planar set has been formed. Anomaly detection is performed on different clustering results, as follows:

[0063] Step 8-1, select a plane Using plane γ as a reference plane, the similarity degree Sim with each plane is calculated sequentially. Taking plane γ as an example, the specific details are as follows:

[0064] A1x+B1y+C1z+D1=0,γ:A2x+B2y+C2z+D2=0

[0065]

[0066] in, γ and γ represent two different planes. This indicates that γ and the reference plane The degree of similarity, i.e., the cosine of the angle between the two planes. It is a plane The angle between γ and (take the smaller value) ), and the rest A, B, C represent constants.

[0067] Step 8-2 involves classifying all obtained Sims and processing them according to the interval in which the Sims are located. There are three cases: abnormal, reasonable, and transition set. Classifying all obtained Sims includes classifying the plane corresponding to the obtained Sim value based on the Sim value (since there are two planes, a reference plane is used). The reference plane is stationary; we are classifying another γ-plane. The appropriate processing involves dividing the plane into one of three categories (abnormal, reasonable, and transition set). Transition set: A new set P is created, and planes different from the reference plane are transferred to this set. The upper limit of the new transition set created using this method needs to be determined based on the specific scenario. Again, taking plane γ as an example, the details are as follows:

[0068]

[0069] Step 8-3: After completing the classification, repeat steps 8-1 and 8-2 within each set, but do not create new sets; only retain the outlier and reasonable results. After calculating all planes, select the reference plane with the highest proportion of reasonable values ​​as the fitting plane for that set. Specifically:

[0070]

[0071] Step 9: After each set selects its corresponding fitting plane, they are combined to form the final clustering result, which is the approximate contour recognition result.

[0072] This invention improves upon the shortcomings of existing methods in processing 3D point cloud segmentation, including but not limited to the inability to identify boundaries, poor performance for data with uneven density, the need for extensive pre-training and computation, and the generation of a large number of outliers.

[0073] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for road vehicle contour recognition based on 3D point cloud data, characterized in that, include: The scanning device scans the road vehicle scene to be identified, resulting in non-uniformly distributed 3D point cloud data. Based on the improved DBSCAN algorithm, contour recognition is performed on non-uniformly distributed 3D point cloud data to determine the contour information of the road vehicle scene to be identified. This paper uses an improved DBSCAN algorithm to perform contour recognition on non-uniformly distributed 3D point cloud data to determine the contour information of the road vehicle scene to be identified. The process includes: querying the EPS neighborhood of n nodes in the non-uniformly distributed 3D point cloud data using a density detection method to determine the coefficient of variation (CV) of each node; segmenting the non-uniformly distributed 3D point cloud data based on the CV to obtain several segmented regions; setting parameters for different segmented regions according to predetermined parameters to obtain several parameter-set regions; clustering these parameter-set regions using the improved DBSCAN algorithm to obtain several different planes; and performing a reasonable verification on each plane to determine the contour information of the road vehicle scene to be identified.

2. The method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, The DBSCAN algorithm includes: 3D local density calculation and 3D point cloud plane segmentation.

3. The method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, The following steps are used to query the EPS neighborhood of n nodes in a non-uniformly distributed 3D point cloud data using a density detection method, and determine the density variation coefficient (CV) of each node: Step 1: Select any node i from the n nodes and calculate the density of node i using the density formula; Step 2: Determine the spherical neighborhood of node i and calculate the density of other points in the spherical neighborhood using the density calculation formula; Step 3: Calculate the average density of the spherical neighborhood based on the density of node i and the density of other points in the spherical neighborhood; Step 4: Calculate the density variance of the neighborhood of node i using the variance formula based on the density of node i and the average density; Step 5: Calculate the density variation coefficient (CV) of node i using the density variation coefficient formula based on the density of node i and the density variance of the neighborhood of node i.

4. The method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, Based on CV, region segmentation is performed on non-uniformly distributed 3D point cloud data to obtain several segmented regions, including: region segmentation of non-uniformly distributed 3D point cloud data based on equal depth block method and CV to obtain several different 3D segmented regions.

5. A method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, The parameters for different segmentation regions are set according to predetermined parameters to obtain several parameter setting regions, including the following steps: Step A: Determine the value of MinPts based on the CV corresponding to each node; where the value of CV is inversely proportional to the value of MinPts; Step B: Calculate the number of nodes in the Eps neighborhood of each node based on the node number calculation formula; Step C: Calculate the radius of each node in each segmentation region based on the Eps calculation formula, according to the number of nodes, MinPts, and the preset Eps of each segmentation region; Step D: Calculate the expected value of the radius of each node in the segmentation region to obtain the EPs of the segmentation region; Step E: Set the corresponding EPs and MinPts parameters for each segmentation region to obtain several parameter setting regions.

6. The method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, The improved DBSCAN algorithm is used to cluster several parameter-defined regions to obtain several different planes, including the following steps: Step 1.1: Obtain the core point and core neighbor points within any parameter-defined region, and select three non-collinear candidate points from the core point and core neighbor points to form a fitting plane; Step 1.2: Perform outlier detection and maximum distance detection on the fitting plane; Step 1.3: After the fitting plane in Step 1.2 has been tested, determine the coplanarity of all other nodes within the current parameter-defined region according to the coplanarity condition, and identify coplanar points; Step 1.4: Perform clustering on all coplanar points and the core point in Step 1.1 to obtain the plane corresponding to the current parameter-defined region.

7. A method for road vehicle contour recognition based on 3D point cloud data according to claim 1, characterized in that, The process involves several steps to determine the contour information of the road vehicle scene to be identified, including: Step 2.1: Select any plane as a reference plane and calculate the similarity Sim between the current plane and other planes based on the similarity formula; Step 2.2: Classify all reference planes corresponding to Sim based on the preset Sim interval to obtain classification results; the classification results include abnormal, reasonable, and transition sets; Step 2.3: After classification, repeat steps 2.1 and 2.2 within each set where the classification result is a transition set, but do not create new sets, retaining only the abnormal and reasonable results; Step 2.4: After calculating all planes, select the reference plane with the highest proportion of reasonable values ​​among all reference planes as the fitting plane for that set; Step 2.5: Combine the fitting planes of each set pair to form the final clustering result as the contour information of the road vehicle scene to be identified.

Images