A target detection method and system based on point cloud segmentation decoupling high-precision map and a medium

Abstract

The application relates to a target detection method and system for decoupling a high-precision map based on point cloud segmentation and a medium, the method comprising the following steps: M1. A vehicle travels on a road, point cloud data information of the road is obtained in real time based on a vehicle-mounted laser radar group, and the point clouds of each laser radar are spliced through calibration parameters to obtain spliced road point cloud data information; M2. Based on the spliced road point cloud data information, a modified Monte Carlo integration algorithm is used to randomly sample the road point cloud to obtain prototype points and structured feature data information of the road point cloud. The application not only improves the accuracy, stability and practicability of target detection, but also reduces the dependence on a high-precision map, decouples the high-precision map through point cloud segmentation, reduces the dependence on the high-precision map, and improves the robustness of an automatic driving system.

Description

Technical Field

[0001] This invention relates to the field of lidar detection technology, and in particular to a target detection method, system and medium based on point cloud segmentation and decoupling of high-precision maps. Background Technology

[0002] High-precision maps typically refer to maps used for autonomous driving assistance. Map elements include detailed lane lines, road signs, traffic signs, traffic lights, lane curvature, slope, and real-time lane-level traffic dynamics. High-precision maps have three main characteristics: first, high precision, reaching centimeter-level accuracy, and being 3D images; second, high timeliness, requiring frequent updates to map information, including dynamic information such as real-time traffic flow and traffic light status, as well as static information such as road changes and maintenance; and third, high richness, including information on road shapes, traffic signs, road markings, and surrounding obstacles. The creation of high-precision maps requires a professional surveying fleet and skilled surveyors who use imagery and sensors such as LiDAR and GPS to conduct detailed surveys of road information and maintain continuous updates.

[0003] In current technology, the creation of high-definition maps is an extremely costly project, and not all automakers possess the qualifications to do so. This is unacceptable for autonomous driving companies, as the update speed of high-definition maps is too slow to keep pace with the development of autonomous driving. Due to the need to collect a large amount of extremely precise data, high-definition maps are generally updated quarterly, while automakers ideally want daily updates. Summary of the Invention

[0004] In view of the above problems, the present invention provides a target detection method, system and medium based on point cloud segmentation and decoupling of high-precision maps. It not only improves the accuracy, stability and practicality of target detection, but also reduces the dependence on high-precision maps. By decoupling high-precision maps through point cloud segmentation, the dependence on high-precision maps is reduced, and the robustness of autonomous driving systems is improved.

[0005] To achieve the above and other related objectives, the present invention provides the following technical solution:

[0006] A target detection method based on point cloud segmentation and decoupling of high-precision maps, the method comprising:

[0007] M1. As the vehicle travels on the road, it acquires point cloud data of the road in real time based on the onboard LiDAR group, and stitches together the point clouds of each LiDAR through calibration parameters to obtain stitched road point cloud data.

[0008] M2. Based on the stitched road point cloud data, an improved Monte Carlo integral algorithm is used to randomly sample the road point cloud to obtain the prototype points and structured feature data information of the road point cloud;

[0009] M3. Input the prototype points and structured feature data of the road point cloud into the trained road point cloud semantic segmentation model for information extraction, and output the semantic data information of the road point cloud;

[0010] M4. Based on the semantic data information of the road point cloud, a clustering algorithm that integrates the features of the angle between normal vectors, spatial location, and semantic information is used to perform clustering analysis on the road targets to obtain the three-dimensional spatial detection data information of the road targets.

[0011] Furthermore, in step M1, the step of stitching together the point clouds of each lidar unit using calibration parameters includes:

[0012] M11. Obtain the calibration parameter data of each lidar in the vehicle-mounted lidar group, and establish the lidar transformation function G.

[0013]

[0014] Where X represents the calibration parameter data of the lidar, and α and β are transformation matrices. The calibration parameters of the lidar are transformed in the same way to obtain the calibration parameter data of the lidar in the same coordinate system.

[0015] M12. Based on the lidar calibration parameter data information in the same coordinate system, the Halcon algorithm is used to stitch together the lidar point clouds at the same time to obtain stitched road point cloud data information.

[0016] Furthermore, the Halcon algorithm is used to stitch together the point clouds of the lidar at the same time to construct a rotation matrix Q based on the feature points of the lidar point clouds.

[0017]

[0018] Where θ is the angle between the point cloud normal vectors of any two lidars, the point clouds are stitched together to obtain the stitched road point cloud data information.

[0019] Furthermore, in step M2, the random sampling of the road point cloud using the improved Monte Carlo integration algorithm includes:

[0020] M21. Based on the stitched road point cloud data, establish a Monte Carlo integral function H.

[0021] H=∫ Ω ρh(x)π(x)dx,

[0022] Where ρ is the integral constant factor, Ω is the sample domain of the road point cloud, h(x) is the distribution function of the road point cloud, π(x) is the distribution function of the road point cloud, and x is the stitched road point cloud data. The distribution of the road point cloud is analyzed to obtain the Monte Carlo integral distribution data of the road point cloud.

[0023] M22. Based on the Monte Carlo integral distribution data of the road point cloud, establish the probability function O of the road point cloud.

[0024]

[0025] Wherein, λ represents the Monte Carlo integral distribution data of the road point cloud, and the probability values ​​of the road point cloud are calculated to obtain the probability value data of the road point cloud.

[0026] M23. Based on the probability value data of the road point cloud, establish a sampling function f for the road point cloud.

[0027]

[0028] Where, η i Here, n is the sampling constant factor for the road point cloud, and w is the sample size. i To obtain the probability value data of the road point cloud, the points in the road point cloud are sampled to obtain the prototype points and structured feature data of the road point cloud.

[0029] Furthermore, the prototype points and structured feature data information of the road point cloud are obtained by sampling the road point cloud to obtain the prototype points, and a structured feature function J of the road point cloud is established.

[0030]

[0031] Where, (ω x ,ω y ,ω z The prototype point coordinate data of the road point cloud is used to extract the structured features of the road point cloud, thus obtaining the structured feature data of the road point cloud.

[0032] Furthermore, in step M3, the trained road point cloud semantic segmentation model includes:

[0033] M31. Based on the structured feature data information of the road point cloud, obtain the semantic label data information, embedded feature data information, and prototype prediction point data information of the road point cloud;

[0034] M32. Based on the prototype prediction point data information and embedded feature data information, establish an uncertainty estimation function L for the road point cloud.

[0035]

[0036] Where r1 is the prototype prediction point data information and r2 is the embedded feature data information, the prototype correction points and covariance matrix data information of the road point cloud are obtained.

[0037] M33. Based on the prototype correction points and covariance matrix data of the road point cloud, the average probability matrix of the road point cloud is obtained, instance labels are assigned to the points in the point cloud data, and point cloud semantic labels are assigned to the points in the point cloud data to obtain the trained road point cloud semantic segmentation model.

[0038] Furthermore, the average probability matrix of the road point cloud is A.

[0039]

[0040] Where N is the number of covariance matrices in the road point cloud, δ j a is the weighting coefficient. j This provides the covariance matrix data information for road point clouds.

[0041] Furthermore, the clustering algorithm that uses the fusion of normal vector angle features, spatial location, and semantic information to perform clustering analysis on road targets involves clustering road targets, with adjacent targets of the same category being clustered into one target. Based on the fusion of normal vector angle features, spatial location, and semantic information of the road point cloud, road foreground instances are segmented to obtain the three-dimensional spatial detection data information of the road targets.

[0042] To achieve the above and other related objectives, the present invention also provides a target detection system based on point cloud segmentation and decoupled high-precision maps, including a computer device programmed or configured to perform the steps of any of the target detection methods based on point cloud segmentation and decoupled high-precision maps described in the present invention.

[0043] To achieve the above and other related objectives, the present invention also provides a computer-readable storage medium storing a computer program programmed or configured to perform any of the target detection methods based on point cloud segmentation and decoupled high-precision maps as described in any one of the claims.

[0044] The present invention has the following positive effects:

[0045] 1. This invention stitches together the point clouds of each lidar by calibrating parameters to obtain stitched road point cloud data. It also uses an improved Monte Carlo integration algorithm to randomly sample the road point cloud data. This not only enables comprehensive analysis of the road point cloud data, but also reduces information redundancy. Decoupling the high-precision map can eliminate unnecessary information, reduce information redundancy, and save storage and transmission costs.

[0046] 2. This invention extracts information from a trained road point cloud semantic segmentation model, outputting semantic data information of the road point cloud. It then combines this with a clustering algorithm that integrates normal vector angle features, spatial location, and semantic information to perform clustering analysis on road targets, obtaining three-dimensional spatial detection data information of the road targets. This not only reduces reliance on high-precision maps by decoupling them through point cloud segmentation, thus improving the robustness of the autonomous driving system, but also enhances target detection efficiency and accuracy. Target detection based on decoupled high-precision maps makes target detection more efficient and accurate. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0048] Figure 2 This is a flowchart illustrating the improved Monte Carlo integration algorithm of the present invention.

[0049] Figure 3 This is a schematic diagram of the structural framework of the trained road point cloud semantic segmentation model of the present invention;

[0050] Figure 4 This is a schematic diagram of the target detection results of the present invention. Detailed Implementation

[0051] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0052] Example 1: As Figure 1 As shown, a target detection method based on point cloud segmentation and decoupling of high-precision maps is proposed, the method comprising:

[0053] M1. As the vehicle travels on the road, it acquires point cloud data of the road in real time based on the onboard LiDAR group, and stitches together the point clouds of each LiDAR through calibration parameters to obtain stitched road point cloud data.

[0054] M2. Based on the stitched road point cloud data, an improved Monte Carlo integral algorithm is used to randomly sample the road point cloud to obtain the prototype points and structured feature data information of the road point cloud;

[0055] M3. Input the prototype points and structured feature data of the road point cloud into the trained road point cloud semantic segmentation model for information extraction, and output the semantic data information of the road point cloud;

[0056] M4. Based on the semantic data information of the road point cloud, a clustering algorithm that integrates the features of the angle between normal vectors, spatial location, and semantic information is used to perform clustering analysis on the road targets to obtain the three-dimensional spatial detection data information of the road targets.

[0057] In this embodiment, step M1, which involves stitching together the point clouds of each lidar unit using calibration parameters, includes:

[0058] M11. Obtain the calibration parameter data of each lidar in the vehicle-mounted lidar group, and establish the lidar transformation function G.

[0059]

[0060] Where X represents the calibration parameter data of the lidar, and α and β are transformation matrices. The calibration parameters of the lidar are transformed in the same way to obtain the calibration parameter data of the lidar in the same coordinate system.

[0061] M12. Based on the lidar calibration parameter data information in the same coordinate system, the Halcon algorithm is used to stitch together the lidar point clouds at the same time to obtain stitched road point cloud data information.

[0062] In this embodiment, the step of stitching together the point clouds of the lidar at the same time using the Halcon algorithm involves constructing a rotation matrix Q based on the feature points of the lidar point clouds.

[0063]

[0064] Where θ is the angle between the point cloud normal vectors of any two lidars, the point clouds are stitched together to obtain the stitched road point cloud data information.

[0065] In this embodiment, as Figure 2 As shown, in step M2, the random sampling of the road point cloud using the improved Monte Carlo integration algorithm includes:

[0066] M21. Based on the stitched road point cloud data, establish a Monte Carlo integral function H.

[0067] H=∫ Ω ρh(x)π(x)dx,

[0068] Where ρ is the integral constant factor, Ω is the sample domain of the road point cloud, h(x) is the distribution function of the road point cloud, π(x) is the distribution function of the road point cloud, and x is the stitched road point cloud data. The distribution of the road point cloud is analyzed to obtain the Monte Carlo integral distribution data of the road point cloud.

[0069] M22. Based on the Monte Carlo integral distribution data of the road point cloud, establish the probability function O of the road point cloud.

[0070]

[0071] Wherein, λ represents the Monte Carlo integral distribution data of the road point cloud, and the probability values ​​of the road point cloud are calculated to obtain the probability value data of the road point cloud.

[0072] M23. Based on the probability value data of the road point cloud, establish a sampling function f for the road point cloud.

[0073]

[0074] Where, η i Here, n is the sampling constant factor for the road point cloud, and w is the sample size. i To obtain the probability value data of the road point cloud, the points in the road point cloud are sampled to obtain the prototype points and structured feature data of the road point cloud.

[0075] In this embodiment, the prototype points and structured feature data information of the road point cloud are obtained by sampling the road point cloud to obtain the prototype points, and a structured feature function J of the road point cloud is established.

[0076]

[0077] Where, (ω x ,ω y ,ω z The prototype point coordinate data of the road point cloud is used to extract the structured features of the road point cloud, thus obtaining the structured feature data of the road point cloud.

[0078] Example 2: Based on the target detection method of high-precision map based on point cloud segmentation decoupling in Example 1, the present invention will be further explained and described below.

[0079] like Figure 1 As shown, a target detection method based on point cloud segmentation and decoupling of high-precision maps is proposed, the method comprising:

[0080] M1. As the vehicle travels on the road, it acquires point cloud data of the road in real time based on the onboard LiDAR group, and stitches together the point clouds of each LiDAR through calibration parameters to obtain stitched road point cloud data.

[0081] M2. Based on the stitched road point cloud data, an improved Monte Carlo integral algorithm is used to randomly sample the road point cloud to obtain the prototype points and structured feature data information of the road point cloud;

[0082] M3. Input the prototype points and structured feature data of the road point cloud into the trained road point cloud semantic segmentation model for information extraction, and output the semantic data information of the road point cloud;

[0083] M4. Based on the semantic data information of the road point cloud, a clustering algorithm that integrates the features of the angle between normal vectors, spatial location, and semantic information is used to perform clustering analysis on the road targets to obtain the three-dimensional spatial detection data information of the road targets.

[0084] In this embodiment, as Figure 3 As shown, in step M3, the trained road point cloud semantic segmentation model includes:

[0085] M31. Based on the structured feature data information of the road point cloud, obtain the semantic label data information, embedded feature data information, and prototype prediction point data information of the road point cloud;

[0086] M32. Based on the prototype prediction point data information and embedded feature data information, establish an uncertainty estimation function L for the road point cloud.

[0087]

[0088] Where r1 is the prototype prediction point data information and r2 is the embedded feature data information, the prototype correction points and covariance matrix data information of the road point cloud are obtained.

[0089] M33. Based on the prototype correction points and covariance matrix data of the road point cloud, the average probability matrix of the road point cloud is obtained, instance labels are assigned to the points in the point cloud data, and point cloud semantic labels are assigned to the points in the point cloud data to obtain the trained road point cloud semantic segmentation model.

[0090] In this embodiment, the average probability matrix of the road point cloud is A.

[0091]

[0092] Where N is the number of covariance matrices in the road point cloud, δ j a is the weighting coefficient. j This provides the covariance matrix data information for road point clouds.

[0093] In this embodiment, the clustering algorithm that uses the fusion of normal vector angle features, spatial location, and semantic information to perform clustering analysis on road targets is to cluster the road targets, with adjacent targets of the same category being clustered into one target. Based on the fusion of normal vector angle features, spatial location, and semantic information of the road point cloud, road foreground instances are segmented to obtain the three-dimensional spatial detection data information of the road targets.

[0094] In this embodiment, the present invention provides a target detection system based on point cloud segmentation and decoupled high-precision maps, including a computer device that is programmed or configured to perform the steps of any of the target detection methods based on point cloud segmentation and decoupled high-precision maps described in the present invention.

[0095] In this embodiment, the present invention also provides a computer-readable storage medium storing a computer program programmed or configured to perform any of the target detection methods based on point cloud segmentation and decoupled high-precision maps as described above.

[0096] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0097] In summary, this invention not only improves the accuracy, stability, and practicality of target detection, but also reduces the reliance on high-precision maps. By decoupling high-precision maps through point cloud segmentation, it reduces dependence on high-precision maps and improves the robustness of autonomous driving systems.

[0098] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A target detection method based on point cloud segmentation and decoupling of high-precision maps, characterized in that, The method includes: M1. As the vehicle travels on the road, it acquires point cloud data of the road in real time based on the onboard LiDAR group, and stitches together the point clouds of each LiDAR through calibration parameters to obtain stitched road point cloud data. M2. Based on the stitched road point cloud data, an improved Monte Carlo integral algorithm is used to randomly sample the road point cloud to obtain the prototype points and structured feature data information of the road point cloud; M3. Input the prototype points and structured feature data of the road point cloud into the trained road point cloud semantic segmentation model for information extraction, and output the semantic data information of the road point cloud; M4. Based on the semantic data information of the road point cloud, a clustering algorithm that integrates the features of the angle between the normal vectors, spatial location, and semantic information is used to perform clustering analysis on the road targets to obtain the three-dimensional spatial detection data information of the road targets; In step M2, the random sampling of the road point cloud using the improved Monte Carlo integration algorithm includes: M21. Based on the stitched road point cloud data, establish a Monte Carlo integral function H. ， Where ρ is the integral constant factor, Ω is the sample domain of the road point cloud, h(x) is the distribution function of the road point cloud, π(x) is the distribution function of the road point cloud, and x is the stitched road point cloud data. The distribution of the road point cloud is analyzed to obtain the Monte Carlo integral distribution data of the road point cloud. M22. Based on the Monte Carlo integral distribution data of the road point cloud, establish the probability function O of the road point cloud. , Wherein, λ represents the Monte Carlo integral distribution data of the road point cloud, and the probability values ​​of the road point cloud are calculated to obtain the probability value data of the road point cloud. M23. Based on the probability value data of the road point cloud, establish a sampling function f for the road point cloud. ， wherein, η i is a sampling constant factor of the road point cloud, n is a sample capacity, w i is probability value data information of the road point cloud, the road point cloud is sampled to obtain prototype points and structured feature data information of the road point cloud; In step M3, the trained road point cloud semantic segmentation model includes: M31. Based on the structured feature data information of the road point cloud, obtain the semantic label data information, embedded feature data information, and prototype prediction point data information of the road point cloud; M32. Based on the prototype prediction point data information and embedded feature data information, establish an uncertainty estimation function L for the road point cloud. ， Where r1 is the prototype prediction point data information and r2 is the embedded feature data information, the prototype correction points and covariance matrix data information of the road point cloud are obtained. M33. Based on the prototype correction points and covariance matrix data of the road point cloud, the average probability matrix of the road point cloud is obtained, instance labels are assigned to the points in the point cloud data, and point cloud semantic labels are assigned to the points in the point cloud data to obtain the trained road point cloud semantic segmentation model.

2. The target detection method based on point cloud segmentation and decoupling high-precision map according to claim 1, characterized in that, In step M1, stitching together the point clouds of each lidar unit using calibration parameters includes: M11. Obtain the calibration parameter data of each lidar in the vehicle-mounted lidar group, and establish the lidar transformation function G. ， Where X represents the calibration parameter data of the lidar, and α and β are transformation matrices. The calibration parameters of the lidar are transformed in the same way to obtain the calibration parameter data of the lidar in the same coordinate system. M12. Based on the lidar calibration parameter data information in the same coordinate system, the Halcon algorithm is used to stitch together the lidar point clouds at the same time to obtain stitched road point cloud data information.

3. The target detection method based on point cloud segmentation and decoupling high-precision map according to claim 2, characterized in that: The method of stitching together the point clouds of the lidar at the same time using the Halcon algorithm involves constructing a rotation matrix Q based on the feature points of the lidar point clouds. ， Where θ is the angle between the point cloud normal vectors of any two lidars, the point clouds are stitched together to obtain the stitched road point cloud data information.

4. The target detection method based on point cloud segmentation and decoupling of high-precision maps according to claim 1, characterized in that: The prototype points and structured feature data of the road point cloud are obtained by sampling the road point cloud to obtain its prototype points, and a structured feature function J of the road point cloud is established. ， Wherein, (ω x ,ω y ,ω z ) is the prototype point coordinate data information of the road point cloud, the structured features of the road point cloud are extracted, and the structured feature data information of the road point cloud is obtained.

5. The target detection method based on point cloud segmentation and decoupled high-precision map according to claim 1, characterized in that: The average probability matrix of the road point cloud is A. ， Where N is the number of covariance matrices in the road point cloud, δ j a is the weighting coefficient. j This provides the covariance matrix data information for road point clouds.

6. The target detection method based on point cloud segmentation and decoupling high-precision map according to claim 1, characterized in that: The clustering algorithm that uses the fusion of normal vector angle features, spatial location, and semantic information to perform clustering analysis on road targets involves clustering road targets, with adjacent targets of the same category being clustered into one target. Based on the fusion of normal vector angle features, spatial location, and semantic information of the road point cloud, road foreground instances are segmented to obtain the three-dimensional spatial detection data information of the road targets.

7. A target detection system based on point cloud segmentation and decoupling high-precision maps, comprising computer equipment, characterized in that, The computer device is programmed or configured to perform the steps of the target detection method based on point cloud segmentation and decoupled high-precision maps as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is programmed or configured to perform the target detection method based on point cloud segmentation and decoupled high-precision maps as described in any one of claims 1 to 6.

Images