Crowdsourced map based lane model construction method and system, and storage medium

Abstract

The application discloses a kind of lane model construction method, system and storage medium based on crowd-sourcing map, comprising: S1: the single-trip data collected by car end is obtained, including semantic data and trajectory data, and respectively into step S2 and step S3;S2: single-frame lane model is constructed based on single-trip data, and enter step S8;S3: feature object in single-trip data is collected, and feature object is merged;S4: the feature object collected in multiple trips is clustered;S5: identify outlier in multiple feature objects;S6: collect all feature objects in multiple-trip data;S7: single-trip data is segmented according to multiple feature objects;S8: according to segmentation result, the data of car end model is cut;S9: each section is matched to output the final lane model.The application solves the problem that lane model construction cannot be realized lane level at present, with high calculation complexity, many preconditions, not enough automation and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-precision maps, and particularly relates to a lane model construction method and system based on a crowdsourcing map and a storage medium. BACKGROUND

[0002] At present, a crowdsourcing-based collection mode for producing high-precision maps is gradually favored by people, but the collected data is affected by the sensing technology and has low precision. How to accurately realize lane-level high-precision map production under the condition of low precision is a very important problem. Map learning can improve the quality of data output to a certain extent, but the algorithm of map learning also has certain limitations, and it is difficult to be compatible with some missing, drifting and error types of data. Therefore, it is urgent to construct a lane model to obtain the relative position relationship between objects, adjust the data input by the vehicle end, and thus provide higher-precision data for map learning and ensure the production of high-precision maps.

[0003] At present, the lane model has many uses, but the expression of its construction mode is less. For example, the lane detection method and device using the lane model disclosed in patent document CN110879961A construct and recognize the lane based on the CNN convolutional neural network of the deep learning model. However, it cannot recognize all lanes and lane-level road elements, and the neural network method needs a large number of parameters and has high computational complexity. For another example, the three-dimensional lane model generation method and system disclosed in patent document CN111047682A needs to know the number of lanes and the width of the lane in advance, which has great limitations and is not automatic enough.

[0004] Therefore, it is necessary to develop a lane model construction method and system based on a crowdsourcing map and a storage medium. SUMMARY

[0005] The application aims to provide a lane model construction method and system based on a crowdsourcing map and a storage medium to solve the problems that the current lane model construction cannot realize lane-level, has high computational complexity, has many preconditions and is not automatic enough.

[0006] In a first aspect, the application provides a lane model construction method based on a crowdsourcing map, which comprises the following steps:

[0007] S1: acquiring single-trip data collected by the vehicle end, including semantic data and trajectory data, and respectively entering step S2 and step S3;

[0008] S2: constructing a single-frame lane model based on the single-trip data, and entering step S8;

[0009] S3: collect feature objects in single-trip data, and merge the feature objects;

[0010] S4: cluster feature objects collected in multiple trips;

[0011] S5: identify outlier objects in multiple-trip feature objects;

[0012] S6: collect all feature objects in multiple-trip data;

[0013] S7: segment single-trip data according to multiple-trip feature objects;

[0014] S8: split data of a lane model according to a segmentation result;

[0015] S9: match each segment to output a final lane model.

[0016] Optionally, in the step S2, the specific steps of making a single-frame lane model are as follows:

[0017] S21: input single-trip semantic data and trajectory data;

[0018] S22: intercept single-frame semantic data and corresponding trajectory data;

[0019] S23: determine trajectory points near semantic objects according to time points and geographical positions, and establish front-back relationships of solid-line and dashed-line objects in a frame;

[0020] S24: add a LINE_ID attribute to semantic objects by relationships of solid lines and dashed lines in a single frame;

[0021] S25: draw a simplified lane marking by the attribute and relationships between objects;

[0022] S26: output a lane marking entity.

[0023] Optionally, in the step S23, the front-back relationships of solid-line and dashed-line objects in a frame are established, i.e., longitudinal relationships between objects are determined, and the specific steps are as follows:

[0024] determine the longitudinal relationship of the rightmost solid line and dashed line, extend the determined object relationship to the left by one lane to determine the object relationship of the second lane, and sequentially determine the relationships of all lanes; after the determination, the objects in the frame are sorted from left to right according to the vertical direction of the driving direction, and are marked with LINE_ID; the relationship between different frames is marked with TRACE_ID; thus, the longitudinal relationships of all lane solid lines and dashed lines in a single frame and the relationships between different frames are found.

[0025] Optionally, in the step S26, the approximate object of the object is determined according to the timestamp, the lane line serial number information in the semantic object, and the tracking ID, the timestamp, and the lane line ID information in the output entity class.

[0026] Optionally, the attribute of the lane marking includes a tracking ID, a lane line serial number, a timestamp, a geometric body, a boundary feature, and a lane line contained object type.

[0027] Optionally, in the step S3, the identified feature object includes a stop line + a zebra crossing, a set of road signs, a set of transverse arrows, a guide belt, and a rod-shaped object.

[0028] Optionally, in the step S4, the feature objects collected together in multiple laps are clustered by using a DBSCAN clustering method, that is, the objects in the same position in multiple laps are clustered together.

[0029] Optionally, the S9 includes:

[0030] S91: down-sampling the observed object on the lane line in the current segment to obtain the topological structure of the lane line and the relative relationship between different lane lines according to the sampling points;

[0031] S92: matching the lane line commonly observed in the multi-lap data according to the type and direction position information of the lane line and in combination with the lane line surrounding feature semantic object;

[0032] S93: if the matching is unsuccessful, expanding the matching range and improving the matching accuracy in combination with the prior accumulated error information, and then performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion; if the matching is successful, performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion;

[0033] S94: judging whether the fusion of the multi-lap same interval segments is completed, and if not, repeating the steps S92-S94 until all the lane information of the current lane is fused and the final lane model is output.

[0034] In a second aspect, a lane model construction system based on a crowd-sourced map is provided, which includes a memory and a controller, the memory stores a computer readable program, and the computer readable program can execute the steps of the lane model construction method based on the crowd-sourced map when called by the controller.

[0035] In a third aspect, a storage medium is provided, which stores a computer readable program, and the computer readable program can execute the steps of the lane model construction method based on the crowd-sourced map when called by the controller.

[0036] The present application has the following advantages: the present application can realize lane-level lane model construction, without preset conditions, and has low computational complexity. The present application effectively optimizes poor data sources of lane-end collected data, such as missing, drift, and errors of road elements. The present application not only has a simple idea, but is also easier to realize industrialization. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0038] Figure 1 The figure is a whole framework of the lane model in the present embodiment.

[0039] Figure 2 The figure is a single-trip lane model flowchart in the present embodiment.

[0040] Figure 3 The figure is a segment and segment matching flowchart in the present embodiment. DETAILED DESCRIPTION

[0041] The present application will be described in detail below with reference to the drawings.

[0042] As shown in the figure, in the present embodiment, a lane model construction method based on crowd-sourced maps, the lane model mainly realizes the matching alignment of the relative positions of multi-trip objects of lane-end collected data, and outputs road elements corrected according to the lane model, effectively improving the output quality of data. The method includes the following steps: Figure 1 S1: acquiring single-trip data collected by lane-end, including semantic data and trajectory data, and entering step S2 and step S3 respectively;

[0043] S2: constructing a single-frame lane model based on single-trip data, and entering step S8;

[0044] S3: collecting feature objects in single-trip data, and merging the feature objects;

[0045] S4: clustering the feature objects collected in multiple trips;

[0046] S5: identifying outlier objects in the multiple-trip feature objects;

[0047] S6: collecting all feature objects in multiple-trip data;

[0048] S7: segmenting single-trip data according to multiple-trip feature objects;

[0049] S8: matching the single-frame lane model and the segmented single-trip data, and outputting road elements corrected according to the lane model.

[0050] S8: segment the data of the lane model according to the segmentation result;

[0051] S9: match each segment to output a final lane model.

[0052] As shown in Figure 2 the embodiment, in step S2, the specific steps of making a single-frame lane model are as follows:

[0053] S21: input single-pass semantic data and trajectory data;

[0054] S22: intercept single-frame semantic data and corresponding trajectory data;

[0055] S23: determine trajectory points near semantic objects according to time points and geographic locations, and establish the front-back relationship of solid and dashed objects in the frame;

[0056] S24: add a lane line serial number LINE_ID attribute to semantic objects through the relationship between solid and dashed lines in a single frame;

[0057] S25: draw a simplified lane marking through the attribute and the relationship between objects;

[0058] S26: output a lane marking entity.

[0059] As shown in Figure 2 the embodiment, in step S23, the front-back relationship of solid and dashed objects in the frame is established, that is, the longitudinal relationship between objects is determined, which is specifically:

[0060] determine the longitudinal relationship of the rightmost solid and dashed objects, extend the determined object relationship to the left by one lane to determine the object relationship of the second lane, and so on until all lane relationships are determined; after the determination, the longitudinal relationship of the objects (solid and dashed) in the frame is sorted from left to right perpendicular to the driving direction and marked with LINE_ID; the relationship between different frames is marked with TRACE_ID; at this point, the longitudinal relationship of all lane solid and dashed lines in a single frame and the relationship between different frames have been found.

[0061] As shown in Figure 2 the embodiment, in step S26, the approximate object of an object is determined according to the timestamp in the semantic object, the lane line serial number information, and the tracking ID, timestamp, and lane line ID information in the output entity class.

[0062] In the embodiment, the attributes of the lane marking include the connection between frames (i.e., tracking ID (TRACE_ID)), the relationship between objects in the same frame, and the relationship between different frames. Figure 2As shown, the lane line sequence number from left to right (i.e., lane line sequence number (LINE_ID)), single frame time (i.e., timestamp), lane line drawn in a single frame (i.e., geometry (LineString)), boundary feature (curb) includes (1: none on left and right; 2: left side; 3: right side), and lane line containing object type (line_type) includes (1: solid line; 2: dashed line; 3: solid and dashed line).

[0063] As shown in the embodiment, in step S3, the identified feature objects include stop line + zebra crossing, a group of road signs, a group of transverse arrows, guide belt and rod-shaped object. Figure 1

[0064] In the embodiment, in step S4, the feature objects collected together in multiple passes are clustered using the DBSCAN clustering method, that is, the objects in multiple passes at the same position are clustered together.

[0065] As shown in the embodiment, S9 includes: Figure 3

[0066] S91: down-sampling the observed objects on the lane lines in the current segment to obtain the topological structure of the lane lines and the relative relationship between different lane lines according to the sampling points;

[0067] S92: matching the lane lines commonly observed in the multi-pass data according to the type and direction position information of the lane lines and in combination with the lane line surrounding feature semantic objects;

[0068] S93: if the matching is unsuccessful, expanding the matching range and improving the matching accuracy in combination with the prior accumulated error information, and then performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion; if the matching is successful, performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion;

[0069] S94: determining whether the fusion of the same interval segments in multiple passes is completed, and if not, repeating steps S92-S94 until all the lane information of the current lane is fused and the final lane model is output.

[0070] In the embodiment, a lane model construction system based on a crowdsourcing map includes a memory and a controller, the memory stores a computer readable program, and the computer readable program can execute the steps of the lane model construction method based on the crowdsourcing map as described in the embodiment when called by the controller.

[0071] In the embodiment, a storage medium stores a computer readable program, and the computer readable program can execute the steps of the lane model construction method based on the crowdsourcing map as described in the embodiment when called by the controller.​​

[0072] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications, etc. made without departing from the spirit and principles of the present application should be equivalent replacement manners and should be included in the protection scope of the present application.

Claims

1. A method for constructing a lane model based on a crowd-sourced map, characterized by, The method comprises the following steps: S1: acquiring single-trip data collected by a vehicle end, including semantic data and trajectory data, and entering steps S2 and S3 respectively; S2: constructing a single-frame lane model based on the single-trip data, and entering step S8; S3: collecting feature objects in the single-trip data and merging the feature objects; S4: clustering feature objects collected in multiple trips; S5: identifying outlier objects in the multiple-trip feature objects; S6: collecting all feature objects in the multiple-trip data; S7: segmenting the single-trip data according to the multiple-trip feature objects; S8: cutting the data of the lane model according to the segmentation result; S9: matching each segment to output a final lane model, specifically: S91: down-sampling observation objects on the lane lines in the current segment to obtain the topological structure of the lane lines and the relative relationship between different lane lines according to the sampling points; S92: matching lane lines commonly observed in the multiple-trip data according to the type and direction position information of the lane lines and the surrounding feature semantic objects of the lane lines; S93: if the matching is unsuccessful, expanding the matching range and improving the matching accuracy by combining prior accumulated error information, and then performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion; if the matching is successful, performing translation and rotation according to the matching relative error between the lane lines to complete the inter-segment fusion; S94: determining whether the fusion of multiple-trip segments in the same interval is completed, and if not, repeating steps S92-S94 until all lane information of the current lane is fused and the final lane model is output. 2.The crowd-sourced map based lane model construction method of claim 1, wherein: In the step S2, the specific steps of making a single-frame lane model are as follows: S21: inputting the semantic data and trajectory data of a single trip; S22: intercepting single-frame semantic data and corresponding trajectory data; S23: determining trajectory points near semantic objects according to time points and geographical positions, and establishing the front-back relationship of solid line and dashed line objects in a frame; S24: adding a lane line serial number LINE_ID attribute to the semantic objects by the relationship between the solid line and the dashed line in a single frame; S25: drawing a simplified lane marking by the attribute and the relationship between the objects; S26: outputting a lane marking entity. 3.The crowd-sourced map based lane model construction method of claim 2, wherein: In the step S23, the front-back relationship of solid line and dashed line objects in a frame is established, that is, the longitudinal relationship between the objects is determined, specifically: The longitudinal relationship of the rightmost side is determined according to the solid line and the dashed line; the object relationship of the second lane is determined by extending the determined object relationship to the left side by one lane, and the same is repeated until the relationships of all lanes are determined; after the determination, the objects in the frame are sorted from left to right according to the vertical direction to the driving direction, and are marked with LINE_ID; the relationship between different frames is marked with TRACE_ID; thus, the longitudinal relationship of all lane solid lines and dashed lines in a single frame, and the relationship between different frames are found.

4. The crowdsourced map based lane model construction method of claim 2 or 3, wherein: In the step S26, the objects are determined according to the timestamp, lane line serial number information in the semantic objects, and the tracking ID, timestamp, lane line ID information in the output entity class.

5. The crowd-sourced map based lane model construction method of claim 4, wherein: The attributes of the lane marking include tracking ID, lane line serial number, time stamp, geometric body, boundary feature, and lane line contained object type.

6. The crowd-sourced map based lane model construction method of claim 5, wherein: In the step S3, the identified feature objects include stop line + zebra crossing, a group of road signs, a group of transverse arrows, guide rail, and pole.

7. The crowd-sourced map based lane model construction method of claim 6, wherein: In the step S4, the feature objects collected in multiple laps are clustered by using DBSCAN clustering method, i.e. the objects in the same position in multiple laps are clustered together. 8.A crowd-sourced map based lane model construction system, comprising: The computer readable program is stored in the memory and can be executed by the controller to perform the steps of the lane model construction method based on the crowd-sourced map according to any one of claims 1 to 7.

9. A storage medium characterized by: The computer readable program is stored in the memory and can be executed by the controller to perform the steps of the lane model construction method based on the crowd-sourced map according to any one of claims 1 to 7.

Images