Intelligent image recognition method for high-precision map

Abstract

The application relates to the technical field of image processing, and discloses an intelligent image recognition method for a high-precision map, which comprises the following steps: collecting various map elements and contour information required by a high-precision map, and recording relevant data of a plurality of attributes of each map element; obtaining a plurality of categories of each attribute of each map element; obtaining a comprehensive basic value of each map element, and constructing a map element basic sequence from large to small; constructing a plurality of scales with the number of map elements from small to large; constructing a map distribution diagram of each scale, and obtaining a chain code vector of each edge in the map distribution diagram; screening a plurality of reserved scales; marking a plurality of nodes; obtaining a path of each node of each reserved scale; screening a plurality of reserved paths of each node, and generating a high-precision map in combination with the map distribution diagram of each reserved scale. The application aims to solve the problem that the matching of a road network and a lane network at an intersection is not accurate enough in the high-precision map production process.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of image processing, in particular to an intelligent image recognition method for high-precision maps. BACKGROUND

[0002] With the popularization of intelligent driving technology, the requirement for map precision is also higher and higher; the traditional map mainly provides the driver with the route to the destination, but there is no more detailed spatial information on the whole route, and the intelligent driving technology needs to be aware of the vehicle-mounted sensors (such as cameras, laser radars, etc.) to perform centimeter-level positioning, so it often needs the service of high-precision maps.

[0003] In the process of making high-precision maps, one of the main tasks is to build the topology of the road network. The topology information describes the connection relationship and traffic rules between map elements. The intelligent transportation system can realize lane-level management and control based on the topology. When building the topology of the road network, the SD road network (traditional navigation map road) and the HD lane network (high-precision map lane element) need to be combined, that is, the key point matching of the SD road network and the HD lane network needs to be performed. Therefore, in order to ensure the accuracy of the subsequent key point matching results and thus ensure the accuracy of the high-precision map, the collected data needs to be preprocessed in the early stage to accurately reflect the information points for matching, especially in complex places such as road intersections in the road network. It is usually difficult to accurately match due to too complex lane elements and information, thereby affecting the accuracy of high-precision map making. SUMMARY

[0004] The application provides an intelligent image recognition method for high-precision maps to solve the problem of insufficient accuracy of matching the road network and the lane network at the intersection in the existing high-precision map making process. The technical solution adopted is as follows:

[0005] The application provides an intelligent image recognition method for high-precision maps to solve the problem of insufficient accuracy of matching the road network and the lane network at the intersection in the existing high-precision map making process. The technical solution adopted is as follows:

[0006] Collecting various map elements and their contour information required for high-precision maps and recording the related data of several attributes of each map element;

[0007] Classifying the related data of each attribute of each map element to obtain several categories of each attribute of each map element; analyzing the distribution relationship of the related data in the categories of different attributes of different map elements in the same geographical position to obtain the basic degree of each attribute of each map element, and then obtaining the comprehensive basic value of each map element and constructing a map element basic sequence from large to small; constructing several scales with the number of map elements from small to large based on the map element basic sequence;

[0008] According to the correlation data of the attributes of the map elements, a map distribution graph of each scale is constructed, and a chain code vector of each edge in the map distribution graph is obtained; a difference in a similarity relationship of the chain code vectors of the nearest edges in adjacent scales is analyzed, and a plurality of reserved scales are screened; and a plurality of nodes are marked based on the distribution relationship of the edges in the map distribution graph of the reserved scale with the least number of map elements.

[0009] Based on the edge distribution in the map distribution graph of each reserved scale and the correlation data of each attribute of each map element, a path of each node in each reserved scale is obtained by combining a Markov model; a difference between the paths of the nodes in adjacent reserved scales is analyzed, a plurality of reserved paths of each node are screened, and a high-precision map is generated by combining the map distribution graph of each reserved scale.

[0010] Optionally, the plurality of categories of each attribute of each map element are obtained by the following specific method:

[0011] For any attribute of any map element, if the attribute is a numerical attribute, a plurality of correlation data of the attribute are clustered to obtain a plurality of clusters, and the plurality of correlation data in any cluster are taken as a category of the attribute.

[0012] If the attribute is a text attribute, text information corresponding to the correlation data of the attribute is numbered, the numbering of the same text information is the same, and the correlation data of the attribute with the same numbering is taken as a category of the attribute.

[0013] Optionally, the base degree of each attribute of each map element is obtained, and then a comprehensive base value of each map element is obtained, and a map element base sequence is constructed from large to small, by the following specific method:

[0014] Based on the latitude and longitude, for the first attribute of any map element, for the plurality of correlation data of any category, a plurality of regions corresponding to the correlation data of the category are taken as distribution regions of the category, a plurality of correlation data of any attribute of any other map element in the distribution regions are obtained, a joint distribution probability of the category of the first attribute and the attribute of the other map element is calculated based on the plurality of correlation data of the two attributes in the distribution regions, an average value of the joint distribution probabilities of the categories of the first attribute and the attributes of the other map element is taken as a base factor of the first attribute of the map element and the other map element, an average value of the base factors of the first attribute of the map element and the other map elements except the map element is taken as a base degree of the first attribute of the map element.

[0015] ​​​​​​The average value of the basic level of each attribute of the map element is taken as the comprehensive basic value of the map element; the map elements are arranged in descending order of comprehensive basic value, and the resulting sequence is denoted as the basic sequence of map elements.

[0016] Optionally, the specific methods for constructing several scales of map feature quantity from small to large based on the basic sequence of map features include:

[0017] Following the order of map elements in the basic sequence of map elements, the first scale is constructed with the first map element, and the second scale is constructed with the first and second map elements, and so on, to obtain several scales.

[0018] Optionally, the specific method for constructing map distribution maps at various scales and obtaining the chain code vectors of each edge is as follows:

[0019] For any given scale, for each map element contained at that scale, based on the relevant data of each attribute of each map element, combined with its spatial contour information and location information, a map distribution map for that scale is constructed.

[0020] Obtain several edges from the map distribution map, and group two edges that are completely parallel and correspond to the same map feature into an edge group; for any edge, obtain several key points on the edge, including the start point, end point and inflection point, and obtain the chain code of the edge based on the key points;

[0021] Obtain the chain code of each edge in the map distribution map at this scale. Based on the chain code of the edge with the most elements in the chain code, obtain the chain code vector of that edge. Obtain the chain code vector of the chain code of other edges. If the number of chain code elements is insufficient, pad with 0 at the end of the chain code.

[0022] Optionally, the specific methods for obtaining several retention scales include:

[0023] The differences in the similarity of chain code vectors at adjacent scales are analyzed to obtain the nearest neighbor variation factor for each edge group at adjacent scales.

[0024] Get the At each scale, the edge groups are in the th... Each scale and the first The nearest neighbor variation factor at the nth scale is used as the mean of all nearest neighbor variation factors as the nth scale. The first scale is relative to the first Information change characteristics at each scale;

[0025] The judgment begins with the information change characteristics of the second scale relative to the first scale. If the information change characteristics are less than the change threshold, the second scale is discarded, and the calculation continues for the information change characteristics of the third scale relative to the first scale. If the third scale is retained, the calculation is then performed based on the information change characteristics of the fourth scale relative to the third scale, and so on. Several scales are discarded based on the information change characteristics, and the remaining scales are retained.

[0026] Optionally, the specific method for obtaining the nearest neighbor variation factor of each edge group at adjacent scales includes:

[0027] For the For any edge group at any scale, if the chain code vectors of two edges in the edge group are identical, and these are used as the chain code vectors of the edge group, the edge closest to the edge group is obtained. Based on the mean Euclidean distance of each keypoint of the edge group to other edges, the edge with the smallest mean Euclidean distance is taken as the closest edge to the edge group. The cosine similarity between the chain code vector of the edge group and the chain code vector of the closest edge to the edge group is obtained, and this similarity is used as the first... Nearest neighbor similarity of the edge group at each scale;

[0028] In the Retrieve the nearest edge to the edge group at each scale, and re-obtain the edge of the first scale. The nearest neighbor similarity of the edge group at each scale is obtained; the absolute value of the difference between the nearest neighbor similarities of the edge group at two adjacent scales is used as the nearest neighbor similarity of the edge group at the [number]th scale. Each scale and the first Nearest neighbor variation factor at each scale.

[0029] Optionally, the specific methods for marking several nodes include:

[0030] The retention scale with the fewest map features among the retention scales is taken as the maximum retention scale; based on the distribution of edges in each edge group in the map distribution map of the maximum retention scale, several combinations of regional edge groups are obtained;

[0031] For any set of region edge groups, the region is divided using this set of region edge groups, and two edge groups constitute a divided region; each edge in the two edge groups is extended until the edges of the two edge groups intersect, and the intersecting part is marked as a node;

[0032] For the original edges in the two edge groups, for any edge group, a preset division ratio is used to obtain several straight lines perpendicular to the two edges. Starting from the starting point of each of the two edges, a vertical straight line is extracted at each division ratio between the starting point and the next key point, which serves as the grid line in the edge group. This results in several grid lines. The intersection of the grid lines with the two edges is marked as a node.

[0033] Optionally, the obtaining of the several combinations of the region edge groups comprises the specific method that:

[0034] For each edge group in the map distribution diagram of the maximum reserved scale, the nearest edge of each edge group is obtained, and based on the edge group corresponding to the nearest edge, the nearest edge group of each edge group is obtained respectively, and several combinations of the nearest edge groups are obtained.

[0035] For any one of the combinations of the nearest edge groups, the direction difference between the two partial edge groups of the two nearest edge groups is obtained, and if the direction difference is greater than or equal to the difference threshold, the combination of the nearest edge groups is taken as a combination of the region edge groups.

[0036] Optionally, the screening of the several reserved paths of each node comprises the specific method that:

[0037] For any node in the two paths of the adjacent reserved scales, the DTW distance of the two paths is obtained through DTW matching, the maximum value of the DTW distance between the paths of the node at all reserved scales is obtained, the ratio of the DTW distance corresponding to the adjacent reserved scale to the maximum value of the DTW distance is taken as a path difference factor of the adjacent reserved scale, a preset path difference threshold is set, if the path difference factor is greater than the path difference threshold, the two paths of the node at the adjacent reserved scale are reserved, if the path difference factor is less than or equal to the path difference threshold, the path of the latter one of the adjacent reserved scales is removed, and the paths of the node at each reserved scale are taken as the several reserved paths of the node.

[0038] The beneficial effects of the present application are: in the process of high-precision map generation, the influence of complex information at intersections on SD road network and HD lane network is considered, wherein through information collection and attribute data analysis of various map elements, the influence relationship between data changes under the attributes of map elements and data changes under other map elements is considered, the basic map elements are screened, and several scales from large to small are constructed according to the number of map elements from small to large, and the comprehensive basic value of the map elements gradually decreases; further, the map distribution diagram is constructed according to the map elements and their attribute data, the difference change of the similarity relationship between the edges under the adjacent scales is analyzed, the scales corresponding to the repeated information are removed, the reserved scales are obtained, and the nodes intersected by the grid lines and the edges are further constructed according to the map distribution diagram, so as to preliminarily obtain the intersection information; the paths of each node are predicted through the historical vehicle trajectory combined with the prediction model, and the reserved paths are further screened to reduce the interference of repeated information on the learning of the prediction model, and finally the matching between the road maps is performed through the nodes and their reserved paths, so as to improve the accuracy of high-precision map generation and recognition. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0040] Figure 1 A high-precision map intelligent image recognition method flow chart provided by an embodiment of the present application;

[0041] Figure 2 A high-precision map schematic diagram. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0043] Please refer to Figure 1 which shows a high-precision map intelligent image recognition method flow chart provided by an embodiment of the present application. The method comprises the following steps:

[0044] Step S001, collect various map elements and their contour information required for high-precision map, and record the relevant data of several attributes of each map element.

[0045] The purpose of the present embodiment is to construct a high-precision map based on road network topology, and the road network contains various map elements. Therefore, the map elements need to be collected and their contour information recorded. In addition, each map element contains several attributes, and the data of each attribute also needs to be recorded as relevant data of each attribute.

[0046] Specifically, the map elements such as road lines, traffic signs, traffic signs, traffic lighting equipment and other traffic facilities are collected. The collection means include but are not limited to vehicle-mounted mobile measurement system, laser radar, oblique photography and other methods. In the collection process, in addition to the map elements themselves, the contour information of each map element is recorded, and the spatial contour information of the map elements is represented in the form of points, lines and surfaces.

[0047] Further, the data of multiple attributes of each map element are also recorded, for example, the attributes included in the road line map element include: line type, line width, line length, line color, turning information, vehicle type setting and speed limit information, and the data corresponding to the map element is collected for each attribute, and each attribute of the same map element includes several data (corresponding to different roads, traffic signs, etc. in the road network), as the related data of each attribute of each map element, stored in SHP format.

[0048] It should be noted that, as shown in Figure 2 The schematic diagram of the road high-precision map is shown, the city high-precision map collection work is provided, the collected elements include but are not limited to highway main road, ramp, interchange, bridge, tunnel, entrance and exit, monitoring equipment, sign and marking, various signs, information board, lighting facility, isolation belt, reflective facility, guardrail, car-road cooperation equipment, etc., the absolute accuracy is 20-50 cm, the relative accuracy is 10-20 cm, which can be applied to high-speed or urban expressway, urban ordinary road, holographic intersection and other scenes, and provides support for intelligent management and control such as congestion governance, safety prevention and control, and fine management and control.

[0049] It should be noted that in the process of making high-precision map, one of the main works is to build the road network topology relationship, and in the process of building the road network topology relationship, the key point matching of SD road network and HD lane network is needed, so as to ensure the accuracy of the subsequent key point matching result, thereby ensuring the accuracy of the high-precision map; the road network is composed of roads and intersections, and the intersection is an important node connecting various roads, which is often the most complex place in the whole road network, if accurate matching is needed, the information at the intersection should be more clear; by analyzing the lane change information at the intersection, the topology relationship at the intersection is determined, which provides conditions for the matching process of SD road and HD lane network.

[0050] Step S002, classify the related data of each attribute of each map element to obtain several categories of each attribute of each map element; analyze the distribution relationship of the related data in the categories of different attributes of different map elements in the same geographical position to obtain the basic degree of each attribute of each map element, and then obtain the comprehensive basic value of each map element, and construct a map element basic sequence from large to small; based on the map element basic sequence, several scales with the number of map elements from small to large are constructed.

[0051] It should be noted that, since the map elements are constructed when collecting data, including: road lines, traffic signs, traffic signs, traffic lighting equipment and other traffic facilities, etc., these map elements are integrated to form the entire road network; the information contained in each map element is different, for example, the information contained in the road lines (lane center line, lane edge line) at different positions is different, the lane center line has description line type, line width, line length, line color, turning information, vehicle type setting, speed limit information, etc., while the lane edge line only has line type, line width, line color, etc., so the information contained is different, and the position attributes of traffic signs, traffic signs, traffic lighting equipment, etc. are determined by the lane line, so the reference map element needs to be determined; the desired reference map element can determine the position and other attributes of other map elements, and can also be used to reflect the topological relationship of the intersection for subsequent matching services.

[0052] Preferably, in an embodiment of the present application, the related data of each attribute of each map element is classified to obtain several categories of each attribute of each map element, including the specific method:

[0053] For any attribute of any map element, if the attribute is a numerical attribute, the related data of the attribute is clustered, this embodiment adopts K-means clustering, and the distance measurement adopts the absolute value of the difference between the related data, a total of 6 clustering centers are set, that is, K=6, a plurality of clusters are obtained, and the plurality of related data in any cluster are taken as a category of the attribute, and then a plurality of categories of the attribute of the map element are obtained.

[0054] Further, if the attribute is a text attribute, the text information corresponding to the related data of the attribute is numbered, and the numbering of the same text information is the same (the numbering has no actual meaning and is only used for classification). The related data of the attribute with the same number is taken as a category of the attribute, and then a plurality of categories of the attribute are obtained.

[0055] It should be further noted that, since the latitude and longitude of various map elements is the most basic geographical position element, comparison is made based on latitude and longitude, that is, distribution comparison is made in a similar latitude and longitude area; a map element contains various attributes, and all attributes reflect all information of the map element. If it is judged whether the map element is a basic map element, the relationship between the attribute and other map elements needs to be combined, and if any attribute of the target map element is the same as the distribution of other map elements, the corresponding map element has a greater degree of basis.

[0056] Preferably, in an embodiment of the present application, the distribution relationship of the related data in the categories of different attributes of different map elements in the similar geographical position is analyzed to obtain the basis degree of each attribute of each map element, and then the comprehensive basis value of each map element is obtained, and a map element basis sequence is constructed from large to small, including the specific method:

[0057] With each 1 second of longitude and latitude to build the region, then for any map element of the first The property, for any category of several related data, the corresponding several regions of the related data of the category are recorded as the distribution area of the category, and the related data of any attribute of any other map element in the distribution area is obtained. Based on the joint distribution probability of the first The average of the joint distribution probability of each category of the first The property and each attribute of the other map element is taken as the basic factor of the first The property of the map element and the average of the basic factors of the first The property of each other map element is taken as the basic degree of the first

[0058] The average of the basic degree of each attribute of the map element is taken as the comprehensive basic value of the map element; the map elements are arranged in order of comprehensive basic value from large to small, and the sequence obtained is recorded as the map element basic sequence.

[0059] Further need to be explained is that different scales contain different information, referring to the main branch of the theory of cartography, different scales have different information expression effects, from simple to complex, and the node change of the possibility is analyzed, wherein in the process of building scale, the small scale is required to be the refinement of the information of the large scale, therefore the first scale is the largest scale, that is, the map element with the largest comprehensive basic value is taken as the largest scale.

[0060] Preferably, in one embodiment of the present application, a plurality of scales are constructed based on the map element basic sequence in order of the number of map elements from small to large, including the specific method:

[0061] According to the order of the map elements in the map element basic sequence, the first map element is used to construct the first scale, and the first and second map elements are used to construct the second scale, and so on, that is, the first several map elements are used to construct the first several scales, then a plurality of scales are obtained, and the number of map elements in the scale is from small to large.

[0062] At this point, a plurality of scales containing different numbers of map elements are obtained.

[0063] Step S003, constructing a map distribution diagram of each scale according to the relevant data of the attributes of the map elements in the scale, and obtaining the chain code vector of each edge; analyzing the difference in the similarity relationship of the chain code vector of the nearest edge in the adjacent scale, and screening to obtain several reserved scales; marking several nodes based on the distribution relationship of the edges in the map distribution diagram of the reserved scale with the least number of map elements.

[0064] It should be noted that, since some map elements are repeated, there will be repetitive content in the map element combination corresponding to the scale, that is, it is not necessary to calculate the repetitive scale when determining the nodes subsequently; for any scale, a map distribution diagram is constructed according to the map elements contained in the scale, combined with the latitude and longitude coordinates and the attributes of the corresponding map elements (similar to a map CAD diagram).

[0065] Preferably, in an embodiment of the present application, constructing a map distribution diagram of each scale according to the relevant data of the attributes of the map elements in the scale, and obtaining the chain code vector of each edge, includes the following specific method:

[0066] For any scale, according to the relevant data of the attributes of each map element, combined with the spatial contour information and the position information (latitude and longitude coordinates), a map distribution diagram of the scale is constructed (reference for constructing a map CAD diagram).

[0067] Further, several edges in the map distribution diagram are obtained, since the contour information of the road line and the traffic facility is usually a set of parallel edges, two edges that are completely parallel and correspond to the same map element (also the same length) are regarded as an edge group; for any edge, several key points including the starting point, the ending point and the inflection point (the turning point where the direction changes, that is, the change point of the edge direction) on the edge are obtained, and the chain code of the edge is obtained based on the key points; the chain code of each edge in the map distribution diagram of the scale is obtained according to the above method, and the chain code vector of the edge is obtained based on the chain code with the most elements, that is, the elements in the chain code form the chain code vector in order, and the number of elements in the chain code vector is equal to the number of elements in the chain code of the edge, then the chain code vector of other edges is also obtained, and 0 is added at the end of the chain code if the number of chain code elements is insufficient.

[0068] It should be further noted that, by analyzing the similarity of the chain code vectors between the edge groups, if the similarity relationship between the same edge group and the nearest edge does not change after the scale is reduced, the scale change does not cause a large change in the map distribution diagram information, and the reduced scale should be discarded to avoid repeated information analysis.

[0069] Preferably, in one embodiment of the present invention, the method for analyzing the differences in the similarity relationships of chain code vectors at adjacent scales and selecting several retainable scales includes:

[0070] For the For any edge group at any scale, if the chain code vectors of two edges in the edge group are identical, and these are used as the chain code vectors of the edge group, the edge closest to the edge group is obtained. Based on the mean Euclidean distance of each keypoint of the two edges in the edge group to other edges (i.e., the mean Euclidean distance of each edge to a keypoint corresponding to any other edge), the edge with the smallest mean Euclidean distance is taken as the closest edge to the edge group. The cosine similarity between the chain code vector of the edge group and the chain code vector of the closest edge to the edge group is obtained, and this similarity is used as the first... The nearest neighbor similarity of the edge group at each scale; similarly, at the first scale... Retrieve the nearest edge to the edge group at each scale, and re-obtain the edge of the first scale. The nearest neighbor similarity of the edge group at each scale is obtained; the absolute value of the difference between the nearest neighbor similarities of the edge group at two adjacent scales is used as the nearest neighbor similarity of the edge group at the [number]th scale. Each scale and the first Nearest neighbor variation factor at each scale.

[0071] Furthermore, obtain the first [item] using the method described above. At each scale, the edge groups are in the th... Each scale and the first The nearest neighbor variation factor at the nth scale is used as the mean of all nearest neighbor variation factors as the nth scale. The first scale is relative to the first Information change characteristics at each scale; a preset change threshold is used, which is 0.2 in this embodiment. The judgment starts from the information change characteristics of the second scale relative to the first scale. If the information change characteristics are less than the change threshold, the second scale is discarded, and the calculation continues for the information change characteristics of the third scale relative to the first scale. If the third scale is retained, the calculation is then performed based on the information change characteristics of the fourth scale relative to the third scale, and so on. Several scales are discarded based on the information change characteristics, and the remaining scales are retained.

[0072] It should be further explained that, since the added nodes are related to the interaction between map elements, the region is divided based on the largest scale among the retained scales, that is, the scale with the fewest map elements. The purpose of dividing the region and marking the nodes is to mark the intersection area through edge extension and subsequent path planning. Therefore, grid lines are generated through the corresponding region, that is, grid construction is carried out based on the width of the corresponding edge group based on the edge group, and node marking is carried out based on the grid.

[0073] Preferably, in one embodiment of the present application, based on the distribution relationship of the edges in the map distribution diagram of the reserved scale with the least number of map elements, a plurality of nodes are marked, including the specific method:

[0074] The reserved scale with the least number of map elements is taken as the maximum reserved scale; for each edge group in the map distribution diagram of the maximum reserved scale, the nearest edge of each edge group is obtained, and based on the edge group corresponding to the nearest edge, the nearest edge group of each edge group is obtained respectively, and then a plurality of nearest edge group combinations are obtained; for any one nearest edge group combination, the directional difference (the included angle of the direction, and the nearest part is the part of the edge group corresponding to the minimum distance between the line segments in the edge group) between the two nearest edge groups of the two edge groups is obtained, a preset difference threshold is set, and in this embodiment, the difference threshold is described as 15 degrees; if the directional difference is greater than or equal to the difference threshold, the nearest edge group combination is taken as a region edge group combination.

[0075] Further, for any one region edge group combination, the region is divided according to the region edge group combination, that is, the two edge groups constitute a division region; each edge in the two edge groups is extended (extended along the direction of the corresponding line segment of the starting point and the ending point), until the edges of the two edge groups intersect, that is, the extensions of the two roads under the road intersection in the road network intersect (since there is no corresponding road line in the road intersection, the extensions need to intersect), and the intersection part is marked as a node; for the original edges in the two edge groups, since the two edges in the same edge group are parallel, for any one edge group, a preset segmentation ratio is set, and in this embodiment, the segmentation ratio is described as 1 / 10; a plurality of straight lines perpendicular to the two edges are obtained, and each straight line is extracted from the starting point of each edge to the next key point (knot or ending point) (if the two edges are completely parallel, the starting points and key points of the two edges are on the same vertical straight line) at an interval of the segmentation ratio (the length of the edge between the starting point and the next key point), as a grid line in the edge group, and then a plurality of grid lines are obtained; the intersection points of the grid lines and the two edges are marked as nodes, and then a plurality of nodes are obtained for the region edge group combination.

[0076] At this point, a plurality of nodes are marked in the map distribution diagram of the maximum reserved scale for subsequent analysis and path planning of the road intersection.

[0077] Step S004, based on the edge distribution in the map distribution diagram of each reserved scale and the related data of each attribute of each map element, the paths of each node in each reserved scale are obtained by combining the Markov model; the difference between the paths of the nodes in adjacent reserved scales is analyzed, a plurality of reserved paths of each node are selected, and the high-precision map is generated by combining the map distribution diagram of each reserved scale.

[0078] Preferably, in an embodiment of the present application, based on the edge distribution in the map distribution diagram of each reserved scale and the relevant data of each attribute of each map element, the path of each node of each reserved scale is obtained by combining the Markov model, including the specific method:

[0079] For a number of nodes of any combination of a group of regional edges in the maximum reserved scale and the corresponding divided region, a large number of vehicle trajectories of vehicles in the divided region in the historical data are obtained, combined with the relevant data of each attribute of each map element (only including the map elements included in the maximum reserved scale) in the divided region, a Markov model is constructed, and the vehicle trajectories and the relevant data are used as input data for training, the nodes of the divided region are judged, and the simulated path of the nodes is generated; the nodes of each group of regional edge combinations in the maximum reserved scale are judged based on the Markov model according to the above method, and the simulated path of each node is generated; wherein the Markov model training and path generation based on vehicle trajectories and relevant data are prior art for prediction models, and this embodiment will not be described again.

[0080] Further, according to the above method based on the map distribution diagram of each reserved scale, for each node marked in the maximum reserved scale, according to the corresponding divided region of each node in the map distribution diagram of each reserved scale in the maximum reserved scale and the relevant data of each attribute of the included map elements, the Markov model is trained, and the simulated path of each node in the corresponding reserved scale is output, then the simulated path of each node in each reserved scale is obtained as the path of each node of each reserved scale; it should be noted that the map distribution diagram of the node in different reserved scales is the same, the corresponding divided region is also the same based on the map distribution diagram of the maximum reserved scale, and the map elements contained in different reserved scales are different, so the training process is affected by the relevant data of the attributes of the map elements in the corresponding divided region.

[0081] Preferably, in an embodiment of the present application, the difference between the paths of the nodes under adjacent reserved scales is analyzed, and a number of reserved paths of each node are screened, and the high-precision map is generated combined with the map distribution diagram of each reserved scale, including the specific method:

[0082] For any node, for two paths of adjacent reserved scales (the reserved scales remaining after discarding part of the scales are still arranged in the original order, here adjacent means sequentially adjacent, and before discarding, they can not be adjacent due to the existence of other discarded scales between them), the DTW distance of the two paths is obtained through DTW matching, the maximum value of the DTW distance between the paths of the node at all reserved scales is obtained, and the ratio of the DTW distance corresponding to the adjacent reserved scale to the maximum value of the DTW distance is taken as the path difference factor of the adjacent reserved scale. A preset path difference threshold is used for description in the embodiment, and the path difference threshold is 0.5. If the path difference factor is greater than the path difference threshold, the two paths of the node at the adjacent reserved scale are reserved. Conversely, if the path difference factor is less than or equal to the path difference threshold, the path of the latter one of the adjacent reserved scales (the reserved scale containing the largest number of map elements among the two reserved scales) is removed. The paths of the node at each reserved scale are taken as a plurality of reserved paths of the node.

[0083] Further, based on the plurality of reserved paths of each node and the plurality of edges in the map distribution diagram corresponding to the reserved scales, the HD lane net and the SD road net are matched. If the topological structures are consistent and the positions are similar, it is considered that the matching is successful. The specific matching method is to match the SD road net and the HD lane net by referring to the existing method, and the embodiment will not be described again. After obtaining the matching result, the topology is hung according to the arrow of the lane and the traffic regulation common sense, and then the high-precision map is generated. The matching of the HD lane net and the SD road net based on the reserved path and the map distribution is prior art, and the embodiment will not be described again. Finally, the generation of the high-precision map is realized.

[0084] Thus, the embodiment is completed.

[0085] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for intelligent image recognition in high-precision mapping, characterized in that, The method includes the following steps: Collect various map elements and their outline information required for high-precision maps, and record relevant data of several attributes of each map element; The relevant data of each attribute of each map element are classified to obtain several categories of each attribute of each map element; the distribution relationship of relevant data in the categories of different attributes of different map elements under similar geographical locations is analyzed to obtain the basic degree of each attribute of each map element, and then the comprehensive basic value of each map element is obtained, and a basic sequence of map elements is constructed from large to small; based on the basic sequence of map elements, several scales of map element quantity are constructed from small to large. Based on the relevant data of the map features at each scale, construct map distribution maps for each scale and obtain the chain code vectors of each edge; analyze the differences in the similarity of the chain code vectors of the nearest edge at adjacent scales, and select several retained scales; based on the distribution relationship of the edges in the map distribution map of the retained scale with the fewest map features, mark several nodes. Based on the edge distribution and related data of each attribute of each map element in the map distribution map at each preservation scale, the path of each node at each preservation scale is obtained by combining the Markov model; the difference between the paths of nodes at adjacent preservation scales is analyzed, several preserved paths of each node are selected, and high-precision maps are generated by combining the map distribution map at each preservation scale. The specific methods for obtaining the basic degree of each attribute of each map element, thereby obtaining the comprehensive basic value of each map element, and constructing a basic sequence of map elements from large to small include: Based on latitude and longitude, a region is constructed. For the i-th attribute of any map element, for several related data of any category, several regions corresponding to the related data of that category are recorded as the distribution region of that category. Several related data of any attribute of any other map element besides that map element are obtained in the distribution region. Based on the several related data of the two attributes in the distribution region, the joint distribution probability of the i-th attribute category and the other map element attribute is calculated. The mean of the joint distribution probability of each category of the i-th attribute and each attribute of the other map element is used as the basic factor of the i-th attribute of the map element and the other map element. The mean of the basic factors of the i-th attribute of the map element and each other map element besides that map element is used as the basic degree of the i-th attribute of the map element. The average value of the basic level of each attribute of the map element is taken as the comprehensive basic value of the map element; the map elements are arranged in descending order of comprehensive basic value, and the resulting sequence is denoted as the basic sequence of map elements.

2. The intelligent image recognition method for high-precision maps according to claim 1, characterized in that, The specific methods for obtaining several categories of attributes for each map element include: For any map feature and any attribute, if the attribute is a numerical attribute, cluster several related data of the attribute to obtain several clusters, and take several related data in any cluster as a category of the attribute. If the attribute is a text attribute, the text information corresponding to the data related to the attribute is numbered. If the text information has the same number, the data related to the attribute with the same number is regarded as a category of the attribute.

3. The intelligent image recognition method for high-precision maps according to claim 1, characterized in that, The specific methods for constructing several scales of map feature quantity from small to large based on the basic sequence of map features include: Following the order of map elements in the basic sequence of map elements, the first scale is constructed with the first map element, and the second scale is constructed with the first and second map elements, and so on, to obtain several scales.

4. The intelligent image recognition method for high-precision maps according to claim 1, characterized in that, The specific methods for constructing map distribution maps at various scales and obtaining the chain code vectors of each edge are as follows: For any given scale, for each map element contained at that scale, based on the relevant data of each attribute of each map element, combined with its spatial contour information and location information, a map distribution map for that scale is constructed. Obtain several edges from the map distribution map, and group two edges that are completely parallel and correspond to the same map feature into an edge group; for any edge, obtain several key points on the edge, including the start point, end point and inflection point, and obtain the chain code of the edge based on the key points; Obtain the chain code of each edge in the map distribution map at this scale. Based on the chain code of the edge with the most elements in the chain code, obtain the chain code vector of that edge. Obtain the chain code vector of the chain code of other edges. If the number of chain code elements is insufficient, pad with 0 at the end of the chain code.

5. The intelligent image recognition method for high-precision maps according to claim 4, characterized in that, The specific methods for obtaining several retention scales are as follows: The differences in the similarity of chain code vectors at adjacent scales are analyzed to obtain the nearest neighbor variation factor for each edge group at adjacent scales. Obtain the nearest neighbor change factors of each edge group at the i-th scale and at the i+1-th scale, and use the mean of all the nearest neighbor change factors as the information change feature of the i+1-th scale relative to the i-th scale. The judgment begins with the information change characteristics of the second scale relative to the first scale. If the information change characteristics are less than the change threshold, the second scale is discarded, and the calculation continues for the information change characteristics of the third scale relative to the first scale. If the third scale is retained, the calculation is then performed based on the information change characteristics of the fourth scale relative to the third scale, and so on. Several scales are discarded based on the information change characteristics, and the remaining scales are retained.

6. The intelligent image recognition method for high-precision maps according to claim 5, characterized in that, The specific method for obtaining the nearest neighbor variation factor of each edge group at adjacent scales includes: For any edge group at the i-th scale, the chain code vectors of the two edges in the edge group are the same and are used as the chain code vector of the edge group. The edge closest to the edge group is obtained. Based on the mean Euclidean distance of the key points of the two edges in the edge group to other edges, the edge with the smallest mean Euclidean distance is taken as the edge closest to the edge group. The cosine similarity between the chain code vector of the edge group and the chain code vector of the edge closest to the edge group is obtained and used as the nearest neighbor similarity of the edge group at the i-th scale. In the (i+1)th scale, re-obtain the nearest edge of the edge group and re-obtain the nearest neighbor similarity of the edge group in the (i+1)th scale; obtain the absolute value of the difference between the nearest neighbor similarities of the edge group in two adjacent scales, and use it as the nearest neighbor change factor of the edge group in the (i)th and (i+1)th scales.

7. The intelligent image recognition method for high-precision maps according to claim 4, characterized in that, The specific method for marking several nodes is as follows: The scale with the fewest map features among the retained scales is taken as the maximum retained scale; Based on the distribution of edges in each edge group in the map distribution map of the maximum preservation scale, several combinations of regional edge groups are obtained; For any set of region edge groups, the region is divided using this set of region edge groups, and two edge groups constitute a divided region; each edge in the two edge groups is extended until the edges of the two edge groups intersect, and the intersecting part is marked as a node; For the original edges in the two edge groups, for any edge group, a preset division ratio is used to obtain several straight lines perpendicular to the two edges. Starting from the starting point of each of the two edges, a vertical straight line is extracted at each division ratio between the starting point and the next key point, which serves as the grid line in the edge group. This results in several grid lines. The intersection of the grid lines with the two edges is marked as a node.

8. The intelligent image recognition method for high-precision maps according to claim 7, characterized in that, The specific method for obtaining several sets of region edge groups is as follows: For each edge group in the map distribution map with the maximum retention scale, the nearest edge of each edge group is obtained, and based on the edge group corresponding to the nearest edge, the nearest edge group of each edge group is obtained, resulting in several combinations of the nearest edge groups; For any nearest edge group combination, obtain the directional difference between the two nearest edge groups. If the directional difference is greater than or equal to the difference threshold, the nearest edge group combination is taken as a set of region edge group combinations.

9. The intelligent image recognition method for high-precision maps according to claim 1, characterized in that, The specific method for filtering the retained paths of each node is as follows: For any node with two paths in an adjacent retention scale, the DTW distance is obtained by matching the two paths. The maximum DTW distance between the node and the paths in all retention scales is obtained. The ratio of the DTW distance corresponding to the adjacent retention scale to the maximum DTW distance is used as the path difference factor of the adjacent retention scale. A path difference threshold is preset. If the path difference factor is greater than the path difference threshold, the two paths of the node in the adjacent retention scale are retained. If the path difference factor is less than or equal to the path difference threshold, the path of the next retention scale in the adjacent retention scale is removed; the paths retained by the node at each retention scale are taken as several retained paths of the node.

Images