KR102477218B1 - Apparatus and method for generating High Definition Map - Google Patents

Abstract

The present application relates to a high-precision map generation device and a high-precision map generation method, and the high-precision map generation method according to an embodiment of the present invention recognizes a road sign of a road included in an aerial image by using an area box detection neural network, , generating an area box representing location information and direction information of the road surface sign; Recognizing road objects included in the aerial image using a semantic extraction neural network, specifying lanes on the road according to the input lane location information, and classifying the type of each lane; and generating a high-precision map corresponding to the aerial image using the road sign and lanes.

Description

High-precision map generating device and high-precision map generating method {Apparatus and method for generating High Definition Map}

The present application relates to a high-precision map generating device and a high-precision map generating method, and more particularly, to a high-precision map generating device and a high-precision map generating method capable of accurately recognizing and classifying various road signs or lanes from aerial images.

Unmanned autonomous vehicles (self-driving vehicles) can largely consist of a step of recognizing the surrounding environment, a step of planning a driving route from the recognized environment, and a step of driving along the planned route. In this process, the most basic What is required is a high definition map (HD-Map).

An HD-Map is a map in which information on the surrounding environment, such as roads, topography, and curvature, is implemented in 3D, and includes various information necessary for driving on the road. For example, the HD-Map includes various types of information necessary for driving on the road, such as road lanes, driving directions, intersections, signs, traffic lights, speed limits, and the like.

Since autonomous vehicles drive on the road while recognizing the surrounding environment based on the HD-Map, the HD-Map can be the most basic in the field of autonomous driving.

This application is intended to provide a high-precision map generating device and a high-precision map generating method capable of accurately recognizing and classifying various road signs or lanes from aerial images.

The present application is intended to provide a high-precision map generating device and a high-precision map generating method capable of more accurately determining the direction of a road sign recognized from an aerial image.

An object of the present application is to provide a high-precision map generation apparatus and method for quickly classifying various types of lanes included in an aerial image and at the same time precisely providing location information of lanes.

A method for generating a high-precision map according to an embodiment of the present invention recognizes a road sign of a road included in an aerial image using an area box detection neural network, and generates an area box indicating location information and direction information of the road surface sign. doing; Recognizing road objects included in the aerial image using a semantic extraction neural network, specifying lanes on the road according to the input lane location information, and classifying the type of each lane; and generating a high-precision map corresponding to the aerial image using the road sign and lanes.

A high-precision map generator according to an embodiment of the present invention recognizes a road surface sign included in an aerial image using an area box detection neural network, and generates an area box indicating location information and direction information of the road surface symbol. a road sign recognition unit; a lane recognition unit that recognizes road objects included in the aerial image using a semantic extraction neural network, identifies lanes on the road according to the input lane location information, and classifies each lane type; and a map generator generating a high-precision map corresponding to the aerial image using the road sign and lanes.

In addition, the solution to the above problem does not enumerate all the features of the present invention. Various features of the present invention and the advantages and effects thereof will be understood in more detail with reference to specific embodiments below.

According to the high-precision map generation apparatus and high-precision map generation method according to an embodiment of the present invention, it is possible to accurately recognize and classify various road signs or lanes from aerial images.

According to the high-precision map generation apparatus and high-precision map generation method according to an embodiment of the present invention, the direction of the name sign recognized from the aerial image is the direction of the road where the road sign is located or the direction of the road sign included in the same group. Since it can be corrected in consideration of , the direction of the road sign can be more accurately calculated.

According to the high-precision map generation apparatus and high-precision map generation method according to an embodiment of the present invention, it is possible to quickly classify various types of lanes included in an aerial image, and at the same time precisely provide location information on the lanes. It is possible.

However, the effects that can be achieved by the high-precision map generating device and the high-precision map generating method according to the embodiments of the present invention are not limited to those mentioned above, and other effects not mentioned are the present invention from the following description. It will be clearly understood by those skilled in the art.

1 is a block diagram showing a high-precision map generator according to an embodiment of the present invention.
2 is a schematic diagram showing road signs and lanes extracted by a high-precision map generator according to an embodiment of the present invention.
3 is a schematic diagram illustrating direction correction for a road sign according to an embodiment of the present invention.
4 is a schematic diagram illustrating lane recognition according to an embodiment of the present invention.
5 is a flowchart illustrating a high-precision map generation method according to an embodiment of the present invention.

Hereinafter, the embodiments disclosed in this specification will be described in detail with reference to the accompanying drawings, but the same or similar elements are given the same reference numerals regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "unit" for components used in the following description are given or used together in consideration of ease of writing the specification, and do not have meanings or roles that are distinct from each other by themselves. That is, the term 'unit' used in the present invention means a hardware component such as software, FPGA or ASIC, and 'unit' performs certain roles. However, 'part' is not limited to software or hardware. A 'unit' may be configured to reside in an addressable storage medium and may be configured to reproduce one or more processors. Thus, as an example, 'unit' refers to components such as software components, object-oriented software components, class components and task components, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and variables. The functionality provided within the components and 'parts' may be combined into a smaller number of elements and 'parts' or further separated into additional elements and 'parts'.

In addition, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of a related known technology may obscure the gist of the embodiment disclosed in this specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, the technical idea disclosed in this specification is not limited by the accompanying drawings, and all changes included in the spirit and technical scope of the present invention , it should be understood to include equivalents or substitutes.

Recently, as interest in self-driving has increased, the need for a high-definition map for self-driving cars has emerged. A high-precision map may include lanes, road signs, etc. necessary for driving an autonomous vehicle, and at this time, very high accuracy is required for location information such as each lane or road sign.

Conventionally, in order to increase the accuracy of a high-precision map, a high-precision map was generated by performing annotation work on road signs and lanes directly from an aerial image by a worker. However, skilled workers are required to accurately position and classify road signs and lanes included in aerial images, and a large number of workers are required to perform tasks in order to create high-precision maps of a large area. have. In other words, there were problems such as requiring a long time and a lot of manpower for the annotation work to create a high-precision map.

Regarding the production of a high-precision map, a plan using an image taken from a vehicle in motion has also been proposed. However, in this case, since the measured road sign or lane position is described centering on the vehicle to be measured, it may be directly used when driving a vehicle, but may not be suitable for creating a high-precision map. That is, in order to find the road sign or the geographic coordinate value of the lane, accurate location information of the vehicle to be measured is required, but the location of the vehicle is in an area where the GPS (Global Positioning System) is inaccurate (for example, an area where vehicles are densely populated). Since it is difficult to accurately grasp, it is also difficult to find out the exact geographic coordinate values of road signs or lanes. In addition, when the vehicle is covered by another vehicle or the angle of view of the photographing device is not displayed, there is a problem in that the road sign or lane position cannot be found regardless of the performance of the detection device.

On the other hand, according to the high-precision map generating apparatus according to an embodiment of the present invention, since it is possible to automate the high-precision map annotation work, it is possible to reduce the time and cost required for the annotation work. In addition, since the high-precision map generator according to an embodiment of the present invention recognizes road signs or lanes based on aerial images, it is possible to solve problems that may occur when using images taken from a vehicle in motion.

That is, since the high-precision map generating device according to an embodiment of the present invention uses an aerial image of which a one-to-one correspondence between image coordinates and geographic coordinates is known in advance, the actual geographic coordinate value of the road sign detected in the aerial image can be easily grasped. It is possible. In addition, in the case of aerial video, it is created by synthesizing images taken at various points of time and from various shooting angles. A corresponding road sign or lane may be displayed in the synthesized aerial image. Therefore, since the conventional obscuring of road signs can be reduced using aerial images, the location of road signs or lanes can be easily found. Hereinafter, a high-precision map generator according to an embodiment of the present invention will be described.

1 is a block diagram showing a high-precision map generator according to an embodiment of the present invention. Referring to FIG. 1 , a high-precision map generating device 100 according to an embodiment of the present invention may include a road sign recognition unit 110, a lane recognition unit 120, and a map generator 130.

The road sign recognition unit 110 may recognize a road sign included in the aerial image (A) by using an area box detection neural network, and may generate an area box indicating location information and direction information of the road surface sign. .

Here, the area box may be an Oriented Bounding Box (OBB) having directionality, as shown in FIG. 2(a), and the area box detection neural network may generate the area box using the SCRDet algorithm. there is.

Specifically, in the case of using the SCRDet algorithm, a feature map for the aerial image (A) can be calculated using a convolution network, and then the feature map is input to the area proposal neural network, Candidate locations (ROIs, Regions Of Interest) in which a road sign may exist in the image A may be calculated. The feature map including the candidate location is again input to the convolutional neural network to perform area box regression and class classification. In this process, correction of each candidate location area and prediction of the direction of the road sign are performed. That is, the area box may be created by determining the location information and direction information of the road sign.

However, when recognizing road signs in the aerial image (A), some road signs may be mistakenly recognized as having different directions even though they are objects having the same direction as shown in FIG. 3(a). there was. To solve this problem, the road sign recognition unit 110 may correct the direction of the road sign so as to accurately recognize the direction.

Specifically, as shown in FIG. 3(b), a road may generally have a directionality, and the directionality of the road may have a high correlation with the direction of a road sign. Accordingly, the road sign recognizing unit 110 may determine the direction of the road included in the aerial image A, and then correct the direction of the road sign in consideration of this, as shown in FIG. 3(c).

Here, the direction information of the road may be generated using a road prediction neural network, and the road prediction neural network may output a matrix including predicted values of direction angles to which each pixel in the aerial image (A) is heading. In this case, the road prediction neural network may be trained to reduce the L2 loss with respect to the ground truth direction for each pixel in the road area using sample aerial images. Accordingly, the direction of the road included in the aerial image A can be checked, and the road sign recognizing unit 110 can correct the direction of the road sign in consideration of the direction of the road.

Depending on the embodiment, it is also possible for the road sign recognition unit 110 to correct the direction of a corresponding road sign in consideration of the direction of the road and the directions of adjacent road signs. That is, as shown in FIG. 3(c), a plurality of road signs may be adjacently arranged for each lane, and they may be grouped into groups of road signs. Specifically, the road sign recognition unit 110 determines that the plurality of road signs belong to the same group when the distance between adjacent road signs is equal to or less than a predetermined reference value and the difference in direction angle between adjacent road signs is equal to or less than a predetermined reference value. can be seen as

Thereafter, representative direction information of each grouped road sign may be generated using an average value of direction information of road signs included in the same group and an average value of road direction information. Thereafter, representative direction information may be set as direction information of each grouped road sign. That is, all road signs included in the same group may be set to have the same direction.

Specifically, as described in Equation 1 below, after obtaining the average value (θ _b ) of direction information of road signs and the average value (θ _r ) of road direction information of roads, the average value (θ _b ) of direction information of road signs ) and the weight (1-λ) for the average value (θ _r ) of the road direction information of the road, the representative direction information (θ) can be calculated.

[Equation 1]

θ = λ*θ _b + (1- λ)* θ _r (where 0 ≤ λ ≤ 1)

Here, the weight (λ) may be a value that varies depending on the characteristics of the aerial image (A), and representative direction information may be calculated while adjusting the weight (λ) accordingly. Furthermore, by constructing a neural network through learning about the weights (λ), it is possible to calculate the optimal representative direction information by applying the optimal weights (λ) according to the input aerial image (A).

The lane recognizing unit 120 recognizes road objects included in the aerial image (A) by using a semantic extraction neural network, and identifies lanes on the road according to the input lane location information (P), thereby specifying the type of each lane. can be classified.

Specifically, as shown in FIG. 4, the lane recognition unit 120 may classify each pixel included in the aerial image A as one of a plurality of road objects using a meaning extraction neural network, A semantic map (M) including objects can be created. Here, the road object may include a background, a road, a lane, a road surface sign, and the like, and other various objects that may be included in the aerial image (A) may also be included in the road object. Depending on embodiments, it is also possible to further subdivide road objects according to lane types.

The meaning extraction neural network calculates the probability that each pixel corresponds to each road object and generates a feature vector for each pixel. can be classified.

Here, the semantic extraction neural network may be pre-learned to generate a semantic map (M) corresponding to an input aerial image (A). Specifically, when a sample aerial image is input, the deep neural network can output a corresponding pseudo semantic map, and the deep neural network can be trained to reduce the error between the pseudo semantic map and the ground truth semantic map. there is. Here, as a loss function for learning, cross-entropy or focal loss may be used. Accordingly, the deep neural network classifies the pixels included in the input aerial image (A) into background, road, lane, road sign, etc., and can give each meaning.

Thereafter, the lane recognizing unit 120 may receive lane location information (P) representing the locations of lanes, set pixels corresponding to the lane location information (P) in the semantic map (M) as lanes, and set the lanes types can be classified.

Here, the lane location information (P) may be created by a worker identifying each lane included in the aerial image (A) and then specifying the location of the corresponding lane appearing in the aerial image (A). That is, in order to increase the accuracy of the lane location, the operator may directly specify the location of each lane to generate the lane location information P.

In this case, the lane location information (P) can be generated separately from the high-precision map generating device 100 by the operator, and the lane recognition unit 120 receives the lane location information (P) for the corresponding aerial image (A) from the outside. can receive However, depending on the embodiment, it is also possible for the high-precision map generating device 100 to provide an interface for generating the lane location map P to the operator. For example, after displaying the input aerial image A to the operator, the operator may provide an interface for specifying the positions of corresponding lanes shown in the aerial image A, or the like.

That is, as shown in FIG. 2(a), it is possible to classify lanes using the semantic map M, but the reliability of location information for pixels classified as lanes in the semantic map M is relatively high. can be as low as Accordingly, the lane recognizing unit 120 may specify the position of a lane within the semantic map M by utilizing the lane location information P input by the operator.

Here, when the lane location information P is provided as a set of points, the lane location information P can be converted into a mask form, and the mask is applied to the semantic map M, so that each lane A set M' of corresponding pixels may be extracted.

Thereafter, the lane recognizing unit 120 may generate lane feature vectors corresponding to respective lanes by sampling feature vectors of the extracted pixels. That is, a set of pixels corresponding to a lane in the semantic map M can be specified using the lane feature vector. Here, the lane recognizing unit 120 may generate the lane feature vector by averaging feature vectors of the extracted pixels or applying various sampling techniques to them.

Here, since the lane feature vectors are generated corresponding to the number of lanes included in the aerial image (A), when the operator generates lane location information (P) for n lanes, a total of n lane feature vectors are generated. It can be.

Thereafter, the lane recognizing unit 120 may discriminate the type of each lane included in the semantic map M by using the lane feature vector. Here, the types of lanes may be variously classified into a center line, a stop line, a course change restriction line, and the like, and the lane recognizing unit 120 may determine to which category each lane corresponding to the lane feature vector corresponds. In this case, the type of lane may be set in advance according to laws and regulations such as the Road Traffic Act.

Meanwhile, depending on the embodiment, n number of lane feature vectors may be input to the lane recognizing unit 120 . For example, if n number of lanes exist in the aerial image A, n number of lane feature vectors may be generated. In this case, the lane recognizing unit 120 outputs n number of lane feature vectors, respectively. It is possible to output a probability vector including a probability distribution that may correspond to the sub-classification of . That is, a probability vector including a probability corresponding to the center line, a probability corresponding to a stop line, a probability corresponding to a course change restriction line, and the like may be output.

Here, the lane recognizing unit 120 may further include a neural network, which may be previously learned for lane classification. That is, when the sample lane feature vectors are input, the corresponding pseudo probability vector is output, and learning is performed to reduce the error between the ground truth probability vector and the pseudo probability vector corresponding to each sample lane feature vector. can Therefore, when the lane feature vector is input, the lane recognizing unit 120 may output a corresponding probability vector, and select a classification having the highest probability value among probabilities corresponding to each classification included in the probability vector for the corresponding lane. can be selected as the final classification.

Additionally, depending on the embodiment, there may be cases in which the lane location information P is not input to the lane recognizing unit 120, and in this case, the lane recognizing unit 120 may specify a lane from the semantic map M. can Specifically, the lane recognizing unit 130 may extract target pixels classified as lanes from the semantic map M, as shown in FIG. 2(b), when lane location information P is not input. Target pixels may be grouped and specified as lanes. That is, a pixel group may be formed by grouping target pixels located within a set distance from among a plurality of target pixels, and a linear lane may be specified by performing vector simplification on the corresponding pixel group. Here, a Douglas-peucker technique or the like can be used for simplification.

After specifying the lane, the method of extracting the lane feature vector and classifying the lane type is the same as described above, so a detailed description thereof is omitted here.

The map generator 130 may generate a high-precision map (B) corresponding to the aerial image (A) by using road signs and lanes.

Here, each road sign and lane may be specific according to the image coordinate system in the aerial image (A). However, since the high-precision map (B) must be provided according to the actual geographic coordinate system, the map generator 130 can convert the coordinate values of each road sign and lanes into the actual geographic coordinate system, and using this, the high-precision map (B ) can be created. Accordingly, the self-driving vehicle or the like can perform precise self-positioning even in a GPS shadow area by using road signs or lanes displayed on the high-precision map (B). For example, when correcting the positioning of an autonomous vehicle, position information of lanes may be used to correct accuracy in the horizontal axis direction, and road signs may be used to correct accuracy in the vertical axis direction.

In addition, since the self-driving vehicle recognizes the current location, even when the recognition device such as the sensor of the self-driving vehicle does not recognize the road surface sign or lane on the road, it uses a high-precision map to detect the surrounding road surface sign or lane. Able to know. Therefore, even in this case, it is possible to drive in compliance with traffic laws.

Additionally, since the road sign recognition unit 110 and the lane recognition unit 120 individually recognize road signs and lanes, road signs and lanes may overlap in the high-precision map. That is, a case may occur in which one of a road surface sign and a lane is erroneously recognized.

In this case, the map generator 130 may leave objects with high reliability among the recognized road signs and lanes, and delete objects with low reliability. Depending on the embodiment, reliability of each recognized road sign and lane may be provided from an area box extraction neural network or a meaning extraction neural network. Accordingly, the map generator 130 may select an object to be deleted by comparing them.

However, since the lane recognition unit 120 specifies the position of the lane by utilizing the lane location information (P) in which the operator precisely identifies the position of the lane, it is generally considered that the reliability of the lane is higher than the reliability of the road surface sign. can Therefore, depending on the embodiment, when a lane and a road sign overlap, it is also possible to set to delete only the road sign while leaving the lane.

Furthermore, depending on the embodiment, the map generator 130 may mix the semantic map and the area box to further improve the reliability of the road sign. Specifically, the map generator 130 may project the area box provided from the road sign recognition unit 110 into the semantic map M. Thereafter, the feature map of the region corresponding to the region box in the semantic map M can be extracted, and the feature vector of each pixel included in the feature map can also be extracted. In this case, the map generator 130 may check whether the road object corresponding to the feature map corresponds to the road surface sign, and if the road object corresponding to the feature map does not correspond to the road surface sign, the road surface sign recognition unit 110 ) can delete the recognized road sign. That is, since it is determined that the corresponding area is not a road sign in the semantic map (M), the area box generating neural network may consider it to be mistakenly recognized and delete it.

5 is a flowchart illustrating a method for generating a high-precision map according to an embodiment of the present invention. Here, each step of the high-precision map generating method may be performed by a high-precision map generating device.

Specifically, the high-precision map generating device may recognize a road sign included in an aerial image using an area box detection neural network, and generate an area box indicating location information and direction information of the road surface sign (S10). Here, the area box may be an Oriented Bounding Box (OBB) having directionality, and the area box detection neural network may use the SCRDet algorithm.

Here, there may be a case where the direction of the road sign is incorrectly recognized, and in this case, the high precision map generating device can correct the direction of the road sign. Specifically, the high-precision map generation device may generate road direction information by recognizing the direction of a road included in an aerial image, and may correct direction information of a road surface sign using the road direction information.

Depending on embodiments, it is also possible to correct the direction of a corresponding road sign in consideration of the direction of the road and the directions of adjacent road signs. That is, the high-precision map generator can group adjacent road signs using location information and direction information of road signs, and grouping using the average value of direction information of each grouped road sign and the average value of road direction information. It is possible to generate representative direction information of each road sign. Thereafter, representative direction information may be set as direction information of each of the grouped road signs.

In addition, the high-precision map generation device recognizes road objects included in the aerial image using a semantic extraction neural network, and can classify the type of each lane by specifying the lanes on the road according to the input lane location information (S20). ).

Specifically, the high-precision map generation device may classify each pixel included in the aerial image as one of a plurality of road objects using a semantic extraction neural network, and may generate a semantic map including the road objects.

Then, when lane location information indicating the location of lanes is received, pixels corresponding to the lane location information may be set as lanes in the semantic map, and types of the set lanes may be classified. Here, the lane location information may be created by a worker identifying each lane included in the aerial image and then specifying the location of the corresponding lane appearing in the aerial image. That is, in order to increase the accuracy of the lane position, the operator may directly specify the position of each lane to generate lane position information.

In this case, the lane location information can be generated separately from the high-precision map generating device by the operator, and the lane location information for the corresponding aerial image can be received from the outside. However, depending on the embodiment, it is also possible for the high-precision map generating device to provide an interface for generating a lane location map to the operator.

Meanwhile, when lane location information indicating the location of lanes is received, feature vectors of pixels corresponding to the lane location information may be extracted from a semantic map, and lane feature vectors corresponding to lanes may be generated by sampling the extracted feature vectors. can Thereafter, the types of lanes may be distinguished using the lane feature vectors.

Additionally, depending on embodiments, lane position information corresponding to some aerial images may not be input. In this case, the high-precision map generator can specify lanes from the semantic map. That is, when lane location information is not input, target pixels classified as lanes may be extracted from the semantic map, and target pixels may be grouped and specified as lanes.

Specifically, among a plurality of target pixels classified into lanes, target pixels positioned within a set distance from each other may be grouped to form a pixel group. Since a plurality of lanes may exist in an aerial image, in order to form a pixel group for each lane, target pixels positioned within a set distance from each other may be grouped into the same pixel group.

Thereafter, each lane may be specified by performing vector simplification on the pixel group. Since the pixel group may not appear as a single straight line, simplification may be performed for straightening, and the Douglas-peucker technique may be used depending on the embodiment.

Meanwhile, although FIG. 5 shows that the step of classifying the type of lane (S20) is performed after the step of generating the area box (S10), the present invention is not limited thereto, and the step of classifying the type of lane is not limited thereto. It is also possible that (S20) is performed first, or that the step of generating area boxes (S10) and the step of classifying lane types (S20) are simultaneously performed in parallel.

Thereafter, the high-precision map generating device may generate a high-precision map corresponding to the aerial image using road signs and lanes (S30). Here, each road sign and lane may be specified using an image coordinate system in an aerial image. However, since the high-precision map must be provided in an actual geographic coordinate system, the high-precision map generator can generate the high-precision map B by converting the coordinate values of each road sign and lanes into the actual geographic coordinate system.

On the other hand, each road sign and lane recognized by the high-precision map generator may overlap in the high-precision map. That is, any one of the road sign and lane may be recognized incorrectly. In this case, the high-precision map generating device may retain objects with high reliability and delete objects with low reliability among the recognized road signs and lanes. Since the reliability of each road sign and lane can be provided from an area box extraction neural network or a meaning extraction neural network, an object to be deleted can be selected by comparing them. However, in the case of a lane, since the operator specifies the location of the lane by using the lane location information (P) that precisely specifies the location of the lane, the reliability of the lane can generally be regarded as higher than the reliability of the road surface sign. Therefore, depending on the embodiment, when a lane and a road sign overlap, it is also possible to set to delete only the road sign while leaving the lane.

Furthermore, depending on the embodiment, it is possible to further improve the reliability of the road sign by mixing the semantic map and the area box. Specifically, a feature map of a region corresponding to the region box may be extracted from the semantic map by projecting the region box into the semantic map. Then, it is possible to check whether the road object corresponding to the feature map corresponds to the road surface sign by extracting the feature vector of each pixel included in the feature map. If the road object corresponding to the feature map does not correspond to the road surface sign, The road sign may be deleted. That is, since it is determined that the corresponding area is not a road sign in the semantic map, the area box generating neural network may consider it to be erroneously recognized and delete it.

The above-described present invention can be implemented as computer readable code on a medium on which a program is recorded. The computer-readable medium may continuously store programs executable by the computer or temporarily store them for execution or download. In addition, the medium may be various recording means or storage means in the form of a single or combined hardware, but is not limited to a medium directly connected to a certain computer system, and may be distributed on a network. Examples of the medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROM and DVD, magneto-optical media such as floptical disks, and ROM, RAM, flash memory, etc. configured to store program instructions. In addition, examples of other media include recording media or storage media managed by an app store that distributes applications, a site that supplies or distributes various other software, and a server. Accordingly, the above detailed description should not be construed as limiting in all respects and should be considered illustrative. The scope of the present invention should be determined by reasonable interpretation of the appended claims, and all changes within the equivalent scope of the present invention are included in the scope of the present invention.

The present invention is not limited by the foregoing embodiments and accompanying drawings. It will be clear to those skilled in the art that the components according to the present invention can be substituted, modified, and changed without departing from the technical spirit of the present invention.

100: high-precision map generator
110: road sign recognition unit
120: lane recognition unit
130: map generator

Claims (13)

Recognizing a road surface included in the aerial image by using an area box detection neural network, and generating an area box indicating location information and direction information of the road surface symbol;
Recognizing road objects included in the aerial image using a semantic extraction neural network, specifying lanes on the road, and classifying the type of each lane; and
Generating a high-precision map corresponding to the aerial image using the road sign and lanes,
The step of creating the area box is
The high-precision map generation method characterized by recognizing the direction of the road included in the aerial image to generate road direction information, and correcting the direction information of the road surface sign using the road direction information.
delete The method of claim 1, wherein the step of generating the area box
Adjacent road surface signs are grouped using the location information and direction information of the road surface sign, and the average value of the direction information of each grouped road surface sign and the average value of the road direction information are used to determine the number of the grouped road signs. A high-precision map generating method characterized in that generating representative direction information and setting the representative direction information as direction information of each of the grouped road signs.
The method of claim 1, wherein the classifying the type of lane
and classifying each pixel included in the aerial image as one of a plurality of road objects using the semantic extraction neural network, and generating a semantic map including the road objects.
The method of claim 4, wherein the classifying the type of lane
The method of generating a high-precision map, characterized in that, when lane location information indicating the location of the lanes is received, pixels corresponding to the lane location information are set as the lanes in the semantic map, and types of the set lanes are classified.
The method of claim 5, wherein the classifying the type of lane
Feature vectors of pixels corresponding to the lane location information are extracted from the semantic map, lane feature vectors corresponding to the lanes are generated by sampling the extracted feature vectors, and types of the lanes are used by using the lane feature vectors. High-precision map generation method, characterized in that for distinguishing.
The method of claim 4, wherein the classifying the type of lane
The method of generating a high-precision map, characterized in that, if lane location information indicating the location of the lanes is not input, target pixels classified as the lanes are extracted from the semantic map, and the target pixels are grouped and specified as the lanes.
The method of claim 7, wherein the classifying the type of the lane
Among a plurality of target pixels, target pixels located within a set distance from each other are grouped to form a pixel group, and vector simplification is performed on the pixel group to specify each lane. Way.
The method of claim 1, wherein generating the high-precision map
The high-precision map generation method of claim 1 , wherein the road-surface sign is deleted if the road-surface sign and the position of the lane overlap each other in the high-precision map.
The method of claim 4, wherein generating the high-precision map
The area box is projected into the semantic map, a feature map of the area corresponding to the area box is extracted, and a road object corresponding to the feature map is obtained by using a feature vector of each pixel included in the feature map. A method for generating a high-precision map, characterized in that for confirming whether it corresponds to a road surface sign.
11. The method of claim 10, wherein generating the high-precision map
and deleting the road surface sign if the road object corresponding to the feature map does not correspond to the road surface sign.
A computer program stored in a medium to execute the high-precision map generation method of any one of claims 1, 3 to 11 in combination with hardware.
a road sign recognition unit for recognizing a road surface included in an aerial image using an area box detection neural network and generating an area box indicating location and direction information of the road surface;
a lane recognition unit for recognizing road objects included in the aerial image using a semantic extraction neural network, specifying lanes on the road, and classifying each lane type; and
It includes a map generator for generating a high-precision map corresponding to the aerial image using the road sign and lanes,
The road surface sign recognition unit
The high-precision map generating device characterized in that for generating road direction information by recognizing the direction of the road included in the aerial image, and correcting the direction information of the road surface sign using the road direction information.

Images