A high-precision map intelligent correction and updating method based on multi-modal data online learning optimization

Abstract

The application discloses a high-precision map intelligent correction and updating method based on multi-modal data online learning optimization, and the application realizes unified sequence representation of map information, improves the organization and processing efficiency of map data by constructing the topology structure of the map reference line and the sequence point set discretization; and the application realizes accurate perception and fusion of the mine road boundary by adopting CNN semantic segmentation and inverse perspective transformation, combining with a pose graph optimization method, and improves the spatial precision of map data; and the application deduces the deviation state of the real-time motion track of the vehicle by using Frenet coordinate system transformation, effectively represents the error amount of the real-time running route of the vehicle and the map reference line, and improves the accuracy of the map updating learning model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a high-precision map intelligent correction and updating method based on multi-modal data online learning optimization. BACKGROUND

[0002] With the rapid development of automatic driving technology, high-precision maps, as an important part of the automatic driving system, play a crucial role in the safety and reliability of the automatic driving system. In specific application scenarios such as open-pit mines, due to the mixed structure and unstructured roads and frequent changes in terrain, higher requirements are placed on the accuracy and update frequency of high-precision maps.

[0003] Traditional high-precision map making and updating methods mainly rely on manual geographic surveys and measurements, which not only have high costs but also have long update cycles, making it difficult to meet the needs of dynamic environments such as open-pit mines. In addition, the harsh environment of the mine area and the safety hazards of manually driving mine trucks further highlight the potential benefits of automatic driving technology in the mine area.

[0004] In recent years, some research has begun to explore map updating techniques based on vehicle GNSS trajectories and data mining methods. For example, Wang et al. proposed a map updating method based on historical vehicle trajectories and geometric features. However, these methods are often only suitable for simple scene map boundary expansion and road network changes when dealing with complex terrain changes in open-pit mines, and the accuracy of the updated map elements is insufficient. SUMMARY

[0005] The purpose of the present application is to overcome the technical defects in the prior art and provide a high-precision map intelligent correction and updating method based on multi-modal data online learning optimization.

[0006] The technical solution adopted to achieve the purpose of the present application is as follows:

[0007] A high-precision map intelligent correction and updating method based on multi-modal data online learning optimization, comprising the following steps:

[0008] Step 1: According to the OpenDrive specification, the road structure, road structure connection relationship, and polynomial ParamPoly3 provided by the initial high-precision map static text are used to complete the topological connection relationship construction of different road reference lines of the initial map, and discrete point sequences (u i , v i ) in the local uv coordinate system are obtained, with the starting point (x0, y0) of each road structure as the discrete point sequence (u i , v iFind the origin of the map and obtain the map reference line sequence (x) in the absolute coordinate system. i y i );

[0009] Step 2: Use the CNN semantic segmentation method to perceive the road boundary lines of the original image and obtain the semantic pixel positions [u, v]. Then, use inverse perspective transformation to backproject the semantic pixel positions [u, v] from the image plane to obtain the road boundary feature positions [x, v] in the vehicle coordinate system. v ,y v ,z v ];

[0010] Step 3: Using a pose graph optimization method combined with vehicle IMU and GPS data, the precise pose of the vehicle is obtained. Based on the precise pose of the vehicle, the road boundary feature sequence in the vehicle coordinate system is calculated. Convert to road boundary sequence in global coordinate system (x i w y i w, , z i w );

[0011] Step 4: Through Frenet coordinate system transformation, the real-time vehicle trajectory sequence (x, y) and the map reference line sequence (x, y) are compared. i y i The path deviation between the two is used to derive the vehicle's real-time motion trajectory sequence offset state quantity (S). i L i );

[0012] Step 5, convert the road boundary sequence (x) in the global coordinate system. i w y i w ,,z i w ) and map reference line sequence (x i y i The resulting map information sequence is then fused with the vehicle's real-time trajectory sequence offset state quantity (S). i L i Sequence fusion is performed to obtain a multimodal sequence vector s. i Multimodal sequence vectors are combined to create a corresponding multimodal information sequence X for the vehicle's real-time motion trajectory sequence. i k ={s i ,s i+1 ,…,s i+k-2 ,s i+k-1};

[0013] Step 6, the dynamic multi-head self-attention Transformer module learning network is trained by using the multi-modal information sequence obtained in step 5, attention mechanism modeling is performed, sequence features are output, and an updated map reference line sequence is obtained and a road boundary sequence, and a loss function is used for accuracy quantification

[0014] Step 7, in an actual running scene of the vehicle, the trained learning network is verified, the structure and parameters of the learning network are adjusted according to the verification result, the map road curve is updated in real time as the initial map for the next time, and the updating cycle is repeated.

[0015] In the technical scheme, in step 1, according to the OpenDrive specification, the initial mine site map is first based on the OpenDrive specification <road>mid- multiple segments <geometry>the start coordinate S of the reference line of each segment is obtained by the element S coordinate, the heading angle hdg and the length length start and the end coordinate S end and is evenly divided and dispersed, and each segment is reconstructed <geometry>sequence S i = S start + i / n - (S end - S start ), where i is the index of the midpoint of the sequence S i and n is the number of segments <geometry>are divided into n discrete points;

[0016] Combining the coefficients aU, bU, cU, dU, aV, bV, cV, dV extracted in the corresponding polynomial ParamPoly3, the discrete point sequence (u i , v i ) of the reference line in the local uv coordinate system is obtained:

[0017]

[0018] wherein w i represents the intermediate parameter point of the S i sequence corresponding to the conversion to the polynomial ParamPoly3 in the local uv coordinate system for calculating u i , v i ;

[0019] and based on the discrete point sequence (u i , v i ) to calculate the intermediate parameter point w <geometry>The sequence of discrete points (x0, y0) of the structure is the origin, and the sequence of absolute coordinate system map reference lines (x i , y i ) is obtained:

[0020]

[0021] In the above technical solution, in step 2, the semantic pixel position [u, v] is mapped from the image plane to the ground plane in the vehicle coordinate system by inverse perspective transformation, to obtain the road boundary feature position in the vehicle coordinate system:

[0022]

[0023] Where, π c (·) represents the distortion and projection model of the camera, is the inverse projection function, which projects the pixel into space, [R c t c ] is the extrinsic matrix of the camera relative to the vehicle center, [u, v] is the semantic pixel position, [x v ,y v ,z v ] is the road boundary feature position in the vehicle coordinate system, λ is a scalar, [] col:i represents taking the i-th column of the matrix, i is 1, 2 and 4.

[0024] In the above technical solution, in step 3, the vehicle maps the collected semantic features from the vehicle coordinate system to the global coordinate system through the optimized pose information, and the global pose of the vehicle at time i is represented by the position p i and the rotation R(q i ), wherein R(q i ) is a function of converting the quaternion q i into a rotation matrix, and the sequence of road boundary features in the vehicle coordinate system can be mapped to the sequence of road boundary features in the global coordinate system by the following transformation:

[0025]

[0026] Through the above transformation, the locally observed features are integrated into a unified global map.

[0027] In the above technical solution, in step 4, the sequence of real-time motion trajectory offset state quantities (S i , L i ) of the vehicle at time i is:

[0028]

[0029] Where, θ is the sequence of map reference lines (x i , y i ) at the tangent angle, α, represents the angle between the current position of the vehicle and the tangent direction of the reference line, and (x, y) is the real-time motion trajectory sequence of the vehicle.

[0030] In the above technical solution, in step 5, the multi-modal sequence is matched with the sequence number of the map reference line, and the multi-modal sequence is truncated according to k steps, and the multi-modal sequence with the starting sequence number i is represented as X i k = {s i ,s i+1 ,…,s i+k-2 ,s i+k-1}, each multi-modal sequence vector s i includes the map reference line sequence (x i , y i ) of the vehicle at the sequence step, the road boundary sequence (x i w , y i w , z i w ) in the global coordinate system, the real-time motion trajectory sequence offset state quantity (S i , L i ) of the vehicle, and s i＝ {x i , y i , x i w , y i w , z i w , S i , L i} is obtained by fusion.

[0031] In the above technical solution, in step 6, the dynamic multi-head self-attention Transformer module includes three independent attention "heads", which respectively learn the semantic expression of the reference line, the boundary, and the road terrain in the sequence, and the outputs of the three independent attention "heads" are combined and spliced to obtain:

[0032] MultiHead(Q,K,V)=Concat(head1,

[0033] ...,head h )W O

[0034] head i =Attention(QW i Q ,KW i K ,VW i V )

[0035]

[0036] Wherein, MultiHead represents multi-head attention, Q, K and V represent query, key and value matrix respectively, Concat represents feature concatenation, head h is the hth attention head, W O is the weight matrix multiplied before the output of the multi-head self-attention, head i is the ith attention head, QW i Q ,KW i K and VW i V are the weight matrices of Q, K and V, Attention is the attention calculation, and softmax is the normalization means, d k is the dimension of the key vector, K T is the transpose of K, QK T is the attention score, which determines the weight of each value when generating the output, and this mechanism enables the dynamic multi-head self-attention Transformer module to process sequences in parallel, significantly improving the computational efficiency.

[0037] In the above technical solution, in step 6, the accuracy of the updated map reference line sequence and the road boundary sequence is quantified by using mean square error (MSE), and the accuracy is evaluated by the following error: Wherein, L map is the loss function of the network, is the updated map reference line sequence, p i is the real coordinate of the vehicle, N is the total number of map discrete points, and sigma is a learnable noise parameter.

[0038] Compared with the prior art, the beneficial effects of the present application are:

[0039] 1. The present application realizes the unified sequence representation of map information by constructing the topological structure of the map reference line and the discrete sequence point set, and improves the organization and processing efficiency of map data.

[0040] 2. The present application realizes accurate perception and fusion of mine road boundaries by using CNN semantic segmentation and inverse perspective transformation combined with pose graph optimization method, and improves the spatial accuracy of map data.

[0041] 3. The application utilizes Frenet coordinate system transformation to deduce the real-time motion trajectory deviation state of the vehicle, effectively represents the error amount of the real-time running route of the vehicle and the reference line of the map, and improves the accuracy of the map updating learning model.

[0042] 4. The application utilizes the learning network of the Transformer, can fully utilize multi-modal data, continuously optimizes the map updating result through online learning, and improves the intelligent level of map updating.

[0043] 5. L map is the loss function of the network, mainly minimizing the error of the predicted map reference line sequence and road boundary sequence and the noise part of the learning input data.

[0044] 6. The application verifies the effectiveness of the algorithm in the actual running scene of the vehicle, realizes high-precision and real-time map updating, and meets the strict requirements of the autonomous vehicle on the high-precision map. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the overall method architecture of the application. DETAILED DESCRIPTION

[0046] The application will be further described in detail below in combination with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0047] A high-precision map intelligent correction and updating method based on multi-modal data online learning optimization, comprising the following steps:

[0048] Step 1, according to the OpenDrive specification, using the road structure provided by the initial mine high-precision map static text, the connection relationship of the road structure and the polynomial, the topological connection relationship construction of different road reference lines of the initial map is completed, and the discrete point sequence (u i , v i ) in the local uv coordinate system is obtained, taking the starting point (x0, y0) of each road structure as the origin of the discrete point sequence (u i , v i ), the map reference line sequence (x i , y i ) in the absolute coordinate system is obtained;

[0049] Specifically, according to the OpenDrive specification, first, based on the initial mine high-precision map text <road>Provided is <link> In <predecessor>and <successor>information, characterizing the binding <road>and <junction>The road link information inside the structure can be used to explicitly construct the topological connection relationship of different roads in the mine. Then, the multi-segment <geometry>in the <type>Get the corresponding reference line line type (using a typical ParamPoly3 as an example), then use the multi-segment <geometry>the start coordinate S of the reference line of each segment is obtained from the element S coordinate, the heading angle hdg and the length length start and the end coordinate S end and is evenly divided and dispersed, and each segment is reconstructed <geometry>sequence S i = S start + i / n - (S end - S start ), where i is the index of the midpoint of the sequence S i and n is the number of segments <geometry>are divided into n discrete points;

[0050] In combination with the coefficients aU, bU, cU, dU, aV, bV, cV, dV extracted in the corresponding polynomial ParamPoly3, the discrete point sequence (u i , v i ) of the reference line in the local uv coordinate system is obtained:

[0051]

[0052] where w i represents the intermediate parameter point in the polynomial ParamPoly3 used to calculate u i , v i converted to the local uv coordinate system by the S i sequence;

[0053] And based on the discrete point sequence (u i , v i ), the map reference line sequence (x i , y i ) in the absolute coordinate system is obtained with the starting point (x0, y0) of each road structure body as the origin:

[0054]

[0055] Step 2, using the CNN semantic segmentation method to perceive the road boundary line in the original image, obtaining the semantic pixel position [u, v], and through inverse perspective transformation, the semantic pixel position [u, v] is back projected from the image plane to obtain the road boundary feature position [x v , y v , z v ] in the vehicle coordinate system;

[0056] Specifically, the semantic pixel position [u, v] is mapped from the image plane to the ground plane in the vehicle coordinate system through inverse perspective transformation, to obtain the road boundary feature position [x v , y v , z v ] in the vehicle coordinate system:

[0057]

[0058] where π c (·) represents the distortion and projection model of the camera, is the back projection function that projects the pixel into space, [R c t c ] is the extrinsic matrix of the camera relative to the vehicle center, [u, v] is the semantic pixel position, [x v , y v , z v The road boundary features in the vehicle coordinate system, where λ is a scalar. col:i This indicates taking the i-th column of the matrix, where i is 1, 2, or 4.

[0059] Step 3: Using a pose graph optimization method combined with vehicle IMU and GPS data, the precise pose of the vehicle is obtained. Based on the precise pose of the vehicle, the road boundary feature sequence in the vehicle coordinate system is calculated. Convert to road boundary sequence in global coordinate system (x i w y i w, , z i w );

[0060] The vehicle uses optimized pose information to map the collected semantic features from the vehicle coordinate system to the global coordinate system, where the road boundary sequence is as follows:

[0061] Suppose that the global pose of the vehicle at time i is determined by position p. i and rotation R(q) i ) represents, where R(q) i ) is the quaternion q i Functions that convert to rotation matrices; road boundary feature sequences in vehicle coordinates. The following transformations can be used to map the road boundary sequence to the global coordinate system:

[0062]

[0063] This allows for the integration of locally observed features into a unified global map.

[0064] Step 4: Through Frenet coordinate system transformation, the real-time vehicle trajectory sequence (x, y) and the map reference line sequence (x, y) are compared. i y i The deviation between the two is used to derive the vehicle's real-time motion trajectory sequence offset state quantity (S). i L i );

[0065] The vehicle's real-time motion trajectory sequence offset state quantity (S) at time i i L i )for:

[0066]

[0067] Where θ is the map reference line sequence (x i y i The tangent angle at the current position of the vehicle, and a represents the included angle between the tangent direction of the reference line and the current position of the vehicle, and (x, y) is the real-time motion trajectory sequence of the vehicle (i.e. the current position of the vehicle), which involves the calculation of the position of the vehicle in the Frenet coordinate system and the tangent direction of the reference line, and the lateral and longitudinal positions of the vehicle relative to the reference line.

[0068] Step 5, fuse the road boundary sequence (x i w , y i w, , z i w ) in the global coordinate system and the map reference line sequence (x i , y i ) to obtain the map information sequence, and then fuse it with the real-time motion trajectory sequence of the vehicle (S i , L i ) and the sequence to obtain the multi-modal sequence vector s i , and the multi-modal sequence vector is combined to create the corresponding multi-modal information sequence X i k = {s i , s i+1 , …, s i+k-2 , s i+k-1} of the real-time motion trajectory sequence of the vehicle.

[0069] Specifically, the multi-modal information sequence corresponds to the map reference line sequence in sequence number, and the multi-modal sequence is truncated according to k steps, and the multi-modal information sequence with the starting sequence number i is represented as X i k = {s i , s i+1 , …, s i+k-2 , s i+k-1}, each multi-modal sequence vector s i includes the map reference line sequence (x i , y i ) of the vehicle in the sequence step, the road boundary sequence (x i w , y i w , z i w ) in the global coordinate system, the real-time motion trajectory sequence offset state (S i , L i ) of the vehicle, thereby obtaining s i＝ {x i , y i , x i w , y i w , z i w , S i , L i}。

[0070] Step 6, train the dynamic multi-head self-attention Transformer module learning network using the multi-modal information sequence obtained in step 5, model through attention mechanism, output sequence features, get updated map reference line sequence and road boundary sequence, and use loss function for accuracy quantification.

[0071] The learning network of dynamic multi-head self-attention Transformer includes three independent attention "heads", which learn the semantic expression of reference line, boundary and road terrain in sequence respectively, and the outputs of the three independent attention "heads" are combined and spliced to obtain:

[0072] MultiHead(Q,K,V)=Concat(head1,

[0073] ...,head h )W O

[0074] head i =Attention(QW i Q ,KW i K ,VW i V )

[0075]

[0076] Where MultiHead represents multi-head attention, Q, K and V represent query, key and value matrices respectively, Concat represents feature splicing, head h is the hth attention head, W O is the weight matrix multiplied before the output of multi-head self-attention, head i is the ith attention head, QW i Q ,KW i K and VW i V are the weight matrices of Q, K and V, Attention is attention calculation, softmax is normalization means, d k is the dimension of the key vector, K T is the transpose of K, QK T For attention score, the weight of each value when generating output is determined, and this mechanism enables the Transformer to process sequences in parallel, significantly improving the computational efficiency; the mean square error (MSE) is used to quantify the accuracy of the updated map reference line sequence and the road boundary sequence, for each map discrete point, the Euclidean distance square sum between the predicted coordinates and the true coordinates is calculated, and a learnable noise parameter sigma is added, so as to obtain the following loss function: Wherein, L map is the loss function of the network, is the updated map reference line sequence, p i is the true coordinates of the vehicle, N is the total number of map discrete points, and sigma is a learnable noise parameter.

[0077] Step 7, in the actual running scene of the vehicle, the trained learning network is tested, the test includes the maximum update error and the average update error of the road centerline and the boundary line, according to the test result, the structure and parameters of the learning network are adjusted, the map road curve is updated in real time, as the initial map for the next time, and the update cycle is repeated.

[0078] Wherein, in the embodiment, the vehicle can be a mining truck.

[0079] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.< / geometry> < / geometry> < / geometry> < / type> < / geometry> < / junction> < / road> < / successor> < / predecessor> < / road> < / geometry> < / geometry> < / geometry> < / geometry> < / road>

Claims

1. A high-precision map intelligent correction and update method based on online learning optimization of multi-modal data, characterized in that, Comprising the following steps: Step 1, according to the OpenDrive specification, using the road structure provided by the initial high-precision map static text, the connection relationship of the road structure and the polynomial ParamPoly3, the topological connection relationship construction of the different road reference lines of the initial map is completed, and the discrete point sequence (u i , v i ) in the local uv coordinate system is obtained. Taking the starting point (x0, y0) of each road structure as the origin of the discrete point sequence (u i , v i ), the map reference line sequence (x i , y i ) in the absolute coordinate system is obtained; Step 2, the road boundary line of the original image is perceived by using a CNN semantic segmentation method to obtain semantic pixel positions [u, v], the semantic pixel positions [u, v] are inversely projected from the image plane to obtain road boundary feature positions [x v ,y v ,z v ] in the vehicle coordinate system through inverse perspective transformation. v v v < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / s> < / Step 3, the accurate pose of the vehicle is obtained by using a pose graph optimization method combined with the IMU and GPS data of the vehicle, and the road boundary feature sequence in the vehicle coordinate system is converted into the road boundary sequence (x , y i , z w ) in the global coordinate system according to the accurate pose of the vehicle i w, . i w ); Step 4, the path deviation between the real-time motion trajectory sequence (x, y) of the vehicle and the map reference line sequence (x i , y i ) is derived by Frenet coordinate system transformation to obtain the offset state quantity (S i , L i ) of the real-time motion trajectory sequence of the vehicle; Step 5, convert the road boundary sequence (x) in the global coordinate system. i w y i w, , z i w ) and map reference line sequence (x i y i The resulting map information sequence is then fused with the vehicle's real-time trajectory sequence offset state quantity (S). i L i Sequence fusion is performed to obtain a multimodal sequence vector s. i Multimodal sequence vectors are combined to create a corresponding multimodal information sequence X for the vehicle's real-time motion trajectory sequence. i k ={s i ,s i+1 ,…,s i+k-2 ,s i+k-1 }; Step 6, train the dynamic multi-head self-attention Transformer module learning network using the multi-modal information sequence obtained in step 5, model through attention mechanism, output sequence features, and obtain the updated map reference line sequence and the road boundary sequence, and adopt a loss function for accuracy quantification; Step 7, in the actual operation scene of the vehicle, the trained learning network is verified, the structure and parameters of the learning network are adjusted according to the verification result, the map road curve is updated in real time as the initial map for the next time, and the update cycle is repeated.

2. The method of claim 1, wherein, In step 1, according to the OpenDrive specification, the initial mine site map is first based on the initial mine site map <road>mid- multiple segments <geometry>the start coordinate S of the reference line of each segment is obtained by the element S coordinate, the heading angle hdg and the length length start and the end coordinate S end and is evenly divided and dispersed, and each segment is reconstructed <geometry>sequence S i = S start + i / n - (S end - S start ), where i is the index of the midpoint of the sequence S i and n is the number of segments <geometry>Is divided into n discrete points;< / geometry> < / geometry> < / geometry> < / road> The discrete point sequence (u i , v i ) of the reference line in the local uv coordinate system is obtained in combination with the coefficients aU, bU, cU, dU, aV, bV, cV, dV extracted in the corresponding polynomial ParamPoly3.

3. The method of claim 2, wherein, the sequence of discrete points (u i , v i ) where w i represents S i equence corresponds to the conversion to the local uv coordinate system under the polynomial ParamPoly3 for calculating the intermediate parameter point u i , v i ; and based on the discrete point sequence (u i , v i ) with each segment <geometry>The sequence of discrete points (x0, y0) of the structure is taken as the origin to obtain the sequence of reference lines (x i , y i ) of the absolute coordinate system map.< / geometry> 4. The method of claim 3, wherein, the sequence of map reference lines (x i , y i ) 5. The method of claim 1, wherein, In step 2, the semantic pixel position [u, v] is mapped from the image plane to the ground plane in the vehicle coordinate system through inverse perspective transformation, and the road boundary feature position in the vehicle coordinate system is obtained: where π c (·) denotes the distortion and projection model of the camera, is the back-projection function that projects a pixel into space, [R c t c ] is the extrinsic matrix of the camera relative to the vehicle center, [u, v] is the semantic pixel position, [x v ,y v ,z v ] is the road boundary feature position in the vehicle coordinate system, λ is a scalar, [] col:i denotes taking the i-th column of the matrix, i being 1, 2 and 4.

6. The method of claim 1, wherein, In step 3, the vehicle maps the collected semantic features from the vehicle coordinate system to the global coordinate system by the optimized pose information. Let the global pose of the vehicle at time i be represented by position p i and rotation R(q i ), where R(q i ) is a function that converts a quaternion q i to a rotation matrix, and the road boundary feature sequence in the vehicle coordinate system can be mapped to the road boundary sequence in the global coordinate system by the following transformation:

7. The method of claim 1, wherein, In step 4, the vehicle real-time motion trajectory sequence offset state quantity (S i , L i ) is: where θ is the tangent angle of the sequence of map reference lines (x i , y i ), α represents the angle between the current position of the vehicle and the tangent direction of the reference line, and (x, y) is the sequence of real-time motion trajectories of the vehicle.

8. The method of claim 1, wherein, In step 5, the multimodal sequence is truncated according to the k-step sequence, and the multimodal sequence with the starting sequence number i is represented as X i k = {s i ,s i+1 ,…,s i+k-2 ,s i+k-1} each multimodal sequence vector s i includes the map reference line sequence (x i , y i ) of the vehicle at the sequence step, the road boundary sequence (x i w , y i w, , z i w ) in the global coordinate system, the vehicle real-time motion trajectory sequence offset state quantity (S i , L i ), and the fusion s i＝ {x i , y i , x i w , y i w , z i w , S i , L i} is obtained.

9. The method of claim 1, wherein, In step 6, the dynamic multi-head self-attention Transformer module includes three independent attention "heads", which respectively learn the semantic expression of the reference line, the boundary and the road terrain in the sequence, and the outputs of the three independent attention "heads" are combined and spliced to obtain: MultiHead(Q, K, V) = Concat(head1,...,head h )W O head i = Attention(QW i Q ,KW i K ,VW i V ) where MultiHead denotes multi-head attention, Q, K and V represent query, key and value matrices respectively, Concat denotes feature concatenation, head h is the i-th attention head, QW O is the weight matrix used to multiply before the output of multi-head self-attention, head i is the i-th attention head, QW i Q is the i-th attention head, QW i K is the i-th attention head, QW i V are the weight matrices for Q, K and V, Attention is the attention calculation, softmax is the normalization means, d k is the dimension of the key vector, K T is the transpose of K, QK T is the attention score.

10. The method of claim 1, wherein, In step 6, the accuracy of the updated map reference line sequence and the road boundary sequence is quantified by the mean square error (MSE) and evaluated by the following error: where L map is the loss function of the network, is the updated map reference line sequence, p i is the real coordinate of the vehicle, N is the total number of map discrete points, and σ is a learnable noise parameter.

Images