Navigation real scene construction method based on lane line model

Abstract

The invention provides a navigation live-action construction method based on a lane line model, and the method comprises the steps: constructing a lane line model which comprises a lane line type, a parameter output network, a heat map network and a lane line segmentation point detection network, the heat map network outputs a heat map containing lane lines, road edges, fences and segmentation points according to the type of the lane lines, and outputs lane line information parameters; training a lane line model: preprocessing the picture data; training a lane line type, a parameter output network and a heat map network in the lane line model; freezing the trained network and then independently training a lane line segmentation point detection network; accelerating and applying the model, and outputting a lane navigation model with a ground guide icon and a fishbone-shaped guidepost; wherein in the construction of the lane line model, a detection aided training branch is set, and the lane line model is trained in an aided manner in combination with an instance segmentation technology, so that the consistency and stability of AR navigation are improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle navigation technology, and more specifically, to a method for constructing a navigation real scene based on a lane line model. Background Art

[0002] AR (augmented reality) navigation technology, as a cutting-edge development of modern intelligent navigation systems, aims to provide drivers with a new navigation experience by integrating the precise path planning and real-time positioning capabilities of traditional navigation with advanced lane-level high-precision positioning technology and visual perception algorithms. Within this technical framework, the driver's perspective becomes the core display window of navigation information. Through the real-time image outside the car captured by the on-board camera, key information such as navigation guidance and warning prompts are directly superimposed on the real scene, realizing the leap of information display from two-dimensional maps to three-dimensional space. The original intention of this design is to simplify the driver's information interpretation process and reduce distractions caused by frequent viewing of navigation equipment, thereby effectively improving driving safety and the accuracy of path guidance. However, AR navigation products widely used in the current market generally face the problem of insufficient integration of virtual information and real environment, that is, the so-called "not real enough". This is mainly reflected in the mismatch between virtual arrows, route signs and actual road signs and traffic flow, which affects the driver's judgment and trust.

[0003] However, the challenges in the implementation of AR navigation technology cannot be ignored. Especially in terms of visual perception and image rendering, the existing AR navigation system faces the dilemma of being "not stable enough". This is mainly reflected in the recognition accuracy and stability of lane lines. Due to factors such as complex and changeable road environments, changing light conditions, and vehicle dynamics, the visual perception system often finds it difficult to continuously and stably capture and accurately identify lane information, resulting in problems such as recognition loss, lane line jitter in the display, or recognition response delays. In addition, the lack of synchronization between image rendering and real-world scenes is also one of the key factors affecting navigation stability. This asynchrony may lead to a misalignment between virtual information and real scenes, further weakening the practicality and user experience of AR navigation. Summary of the invention

[0004] The present invention provides a method for constructing a navigation real scene based on a lane line model, aiming to solve the problem that the real scene of AR navigation is difficult to provide correct navigation instructions for complex road conditions, and to improve the consistency and stability of AR navigation.

[0005] To achieve the above object, the present invention provides a method for constructing a navigation real scene based on a lane line model, comprising:

[0006] Constructing a lane line model, the lane line model includes a lane line type, a parameter output network, a heat map network, and a lane line segmentation point detection network, the heat map network outputs a heat map containing lane lines, curbs, fences, and segmentation points according to the lane line type, and outputs lane line information parameters;

[0007] Train the lane line model, including:

[0008] Preprocess the image data;

[0009] Training the lane line type, parameter output network and heat map network in the lane line model;

[0011] Freeze the trained network and then train the lane segmentation point detection network separately.

[0012] Accelerate and apply the model to output a lane navigation model with ground guidance icons and fishbone-shaped road signs;

[0013] In constructing the lane line model, a detection auxiliary training branch is set up, and instance segmentation technology is combined to assist in training the lane line model.

[0014] In one embodiment, constructing a lane line model specifically includes:

[0015] Extracting image features, the image features include the starting point, the ending point, the segmentation point of the lane line, and the types of lanes before and after the lane line at the segmentation point;

[0016] Enhance and decode image features, and decode lane line information;

[0017] The decoded lane line information parameters are output, wherein the lane line information parameters include category information, segmentation point coordinates, and lane line shape parameters.

[0018] In one embodiment, a deep convolutional neural network is used to extract image features;

[0019] Transformer is used to enhance and decode the extracted image features;

[0020] The head is used to decode the enhanced and decoded feature maps and output the category information, segmentation point coordinates, and lane line shape parameters;

[0021] The lane line shape parameters include a starting point, an end point and cubic curve fitting parameters.

[0022] In one embodiment, the output size of the category information is 16*18, where 16 is the number of lane lines, 18 is the 9 probability values ​​of the linearity before the lane line segmentation point and the 9 probability values ​​of the linearity after the segmentation point, and the linearity includes none, single solid line, single dashed line, double solid line, double dashed line, left solid and right dashed line, left dashed and right solid line, curb and fence road conditions;

[0023] The output size of the segmentation point coordinates is 16*7, where 16 is the number of lane lines and 7 is the regression coefficient of each lane line;

[0024] The output size of the lane line shape parameter is 16*2, where 16 is the number of lane lines, and 2 is the probability value of the existence of the segmentation point and the ordinate value of the segmentation point;

[0025] The size of the heat map output is 4*96*512, where 96*512 is the pixel size of the output heat map, and 4 is the four heat map types of four different channels, including lane line heat map, curb heat map, fence heat map and segmentation point heat map.

[0026] In one embodiment, step S2011 specifically includes:

[0027] Unify the image size;

[0028] Use skyline to crop the image to 960*192;

[0029] Generates the scaled json file and the heatmap true value file.

[0030] In one embodiment, applying the model specifically includes: using the trained lane line type, parameter output network, heat map network and lane line segmentation point detection network to extract image features and generate a ground indicator icon corresponding to the lane line type.

[0031] In one embodiment, the ground indication icons include two-way guidance, three-way guidance, main and auxiliary road guidance, roundabout guidance, and lane change guidance. The ground icons include two-way guidance icons, three-way guidance icons, main and auxiliary road guidance icons, roundabout guidance icons, and lane change guidance icons.

[0032] In one embodiment, the fishbone-shaped signpost is suspended on the AR navigation display page, and the fishbone-shaped signpost is combined with SD+ map data to point to the direction of the intersection to exit the lane.

[0033] In one embodiment, when the road type between the starting point and the ending point of the road is consistent, no segmentation point is arranged;

[0034] When the road types between the starting point and the end point are inconsistent, a split point is arranged to split the roads of different types.

[0035] In one embodiment, SNPE is used to perform quantization acceleration on the trained lane line model.

[0036] The present invention has the following beneficial effects:

[0037] 1. Consistency: In AR navigation display, lane-level SD+ map data is better integrated, and the recognition results can be used to draw according to the shape of the intersection, which is closer to the real scene. In addition, the present invention indicates the direction of the intersection through the aerial fishbone, giving clear and intuitive direction guidance of the real scene of the intersection.

[0038] 2. Accuracy: The model network constructed by the present invention adopts a lightweight backbone. In order to improve the recognition accuracy of the model, on the basis of retaining the lane line recognition head, an additional detection auxiliary training branch is added for training to further improve the model's recognition accuracy of linear road signs, guardrails and segmentation points. Through a reasonable loss coefficient ratio, the model takes into account both high "accuracy" and "speed". BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of a method for constructing a navigation real scene based on a lane line model according to an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a process of constructing a lane line model in a method for constructing a navigation real scene based on a lane line model according to an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of the process of an auxiliary detection training branch based on lane line model training according to an embodiment of the present invention;

[0042] Figure 4 It is a schematic diagram of the structure of displaying map data using SD+ according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the embodiment of this application will be described in more detail below in conjunction with the drawings in the embodiment of this application. In the drawings, the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions. The described embodiments are part of the embodiments of this application, not all of them. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be construed as limitations on this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0044] Figure 1The flowchart of a method for constructing a real navigation scene based on a lane line model according to an embodiment of the present invention is as follows:

[0045] Step S101, constructing a lane line model, wherein the lane line model includes a lane line type, a parameter output network, a heat map network, and a lane line segmentation point detection network, wherein the heat map network outputs a heat map containing lane lines, curbs, fences, and segmentation points according to the lane line type, and outputs lane line information parameters;

[0046] Step S201, training lane line model:

[0047] Step S2011, pre-processing the image data;

[0048] Step S2012: training the lane line type and parameter output network in the lane line model

[0049] Networks and heatmap networks;

[0050] Step S2013, freezing the trained network and then training the lane line segmentation point detection network separately;

[0051] Step S301, accelerating and applying the model, and outputting a lane navigation model with ground guidance icons and fishbone-shaped road signs;

[0052] In constructing the lane line model, a detection auxiliary training branch is set up, and instance segmentation technology is combined to assist in training the lane line model.

[0053] Specifically, the lane line segmentation point detection network accurately segments each lane line from the image and distinguishes different lane lines to handle multiple lane lines in complex scenes.

[0054] Furthermore, during the training process, the heat map is used as a supervision signal to train the lane line model.

[0055] Furthermore, the detection-assisted training branch adds additional paths or sub-networks to the network to assist the main object detection task, which can provide additional supervision signals to help the network better learn feature representations.

[0056] Use the cubic equation to approximate the lane line. The cubic equation is:

[0057] X=kZ 3 +mZ 3 +nZ+b

[0058] Convert the cubic equation to the plane, and the conversion formula is:

[0059]

[0060] Simplifying the formula we get:

[0061]

[0062] Where, f″=f sinφ, m″=m′cosφ, k″=k′×cos 2 φ,υ′ are the ordinate values, and the network output parameters include f”, m”, b”, and k”.

[0063] After outputting f”, m”, b”, and k”, the neural network will extract, enhance, and decode these parameters to output category information, segmentation point coordinates, and lane line morphology parameters or heat maps.

[0064] Further, f" is related to the component of the curvature or bending degree of the lane line in the longitudinal direction. m" is related to the offset of the lane line. b" is related to the slope of the lane line. k" is related to the curvature of the lane line.

[0065] These variables play a key role in the lane line fitting process. Through these conversions and definitions, the cubic equation can be converted from three-dimensional space to a plane, which is convenient for subsequent calculations and processing.

[0066] Furthermore, the display positions of the ground guidance icons and the fishbone-shaped road signs are fixed at a certain distance in front of the vehicle to present a propulsion effect and a sense of reality of following the vehicle.

[0067] In one embodiment, constructing a lane line model specifically includes:

[0068] Extracting image features, the image features include the starting point, the ending point, the segmentation point of the lane line, and the types of lanes before and after the lane line at the segmentation point;

[0069] Enhance and decode image features, and decode lane line information;

[0070] The decoded lane line information parameters are output, wherein the lane line information parameters include category information, segmentation point coordinates, and lane line shape parameters.

[0071] Specifically, Figure 2 As shown, it consists of feature extraction, feature fusion enhancement, output decoding, and heat map feature decoding.

[0072] The network feature extraction is based on the backbone of the Resnet module, and the feature resolution after extraction is 1 / 32 of the resolution before extraction.

[0073] Transformer is used to enhance and decode the extracted image features. The dimension of the feature map after enhancement and decoding is 16*32. The features extracted from the backbone part are enhanced and queried and decoded using the self-attention mechanism, and the category information, segmentation point coordinates, and lane line morphology parameters are output.

[0074] The network output head decodes the 16*32 feature map after Transformer feature enhancement, and outputs category information, segmentation point coordinates, and lane line shape parameters respectively.

[0075] The extracted image features are converted into heat maps, and the resolution of the network heat map output is 1 / 2 of the original image, which assists in fitting the parameter lane line output.

[0076] In one embodiment, a deep convolutional neural network is used to extract image features;

[0077] Transformer is used to enhance and decode the extracted image features;

[0078] The head is used to decode the enhanced and decoded feature maps and output the category information, segmentation point coordinates, and lane line shape parameters;

[0079] The lane line shape parameters include a starting point, an end point and cubic curve fitting parameters.

[0080] Furthermore, the trained model can be selected for enhanced training, the head uses a regularization method, and the resnet is replaced with a PAN (path aggregation network) for training.

[0081] Furthermore, two loops can be added to the backbone part to extract lane lines of different sizes, thereby increasing the number of backbone parameters.

[0082] Set up a detection-assisted training branch and use instance segmentation to accurately extract features of each layer of the backbone.

[0083] Combining the detection-assisted training branch with instance segmentation enables the trained network to not only identify the object categories in the image, but also generate precise bounding boxes or pixel-level masks for each object instance to achieve accurate segmentation.

[0084] In one embodiment, the output size of the category information is 16*18, where 16 is the number of lane lines, 18 is the 9 probability values ​​of the linearity before the lane line segmentation point and the 9 probability values ​​of the linearity after the segmentation point, and the linearity includes none, single solid line, single dashed line, double solid line, double dashed line, left solid and right dashed line, left dashed and right solid line, curb and fence road conditions;

[0085] The output size of the segmentation point coordinates is 16*7, where 16 is the number of lane lines, 7 is the regression coefficient of each lane line, the first two numbers represent the starting point and the end point, and the last five numbers represent the five parameters of the regression curve equation;

[0086] The output size of the lane line shape parameter is 16*2, where 16 is the number of lane lines, and 2 is the probability value of the existence of the segmentation point and the ordinate value of the segmentation point;

[0087] The size of the heat map output is 4*96*512, where 96*512 is the pixel size of the output heat map, and 4 is the four heat map types of four different channels, including lane line heat map, curb heat map, fence heat map and segmentation point heat map.

[0088] The network output results include a heat map containing lane lines, curbs, fences and segmentation points, and the auxiliary detection training branch outputs lane line information parameters.

[0089] like Figure 3 As shown, out_conv1 is the convolution layer that outputs category information. The size of the output category information is 16*18, which means that there are 16 lane line types in the output, and each type corresponds to 18 output values.

[0090] Furthermore, 16 represents the number of lane lines, 18 represents 9 probability values ​​of the linear line before the lane line segmentation point and 9 probability values ​​of the linear line after the segmentation point, and the 9 linear lines represent: none, single solid line, single dashed line, double solid line, double dashed line, left solid and right dashed, left dashed and right solid, curb and forked solid line.

[0091] out_conv2 is the convolution layer that outputs the segmentation point coordinates. The output segmentation point coordinate size is 16*7. 16 means the number of lane lines is 16, which is consistent with the 16 lane lines output by the out_conv1 channel. 7 represents the regression coefficient of each lane line. The first two numbers in the regression coefficient represent the starting point and the end point of the lane line, and the last five numbers represent the five parameters of the regression curve equation.

[0092] out_conv3 is the convolution layer that outputs lane line shape parameters. The size of the output lane line shape parameters is 16*2. 16 means that there are 16 lane lines, which is consistent with the 16 lane lines output by the out_conv1 channel. 2 represents the probability value of the existence of the segmentation point and the ordinate value of the segmentation point.

[0093] out_heatmap is the channel of the output heatmap. The output heatmap size is 4*96*512, 96*512 represents the pixel size of the output heatmap, and 4 represents the heatmap types of four different channels, namely lane line, curb, fence and segmentation point heatmap.

[0094] In one embodiment, step S2011 specifically includes:

[0095] Unify the image size;

[0096] Use skyline to crop the image to 960*192;

[0097] Generates scaled json (object notation) files and heatmap truth files.

[0098] Specifically, the heatmap truth file is used to guide model training. During the training process, the model generates a predicted heatmap, which is then compared with the heatmap in the heatmap truth file. After the model training is completed, the model can be tested using the data in the heatmap truth file. By comparing the heatmap output by the model with the heatmap in the heatmap truth file, such as calculating the similarity index between them, it is possible to determine whether the model has accurately learned the rules for heatmap generation, as well as the accuracy and reliability of the model.

[0099] In one embodiment, applying the model specifically includes: using the trained lane line type, parameter output network, heat map network and lane line segmentation point detection network to extract image features and generate a ground indicator icon corresponding to the lane line type.

[0100] In one embodiment, the ground indication icons include two-way guidance, three-way guidance, main and auxiliary road guidance, roundabout guidance, and lane change guidance. The ground icons include two-way guidance icons, three-way guidance icons, main and auxiliary road guidance icons, roundabout guidance icons, and lane change guidance icons.

[0101] In one embodiment, the fishbone-shaped signpost is suspended on the AR navigation display page, such as Figure 4 As shown, the fishbone-shaped road sign is combined with SD+ map data to point to the direction of the intersection of the exit lane.

[0102] In one embodiment, when the road type between the starting point and the ending point of the road is consistent, no segmentation point is arranged;

[0103] When the road types between the starting point and the end point are inconsistent, a split point is arranged to split the roads of different types.

[0104] In one embodiment, the trained lane line model is quantized and accelerated using SNPE (Snapdragon Neural Processing Engine).

[0105] When the vehicle approaches a complex intersection, AR navigation will fix a ground indicator icon in front of the vehicle to ensure that the driver can intuitively see the upcoming turn instructions. The display position of the icon will be dynamically adjusted according to the recognition of the lane line. If the system successfully recognizes the lane line, the icon will be fixed in the middle of the lane line; if the lane line is unclear or cannot be recognized, the icon will be fixed in the middle of a predefined shape by default and move forward as the vehicle moves. At the same time, in order to better fit the real scene, the display content of the ground icon is subdivided into two-way guidance, three-way guidance, main and auxiliary road guidance, roundabout guidance, lane change guidance, etc. according to different scenarios to help drivers make correct decisions in complex intersection environments.

[0106] Furthermore, the display position and shape of the arrow guidance in the air are optimized. In terms of display position: the guidance is not fixed at the turning point, but fixed at a certain distance in front of the car, showing a propulsion effect, which can give a more realistic feeling of following the car. In terms of display shape: the map data is better integrated, the lane-level SD+ data is used, and the recognition results are used to draw according to the shape of the intersection, which is closer to the real scene.

[0107] The present invention processes the input image by randomly flipping horizontally and filling 32 zero pixels on the left and right sides of the image.

[0108] Initialize the model using a learnable Laplacian Gaussian filter

[0109] Limiting the range of model learning parameter values ​​can prevent the model backbone from failing to quantize due to excessively large parameter value ranges during the quantization process.

[0110] In one embodiment, the learning rate reduction strategy adopts a OneCycle approach.

[0111] The present invention has the following beneficial effects:

[0112] 1. Consistency: In AR navigation display, lane-level SD+ map data is better integrated, and the recognition results can be used to draw according to the shape of the intersection, which is closer to the real scene. In addition, the present invention indicates the direction of the intersection through the aerial fishbone, giving clear and intuitive direction guidance of the real scene of the intersection.

[0113] 2. Accuracy: The model network constructed by the present invention adopts a lightweight backbone. In order to improve the recognition accuracy of the model, on the basis of retaining the lane line recognition head, an additional detection auxiliary training branch is added for training to further improve the model's recognition accuracy of linear road signs, guardrails and segmentation points. Through a reasonable loss coefficient ratio, the model takes into account both high "accuracy" and "speed".

[0114] In the description of the present application, it should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. For ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The technology, methods and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be regarded as part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary, rather than as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0115] It should be noted that, in the present application, the terms "comprises", "includes" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. It should also be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0116] The above embodiments are provided for persons familiar with the art to implement or use the present application. Personnel familiar with the art can make various modifications or changes to the above embodiments without departing from the application concept of the present application. Therefore, the protection scope of the present application is not limited to the above embodiments, but should be the maximum scope of the innovative features mentioned in the claims.

Claims

1. A method for constructing a navigation scene based on a lane line model, characterized in that: include: Step S101, constructing a lane line model, wherein the lane line model includes a lane line type, a parameter output network, a heat map network, and a lane line segmentation point detection network, wherein the heat map network outputs a heat map containing lane lines, curbs, fences, and segmentation points according to the lane line type, and outputs lane line information parameters; Step S201, training lane line model: Step S2011, pre-processing the image data; Step S2012, training the lane line type, parameter output network and heat map network in the lane line model; Step S2013, freezing the trained network and then training the lane line segmentation point detection network separately; Step S301, accelerating and applying the model, and outputting a lane navigation model with ground guidance icons and fishbone-shaped road signs; In constructing the lane line model, a detection auxiliary training branch is set up, and instance segmentation technology is combined to assist in training the lane line model.

2. The method for constructing a navigation real scene based on a lane line model according to claim 1, characterized in that: Building a lane line model specifically includes: Extracting image features, the image features include the starting point, the ending point, the segmentation point of the lane line, and the types of lanes before and after the lane line at the segmentation point; Enhance and decode image features, and decode lane line information; The decoded lane line information parameters are output, wherein the lane line information parameters include category information, segmentation point coordinates, and lane line shape parameters.

3. The method for constructing a navigation real scene based on a lane line model according to claim 2, characterized in that: Use deep convolutional neural network to extract image features; Transformer is used to enhance and decode the extracted image features; The head is used to decode the enhanced and decoded feature maps and output the category information, segmentation point coordinates, and lane line shape parameters; The lane line shape parameters include a starting point, an end point and cubic curve fitting parameters.

4. The method for constructing a navigation real scene based on a lane line model according to claim 2, characterized in that: The output size of the category information is 16*18, where 16 is the number of lane lines, 18 is the 9 probability values ​​of the linearity before the lane line segmentation point and the 9 probability values ​​of the linearity after the segmentation point. The linearity includes none, single solid line, single dashed line, double solid line, double dashed line, left solid and right dashed line, left dashed and right solid line, curb and fence road conditions; The output size of the segmentation point coordinates is 16*7, where 16 is the number of lane lines and 7 is the regression coefficient of each lane line; The output size of the lane line shape parameter is 16*2, where 16 is the number of lane lines, and 2 is the probability value of the existence of the segmentation point and the ordinate value of the segmentation point; The size of the heat map output is 4*96*512, where 96*512 is the pixel size of the output heat map, and 4 is the four heat map types of four different channels, including lane line heat map, curb heat map, fence heat map and segmentation point heat map.

5. The method for constructing a navigation real scene based on a lane line model according to claim 1, characterized in that: Step S2011 specifically includes: Unify the image size; Use skyline to crop the image to 960*192; Generates the scaled json file and the heatmap true value file.

6. The method for constructing a navigation real scene based on a lane line model according to claim 1, characterized in that: The specific application of the model includes: using the trained lane line type, parameter output network, heat map network and lane line segmentation point detection network to extract image features and generate ground indicator icons corresponding to the lane line type.

7. The method for constructing a navigation real scene based on a lane line model according to claim 6, characterized in that: The ground indication icons include two-way road indication, three-way road indication, main and auxiliary road indication, roundabout indication, and lane change indication. The ground indication icons include two-way road indication icons, three-way road indication icons, main and auxiliary road indication icons, roundabout indication icons, and lane change indication icons.

8. The method for constructing a navigation real scene based on a lane line model according to claim 7, characterized in that: The fishbone-shaped signpost is suspended on the AR navigation display page. The fishbone-shaped signpost is combined with SD+ map data to point to the direction of the intersection of the exit lane.

9. The method for constructing a navigation real scene based on a lane line model according to claim 2, characterized in that: When the road type is the same from the starting point to the end point of the road, no split point is arranged; When the road types between the starting point and the end point are inconsistent, a split point is arranged to split the roads of different types.

10. The method for constructing a navigation real scene based on a lane line model according to claim 1, characterized in that: SNPE is used to quantize and accelerate the trained lane line model.

Images