Test road selection method and device and terminal equipment

Abstract

The invention relates to the technical field of automobiles, and discloses a test road selection method and device and terminal equipment. The method comprises the following steps: acquiring a map of a target area and problem data of historical vehicles on each candidate road in the target area; performing image matching on the map to obtain scene matching degrees among the candidate roads; performing aggregation analysis on the problem data, and determining problem coverage among the candidate roads; and processing the scene matching degree and the problem coverage degree through a road priority scoring model to obtain a target road combination screened from the candidate roads. By adopting the method, one or a group of'high-value 'test roads can be scientifically selected from a wide road network, so that the test roads can cover driving scenes as diversified as possible, and problems can be exposed to the greatest extent.

Description

Technical Field

[0001] This application relates to the field of automotive technology, and in particular to a test road selection method, apparatus, and terminal equipment. Background Technology

[0002] In the research and development and verification process of automobiles, real-vehicle road testing is a key step in verifying the performance of vehicle perception, decision-making and control algorithms, as well as exposing potential problems in the system.

[0003] However, the real-world road network is vast and the scenarios are complex and varied. Under limited time and cost constraints, it is unrealistic to complete exhaustive testing of all roads. Therefore, how to scientifically select one or a group of "high-value" test roads from the vast road network, so as to cover as many driving scenarios as possible and expose problems to the greatest extent, has become an important technical problem that those skilled in the art are trying to solve. Summary of the Invention

[0004] In view of this, embodiments of this application provide a test road selection method, apparatus, and terminal device, which can effectively solve the technical problem that it is unrealistic to complete exhaustive testing of all roads under limited time and cost constraints due to the large scale and complex and ever-changing nature of real-world road networks.

[0005] In a first aspect, embodiments of this application provide a method for selecting a test road. The method includes: Obtain a map of the target area, as well as historical data on problems encountered by vehicles on each candidate road within the target area; Image matching is performed on the map to obtain the scene matching degree between each of the candidate roads; The problem data is aggregated and analyzed to determine the problem coverage among the candidate roads; The scenario matching degree and the problem coverage are processed by the road priority scoring model to obtain the target road combination selected from the candidate roads.

[0006] Secondly, embodiments of this application provide a test road selection device, the device comprising: The acquisition module is used to acquire a map of the target area, as well as historical data on problems that vehicles have encountered on roads within the target area. The matching module is used to perform image matching on the map to obtain the scene matching degree between each candidate road; The determination module is used to perform aggregate analysis on the problem data to determine the problem coverage among the candidate roads; The processing module is used to process the scene matching degree and the problem coverage through a road priority scoring model to obtain the target road combination selected from the candidate roads.

[0007] Thirdly, embodiments of this application provide a terminal device, including a processor and a memory, wherein the memory stores a computer program, and the processor is used to execute the computer program to perform the following steps: Obtain a map of the target area, as well as historical data on problems encountered by vehicles on each candidate road within the target area; Image matching is performed on the map to obtain the scene matching degree between each of the candidate roads; The problem data is aggregated and analyzed to determine the problem coverage among the candidate roads; The scenario matching degree and the problem coverage are processed by the road priority scoring model to obtain the target road combination selected from the candidate roads.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the following steps: Obtain a map of the target area, as well as historical data on problems encountered by vehicles on each candidate road within the target area; Image matching is performed on the map to obtain the scene matching degree between each of the candidate roads; The problem data is aggregated and analyzed to determine the problem coverage among the candidate roads; The scenario matching degree and the problem coverage are processed by the road priority scoring model to obtain the target road combination selected from the candidate roads.

[0009] The embodiments of this application have the following beneficial effects: This application constructs an objective, quantifiable, and reproducible evaluation system by integrating map matching and historical problem data aggregation analysis. It upgrades subjective experience-based judgment to an intelligent scoring mechanism based on multi-dimensional data fusion, significantly improving the scientific rigor and reliability of test road selection. Furthermore, with limited testing resources, it maximizes the probability of discovering potential problems, shortens the testing cycle, and reduces road testing costs. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of a test road selection method according to an embodiment of this application is shown; Figure 2 Another flowchart of the test road selection method according to an embodiment of this application is shown; Figure 3 This diagram illustrates the output of the map extraction model according to an embodiment of this application. Figure 4 A schematic diagram of a test road selection device according to an embodiment of this application is shown. Detailed Implementation

[0012] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0013] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0014] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0015] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in a generally used dictionary) shall be interpreted as having the same meaning as in the context of the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0016] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0017] The following examples illustrate the method for selecting test roads.

[0018] Figure 1 A flowchart illustrating a test road selection method according to an embodiment of this application is shown. Exemplarily, the test road selection method includes the following steps: Step S102: Obtain a map of the target area and historical data on problems that vehicles have encountered on each candidate road in the target area.

[0019] The target area roads refer to a set of candidate roads located within a preset geographical area, used for on-site testing of autonomous driving or intelligent driving assistance systems. Optionally, the target area can be divided according to urban administrative divisions, traffic function zones (such as main road areas, school surrounding areas, and densely populated commercial areas), climate and environmental characteristics, or areas with a high incidence of historical accidents.

[0020] Understandably, each road has an identifiable spatial coordinate range and topological connectivity, serving as the basic unit for subsequent image matching and problem data analysis.

[0021] A map is a digital image or layer data that represents the spatial layout and visual scene information of a target area, including at least one or more of the following in a high-precision map: lane-level geometry, traffic signs and markings, intersection morphology, and surrounding environment outline.

[0022] Optionally, the map can be stored as a raster or vector image, which can be used to perform image-level similarity comparison operations. Furthermore, the map undergoes uniform coordinate system projection and scale normalization to ensure the comparability of image matching results between different road segments.

[0023] Problem data refers to log information related to anomalies, control failures, perception errors, or human-machine interaction interventions recorded during the actual driving of vehicles in the target area. Optionally, this problem data may include at least the event type, location, timestamp, vehicle status (speed, acceleration), environmental conditions (weather, lighting), and the corresponding index of the original sensor data.

[0024] Specifically, the target area of ​​roads to be evaluated is first determined. For example, the target area might be the main urban area of ​​a first-tier city, with its geographical boundaries defined by polygonal vector fences in a predefined coordinate system. All open public roads within this target area are included in the candidate road set, totaling approximately 3,200 kilometers. Each candidate road is uniquely identified by its start and end node numbers, road segment ID, and number of lanes, and is further divided into several sub-areas according to urban functional zones, such as commercial center areas, school perimeter areas, and highway entrance / exit connecting sections, for subsequent differential analysis.

[0025] Then, a map corresponding to the roads in the target area is obtained. This map contains the geometric information of each road segment, lane topology, traffic sign locations, and 3D environment contours. Optionally, to support image matching operations, the original vector map is rendered into a raster image of a uniform format: image resolution is 0.1 meters / pixel; UTM (Transverse Mercator) projection is used to eliminate curvature distortion; each image covers a continuous road segment of 500 meters in length, with an overlap rate set to 20% to ensure continuity; image channels include lane line binary maps, static obstacle distribution heatmaps, and intersection complexity marker layers. All images are normalized and stored in a distributed file for use in subsequent steps.

[0026] As can be understood, UTM projection refers to Universal Transverse Mercator projection, a map projection method widely used in geographic information systems (GIS), surveying, and navigation. Its main function is to convert the latitude and longitude coordinates of the Earth's approximately spherical surface into two-dimensional coordinates (X, Y) in a Cartesian coordinate system, so as to facilitate spatial calculations such as distance, area, and direction, and avoid geometric distortion caused by the Earth's curvature.

[0027] Simultaneously, historical vehicle issue data is collected. For example, issue data originates from onboard data records deployed on 1,000 intelligent connected test vehicles, spanning the past six months, continuously recording abnormal event logs during the operation of the autonomous driving function. Issue data includes, but is not limited to, the following fields: Event type: such as automatic emergency braking false trigger, lane keeping assist failure, target recognition error (false detection / missed detection), path planning anomaly, etc.; timestamp of occurrence; GPS location information; vehicle status: vehicle speed, acceleration, steering angle; environmental conditions: weather (sunny / rainy / foggy), light intensity (day / dusk / night); corresponding sensor data index (pointing to the original camera, LiDAR or millimeter-wave radar data segment); function mode: NOA highway navigation, city driving, parking assist, etc.

[0028] Optionally, all original anomaly logs are anonymized and uploaded to a cloud data center for cleaning and structure transformation. Subsequently, each problem event is precisely mapped to a corresponding candidate road segment based on GPS location (road segment matching accuracy exceeds 98%). This results in a "problem event statistics table" organized by road segment, serving as the basic input for subsequent aggregate analysis.

[0029] This application achieves comprehensive collection and preprocessing of road maps and problem data in the target area through multi-source heterogeneous data fusion, ensuring that the data foundation for calculating scene matching degree and problem coverage in subsequent steps is real, accurate, and engineering-operable.

[0030] Step S104: Perform image matching on the map to obtain the scene matching degree between each candidate road.

[0031] Among them, candidate roads refer to several continuous road segments within the target area that are eligible to be selected into the test road combination; each candidate road has clear spatial boundaries (start and end coordinates), road topology attributes (number of lanes, direction, connection relationship) and functional classification labels (such as urban main roads, roads around schools, highway ramps, intersections without traffic lights, etc.), and serves as the basic analysis unit for image matching, problem data aggregation and priority scoring.

[0032] Scene matching degree refers to the quantitative index of the similarity between two candidate roads in terms of visual structure, environmental layout and driving situation features. It is used to measure whether one road can represent the perceptual challenge of another road. The index is generated based on image-level comparison of its corresponding map. The higher the value, the more similar the road scenes of the two are and the stronger the representativeness.

[0033] Specifically, the first step involves acquiring the map corresponding to the target area roads output from the above steps, along with the standardized map datasets for each candidate road. For example, each candidate road corresponds to a rasterized top-down view of size A pixels and resolution B. The image content originates from a high-precision map rendering engine and includes key visual features such as lane line geometry, traffic sign locations, intersection topology, and the distribution of surrounding static obstacles. All images have been uniformly projected onto the UTM coordinate system and have undergone brightness normalization and rotation correction to ensure directional consistency.

[0034] Then, a two-stage image matching process is used to calculate the scene matching degree between any two candidate roads: The first stage is feature extraction, which involves extracting structured features for each map using one of two methods: The first is the traditional image feature method, which uses the SIFT (Scale-Invariant Feature Transform) algorithm to detect key points in the image and generate a 128-dimensional descriptor vector; the feature point density per unit area is used as a structural complexity indicator. The second is the deep learning embedding method, which involves inputting the image into a pre-trained convolutional neural network, removing the last classification layer, and extracting the 512-dimensional feature vector output from the penultimate layer as the "scene fingerprint" of the road.

[0035] The second stage is similarity measurement, which involves calculating cosine similarity based on the extracted feature vectors. This cosine similarity reflects the similarity score between two candidate roads, i.e., the scene matching degree. Optionally, the scene matching degree, after normalization, ranges from [0, 1]. The closer the value is to 1, the more similar the two roads are in visual structure and spatial layout, and the more likely they are to trigger similar response behaviors from the autonomous driving perception module.

[0036] Step S106: Perform aggregate analysis on the problem data to determine the problem coverage among the candidate roads.

[0037] Among them, problem coverage refers to the comprehensive quantitative index of the breadth and intensity of defects in autonomous driving or intelligent driving assistance technology exposed by a candidate road in its historical operation. It is used to characterize the road's "detection ability" or "fault disclosure ability" for potential problems. Its value is calculated based on aggregated historical problem data. The higher the value, the wider the types of problems involved, the higher the frequency of occurrence, or the greater the severity of the problems, and therefore the higher the testing priority.

[0038] Specifically, the first step is to obtain the raw problem dataset output from the above steps. For example, this data comes from the vehicle fault records deployed on 500 L3-level intelligent connected test vehicles, spanning the past 6 months, with approximately 127,000 valid event records collected. Each record contains the following fields: event type (e.g., target missed detection, false braking, lane departure warning without warning, path planning oscillation, etc.); GPS location information; timestamp; function mode (NOA navigation assist, automatic parking, traffic jam following, etc.); vehicle status (vehicle speed, steering angle, acceleration); environmental conditions (day / night, sunny / rainy); severity level label (labeled by the backend based on the degree of safety impact, rated as low / medium / high).

[0039] Then, spatial mapping and road segment aggregation are performed. Using the road topology database in the high-precision map, each problem event is associated with its corresponding candidate road segment using a nearest neighbor matching algorithm. Specific steps include: 1. Extract the latitude and longitude coordinates of the event occurrence point; 2. Search the candidate road set for the road centerline closest to the point, with a distance not exceeding 5 meters; 3. If a match is successful, the event will be recorded in the problem statistics table for that candidate road; 4. For consecutive events that cross road segments (such as long-distance lane keeping failure), the events are split and assigned to multiple related road segments according to the actual trajectory.

[0040] After the above processing, a "problem event distribution matrix" is formed, with each candidate road as the unit. Each row represents a candidate road, each column represents a problem type, and the cell value indicates the frequency of occurrence of that type of problem on that road. Furthermore, through multi-dimensional weighted aggregation analysis, the problem coverage of each candidate road is calculated.

[0041] Step S108: The scene matching degree and problem coverage are processed by the road priority scoring model to obtain the target road combination selected from each candidate road.

[0042] The road priority scoring model is an algorithm or mathematical model used to comprehensively evaluate the importance of candidate roads in terms of testing value. It receives multiple input indicators (such as scene matching degree and problem coverage), and outputs a "priority score" for each road or combination of roads through weighting, normalization, fusion and other processing methods, thereby supporting subsequent road selection decisions.

[0043] Understandably, the inputs to the road priority scoring model are scenario matching and problem coverage. Scenario matching refers to the similarity between the visual / environmental features of the current candidate road and preset typical test scenarios (such as school zones, tunnels, and intersections). Problem coverage refers to the number of typical problems (such as perception failure, planning anomalies, and control jitter) that have occurred on this road historically, and their frequency. The road priority scoring model processes the input in the following ways: normalizing data from different dimensions; setting weight coefficients (e.g., if the current project focuses on safety, then problem coverage has a higher weight); using machine learning models (such as regression models and neural networks) or rule engines for fusion scoring; and outputting a comprehensive score for each candidate road.

[0044] The target road combination refers to a set of optimal or multiple high-value roads selected from all candidate roads based on road priority scoring results, which are used as key routes for subsequent real vehicle testing, simulation injection or data collection.

[0045] Specifically, we first obtain the two types of key metrics output from the previous steps: The first category is the scene matching degree Mi∈[0,1] of each candidate road, which represents the degree of similarity between the i-th road and the preset typical driving scene (such as school area, main road intersection, tunnel, low light road section at night, etc.).

[0046] The second category is the problem coverage Ci∈[0,1] of each candidate road, which represents the comprehensive quantitative result of the richness and frequency intensity of problem types that have occurred during the historical operation of vehicles on the i-th road. For example, if a road has repeatedly experienced problems such as target miss detection, trajectory jitter, and false emergency braking, then its problem coverage is high.

[0047] Then, a configurable road priority scoring model is constructed, the basic form of which is a weighted linear fusion function. Understandably, the output of this road priority scoring model is the comprehensive priority score of the Xth candidate road; the weight coefficients of this road priority scoring model can be dynamically adjusted according to the test objective to cover more typical scenarios.

[0048] Furthermore, to avoid the selected roads being too concentrated in a certain region or scene type, a diversity constraint mechanism is introduced. For example, in the ranked set of high-scoring roads, a clustering method is used to ensure that the final selected roads are spatially and semantically representative.

[0049] Finally, target road combinations are generated based on the set filtering strategy. For example: Strategy 1: Select the top K=10 roads with the highest priority scores to form the initial combination; Strategy 2: Use the path planning module to connect the selected roads into one or more continuous drivable closed loop routes, forming a complete task route package that can be used for real vehicle road testing.

[0050] In one embodiment, the map is converted into a top-down view image; the image feature vector of each top-down view image is extracted using a map extraction model; the Euclidean distance between different image feature vectors is calculated; and the scene matching degree between each candidate road is determined based on the Euclidean distance between different image feature vectors.

[0051] In this context, a top-down view image refers to a two-dimensional rasterized image that presents map data of the target area from a top-down perspective. It is used to simulate the road structure layout perceived by vehicle sensors or maps. Understandably, top-down view images can effectively preserve road topology (number of lanes, turning radius, intersection shape, etc.), facilitating subsequent feature extraction.

[0052] Map extraction model refers to a machine learning model used to automatically identify and encode semantic information of road environment in top view images. Its output is a compact vector (i.e., "image feature vector") that can characterize the essential features of the road scene.

[0053] Image feature vectors are low-dimensional numerical vectors generated by map extraction models, used to mathematically represent the overall visual and structural features of the area where a candidate road is located.

[0054] Specifically, firstly, using the latitude and longitude trajectory points of each candidate road segment on the map as a reference, a fixed step size is set (e.g., sampling a viewpoint center every 5 meters), and the map data of the corresponding geographic coordinate area is rendered as a top-down view image at maximum zoom level. The image size is uniformly set to 256×256 pixels, and the color mode is either RGB or grayscale to ensure complete coverage of road geometry and lane-level structural features. Understandably, this step transforms abstract road vector data into a two-dimensional image with visual representation capabilities, laying the foundation for subsequent feature extraction based on deep learning.

[0055] Then, for each generated top-down view image, an image feature vector is automatically extracted using a map feature extraction model based on the D-LinkNet (semantic segmentation) network structure. The D-LinkNet network structure is an improved semantic segmentation network, consisting of an encoder, a central connection module, and a decoder. In the encoder, dilated convolutions are introduced to replace traditional pooling layers, maintaining the receptive field while avoiding spatial information loss. The utilization of the output feature maps from the second and third feature extraction layers of the encoder—the two intermediate layers—is enhanced, as they correspond to mid-level semantic information (such as lane line direction) and high-level semantic information (such as intersection type), respectively.

[0056] In one example, after inputting a 256×256 top-view image, it is forward-propagated to the second feature extraction layer of the encoder to obtain a feature map of size 128×32×32, and then to the third feature extraction layer of the encoder to obtain a feature map of size 256×16×16. The feature maps of these two levels are flattened and stitched together to form a high-dimensional feature vector, which serves as the digital scene representation of the road segment.

[0057] Finally, for the set of image feature vectors corresponding to all candidate roads, the Euclidean distance between any two pairs is calculated as the initial similarity metric. Furthermore, all pairwise distances are normalized to the interval [0, 1] to define the scene matching degree. This allows the construction of a multi-dimensional scene similarity matrix, where each element reflects the degree of similarity between the corresponding two candidate roads at the micro-driving scene level.

[0058] In one embodiment, the training steps of the map extraction model include: The unlabeled map is converted to grayscale to obtain a grayscale image; an initial edge image is generated based on the grayscale value of each pixel in the grayscale image and a preset grayscale threshold; the initial edge image is repaired for broken lines and noise is removed to obtain a binarized image; the binarized image is used as input and the grayscale image is used as output to train the pre-trained model to obtain the map extraction model.

[0059] Unlabeled maps refer to raw road data from high-precision navigation map databases or electronic map platforms. These data are renderable vector or raster images containing geographic information such as road topology, lane types, and traffic signs, but have not undergone manual annotation or semantic segmentation labeling. This map serves as the data input source for model training, allowing for the construction of training samples through self-supervised or weakly supervised methods without pixel-level annotation, thus reducing data preparation costs.

[0060] Optionally, unlabeled maps refer specifically to road top-view images generated by rendering at a fixed zoom level and perspective.

[0061] A grayscale image is the result of converting an original color map image (such as in RGB format) into a single-channel brightness image, where each pixel has a value ranging from 0 to 255, representing gray levels from black to white. In this application, the grayscale image is a base image obtained by grayscale conversion of an unlabeled map, used for subsequent edge extraction and feature enhancement operations.

[0062] The preset grayscale threshold refers to an empirical numerical range used to filter specific pixel areas. Optionally, in this application, it is set to grayscale values ​​[254, 255]. Since lane markings (white or yellow solid / dashed lines) typically appear close to pure white after map rendering, the grayscale values ​​of their corresponding pixels are concentrated in the high-level range. By setting this threshold, only pixels with grayscale values ​​within this range are retained, which can effectively extract clear and continuous lane line outlines, eliminate interference information such as background road surfaces, green belts, and buildings, and form preliminary edge candidate areas.

[0063] The initial edge image refers to the preliminary edge result image extracted based on the grayscale image and a preset grayscale threshold. It is represented as a binary image, where pixels that meet the grayscale conditions are set as foreground (value 1), and the rest are background (value 0). This initial edge image reflects the approximate location and direction of the lane lines, but due to problems such as breaks, burrs, and isolated noise, it does not yet have complete geometric connectivity and needs further morphological optimization processing as the basis for generating high-quality training targets.

[0064] A binarized image refers to a clean edge image obtained after performing line break repair and noise removal on an initial edge image. It remains a single-channel binary image (pixel values ​​are only 0 or 1). In this application, the binarized image undergoes skeletonization processing through a two-stage iterative thinning algorithm: first, spatial filtering is used to eliminate isolated points; then, a conditional judgment function based on 8-neighborhood connectivity is applied to progressively remove redundant boundary pixels; finally, a continuous road centerline or marking skeleton image with a width of approximately one pixel and no branch breaks is output, i.e., the binarized image.

[0065] A pre-trained model refers to the parameters of a deep neural network model that has been trained on a large-scale general image dataset or a public road segmentation dataset. Specifically, in this application, it refers to the encoder-decoder structure and weight parameters of the D-LinkNet network. This pre-trained model has basic image feature extraction capabilities, and is particularly good at capturing linear structures and scene layouts.

[0066] In this application, the pre-trained model is used as a starting point, its input layer is adapted to the map image size of this task, and the "unlabeled map" is used as input and the "binarized image corresponding to the grayscale image" is used as the expected output. Transfer learning and fine-tuning are performed to quickly converge to obtain a dedicated map extraction model suitable for lane-level map feature extraction.

[0067] For details, please refer to the following: Figure 2 First, S202 performs grayscale processing on the unlabeled map to obtain a grayscale image. This includes acquiring a batch of original high-precision navigation map images from the target area. Each image corresponds to a top-down view screenshot of a candidate road segment, with a uniform size of 256×256 pixels and an RGB color image format. Since these images are only used for training and do not require manual labeling, they are called "unlabeled maps".

[0068] refer to Figure 3 The images are processed frame by frame to convert the three-channel color images into single-channel grayscale images, where the grayscale value of each pixel (x, y) ranges from [0, 255]. This grayscale image retains the main visual information of the road geometry and marking distribution, serving as the basis for subsequent edge extraction.

[0069] Next, S204 generates an initial edge image based on the grayscale value of each pixel in the grayscale image and a preset grayscale threshold. This includes setting a preset grayscale threshold range [254, 255] for the obtained grayscale image. The selection of this range is based on the following: In the rendering rules of mainstream electronic map platforms, key traffic markings such as lane center lines, boundary lines, and directional arrows are usually displayed in near-pure white, and the grayscale values ​​of their corresponding pixels are concentrated in the high range, while road surfaces, green belts, buildings, and other features are presented in lower grayscale. Understandably, this preset grayscale threshold range can also be adjusted according to the actual situation of the application scenario and requirements. Therefore, each pixel in the grayscale image is traversed. If its grayscale value falls within the range of [254, 255], it is set as the foreground (value 1); otherwise, it is set as the background (value 0), thereby generating a binary image, i.e., the initial edge image. Understandably, this initial edge image initially outlines the approximate location and direction of the lane lines, but due to problems such as broken lines, burrs, and isolated noise, it cannot be directly used as an ideal learning target and needs further optimization.

[0070] Next, S206 performs line break repair and noise removal on the initial edge image to obtain a binarized image. This includes: to improve the quality of the initial edge image, a series of morphological processing operations are performed on the initial edge image, ultimately obtaining a continuous, single-pixel-width, topologically correct road skeleton image, denoted as the "binarized image". This process employs a two-stage iterative thinning algorithm, specifically including: Phase 1 (Sub-iteration 1): In this sub-step, a foreground pixel P1 will be marked as to be deleted if and only if it meets the following 4 conditions (but it will not be deleted immediately; it will be processed uniformly after all pixels have been judged).

[0071] Condition 1: 2≤B(P1)≤6 (ensure that P1 is a boundary point, not an interior point or an isolated endpoint).

[0072] Condition 2: A(P1) = 1 (ensuring that deleting P1 will not disrupt the connectivity of the region).

[0073] Condition 3: P2×P4×P6=0 (Check the east, south, and north directions: at least one of them is the background).

[0074] Condition 4: P4×P6×P8=0 (Check the south, west, and north directions: at least one of them is the background).

[0075] Second stage (sub-iteration 2): Traverse all foreground pixels again on the updated image, with judgment conditions similar to those in the first stage, except that conditions 3 and 4 change.

[0076] Condition 1: 2≤B(P1)≤6 (ensure that P1 is a boundary point, not an interior point or an isolated endpoint).

[0077] Condition 2: A(P1) = 1 (ensuring that deleting P1 will not disrupt the connectivity of the region).

[0078] Condition 3: P2×P4×P8=0 (Check the east, north, and west directions: at least one of them is the background).

[0079] Condition 4: P2×P6×P8=0 (Check the east, south, and west directions: at least one of them is the background).

[0080] The two sub-iterations are executed alternately until the image no longer changes, and the output is the thinned binarized image. This image represents the central skeleton of the road markings, has good continuity and structural integrity, and is suitable as the target output of a deep learning model.

[0081] Finally, S208 uses the binarized image as input and the grayscale image as output to train the pre-trained model, resulting in a map extraction model. This includes training the pre-trained model using the binarized image as input and the grayscale image as output. For example, this application innovatively constructs a reverse supervised training paradigm: using the processed binarized image as input to the network and the original grayscale image as the desired output, a neural network with an encoder-decoder structure is trained, enabling it to learn to reconstruct the original visual image from a simplified structural representation.

[0082] The pre-trained model uses a network architecture based on the D-LinkNet model, which has been pre-trained on a large-scale dataset and has good feature extraction capabilities, hence the name "pre-trained model". This step involves fine-tuning it, with the following specific configuration: the input layer receives a binarized image of size 256×256; the encoder extracts structural features, and the decoder fuses multi-layer features through skip connections, progressively upsampling to restore the original resolution; the output layer generates a reconstructed image of the same size as the original grayscale image; the loss function is a weighted combination of absolute difference loss and perceptual loss.

[0083] During training, the Adam (Adaptive Moment Estimator) optimizer was used with an initial learning rate of 1e-4 and a batch size of 16, for a total of 100 training epochs. Finally, the optimal model parameters were saved to obtain a map extraction model specifically for lane-level map feature extraction.

[0084] In one embodiment, the pixels in the initial edge image are traversed. When a first target pixel that meets a first preset condition is encountered, the first target pixel is marked as a first pixel to be deleted. Based on the first pixel to be deleted, an iterative edge image is generated. The pixels in the iterative edge image are traversed. When a second target pixel that meets a second preset condition is encountered, the second target pixel is marked as a second pixel to be deleted. Based on the second pixel to be deleted, a binarized image is generated.

[0085] The first target pixel refers to the foreground pixel identified as safe for deletion during the first stage of thinning operation on the initial edge image. These pixels are typically located on small branches of the edge or isolated noise points; deleting them helps simplify the image structure without affecting the connectivity of the main body.

[0086] The iterative edge image refers to the updated edge image obtained after the first stage of pixel marking and deletion operations. It is the intermediate result image formed after all marked "first pixels to be deleted" are set from the foreground (1) to the background (0). This image serves as the input for the next stage of processing, entering the second stage of the thinning process. Because the algorithm employs an alternating iterative strategy, the "iterative edge image" is not the final output, but rather a transitional state image used to support subsequent processing steps.

[0087] The second target pixel refers to the foreground pixel identified as eligible for further deletion during the second-stage thinning operation on the iterative edge image. Understandably, the judgment conditions of the second stage complement those of the first stage, respectively cleaning up redundant structures in different spatial directions. Alternating execution of the two stages allows for more comprehensive removal of edge burrs while preserving critical connection paths.

[0088] Specifically, the input is the initial edge image generated in the previous step, which is a binary image with a size of 256×256, where the foreground pixel value is 1 (representing the lane line area) and the background pixel value is 0. The image is scanned pixel by pixel. For each foreground pixel X, an 8-neighborhood is defined. A foreground pixel X is marked as the "first target pixel" and classified as the "first pixel to be deleted" if and only if it simultaneously satisfies the four conditions of the first stage mentioned above. Understandably, all pixels that satisfy the above conditions are only "marked" and the image data is not modified temporarily; they will be deleted uniformly after the entire image has been traversed.

[0089] All marked "first pixels to be deleted" are set from the foreground (1) to the background (0), resulting in an updated image called the "iterative edge image". This image is used as input for the next stage of processing and enters the second sub-iterative process. Understandably, this operation achieves effective removal of redundant pixels in a specific direction, but if the same rule is applied again immediately, it may cause over-erosion or breakage. Therefore, a complementary second-stage processing strategy is introduced.

[0090] The iterative edge image is then traversed again, and each foreground pixel is checked to see if it meets the judgment conditions in the second stage of the above steps. If and only if all four conditions are met simultaneously, the pixel is marked as the "second target pixel" and recorded as the "second pixel to be deleted". This set of conditions complements the first stage, focusing on removing burrs and false edges existing in different spatial orientations, and avoiding directional deviations caused by a single rule.

[0091] Finally, all "second target pixels" are uniformly set to the background value (0), completing the second sub-iteration of this round and obtaining a new image. Subsequently, this core image is used as the new input, and sub-iteration 1 (first target pixel identification) and sub-iteration 2 (second target pixel identification) are repeated alternately until no pixel is marked as to be deleted in a certain round. At this point, the algorithm converges and outputs the final result image, which is the "binarized image".

[0092] In one embodiment, associated problem data in each candidate road is determined; the associated problem data is deduplicated by a preset duplicate determination condition, which includes considering problem data with the same scene description and the same fault phenomenon as the same problem data; the ratio between the number of deduplicated problem data and the number of problem data that have occurred on the target area roads is used as the problem coverage between each candidate road.

[0093] The pre-defined duplicate detection criteria refer to the standard basis used to identify and determine whether different problem data recorded by historical vehicles during the testing process belong to the same type of fault event. Its core purpose is to effectively deduplicate massive, multi-source problem reports during the aggregation analysis phase, avoid the same problem being counted repeatedly, and thus more accurately assess the ability of each candidate road to expose weaknesses.

[0094] Specifically, each issue record (i.e., issue data) is bound to its corresponding physical road segment using a spatial matching algorithm (optionally, geofence-based point inclusion judgment or polygon matching). All issue records belonging to the same candidate road are aggregated into a "related issue dataset" for that road. For example, if candidate road R001 is "Changchun Renmin Street from Jiefang Avenue to Ziyou Avenue", then all issue records whose GPS locations fall within this range are marked as associated with R001.

[0095] Because autonomous driving systems may repeatedly exhibit the same type of malfunction in identical or highly similar driving scenarios (e.g., multiple instances of ACC (Adaptive Cruise Control) erroneously slowing down at a ramp curve), directly counting the original number of problems would overestimate the "problem exposure value" of that road. Therefore, it is necessary to deduplicate related problem data. Optionally, automated judgment can be achieved through field comparison or natural language similarity calculation.

[0096] In one example, if both issue record A and issue record B occur under the conditions of "urban expressway, four lanes, dashed line separation, sunny daytime," and both exhibit the characteristic of "adaptive cruise control (ACC) misjudging the distance to the vehicle in front, causing sudden deceleration," then they are considered duplicate issues and merged into a single issue entry. Understandably, this deduplication mechanism effectively avoids statistical bias caused by differences in test frequency, ensuring that "issue coverage" reflects the road's ability to trigger unique failure modes, rather than simply a stacking of issue numbers.

[0097] After deduplication, the number of unique problem data associated with each candidate road (representing the total number of unique faults that each candidate road can expose) and the total number of global problems after deduplication of all candidate roads in the entire target area are calculated, i.e., the union of all unique problems. The ratio between the two is calculated, which reflects the contribution of the road to the overall problem exposure capability, and the value ranges from [0, 1]. The higher the value, the more typical or critical failure scenarios the road can represent.

[0098] Optional weighted expansion allows for assigning different weights based on the severity level of the problem (e.g., a weight of 2 for a serious fault and a weight of 1 for a general warning), with the weighted number of problems participating in the calculation to further improve the rationality of the scoring.

[0099] In one embodiment, multi-dimensional road condition features are extracted from driving data in the city where the test area is located, and a local scene feature vector for the city where the test area is located is constructed based on the multi-dimensional road condition features. The local scene feature vector is matched with the standard scene feature vector of the city where the vehicle is mainly sold to obtain the scene matching degree. If the scene matching degree is less than a preset matching degree threshold, the priority weights of roads corresponding to highly differentiated scenes in the city where the test area is located are adjusted. Based on the problem heatmap, the problem coverage potential value and accident risk coefficient corresponding to each road in the city where the test area is located are determined. The problem heatmap includes the occurrence locations of various historical road condition problems. Based on the problem coverage potential value, accident risk coefficient, and priority weights of each road in the city where the test area is located, a fusion scoring model is constructed. Based on the fusion score of the fusion scoring model, the priority weights of each road are adjusted under the constraint of a preset total test mileage, and test path combinations are generated under the constraint. Based on the test path combinations and each path label, the loss function of the fusion scoring model is iteratively adjusted to obtain a trained road priority scoring model.

[0100] The test area refers to the target geographical area where real-vehicle verification will be conducted, typically a set of road networks with clearly defined boundaries, such as a city, a specific administrative region, or a highway network. The test area is the specific target for data collection and route planning in this method, and its map data and historical problem records will serve as the basic resources for model input.

[0101] Multidimensional road condition features refer to a combination of technical parameters extracted from driving data that characterize the complexity of the road environment, used to quantitatively describe typical traffic and road conditions within a certain area. In this application, multidimensional road condition features include at least the following categories: road geometry features, traffic flow features, environmental perception challenges, and infrastructure density.

[0102] The local scene feature vector refers to a fixed-length numerical vector encoded by normalizing the multi-dimensional road condition features within the test area, which is used to represent the unique driving scene patterns of that area.

[0103] The main sales cities for vehicles refer to the cities where the target model has the highest sales volume and the most concentrated user base. These cities represent the market environment in which the product primarily serves, and their road characteristics and user behaviors are typical and representative. In this application, the main sales cities are selected to construct a "standard scenario feature library" as a benchmark for comparison with other test areas.

[0104] Standard scenario feature vectors refer to a standardized set of feature vectors representing mainstream driving environments, constructed based on multi-dimensional road condition features of major vehicle sales cities. This vector set can be viewed as the distribution of the main scenarios expected to be addressed during autonomous driving design, and typically consists of multiple sub-vectors (each corresponding to a typical scenario), forming a "standard scenario feature library". Optionally, the standard scenario feature vectors can be constructed by extracting N typical scenario cluster centers as standard feature vectors through data clustering analysis of major sales cities.

[0105] The preset matching threshold is an empirical numerical threshold used to determine whether the scene similarity between the test area and the main sales city is sufficient. In this application, cosine similarity, Euclidean distance, or other similarity measurement methods are used to calculate the matching degree between the local scene feature vector and the standard scene feature vector. When this value is lower than the preset matching threshold (e.g., 0.75), it is determined that there is a significant difference.

[0106] A problem heatmap is a spatial distribution map generated based on historical vehicle problem data. The color intensity or numerical value of each road segment reflects the density of fault events occurring on that segment. The problem heatmap is based on a GIS map, overlaid with unique, deduplicated problem records, and visualized according to the frequency of problems per unit mileage or a weighted severity level.

[0107] Problem coverage potential refers to the estimated probability that a road may expose new problems in future testing, based on the growth trend of the number of currently discovered problems. Specifically, it can be estimated by fitting a "problem discovery cumulative curve" and calculating its slope, or by using a Poisson regression model to estimate the remaining potential problems.

[0108] The accident risk coefficient is a quantitative indicator of the high-risk attributes of a road due to traffic complexity or historical accident records. This coefficient comprehensively considers the following factors: the frequency of accidents in official traffic accident statistics; the density of high-risk scenarios (such as unprotected left turns and pedestrian crossings); the number of days affected by severe weather; and the proportion of road sections with insufficient nighttime lighting. Finally, a weighted summation method is used to generate a single value to reflect the objective degree of danger of the road.

[0109] The fusion scoring model is a multi-factor weighted decision-making model used to comprehensively evaluate the combined value of each candidate road in terms of scene representativeness and problem exposure capability, and output a comparable numerical score. This model is the core computational unit of the road priority scoring system. Optionally, the fusion scoring model can take the form of a linear weighted function, a neural network model, or a gradient boosting tree, etc., and its input variables include at least: scene matching degree (from map image feature analysis); problem coverage potential value (based on historical problem growth trends); accident risk coefficient (from traffic accident statistics); and differential scene enhancement weights (triggered by differences between local and main city scenes); its output is the fusion score for each road.

[0110] The preset total test mileage refers to the total target test mileage set in advance based on product development plans, regulatory requirements, or market user research results. This parameter serves as a hard constraint in the road selection process, used to control the length of the final generated test path to not exceed the resource budget.

[0111] In one example, the preset total test mileage can be determined as follows: based on the average intelligent driving mileage during the user's life cycle × the number of years the user changes vehicles × the safety redundancy factor. A test path combination refers to one or more sets of continuous drivable routes formed by connecting multiple candidate road segments, constituting a complete real-vehicle test task plan. This path combination is the optimal set generated after sorting and selecting the scores of each road according to the fusion scoring model, under the premise of satisfying the "preset total test mileage" constraint.

[0112] Each path label refers to the semantic annotation information attached to each generated test path combination, used to describe its technical value and test objective at the functional verification level. These labels participate as supervision signals in the training loop of the fusion scoring model, realizing continuous evolution of "practice-feedback-optimization".

[0113] Specifically, a large amount of driving data is first collected from actual vehicles operating in the test city (i.e., the test area), and various key information reflecting road complexity is extracted from it, such as the density of curves, gradient changes, and lane number distribution. These are collectively referred to as "multi-dimensional road condition features." Based on these features, a "local scene feature vector" that can represent the overall driving environment characteristics of the city is constructed to digitally characterize its unique road conditions.

[0114] Next, the local scene feature vector of the city is compared with the "standard scene feature vector" corresponding to the identified main sales cities, and the similarity between the two is calculated to obtain a "scene matching degree". If it is found that the area where a road is located differs significantly from the typical scene of the main sales city, that is, the scene matching degree is lower than the pre-set reasonable range (preset matching degree threshold), it indicates that this road may contain some atypical, edge-of-the-road driving conditions. Although such roads are not common, they may be the key road sections that expose boundary problems. Therefore, the priority weight of roads corresponding to these highly differentiated scenes will be automatically increased, making them more likely to be selected in subsequent screening.

[0115] Simultaneously, a "problem heat map" is referenced. This map displays the location and concentration of various functional anomalies throughout history. Based on this map, it can be determined whether each road has frequently experienced malfunctions in the past, thereby assessing its likelihood of exposing new problems in the future; this indicator is called the "problem coverage potential value." Furthermore, by combining traffic accident statistics released by the traffic police department, accident-prone road sections are identified and assigned a higher "accident risk coefficient," indicating that these roads have higher safety verification value.

[0116] Next, after obtaining the above indicators, they are integrated to construct a "fusion scoring model." This model comprehensively considers the uniqueness of a road's scenario, the likelihood of problem exposure, and the level of accident risk, providing a comprehensive score as the core basis for measuring its testing priority. Under the hard constraint of not exceeding the predetermined "total test mileage," the roads with the highest scores are selected according to their ranking, ensuring that they can be connected to form one or more continuous and feasible driving routes, ultimately generating an optimal set of "test path combinations." After these test paths are actually executed, the verification effect of each path is summarized, such as recording whether it successfully triggered the expected functional problem or discovered new boundary cases. These results are compiled into "path labels," marking the actual technical value of each path.

[0117] Finally, this real-world test feedback is used to check the accuracy of the original fusion scoring model's predictions. If certain roads that should be important are not recommended with high scores, or inefficient paths are incorrectly overestimated, the model will automatically adjust its internal parameters and correct its scoring logic. Through round after round of "path generation—real-world verification—feedback optimization," the road priority scoring model continuously improves itself, gradually enhancing the efficiency of test resource utilization.

[0118] In one embodiment, user driving data from the main sales cities of the vehicles is obtained; based on the user driving data, the average annual mileage and average annual vehicle replacement cycle of each vehicle in the main sales cities are determined; based on the average mileage, average vehicle replacement cycle and preset safety factor, a preset total test mileage is determined.

[0119] Specifically, the first step is to acquire user driving data for the target vehicle model in major sales cities. This data comes from information transmitted back from the vehicle networking platforms of vehicles already on the market, covering regions with the highest sales volume and the highest concentration of users, such as Beijing, Shanghai, Guangzhou, and Shenzhen. The data includes long-term operational records for each vehicle, such as daily mileage, runtime, and function activation status, typically spanning at least one year to ensure the representativeness of the statistical results.

[0120] Next, statistical analysis was conducted on these user driving data to calculate two key indicators: first, the average annual mileage of all vehicles in major sales cities, i.e., "average annual mileage"; and second, the average number of years a car owner uses the vehicle from purchase to replacement, i.e., "average annual replacement cycle". For example, data analysis revealed that users of this model drive an average of approximately 12,000 kilometers per year and replace their vehicles on average every 6 years. This suggests that a single vehicle's cumulative mileage over its entire lifespan is approximately 72,000 kilometers.

[0121] Furthermore, a "preset safety factor" is introduced to compensate for the differences between the test environment and real roads, thereby improving the adequacy of the verification. This preset safety factor is set by comprehensively considering the complexity of autonomous driving, functional safety level requirements, and corporate quality control standards, and is generally between 1.2 and 2.0. For example, if the safety factor is set to 1.5, the final preset total test mileage is: the user's lifetime driving mileage multiplied by this preset safety factor, ensuring that the test intensity is higher than the actual usage level.

[0122] The resulting preset total test mileage not only reflects real market demand but also includes sufficient redundancy to effectively cover various boundary conditions and low-probability, high-risk events, thereby comprehensively verifying the stability and safety of intelligent driving. In this application, the preset total test mileage is used as a hard constraint to guide the subsequent road selection process. When generating test path combinations, the cumulative length of the selected roads will be automatically controlled to not exceed this preset value, avoiding resource waste or over-testing and achieving the goal of "precise investment and efficient verification".

[0123] Figure 4 A schematic diagram of a test road selection device 400 according to an embodiment of this application is shown. Exemplarily, the test road selection device 400 includes: The acquisition module 402 is used to acquire a map of the target area and historical data on problems that have occurred on roads in the target area. The matching module 404 is used to perform image matching on the map to obtain the scene matching degree between each candidate road; Module 406 is used to perform aggregate analysis on the problem data and determine the problem coverage among the candidate roads. The processing module 408 is used to process the scene matching degree and problem coverage through the road priority scoring model to obtain the target road combination selected from each candidate road.

[0124] It is understood that the apparatus in this embodiment corresponds to the test road selection method in the above embodiment, and the options in the above embodiment are also applicable to this embodiment, so they will not be described again here.

[0125] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to enable the terminal device to perform the functions of the various modules in the above-described test road selection method or the above-described test road selection device.

[0126] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0127] Memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). Memory is used to store computer programs, and the processor can execute these programs upon receiving execution instructions.

[0128] This application also provides a computer-readable storage medium for storing computer programs used in the aforementioned terminal devices. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0129] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0130] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0131] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0132] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for selecting test roads, characterized in that, The method includes: Obtain a map of the target area, as well as historical data on problems encountered by vehicles on each candidate road within the target area; Image matching is performed on the map to obtain the scene matching degree between each candidate road; The problem data is aggregated and analyzed to determine the problem coverage among the candidate roads; The scenario matching degree and the problem coverage are processed by the road priority scoring model to obtain the target road combination selected from the candidate roads.

2. The method according to claim 1, characterized in that, The step of performing image matching on the map to obtain the scene matching degree between each of the candidate roads includes: Convert the map into a top-down view image; The image feature vector of each top-view image is extracted using a map extraction model. Calculate the Euclidean distance between the different image feature vectors; The scene matching degree between each candidate road is determined based on the Euclidean distance between the different image feature vectors.

3. The method according to claim 2, characterized in that, The training steps of the map extraction model include: The unlabeled map is converted to grayscale to obtain a grayscale image; An initial edge image is generated based on the grayscale value of each pixel in the grayscale image and a preset grayscale threshold; The initial edge image is subjected to line break repair and noise removal to obtain a binarized image; The pre-trained model is trained by using the binarized image as input and the grayscale image as output to obtain the map extraction model.

4. The method according to claim 3, characterized in that, The initial edge image is subjected to line break repair and noise removal to obtain a binarized image, including: Traverse the pixels in the initial edge image, and when a first target pixel that meets the first preset condition is encountered, mark the first target pixel as the first pixel to be deleted; Based on the first pixel to be deleted, an iterative edge image is generated; Traverse the pixels in the iterative edge image. When a second target pixel that meets the second preset condition is encountered, mark the second target pixel as the second pixel to be deleted. A binarized image is generated based on the second pixel to be deleted.

5. The method according to claim 1, characterized in that, The aggregation analysis of the problem data to determine the problem coverage among the candidate roads includes: Identify the associated problem data in each of the candidate roads; The related problem data is deduplicated by using preset duplicate determination conditions. The preset duplicate determination conditions include considering problem data with the same scenario description and the same fault phenomenon as the same problem data. The ratio between the number of deduplicated problematic data and the number of problematic data that have occurred on roads in the target area is used as the problem coverage among the candidate roads.

6. The method according to claim 1, characterized in that, The training steps of the road priority scoring model include: Multidimensional road condition features are extracted from driving data in the city where the test area is located, and a local scene feature vector of the city where the test area is located is constructed based on the multidimensional road condition features. The local scene feature vector is matched with the standard scene feature vector of the main city where the vehicle is sold to obtain the scene matching degree; If the scene matching degree is less than the preset matching degree threshold, the priority weight of the roads corresponding to the high-discrepancy scenes in the city where the test area is located will be adjusted. Based on the problem heatmap, the problem coverage potential value and accident risk coefficient of each road in the city where the test area is located are determined. The problem heatmap includes the location of various historical road condition problems. Based on the problem coverage potential value, the accident risk coefficient, and the priority weights of each road in the city where the test area is located, a fusion scoring model is constructed. Based on the fusion score of the fusion scoring model, the priority weight of each road is adjusted under the constraint of the preset total test mileage, and a combination of test paths is generated under the constraint. Based on the test path combination and each path label, the loss function of the fusion scoring model is iteratively adjusted to obtain a trained road priority scoring model.

7. The method according to claim 6, characterized in that, The method for determining the preset total test mileage includes: Obtain user driving data from the main cities where the vehicles are sold; Based on the user driving data, the average annual mileage and average annual vehicle replacement cycle of each vehicle in the main sales cities of the vehicles are determined. Based on the average mileage, the average vehicle replacement cycle, and the preset safety factor, the preset total test mileage is determined.

8. A test road selection device, characterized in that, include: The acquisition module is used to acquire a map of the target area, as well as historical data on problems that vehicles have encountered on each candidate road in the target area. The matching module is used to perform image matching on the map to obtain the scene matching degree between each of the candidate roads; The determination module is used to perform aggregate analysis on the problem data and determine the problem coverage among the candidate roads; The processing module is used to process the scene matching degree and the problem coverage through a road priority scoring model to obtain the target road combination selected from the candidate roads.

9. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the test road selection method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements the test road selection method according to any one of claims 1-7.

Images