A method for updating lane-level dynamic information in high-precision maps based on crowdsourced data

Abstract

This invention proposes a method for updating lane-level dynamic information in high-precision maps based on crowdsourced data. The method includes: collecting trajectory information and road image data from multiple sources, with a high-precision map embedded within; constructing a single-stage line-anchor detection model using a convolutional neural network; performing lane line perception detection on the road image data based on the single-stage line-anchor detection model; outputting lane line perception information; using an image alignment compensation algorithm to register and fuse road image data from consecutive frames during vehicle movement, and inputting this data into a target detection model for recognition, outputting traffic cone perception information; and projecting the lane line perception information and traffic cone perception information onto the high-precision map's geographic coordinate system using camera parameters. This method effectively suppresses missed and false detections in single-vehicle perception by combining low-cost crowdsourced data with a multi-vehicle evidence fusion mechanism, improving the accuracy and real-time performance of lane-level dynamic information updates, and meeting the high-precision requirements of autonomous driving for high-precision maps.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for updating lane-level dynamic information of high-precision maps based on crowdsourced data. Background Technology

[0002] High-precision maps, as the core infrastructure of autonomous driving, carry key information such as high-precision road geometry, lane attributes, and traffic facilities, providing indispensable support for vehicle positioning, environmental perception, and path planning. As autonomous driving technology evolves to higher levels, the requirements for the freshness of high-precision maps are becoming increasingly stringent. Real-time updates of lane-level dynamic information such as temporary lane closures, road construction, and traffic cone control are directly related to driving safety and traffic efficiency, becoming one of the core factors affecting the reliability of autonomous driving systems.

[0003] Most existing high-precision map dynamic update methods are based on single-vehicle data, which are easily affected by environmental factors such as changes in lighting and vehicle posture fluctuations. The detection accuracy and stability are not good. The perception task attributes of lane lines and traffic cones are very different, and a single model cannot take into account the accuracy of both. Moreover, single-vehicle perception has inherent problems of missed detection and false detection, which makes the reliability of dynamic information updates unable to meet the high-precision requirements of autonomous driving for high-precision maps. Summary of the Invention

[0004] In view of this, the present invention proposes a method for updating lane-level dynamic information of high-precision maps based on crowdsourced data. By combining low-cost crowdsourced data with a multi-vehicle evidence fusion mechanism, it can effectively suppress missed detections and false detections in single-vehicle perception, improve the accuracy, reliability and real-time performance of lane-level dynamic information updates, and meet the high-precision requirements of autonomous driving for high-precision maps.

[0005] The technical solution of this invention is implemented as follows: Firstly, this invention provides a method for updating lane-level dynamic information in high-precision maps based on crowdsourced data, comprising the following steps: S1, collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information. S2 uses a convolutional neural network to construct a single-stage detection model for lane anchors. Based on the single-stage detection model for lane anchors, lane lines in road image data are detected and perceived. The start and end coordinates and slope of the lane lines in the image coordinate system are output as lane line perception information. S3 uses an image alignment compensation algorithm to register road image data of consecutive frames during vehicle travel, calculates motion difference maps of consecutive frames, and performs multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The data is then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information. S4 projects lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system through camera parameters; S5. Based on the collected vehicle trajectory data, the hidden Markov model is used to determine the road segment where the vehicle is located, the candidate lane set is filtered, and the lane width information and lane slope information are combined to calculate the vehicle's lateral absolute position in the current lane and determine the current lane of the vehicle. S6, based on evidence theory, constructs the perception information of multiple vehicles on the same lane into three types of evidence vectors with probabilistic quality, including lane closure state, lane normal traffic state, and observation uncertainty state. The evidence vectors of multiple vehicles are accumulated and fused through evidence combination rules, and the high-precision map is updated after deciding the lane dynamic state according to a preset threshold.

[0006] Based on the above technical solutions, preferably, step S2 includes the following sub-steps: S21, a single-stage detection model for line anchors is constructed using a convolutional neural network. The backbone network of the convolutional neural network is used to extract features from the road image. The backbone network of the convolutional neural network includes multiple downsampling convolutional layers connected in sequence, which are used to output high-level feature maps and shallow feature maps respectively. S22, a set of radiating line anchors emanating from the bottom to the top of the image are predefined on the high-level feature map. Each line anchor has a preset starting point coordinate, direction angle and length. The starting point coordinate is fixed at the bottom edge of the image, and the direction angle is selected within a preset angle range. S23, the feature vectors of the areas covered by each anchor are mapped to fixed-length feature vectors through the region of interest alignment operation, and then input to the classification branch and the coarse regression branch respectively. The classification branch outputs the confidence score that the anchor belongs to the lane line, and the coarse regression branch outputs the coordinate offset between the start point of the anchor and the start point of the real lane line and the coordinate offset between the end point of the anchor and the end point of the real lane line. S24. Select the top K line anchors as candidate line anchors based on the confidence scores from high to low, and perform coarse correction on the start and end coordinates of the candidate line anchors based on the output coordinate offset of the coarse regression branch to obtain the rough start and end coordinates of the lane line on the high-level feature map. S25, map the coarse start and end point coordinates of the candidate line anchors from the high-level feature map to the shallow feature map to obtain the corresponding mapping region on the shallow feature map, extract the local feature vector of the mapping region, input it into the fine regression branch, and the fine regression branch outputs the fine offset to correct the coarse start and end point coordinates to obtain the fine start and end point coordinates of the lane line in the image coordinate system. S26, calculate the lane slope based on the fine start and end point coordinates, and output the start coordinates, end coordinates and slope of the left lane line, as well as the start coordinates, end coordinates and slope of the right lane line.

[0007] Based on the above technical solutions, preferably, step S3 includes the following sub-steps: S31, acquire road image data of consecutive frames during vehicle travel, and register adjacent frame images using an image alignment compensation algorithm to obtain the aligned and compensated consecutive frame images; S32, calculate the pixel-by-pixel motion difference based on the aligned and compensated consecutive frame images, and construct the motion difference map of the consecutive frames; S33, the original road image features of the corresponding frame are stitched and fused with the motion difference map in the channel dimension to obtain a multimodal feature map; S34. Input the multimodal feature map into the target detection model for recognition, output the coordinates of the traffic cone center point and the confidence score in the image coordinate system, and filter the traffic cone detection results with a confidence score greater than the confidence threshold as traffic cone perception information.

[0008] Based on the above technical solutions, preferably, step S4 includes the following sub-steps: S41, obtain the camera's intrinsic parameter matrix, and use the intrinsic parameter matrix to convert the lane line perception information and traffic cone perception information in the image coordinate system into three-dimensional coordinates in the camera coordinate system; S42, obtain the camera's extrinsic parameter matrix, which includes a rotation matrix and a translation vector. Use the extrinsic parameter matrix to convert the three-dimensional coordinates in the camera coordinate system into three-dimensional coordinates in the vehicle coordinate system. S43, acquire the vehicle's positioning information, which includes the global navigation satellite system position and the vehicle's attitude angle. Use the positioning information to transform the three-dimensional coordinates in the vehicle coordinate system to the geographic coordinate system of the high-precision map, and output the projection results of lane lines and traffic cones in the geographic coordinate system of the high-precision map.

[0009] Based on the above technical solutions, preferably, step S5, which involves using a hidden Markov model to determine the road segment where the vehicle is located based on the collected vehicle trajectory data and filtering the candidate lane set, includes the following sub-steps: S51, the road segments adjacent to the spatial location of the vehicle trajectory in the high-precision map are taken as the set of hidden states of the hidden Markov model, and the trajectory points of the trajectory data collected during the vehicle's driving process are taken as the observation state sequence. S52, for any trajectory point and candidate road segment, calculate the minimum perpendicular distance from the trajectory point to the centerline of the candidate road segment, and calculate the probability of observing each trajectory point under the condition that the vehicle is located in the candidate road segment based on the Gaussian distribution, and obtain the observation probability; S53, calculate the observed displacement distance between two adjacent trajectory points, and calculate the shortest path distance between candidate road segments based on the candidate road segments corresponding to the two adjacent trajectory points. Calculate the state transition probability between the two candidate road segments based on the difference between the observed displacement distance between trajectory points and the shortest path distance between candidate road segments. S54. Based on the spatial distance between the starting trajectory point of the trajectory sequence and each candidate road segment, calculate the probability that the vehicle's initial state is located in each candidate road segment, and obtain the initial state probability. S55. Based on the observation probability, state transition probability and initial state probability, the Viterbi algorithm is used to solve the road segment sequence corresponding to the maximum posterior probability to obtain the optimal matching road segment sequence. S56. Based on the frequency of occurrence of each road segment in the optimal matching road segment sequence, calculate the posterior probability of each road segment, filter road segments with posterior probabilities higher than a preset probability threshold, and construct a candidate road segment set.

[0010] Based on the above technical solutions, preferably, step S5, which combines lane width information and lane slope information to calculate the vehicle's lateral absolute position in the current lane and determine the vehicle's current lane, includes the following sub-steps: S57, obtain the slope of the left lane line and the slope of the right lane line, and calculate the ratio of the distance from the vehicle to the left lane line to the total lane width based on the difference between the slopes of the left lane line and the right lane line and the slope of the right lane line. S58: Obtain the width information of the current lane from the high-precision map, multiply the current lane width by a ratio, and obtain the lateral absolute position of the vehicle relative to the left boundary of the current lane. S59: Obtain all lane information of the road where the vehicle is located in the horizontal direction from the high-precision map, take the leftmost boundary of the road as the starting position in the horizontal direction, arrange them in the horizontal order of the road, and accumulate them according to the width of each lane to construct the cumulative width range of each lane in the horizontal direction of the road. S510: Based on the vehicle's lateral absolute position relative to the left boundary of the current lane and the sequential number of the current lane in the road, the widths of all lanes before the current lane are accumulated, and then added to the vehicle's lateral absolute position in the current lane to obtain the vehicle's cumulative position in the lateral direction of the entire road. S511: Based on the width information of each lane in the road, the lateral space of the road is divided into multiple lane intervals. Each lane interval corresponds to a lane number. When the lateral cumulative position of a vehicle is located in a certain lane interval, the lane number where the vehicle is currently located is determined to be the lane corresponding to that lane interval, and the lane-level matching result is output.

[0011] Based on the above technical solutions, preferably, step S6, which involves constructing evidence vectors of three probabilistic quality—lane closure state, lane normal traffic state, and observation uncertainty state—based on evidence theory, includes the following sub-steps: S61, Define a set of lane state assumptions, which includes lane closure state, lane normal traffic state, and observation uncertainty state; S62, construct an evidence vector based on the probability mass of lane closure state, lane normal traffic state and observation uncertainty state, wherein the sum of the probability mass of lane closure state, lane normal traffic state and observation uncertainty state is 1; S63, Obtain the perception results of a single vehicle on the target lane. The perception results include traffic cone perception information, lane line perception information, and matching error between the lane line and the high-precision map. The matching error includes lane line start and end point position deviation and lane line slope deviation. S64, dynamically assign values ​​to each component of the evidence vector based on the matching quality of the perception results, and output the evidence vector of a single vehicle to the target lane, including: When a traffic cone is detected, and the deviations in the starting and ending points and the slope between the lane line perception information and the high-precision map are both less than the preset deviation threshold, and the confidence level of the traffic cone is greater than the preset confidence threshold, the perception quality is determined to be high confidence. The probability quality of the lane closure state in the evidence vector is set to be no less than the first preset threshold, and the remaining probability quality is allocated to the observation uncertainty state. When no traffic cone is detected, and the deviations in the starting and ending points and the slope between the lane line perception information and the high-precision map are both less than the preset deviation threshold, and the observation is stable for multiple consecutive frames, the lane is determined to be in a normal traffic state. The probability mass of the normal traffic state of the lane in the evidence vector is set to no less than the second preset threshold, and the remaining probability mass is allocated to the observation uncertainty state. When the deviation of the lane line perception information from the start and end point positions or the slope deviation from the high-precision map is greater than the preset deviation threshold, or the lane line detection confidence is lower than the preset lane line confidence threshold, or there is image occlusion or abnormal lighting interference, the perception quality is judged to be low confidence. The probability quality of the uncertain observation state in the evidence vector is set to no less than the third preset threshold, and the remaining probability quality is allocated to the lane closure state and the lane normal traffic state.

[0012] Based on the above technical solutions, preferably, step S6, which involves accumulating and fusing evidence vectors from multiple vehicles using evidence combination rules and updating the high-precision map after determining the lane dynamic state according to a preset threshold, includes the following sub-steps: S66, obtain the evidence vector corresponding to the observation results of multiple vehicles in the same lane; S67. The Dempster evidence combination rule is used to fuse the evidence vectors of multiple vehicles. The fusion process includes: for any proposition in the lane closure state, lane normal passage state, and observation uncertainty state, the confidence of the corresponding proposition in the evidence vectors of multiple vehicles is multiplied together, and the multiplication results of all intersections equal to the proposition are summed. Then, the normalization is performed by normalization factor, and the confidence assignment value of each state after fusion is calculated. S68. Calculate a normalization factor based on the evidence vectors of multiple vehicles. The normalization factor is obtained by multiplying the confidence scores of all propositions whose intersection is an empty set, summing the results, and then subtracting the sum from 1. The normalization factor is used to measure the degree of conflict between the evidence. When the normalization factor is less than a preset conflict threshold, the evidence is highly conflicting. A weighted average fusion strategy is adopted to arithmetically average the confidence scores of each state in the evidence vectors of multiple vehicles to obtain the fused confidence allocation value. S69, based on the fused confidence assignment value, set a decision threshold and make a lane state decision: When the confidence assignment value for the lane closure status is greater than or equal to the decision threshold, the lane is determined to be in a closed state. When the confidence assignment value of the lane's normal traffic status is greater than or equal to the decision threshold, the lane is determined to be in a normal traffic status. When the confidence assignment value of the observed uncertain state is greater than or equal to the decision threshold, or when the confidence assignment values ​​of both the lane closure state and the lane normal traffic state are less than the decision threshold, the lane state is determined to be uncertain and the map is not updated for the time being. S610, when the lane status is determined to be closed, or changes from closed to normal traffic status, update information is written to the dynamic layer of the high-precision map. The update information includes the closed lane number, the geographical coordinates of the starting point of the closed range, the geographical coordinates of the ending point of the closed range, the status update timestamp, and the confidence score. S611, Set the update frequency of the high-precision map dynamic layer to a preset frequency value, and push the updated map information to vehicles in the area in real time according to the update frequency, so as to provide the latest road condition data for route planning and driving decisions.

[0013] Secondly, the present invention also provides a high-precision map lane-level dynamic information update system based on crowdsourced data, implemented using a high-precision map lane-level dynamic information update method based on crowdsourced data, including: The crowdsourced data acquisition module is used to collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information. The lane line perception module is used to construct a single-stage detection model of the line anchor using a convolutional neural network. Based on the single-stage detection model of the line anchor, the module performs perception and detection of lane lines in road image data and outputs the start and end coordinates and slope of the lane lines in the image coordinate system as lane line perception information. The traffic cone perception module is used to register road image data of consecutive frames during vehicle travel using an image alignment compensation algorithm, calculate motion difference maps of consecutive frames, and perform multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The data is then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information. The perception information projection module is used to project lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system using camera parameters; The lane-level matching and positioning module is used to determine the road segment where the vehicle is located based on the collected vehicle trajectory data, using a hidden Markov model, filtering the candidate lane set, and combining lane width information and lane slope information to calculate the vehicle's lateral absolute position in the current lane and determine the current lane of the vehicle. The multi-vehicle evidence fusion and update module is used to construct evidence vectors of three probabilistic quality based on evidence theory, which include lane closure status, lane normal traffic status and observation uncertainty status, by accumulating and fusing the evidence vectors of multiple vehicles through evidence combination rules, and then updating the high-precision map after deciding the lane dynamic status according to a preset threshold.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing a method program for updating lane-level dynamic information of a high-precision map based on crowdsourced data. When the method program is executed, it implements the method for updating lane-level dynamic information of a high-precision map based on crowdsourced data.

[0015] The high-precision map lane-level dynamic information update method based on crowdsourced data of the present invention has the following advantages over the prior art: By collecting crowdsourced vehicle trajectory and road image data, lane line and traffic cone information are accurately perceived using a line-anchor single-stage detection model and a multimodal fusion method. Lane-level matching is achieved by combining a hidden Markov model and lateral position calculation. Based on evidence theory, multi-vehicle perception information is fused to make decisions on lane dynamic status, which improves the accuracy, reliability and real-time performance of lane-level dynamic information updates, while reducing update costs. This provides accurate and reliable high-precision map data support for autonomous vehicles, effectively improving driving safety and traffic efficiency. By using a multi-level feature fusion and fine regression mechanism in the single-stage detection model of the line anchor, the accuracy and anti-interference ability of lane line detection are significantly improved. At the same time, by using continuous frame motion difference map and multi-modal feature fusion method, the robustness and accuracy of traffic cone detection are effectively enhanced, realizing high precision and high reliability of lane-level dynamic information perception. By combining the Hidden Markov Model with the road topology of high-precision maps and vehicle trajectory data, accurate matching and candidate filtering of the road segment where the vehicle is located were achieved. At the same time, by using lane slope and width information to construct the cumulative lane width range, millimeter-level positioning of the vehicle's lateral position and lane-level attribution determination were achieved, improving the accuracy and robustness of vehicle positioning under complex road conditions. By constructing multi-state evidence vectors using evidence theory, combining them with dynamic allocation of confidence based on perception quality, and employing Dempster combination rules and conflict resolution mechanisms to achieve reliable fusion of multi-vehicle perception information, the system effectively solves the ambiguity and uncertainty problems of single-vehicle perception, improves the accuracy and robustness of lane dynamic state decision-making, and realizes high-confidence updates and real-time push of lane-level dynamic information for high-precision maps. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the high-precision map lane-level dynamic information update method based on crowdsourced data according to the present invention; Figure 2 This is a schematic diagram of lane line perception and lane-level matching in the high-precision map lane-level dynamic information update method based on crowdsourced data of the present invention. Figure 3This is a schematic diagram illustrating the geometric relationship between multi-frame sensing data and high-precision map matching in the high-precision map lane-level dynamic information update method based on crowdsourced data of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1-3 As shown, the present invention provides a method for updating lane-level dynamic information of high-precision maps based on crowdsourced data, comprising the following steps: S1, collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information.

[0020] In this embodiment, step S1 specifically includes the following operations: By selecting public vehicles as the data collection carrier, which are ordinary civilian vehicles including but not limited to passenger cars, commercial vehicles, or motor vehicles with legal road driving qualifications, there is no need to equip them with professional high-precision collection equipment, thereby effectively reducing data collection costs and meeting the needs of large-scale promotion and application.

[0021] A positioning and visual perception device is installed inside the windshield of the Zhongyuan vehicle. This device is an integrated unit and specifically includes: The positioning module uses a low-cost global navigation satellite system receiver to receive satellite signals in real time and calculate the vehicle's geographical location information. This receiver supports multiple satellite navigation systems such as GPS, Beidou, and GLONASS, and can provide meter-level positioning accuracy in open road environments. The positioning module outputs the vehicle's trajectory point information at a fixed frequency. The trajectory point information includes the horizontal and vertical coordinates in the geographic coordinate system and the collection timestamp.

[0022] The visual perception module uses an industrial camera to collect real-time road environment image data during vehicle operation. The industrial camera features high dynamic range, low noise, and stable frame rate, and can adapt to image acquisition needs under various lighting conditions such as daytime, nighttime, and rain. The camera lens faces the road ahead, and the field of view covers the current lane and adjacent lanes, ensuring clear capture of road elements such as lane lines and traffic cones. The camera outputs continuous frames of road image data at a fixed frame rate, with each frame carrying a corresponding timestamp.

[0023] The positioning module and the visual perception module synchronize time through an internal clock to ensure that the trajectory point information and road image data correspond one-to-one in time. The collected data is stored in the built-in memory of the vehicle electronic device in real time, or uploaded to the cloud server through the wireless communication module.

[0024] Secondly, high-precision maps are pre-installed in the vehicle's electronic devices. These high-precision maps are vector maps with centimeter-level accuracy, containing multi-layered, detailed information about the road structure. Specifically, the high-precision map includes lane group information, lane marking information, lane width information, lane centerline information, and road boundary information. Lane group information describes the structure of lane groups composed of multiple lanes on the road, including the topological relationships, connectivity relationships, and the number and arrangement order of lanes within each lane group. Lane marking information describes the type of lane lines (such as solid lines, dashed lines, double solid lines, double dashed lines, etc.) and their color (white, yellow). The information includes: position coordinates and geometry, serving as the benchmark for subsequent lane detection result matching and verification; lane width information, used to record the width value of each lane in the lateral direction of the road, in meters, providing key parameters for calculating the lateral position of vehicles in the fine matching stage; lane centerline information, used to describe the centerline coordinate sequence of each lane along the driving direction, serving as a reference benchmark for calculating the distance from the trajectory point to the road in the coarse matching stage of the Hidden Markov Model; and road boundary information, used to describe the physical boundary positions of the leftmost and rightmost parts of the road, including the coordinate information of boundary elements such as curbs, guardrails, and shoulders, used to constrain the vehicle's driving range.

[0025] The aforementioned high-precision map data is organized in a hierarchical structure, including a static layer and a dynamic layer. The static layer stores the inherent attribute information of the road, such as lane groups, lane markings, lane width, lane center line, and road boundaries. The dynamic layer is used to write dynamic event information such as lane closure and road construction in subsequent steps to realize the real-time updating of the high-precision map.

[0026] During vehicle operation, positioning and visual perception devices continuously collect data at preset frequencies. The positioning module outputs trajectory point information at a frequency of 1Hz to 10Hz, while the visual perception module collects road image data at a frequency of 10Hz to 30Hz. After aligning the collected trajectory point information and road image data according to timestamps, a crowdsourced data stream containing location and visual information is formed, providing basic data support for subsequent lane line detection, traffic cone detection, and lane-level matching.

[0027] S2 uses a convolutional neural network to construct a single-stage detection model for lane anchors. Based on the single-stage detection model for lane anchors, lane lines in road image data are detected and perceived. The start and end coordinates and slope of the lane lines in the image coordinate system are output as lane line perception information.

[0028] Step S2 includes the following sub-steps: S21, a single-stage detection model for line anchors is constructed using a convolutional neural network. The backbone network of the convolutional neural network is used to extract features from the road image. The backbone network of the convolutional neural network includes multiple downsampling convolutional layers connected in sequence, which are used to output high-level feature maps and shallow feature maps respectively. It should be noted that a single-stage detection model for line anchors is constructed using a convolutional neural network. The backbone network of the convolutional neural network uses a ResNet-50, ResNet-101, or similar architecture, and consists of multiple sequentially connected downsampling convolutional layers. The backbone network extracts multi-level feature maps of the road image through layer-by-layer downsampling.

[0029] Specifically, it includes high-level semantic feature maps and shallow feature maps. High-level semantic feature maps are output from the deep layers of the backbone network. Their spatial resolution is 1 / 16 of the input image. They have a large receptive field and rich semantic information, and can perceive the overall direction and context of lane lines, but their spatial positioning accuracy is low. Shallow feature maps are output from the shallow layers of the backbone network. Their spatial resolution is 1 / 4 of the input image. They have a small receptive field and rich spatial detail information, and can accurately depict the local geometric features of lane lines, such as edges and endpoints.

[0030] S22, a set of radiating line anchors emanating from the bottom to the top of the image are predefined on the high-level feature map. Each line anchor has a preset starting point coordinate, direction angle and length. The starting point coordinate is fixed at the bottom edge of the image, and the direction angle is selected within a preset angle range. It should be noted that a set of line anchors is predefined on the high-level semantic feature map. The line anchors adopt a ray-like structure radiating from the bottom to the top of the image to simulate the perspective projection characteristics of lane lines in the image. Each line anchor corresponds to a preset start point coordinate, direction angle, and length. The start point coordinates of each line anchor are fixed at the bottom edge of the image, the horizontal coordinates are evenly distributed within the image width with a preset spacing, and the vertical coordinate is fixed at the image height. The direction angle of each line anchor is selected within a preset angle range, such as between -15° and +15°, with a direction angle set every 5° to form multiple directions of line anchors to cover lane lines with different curvatures and directions. The length of each line anchor is a fixed pixel length to ensure that the line anchors can cover the lane line area from the near field to the far field. In this way, a total of M×N line anchors are generated, where M is the number of start points and N is the number of direction angles, forming a set of line anchors covering the entire image.

[0031] S23, the feature vectors of the areas covered by each anchor are mapped to fixed-length feature vectors through the region of interest alignment operation, and then input to the classification branch and the coarse regression branch respectively. The classification branch outputs the confidence score that the anchor belongs to the lane line, and the coarse regression branch outputs the coordinate offset between the start point of the anchor and the start point of the real lane line and the coordinate offset between the end point of the anchor and the end point of the real lane line. It should be noted that for each anchor, the feature vector of its covered area is mapped to a fixed-length feature vector through a region of interest alignment operation. The region of interest alignment operation is used to solve the problem of misalignment between the anchor region and the feature map grid. Through bilinear interpolation and other methods, the irregular region corresponding to the anchor is sampled into a fixed-size feature vector. The extracted fixed-length feature vectors are then input into the classification branch and the coarse regression branch, respectively. The classification branch is composed of a fully connected network or convolutional layers and outputs a confidence score between 0 and 1, representing the probability that the anchor belongs to the lane line. The coarse regression branch is composed of a fully connected network and outputs a four-dimensional offset, representing the horizontal and vertical offsets between the anchor start point and the actual lane line start point, and the horizontal and vertical offsets between the anchor end point and the actual lane line end point, respectively.

[0032] S24. Select the top K line anchors as candidate line anchors based on the confidence scores from high to low, and perform coarse correction on the start and end coordinates of the candidate line anchors based on the output coordinate offset of the coarse regression branch to obtain the rough start and end coordinates of the lane line on the high-level feature map. It should be noted that, based on the confidence scores output by the classification branch, all line anchors are sorted from highest to lowest score, and the top K line anchors are selected as candidate line anchors. The value of K is set according to the actual scenario to ensure that the candidate line anchors cover all areas that may contain lane lines; For each candidate line anchor, based on the four-dimensional offset output by the coarse regression branch, its preset start and end coordinates are coarsely corrected. The preset start coordinates are added to the offset to obtain the approximate start coordinates; the preset end coordinates are added to the offset to obtain the approximate end coordinates. After correction, the approximate start and end coordinates of the lane line on the high-level feature map are obtained.

[0033] S25, map the coarse start and end point coordinates of the candidate line anchors from the high-level feature map to the shallow feature map to obtain the corresponding mapping region on the shallow feature map, extract the local feature vector of the mapping region, input it into the fine regression branch, and the fine regression branch outputs the fine offset to correct the coarse start and end point coordinates to obtain the fine start and end point coordinates of the lane line in the image coordinate system. It should be noted that mapping the approximate start and end coordinates of candidate line anchors from the high-level feature map to the shallow feature map requires coordinate scaling due to the different spatial resolutions of the high-level and shallow feature maps. For example, if the resolution of the high-level feature map is 1 / 16 of the input image and the resolution of the shallow feature map is 1 / 4 of the input image, then the coordinates on the high-level feature map can be mapped to the shallow feature map by multiplying by 4.

[0034] On the shallow feature map, local feature vectors are extracted from the region corresponding to the candidate line anchor. This region is determined based on the rough start and end point coordinates and covers the possible range of lane lines. The extracted local feature vectors are then input into the fine regression branch.

[0035] The fine regression branch is also constructed using a fully connected network, outputting fine offsets, including fine offsets of the start coordinates and fine offsets of the end coordinates. The coarse start and end coordinates are added to the fine offsets to obtain the fine start and end coordinates of the lane line in the image coordinate system.

[0036] S26, calculate the lane slope based on the fine start and end point coordinates, and output the start coordinates, end coordinates and slope of the left lane line, as well as the start coordinates, end coordinates and slope of the right lane line.

[0037] It should be noted that for the left lane line, its starting coordinates are set as ( x l1 , y l1 The endpoint coordinates are ( ). x l2 , y l2 For the right lane line, let its starting coordinates be ( ); x r1 , y r1 The endpoint coordinates are ( ). x r2 , y r2 ); Slope of the left lane line k l and the slope of the right lane line k r The calculation expression is: ; Finally, the starting coordinates, ending coordinates, and slope of the left lane line, as well as the starting coordinates, ending coordinates, and slope of the right lane line, are output as lane line perception information for projection and matching in subsequent steps.

[0038] This embodiment employs a single-stage detection model based on line anchors. Taking into account the geometric characteristics of lane lines as linear continuous instances, ray-shaped line anchors radiating from the bottom to the top of the image are predefined on the high-level feature map. A coarse-fine two-stage regression strategy is used to achieve high-precision lane line detection. Specifically, the coarse stage uses the high-level semantic feature map to perform preliminary classification and regression of the line anchors, quickly filtering candidate regions and obtaining coarse geometric parameters. The fine stage maps the candidate regions to a high-resolution shallow feature map for secondary fine regression, accurately correcting the start and end point coordinates and slope, thus effectively balancing global semantic understanding and local detail depiction. A region of interest alignment operation maps the line anchor regions to fixed-length feature vectors, ensuring the accuracy of feature extraction. Simultaneously, the collaborative processing of the classification branch and the coarse regression branch, combined with confidence scores to filter the top K candidate line anchors, significantly reduces computational complexity. Finally, the start and end point coordinates and slopes of the lane lines on both sides are output, providing accurate geometric parameters for subsequent projection and lane-level matching, improving the robustness and positioning accuracy of lane line detection in complex scenarios.

[0039] S3 employs an image alignment compensation algorithm to register road image data of consecutive frames during vehicle travel, calculates motion difference maps of consecutive frames, and performs multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The results are then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information.

[0040] Step S3 includes the following sub-steps: S31, acquire road image data of consecutive frames during vehicle travel, and register adjacent frame images using an image alignment compensation algorithm to obtain the aligned and compensated consecutive frame images; It should be noted that the acquisition of road image data in consecutive frames during vehicle movement includes the first consecutive frame. t Frame image and the first t -1 frame image; Due to the vehicle's self-motion during travel, the background in consecutive frames is displaced, directly affecting the detection accuracy of moving targets. To eliminate this effect, an image alignment compensation algorithm is used to register adjacent frames.

[0041] Specifically, the image alignment compensation algorithm adopts an image alignment method based on feature point matching; firstly, it extracts the first... t -1 frame image and the t Feature points in the frame image are identified, including corner points, edge points, or scale-invariant feature transform points. Then, a feature point matching algorithm is used to establish the correspondence between two frames, calculating the homography matrix or affine transformation matrix. Finally, the transformation matrix is ​​used to analyze the first frame image. t Perform a geometric transformation on the frame image to make it consistent with the first frame. t-1 frames of images are aligned in spatial position to compensate for background shift caused by camera motion.

[0042] Through the above image alignment compensation operation, we obtain consecutive frame images after alignment compensation. In the consecutive frame images after alignment compensation, the pixel positions of the static background area remain consistent, while the dynamic moving targets show significant differences.

[0043] S32, calculate the pixel-by-pixel motion difference based on the aligned and compensated consecutive frame images, and construct the motion difference map of the consecutive frames; It should be noted that, based on the aligned and compensated consecutive frame images, pixel-by-pixel motion differences are calculated to construct motion difference maps for the consecutive frames; specifically, for the ... t Original RGB image after frame alignment compensation I t and the t -1 frame alignment compensation of the original RGB image I t-1 The absolute value of the difference between the two is calculated pixel by pixel to obtain the motion difference map. M t The expression is: ; In the formula, M t For the first t Motion difference map of frames, I t For the first t The original RGB image after frame alignment compensation. I t-1 For the first t -1 frame is the original RGB image after alignment compensation, and |.| is the absolute value operation for each pixel.

[0044] Motion difference maps are used to highlight independent moving targets that are not in the background, while suppressing static background and pseudo-motion noise. In motion difference maps, the pixel values ​​of static background areas approach zero, while the pixel values ​​of moving target areas show significant positive or negative values, effectively enhancing the contrast between dynamic targets such as traffic cones and the background.

[0045] S33, the original road image features of the corresponding frame are stitched and fused with the motion difference map in the channel dimension to obtain a multimodal feature map.

[0046] It should be noted that the fusion is performed by concatenating the channels. First, a feature map of the original road image is extracted using a convolutional neural network. This feature map contains rich spatial appearance information, such as color, texture, and edges. At the same time, the motion difference map is used as an independent channel input, which contains temporal motion information and can reflect the dynamic characteristics of the target. The original image feature map and the motion difference map are concatenated along the channel dimension to form a multimodal feature map.

[0047] Multimodal feature maps contain both spatial appearance information and temporal motion information, giving them a stronger ability to discriminate targets. Compared with simply relying on the original image features, fusing motion difference maps can effectively reduce the false detection rate of static interference objects and enhance the robustness of detecting sparse small target traffic cones.

[0048] S34. Input the multimodal feature map into the target detection model for recognition, output the coordinates of the traffic cone center point and the confidence score in the image coordinate system, and filter the traffic cone detection results with a confidence score greater than the confidence threshold as traffic cone perception information.

[0049] It should be noted that the object detection model uses a single-stage object detection network, such as YOLO, SSD, or RetinaNet. This network takes multimodal feature maps as input and outputs the class probability and bounding box coordinates of each detected object through convolutional layers, classification layers, and regression layers.

[0050] For traffic cone detection, the target detection model outputs the coordinates of the center point of each detection box. x c , y c And the confidence score of the detection box belonging to the traffic cone. conf The confidence score ranges from 0 to 1, with a higher value indicating a more reliable test result.

[0051] The detection results are filtered based on a preset confidence threshold, which is set according to the actual application scenario. Only detection results with a confidence score greater than or equal to the confidence threshold are retained as valid traffic cone perception information; detection results with a confidence score lower than the threshold are considered false detections and are discarded.

[0052] Output the center point coordinates and confidence scores of the filtered traffic cones as traffic cone perception information for subsequent projection and fusion steps.

[0053] This embodiment addresses the issue of traffic cones, as sparse small targets, being easily confused with noise in complex motion backgrounds. It introduces temporal motion features as a supplementary modality and employs image alignment technology to register consecutive frames, effectively compensating for camera self-motion caused by vehicle movement and eliminating visual shifts in the background due to vehicle movement. This ensures that the motion difference only reflects real, independent moving targets in the scene. A motion difference map is constructed by calculating pixel-by-pixel motion differences through aligned and compensated consecutive frame images. This map highlights independent moving targets outside the background and suppresses static backgrounds and pseudo-motion noise. Furthermore, the original RGB image features are fused with the motion difference map using a multimodal approach, constructing a fused feature that simultaneously includes spatial appearance information and temporal motion information, significantly enhancing target discrimination capabilities. Finally, the fused feature is input into the target detector, outputting the center point coordinates and confidence score of the traffic cone in the image coordinate system. Valid detection results are then filtered using a confidence threshold. This method effectively solves the problem of missed or false detections of sparse small targets in complex motion backgrounds using traditional single-modal detection, improving the accuracy and robustness of traffic cone detection and providing reliable data support for accurate perception of lane-level dynamic information.

[0054] S4 projects lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system through camera parameters; Step S4 includes the following sub-steps: S41, obtain the camera's intrinsic parameter matrix, and use the intrinsic parameter matrix to convert the lane line perception information and traffic cone perception information in the image coordinate system into three-dimensional coordinates in the camera coordinate system; It should be noted that the camera's intrinsic parameter matrix is ​​obtained. This intrinsic parameter matrix consists of inherent parameters of the camera itself, obtained through camera calibration. Its mathematical form is a 3×3 matrix, containing parameters such as the camera's focal length and principal point coordinates. Specifically, the intrinsic parameter matrix describes the mapping relationship between the image coordinate system and the camera coordinate system.

[0055] For lane line perception information, the output format is the start and end point coordinates in the image coordinate system, including the start and end coordinates of the left lane line and the start and end coordinates of the right lane line. For traffic cone perception information, the output format is the center point coordinates in the image coordinate system. Using the camera intrinsic parameter matrix, the two-dimensional coordinates in the image coordinate system are converted into three-dimensional coordinates in the camera coordinate system. The camera coordinate system has the camera optical center as the origin, the optical axis as the Z-axis, and the X-axis and Y-axis are parallel to the u-axis and v-axis of the image coordinate system, respectively.

[0056] During the transformation process, for any point in the image coordinate system ( u , v First, it is converted to homogeneous coordinate form, and then transformed using the inverse of the intrinsic parameter matrix to obtain the three-dimensional coordinates in the camera coordinate system. Xc , Y c , Z c );in, Zc The depth information representing the target point from the camera's optical center can be obtained through monocular vision depth estimation methods or approximated based on the road plane assumption. S42, obtain the camera's extrinsic parameter matrix, which includes a rotation matrix and a translation vector. Use the extrinsic parameter matrix to convert the three-dimensional coordinates in the camera coordinate system into three-dimensional coordinates in the vehicle coordinate system. It should be noted that the extrinsic parameter matrix of the camera is obtained. The extrinsic parameter matrix describes the camera's mounting position and attitude relative to the vehicle, including the rotation matrix and translation vector. The extrinsic parameter matrix is ​​obtained in advance through the rigid connection between the camera and the vehicle. The rotation matrix is ​​a 3×3 orthogonal matrix used to describe the rotation relationship between the camera coordinate system and the vehicle coordinate system. The translation vector is a three-dimensional vector used to describe the translation relationship between the camera's optical center and the origin of the vehicle coordinate system.

[0057] The 3D coordinates in the camera coordinate system are converted to 3D coordinates in the vehicle coordinate system using an extrinsic parameter matrix; the vehicle coordinate system has its origin at the center of the rear axle of the vehicle. X The axis points forward of the vehicle. Y The axis points to the left side of the vehicle. Z The axis points upwards from the vehicle; during the transformation, the coordinates in the camera coordinate system are multiplied by the rotation matrix and then added to the translation vector to obtain the three-dimensional coordinates in the vehicle coordinate system.

[0058] This transformation eliminates the influence of camera mounting position and orientation on the perception results, unifying the perceived information into the vehicle's own reference frame.

[0059] S43, acquire the vehicle's positioning information, which includes the global navigation satellite system position and the vehicle's attitude angle. Use the positioning information to transform the three-dimensional coordinates in the vehicle coordinate system to the geographic coordinate system of the high-precision map, and output the projection results of lane lines and traffic cones in the geographic coordinate system of the high-precision map.

[0060] It should be noted that the vehicle's positioning information is obtained, which includes the Global Navigation Satellite System (GNSS) position and the vehicle's attitude angles. The GNSS position is obtained in real time through the onboard GNSS receiver, including the horizontal and vertical coordinates in the geographic coordinate system. The vehicle attitude angles are obtained through the inertial measurement unit or integrated navigation system, including the heading angle, pitch angle, and roll angle, which are used to describe the vehicle's orientation and attitude during driving.

[0061] Using positioning information, the three-dimensional coordinates in the vehicle coordinate system are transformed to the geographic coordinate system of a high-precision map. The geographic coordinate system uses the universal transverse Mercator projection or the Gauss-Kruger projection, with the Earth's surface as the reference plane, and the unit is meters. During the transformation, firstly, based on the vehicle's heading angle, the coordinates in the vehicle coordinate system are rotated to align with the north direction of the geographic coordinate system; then, based on the vehicle's global navigation satellite system position, the rotated coordinates are translated to their absolute position in the geographic coordinate system; finally, the three-dimensional projected coordinates of lane lines and traffic cones in the high-precision map's geographic coordinate system are obtained.

[0062] The output shows the projection results of lane lines and traffic cones in the geographic coordinate system of a high-precision map. These projection results include the x-coordinate and y-coordinate of the geographic coordinate system, as well as optional height information. This projection result provides perception information within a unified coordinate framework for subsequent lane-level map matching, ensuring accurate alignment between the perception data and the high-precision map.

[0063] In this embodiment, a complete coordinate system cascade transformation is constructed using the camera intrinsic and extrinsic parameter matrices and vehicle positioning information. The coordinates of the start and end points of lane lines and the center point of traffic cones in the image coordinate system are sequentially converted into the camera coordinate system and the vehicle coordinate system, and finally unified into the geographic coordinate system of the high-precision map. This process makes full use of the camera calibration parameters to accurately describe the mapping relationship between image pixels and three-dimensional space. The influence of camera installation position and attitude on the perception results is corrected by rotation matrix and translation vector. The local perception data is accurately aligned to the absolute geographic coordinates by combining the position of the global navigation satellite system and the vehicle attitude angle. The entire projection process only involves lightweight operations such as matrix multiplication and coordinate translation, which has high computational efficiency and can meet the real-time processing requirements of the vehicle terminal. By unifying the multi-source perception information into the coordinate framework of the high-precision map, a spatially consistent data foundation is provided for subsequent lane-level map matching, ensuring the accurate alignment of perception results with map prior information and effectively supporting the accurate fusion and updating of lane-level dynamic information.

[0064] S5, based on the collected vehicle trajectory data, uses a hidden Markov model to determine the road segment where the vehicle is located, filters the candidate lane set, and combines lane width information and lane slope information to calculate the vehicle's lateral absolute position in the current lane, thus determining the current lane of the vehicle.

[0065] In this embodiment, step S5 adopts a two-stage strategy of coarse matching and fine matching, which performs lane-level matching between the perceived information and the high-precision map. The coarse matching stage determines the road segment where the vehicle is located and filters the candidate lane set based on the hidden Markov model, while the fine matching stage combines lane width information and lane slope information to determine the lane where the vehicle is currently located.

[0066] Step S5, which involves using a hidden Markov model to determine the road segment where the vehicle is located based on the collected vehicle trajectory data and filtering the candidate lane set, includes the following sub-steps: S51, the road segments adjacent to the spatial location of the vehicle trajectory in the high-precision map are taken as the set of hidden states of the hidden Markov model, and the trajectory points of the trajectory data collected during the vehicle's driving process are taken as the observation state sequence. It should be noted that road segments adjacent to the spatial location of vehicle trajectories in the high-precision map are used as the set of hidden states in the Hidden Markov Model. Specifically, based on the spatial distribution of vehicle trajectory points, all road segments within a certain distance from the trajectory points are retrieved from the high-precision map to form the set of hidden states. S ={ s 1, s 2,…, s M}, where each hidden state s j This represents a candidate road segment. M This indicates the number of candidate road segments.

[0067] The trajectory points collected during the vehicle's movement are used as the observation state sequence. Assume the trajectory points collected by the vehicle are arranged in chronological order to form the observation state sequence. O ={ o 1, o 2,…, o T}, where each observation state o t Includes the x-coordinate and y-coordinate in a geographic coordinate system. T This represents the total number of trajectory points.

[0068] S52, for any trajectory point and candidate road segment, calculate the minimum perpendicular distance from the trajectory point to the centerline of the candidate road segment, and calculate the probability of observing each trajectory point under the condition that the vehicle is located in the candidate road segment based on the Gaussian distribution, and obtain the observation probability; It should be noted that for any trajectory point o t With candidate road segments s j The minimum vertical distance from the trajectory point to the centerline of the candidate road segment is calculated. This distance reflects the geometric proximity between the trajectory point and the road segment. The smaller the distance, the more likely the trajectory point is to be located on the road segment.

[0069] Based on Gaussian distribution calculations, the vehicle is located in the candidate road segment. s j Trajectory points observed under the conditions o tThe probability of observation is calculated using a Gaussian distribution, with the vertical distance from the trajectory point to the centerline of the road segment as the input variable. When the vertical distance is small, the observation probability is large; as the vertical distance increases, the observation probability decreases exponentially. The standard deviation parameter in the Gaussian distribution is set according to the actual statistical characteristics of the vehicle positioning error and is used to adjust the degree of influence of distance error on the observation probability.

[0070] The expression for the probability of observation is: ; In the formula, P( o t | s j (For vehicles located on road sections) s j Under these conditions, the trajectory points were observed. o t The probability, d t,j For the first t The trajectory point to the first j The perpendicular distance between the centerlines of the candidate road segments, where σ is the standard deviation of the vehicle positioning error, is used to adjust the degree of influence of distance error on the observation probability. S53, calculate the observed displacement distance between two adjacent trajectory points, and calculate the shortest path distance between candidate road segments based on the candidate road segments corresponding to the two adjacent trajectory points. Calculate the state transition probability between the two candidate road segments based on the difference between the observed displacement distance between trajectory points and the shortest path distance between candidate road segments. It should be noted that, based on two adjacent trajectory points o t-1 and o t The observed displacement distance between trajectory points is calculated, which is the Euclidean distance between two trajectory points, reflecting the actual distance traveled by the vehicle in continuous time.

[0071] The expression for the observed displacement distance is: ; In the formula, ( x t , y t )and( x t-1 , y t-1 ) are the position coordinates of two adjacent trajectory points.

[0072] Based on the candidate road segments corresponding to two adjacent trajectory points s i and sj The shortest path distance between candidate road segments is calculated. This distance is obtained from the road topology of a high-precision map, reflecting the path taken by vehicles from road segments. s i Driving to the section of road s j The shortest path length required.

[0073] The state transition probability between two candidate road segments is calculated based on the difference between the observed displacement distance between trajectory points and the shortest path distance between candidate road segments. When the difference between the observed displacement distance and the shortest path distance is small, it indicates that the vehicle's actual driving trajectory is highly consistent with the map topology, and the state transition probability is relatively high; when the difference is large, the state transition probability is relatively low. This calculation uses an exponential function, with a smoothing coefficient controlling the degree of penalty imposed on the transition probability by the difference.

[0074] The expression for the state transition probability is: ; In the formula, P( s j | s i (From road section) s i Transfer to road section s j The state transition probability, In the high-precision map topology, from road segments s i to road section s j The shortest path distance; β This is a smoothing coefficient for distance differences, used to control the degree of penalty for transition probability when the trajectory distance is inconsistent with the map distance.

[0075] S54. Based on the spatial distance between the starting trajectory point of the trajectory sequence and each candidate road segment, calculate the probability that the vehicle's initial state is located in each candidate road segment, and obtain the initial state probability. It should be noted that, based on the starting trajectory point of the trajectory sequence o 1. Calculate the probability that the vehicle's initial state is located in each candidate road segment based on the spatial distance between the vehicle and each candidate road segment. The initial state probability is inversely proportional to the vertical distance from the starting trajectory point to the centerline of the candidate road segment; the smaller the distance, the higher the initial probability. To ensure that the sum of all initial state probabilities is 1, a normalization coefficient is used for normalization.

[0076] The expression for the initial state probability is: ; In the formula, P( sj The initial state of the vehicle is located on the road segment. s j The probability, d 1,j From the first trajectory point to the... j The distance between the centerlines of the candidate road segments; Z These are normalization coefficients used to ensure that the sum of the probabilities of all initial states is 1. σ This represents the standard deviation of the vehicle positioning error.

[0077] S55. Based on the observation probability, state transition probability and initial state probability, the Viterbi algorithm is used to solve the road segment sequence corresponding to the maximum posterior probability to obtain the optimal matching road segment sequence. It should be noted that, based on the observation probability, state transition probability, and initial state probability, the Viterbi algorithm is used to solve for the road segment sequence corresponding to the maximum posterior probability. The Viterbi algorithm is a dynamic programming algorithm that recursively calculates the maximum probability path for each hidden state corresponding to each trajectory point, and finally obtains the globally optimal road segment sequence. This optimal sequence reflects the road segments actually traveled by the vehicle, achieving accurate matching between the trajectory and the road.

[0078] The expression is: ; In the formula, This is the optimal matching sequence of road segments. S Represents any possible sequence of road segment states. s t For the first t The road segment status corresponding to each trajectory point T This represents the total number of trajectory points.

[0079] S56. Based on the frequency of occurrence of each road segment in the optimal matching road segment sequence, calculate the posterior probability of each road segment, filter road segments with posterior probabilities higher than a preset probability threshold, and construct a candidate road segment set.

[0080] It should be noted that the posterior probability of each road segment is calculated based on the frequency of occurrence of each road segment in the optimally matched road segment sequence. The posterior probability integrates observation information and transfer information, reflecting the likelihood of a vehicle passing through each road segment given the entire trajectory observation sequence. Road segments with posterior probabilities higher than a preset probability threshold are selected to construct a candidate road segment set. This candidate set narrows the search scope in the subsequent fine matching stage, reducing the matching task from the entire road network to a few candidate road segments, thus reducing computational complexity. The preset probability threshold is set according to actual application requirements.

[0081] The expression for constructing the candidate road segment set is: ; In the formula, S cand For the set of candidate road segments, P ( s j | O ) is the observation sequence of a given trajectory O Under these conditions, road section s j The posterior probability, τ The threshold for selecting candidate road segments.

[0082] In this embodiment, a Hidden Markov Model (HMM) is used to achieve coarse matching between trajectories and road segments. Road segments adjacent to the spatial location of the vehicle trajectory in the high-precision map are taken as the set of hidden states, and trajectory points are taken as the sequence of observed states. The minimum vertical distance from the trajectory point to the centerline of the road segment is calculated and the observation probability is obtained based on the Gaussian distribution. The state transition probability is calculated by combining the difference between the observed displacement distance between adjacent trajectory points and the shortest path distance between candidate road segments. The Viterbi algorithm is used to solve for the optimal road segment sequence corresponding to the maximum posterior probability. Finally, the candidate road segment set is selected based on the posterior probability of each road segment. This coarse matching strategy makes full use of the road topology information of the high-precision map and the spatial geometric features of the vehicle trajectory. While effectively dealing with GPS positioning errors and trajectory noise, it quickly reduces the matching problem of the entire road network to a few candidate road segments, reducing the computational complexity of the subsequent fine matching stage. It provides an efficient and robust search space for lane-level positioning and ensures the real-time performance and feasibility of the entire two-stage matching strategy.

[0083] Step S5, which combines lane width information and lane slope information to calculate the vehicle's lateral absolute position within the current lane and determine the vehicle's current lane, includes the following sub-steps: S57, obtain the slope of the left lane line and the slope of the right lane line, and calculate the ratio of the distance from the vehicle to the left lane line to the total lane width based on the difference between the slopes of the left lane line and the right lane line and the slope of the right lane line. It should be noted that when obtaining the slopes of the left and right lane lines, in the image coordinate system, the slope of the lane line reflects the degree of inclination of the lane line in the image. Due to the perspective projection effect, there is a definite geometric relationship between the lateral position of the vehicle in the lane and the slopes of the left and right lane lines.

[0084] Based on the difference between the slopes of the left and right lane lines, and the slope of the right lane line, calculate the ratio of the distance from the vehicle to the left lane line to the total lane width. This ratio is a value between 0 and 1. When the vehicle is in the center of the lane, the ratio is approximately 0.5; when the vehicle is close to the left lane line, the ratio approaches 0; and when the vehicle is close to the right lane line, the ratio approaches 1.

[0085] The specific formula is as follows: ; In the formula, C i It is the ratio of the distance from the vehicle to the left lane line to the total lane width. k l The slope of the lane line on the left side of the vehicle. k r This represents the slope of the lane line on the right side of the vehicle.

[0086] S58: Obtain the width information of the current lane from the high-precision map, multiply the current lane width by a ratio, and obtain the lateral absolute position of the vehicle relative to the left boundary of the current lane. It should be noted that the width information of the current lane is obtained from a high-precision map. This lane width information is pre-stored in the high-precision map, in meters, recording the actual physical width of each lane. The current lane width is multiplied by the ratio calculated in step S57 to obtain the vehicle's lateral absolute position relative to the left boundary of the current lane. This lateral absolute position, in meters, represents the vertical distance from the vehicle's center point to the left boundary line of the lane. This calculation converts the proportional relationship obtained from the image into an absolute distance in actual physical space.

[0087] The specific expression is: ; In the formula, L w The current lane width recorded in the high-precision map. d Oi This represents the vehicle's absolute lateral position relative to the left edge of the current lane.

[0088] S59: Obtain all lane information of the road where the vehicle is located in the horizontal direction from the high-precision map, take the leftmost boundary of the road as the starting position in the horizontal direction, arrange them in the horizontal order of the road, and accumulate them according to the width of each lane to construct the cumulative width range of each lane in the horizontal direction of the road. It should be noted that the information is obtained from a high-precision map, which contains all lane information of the road in the lateral direction. The road contains multiple lanes, and the high-precision map records the width of each lane and the order in which the lanes are arranged on the road cross-section. Taking the leftmost boundary of the road as the starting position in the lateral direction, the lanes are arranged from left to right according to the lateral order of the road. The width of each lane is accumulated sequentially to construct the cumulative width interval of each lane in the lateral direction of the road. For example, the interval corresponding to the first lane is from 0 to the width of the first lane, the interval corresponding to the second lane is from the width of the first lane to the sum of the widths of the first and second lanes, and so on. These cumulative width intervals divide the entire lateral space of the road into continuous segmented intervals, and each interval uniquely corresponds to one lane.

[0089] S510: Based on the vehicle's lateral absolute position relative to the left boundary of the current lane and the sequential number of the current lane in the road, the widths of all lanes before the current lane are accumulated, and then added to the vehicle's lateral absolute position in the current lane to obtain the vehicle's cumulative position in the lateral direction of the entire road. It should be noted that the cumulative position of the vehicle in the entire lateral direction is calculated based on the vehicle's absolute lateral position relative to the left boundary of the current lane and the sequential number of the current lane in the road. Specifically, the widths of all lanes preceding the current lane are summed to obtain the position of the left boundary of the current lane on the entire road cross-section. This summation result is then added to the vehicle's absolute lateral position within the current lane to obtain the cumulative absolute position of the vehicle's center point in the entire lateral direction of the road. This cumulative position reflects the vehicle's absolute coordinates on the road cross-section, with the leftmost boundary of the road as the origin.

[0090] The specific expression is: ; In the formula, X This represents the cumulative absolute position of the vehicle in the lateral direction of the road. w m Indicates the first m The width of each lane i This is the sequential number of the current lane in the road.

[0091] S511: Based on the width information of each lane in the road, the lateral space of the road is divided into multiple lane intervals. Each lane interval corresponds to a lane number. When the lateral cumulative position of a vehicle is located in a certain lane interval, the lane number where the vehicle is currently located is determined to be the lane corresponding to that lane interval, and the lane-level matching result is output.

[0092] It should be noted that, based on the width information of each lane in the road, the lateral space of the road is divided into multiple lane intervals, each lane interval corresponding to a lane number. These lane intervals are consistent with the constructed cumulative width intervals; No. j The lateral section corresponding to each lane is defined as follows: ; When the vehicle's lateral cumulative position X satisfy: X∈Ω j Then the lane number where the vehicle is currently located is determined to be the first lane. j One lane.

[0093] The calculated lateral cumulative position of the vehicle is compared with each lane interval. When the lateral cumulative position of the vehicle is located in a certain lane interval, it is determined that the lane number where the vehicle is currently located is the lane corresponding to that lane interval. The lane-level matching result is output, which includes the specific lane number where the vehicle is currently located, as well as the optional lateral position information of the vehicle in that lane. This matching result provides accurate lane attribution information for subsequent multi-vehicle evidence fusion, ensuring that the perception data of each vehicle can be aligned and fused according to the correct lane.

[0094] In this embodiment, based on the candidate road segments selected by coarse matching, the ratio of the distance from the vehicle to the left lane line to the total lane width is calculated using the geometric relationship between the slopes of the left and right lane lines. This ratio is then converted into the vehicle's lateral absolute position relative to the left boundary of the lane by combining the lane width information provided by the high-precision map. Furthermore, a cumulative width interval is constructed based on the total lane width information. By adding the widths of all lanes before the current lane to the lateral absolute position within the current lane, the cumulative position of the vehicle in the lateral direction of the entire road is obtained. Finally, the specific lane number of the vehicle is determined by the lane interval into which the cumulative position falls. This fine matching strategy fully utilizes the slope parameters in the lane line perception information and the prior lane width information from the high-precision map. It achieves centimeter-level lateral positioning within the lane using only simple arithmetic operations and interval judgments, without the need for complex iterative optimization or geometric calculations, resulting in minimal computational load. This makes it suitable for crowdsourced data processing scenarios involving ordinary vehicles with low computing power. Simultaneously, it forms a complete coarse-fine two-stage matching link with the Hidden Markov Model in the coarse matching stage, ensuring both the accuracy of lane-level positioning and effectively controlling computational complexity, providing accurate lane attribution information for subsequent multi-vehicle evidence fusion.

[0095] S6, based on evidence theory, constructs the perception information of multiple vehicles on the same lane into three types of evidence vectors with probabilistic quality, including lane closure state, lane normal traffic state, and observation uncertainty state. The evidence vectors of multiple vehicles are accumulated and fused through evidence combination rules, and the high-precision map is updated after deciding the lane dynamic state according to a preset threshold.

[0096] Step S6, based on evidence theory, constructs three types of probability quality evidence vectors from the perception information of multiple vehicles on the same lane, including lane closure state, lane normal traffic state, and observation uncertainty state. This includes the following sub-steps: S61, Define a set of lane state assumptions, which includes lane closure state, lane normal traffic state, and observation uncertainty state; It should be noted that a lane state hypothesis set Θ={} is defined to describe the various states that a lane may be in; the lane state hypothesis set includes three mutually exclusive hypotheses: θ c The lane is closed, indicating that the lane is temporarily closed due to construction, an accident, or other reasons, and vehicles are prohibited from passing through it. θ n The lane is in normal traffic condition, meaning that the lane is unobstructed and vehicles can travel normally. θ u The observation is in an uncertain state, indicating that the current observation information is insufficient to determine whether the lane is closed or normal, and there may be situations such as perceived noise, obstruction, or insufficient data.

[0097] S62, construct an evidence vector based on the probability mass of lane closure state, lane normal traffic state and observation uncertainty state, wherein the sum of the probability mass of lane closure state, lane normal traffic state and observation uncertainty state is 1; It should be noted that the expression for the evidence vector is: ; in, Each component represents the first... i The vehicle observes the probability quality of lane closure, normal, and uncertain.

[0098] In this embodiment, the construction of the evidence vector quantifies the uncertainty of the perception results of a single vehicle. When the reliability of the perception results is high, the probabilistic quality of the deterministic state (closed or normal) dominates; when the reliability of the perception results is low, the probabilistic quality of the uncertain observation state dominates.

[0099] S63, Obtain the perception results of a single vehicle on the target lane. The perception results include traffic cone perception information, lane line perception information, and matching error between the lane line and the high-precision map. The matching error includes lane line start and end point position deviation and lane line slope deviation. It should be noted that traffic cone perception information includes whether traffic cones were detected, the number of traffic cones detected, and the confidence score of each traffic cone, reflecting the presence of temporary control facilities in the lane; lane line perception information includes the output coordinates of the start and end points of the lane line, the slope, and the lane line detection confidence score, reflecting the geometric shape of the lane line; the matching error between the lane line and the high-precision map includes the deviation of the lane line start and end point positions and the deviation of the lane line slope. The deviation of the start and end point positions is the Euclidean distance between the coordinates of the detected lane line start and end points and the corresponding coordinates of the lane line start and end points in the high-precision map; the slope deviation is the absolute value of the difference between the slope of the detected lane line and the slope of the corresponding lane line in the high-precision map; the matching error comprehensively reflects the degree of consistency between the perception result and the prior information of the map.

[0100] S64, dynamically assign values ​​to each component of the evidence vector based on the matching quality of the perception results, and output the evidence vector of a single vehicle to the target lane, including: When a traffic cone is detected, and the deviations in the start and end points and the slope between the lane line perception information and the high-precision map are both less than a preset deviation threshold, and the confidence level of the traffic cone is greater than a preset confidence threshold, the perception quality is determined to be high confidence. The probability quality of the lane closure state in the evidence vector is set to no less than a first preset threshold (e.g., 0.6), and the remaining probability quality is allocated to the observation uncertainty state. This assignment strategy reflects the strong support of highly reliable observations with good matching of traffic cones and lane lines for the lane closure state.

[0101] When no traffic cones are detected, and the deviations in the start and end points and slopes between the lane line perception information and the high-precision map are both less than the preset deviation thresholds, and the observations remain stable for multiple consecutive frames (e.g., 5 to 10 frames), the lane is determined to be in a normal traffic state. The probability mass of the normal traffic state in the evidence vector is set to no less than the second preset threshold (e.g., 0.6), and the remaining probability mass is allocated to the observation uncertainty state. This assignment strategy reflects the strong support of highly reliable observations with no traffic cones, good lane line matching, and stable observations for the normal traffic state of the lane.

[0102] When the deviation or slope between the lane line perception information and the start and end point positions of the high-precision map is greater than a preset deviation threshold, or the lane line detection confidence is lower than a preset lane line confidence threshold, or there is image occlusion or abnormal lighting interference, the perception quality is judged as low confidence. The probability quality of the uncertain observation state in the evidence vector is set to no less than a third preset threshold (e.g., 0.4), and the remaining probability quality is allocated to the lane closure state and the lane normal traffic state. This assignment strategy reflects the conservative handling of low-quality observations and avoids misjudgment due to perception errors.

[0103] Output the evidence vector of a single vehicle to the target lane. The evidence vector includes the probability mass of the lane closure state, the probability mass of the lane normal traffic state, and the probability mass of the observation uncertainty state, which are used for subsequent multi-vehicle evidence fusion.

[0104] This embodiment constructs an uncertainty quantification mechanism for single-vehicle perception results based on evidence theory. First, it defines three mutually exclusive hypotheses: lane closure, normal lane passage, and observation uncertainty, constructing an evidence vector containing these three probability qualities, summing to 1. Then, it dynamically evaluates the perception quality based on traffic cone detection results, lane line perception information, and the matching error between lane lines and the high-precision map. When a traffic cone is detected, the matching error is less than a threshold, and the traffic cone confidence is high, the probability quality of the lane closure state is set to no less than a first preset threshold. When no traffic cone is detected, the matching error is less than a threshold, and observations are stable across multiple frames, the probability quality of the normal lane passage state is set to no less than a second preset threshold. When the matching error is large, the lane line detection confidence is low, or there is occlusion interference, the probability quality of the observation uncertainty state is set to no less than a third preset threshold. This dynamic assignment strategy fully reflects the impact of single-vehicle perception quality on evidence credibility, transforming traditional hard decisions into soft decisions containing uncertainty information. This provides a reliable and adaptive input for subsequent multi-vehicle evidence fusion, effectively solving the problem of single-vehicle perception being susceptible to environmental interference and insufficient reliability due to missed or false detections.

[0105] Step S6, which involves accumulating and fusing evidence vectors from multiple vehicles using evidence combination rules and updating the high-precision map after determining the lane dynamic state based on a preset threshold, includes the following sub-steps: S66, obtain the evidence vector corresponding to the observation results of multiple vehicles in the same lane; It should be noted that evidence vectors are obtained from the independent observations of multiple vehicles on the same lane. These evidence vectors come from observations by different vehicles, at different times, and from different perspectives. Each evidence vector contains three types of probability mass: lane closure, lane normal traffic, and observation uncertainty. The sum of the probability mass of each vehicle's evidence vector is 1. The acquisition of evidence vectors from multiple vehicles realizes the aggregation of crowdsourced data, providing multi-source and multi-perspective observation information for subsequent fusion.

[0106] S67. The Dempster evidence combination rule is used to fuse the evidence vectors of multiple vehicles. The fusion process includes: for any proposition in the lane closure state, lane normal passage state, and observation uncertainty state, the confidence of the corresponding proposition in the evidence vectors of multiple vehicles is multiplied together, and the multiplication results of all intersections equal to the proposition are summed. Then, the normalization is performed by normalization factor, and the confidence assignment value of each state after fusion is calculated. For the proposition The fused confidence assignment function is: ; In the formula, m fusion (A) is the proposition after fusion. A The confidence assignment function, A Denotes any subset of the hypothesis set Θ, which can be a single hypothesis (such as θ). c ) or hypothetical combinations (such as {θ) c ,θ n}), A i Indicates the first i Vehicles on Proposition A i Reliability allocation, ∩ A i = A The intersection of the confidence assignments of all vehicles is equal to... A , N The number of vehicles participating in the integration. K This represents the normalization coefficient, used to eliminate conflicts between pieces of evidence.

[0107] S68. Calculate a normalization factor based on the evidence vectors of multiple vehicles. The normalization factor is obtained by multiplying the confidence scores of all propositions whose intersection is an empty set, summing the results, and then subtracting the sum from 1. The normalization factor is used to measure the degree of conflict between the evidence. When the normalization factor is less than a preset conflict threshold, the evidence is highly conflicting. A weighted average fusion strategy is adopted to arithmetically average the confidence scores of each state in the evidence vectors of multiple vehicles to obtain the fused confidence allocation value. The formula for calculating the normalization factor K is: ; In the formula, This indicates that the intersection of the confidence assignments of all vehicles is an empty set, meaning the evidence is completely conflicting. This indicates a product of the confidence scores of N vehicles.

[0108] K The value reflects the degree of consistency among pieces of evidence. K When the value is close to 1, it indicates that the observations from multiple vehicles are highly consistent; when K When the value is small (e.g., K < 0.1), it indicates a severe conflict; when a high conflict is detected, such as... Traditional Dempster rules can produce unreasonable fusion results. In this case, a weighted average fusion strategy is used as an alternative, expressed as: ; In the formula, 1 / N indicates that each vehicle has an equal weight. Indicates to N The car is about the proposition AThe reliability is summed; this strategy has better robustness when the evidence is highly conflicting, and avoids distortion of the fusion results due to individual anomalous observations.

[0109] S69. Based on the fused confidence assignment value, a decision threshold is set to make a lane state decision. The decision threshold is a preset value, for example, set to 0.7, to determine whether the probability quality after fusion is sufficient to support deterministic decision-making. The specific decision rules are as follows: When the confidence assignment value for lane closure is greater than or equal to the decision threshold, the lane is determined to be closed; this determination indicates that the multi-vehicle observation information highly consistently supports lane closure and has sufficient confidence.

[0110] When the confidence assignment value of the lane's normal traffic status is greater than or equal to the decision threshold, the lane is determined to be in a normal traffic status; this determination indicates that the multi-vehicle observation information highly consistently supports the normal lane status, and the confidence level is high enough.

[0111] When the confidence assignment value for an uncertain observation state is greater than or equal to the decision threshold, or when the confidence assignment values ​​for both lane closure and normal lane traffic states are less than the decision threshold, the lane state is determined to be uncertain, and the map is not updated temporarily. This determination indicates that the current observation information is insufficient to make a definitive judgment, and possible reasons include insufficient observation data, conflicts between multi-vehicle observations, or generally low perception quality. In this case, the map is not updated temporarily, and more observation data is awaited.

[0112] S610, when the lane status is determined to be closed, or changes from closed to normal traffic status, update information is written to the dynamic layer of the high-precision map. The update information includes the closed lane number, the geographical coordinates of the starting point of the closed range, the geographical coordinates of the ending point of the closed range, the status update timestamp, and the confidence score. S611, Set the update frequency of the high-precision map dynamic layer to a preset frequency value, and push the updated map information to vehicles in the area in real time according to the update frequency, so as to provide the latest road condition data for route planning and driving decisions.

[0113] It should be noted that the update frequency of the high-precision map's dynamic layer is set to a preset value, such as once per minute. This frequency setting balances the real-time nature of dynamic information with the system's computational load, ensuring information freshness while avoiding waste of system resources due to overly frequent updates. According to the set update frequency, the updated map information is pushed to vehicles within the area in real time. The pushed information includes the latest lane dynamic status, closure range, and update time, which vehicles receive for route planning and driving decisions. Through real-time push notifications, autonomous vehicles can anticipate sudden road conditions such as temporary lane closures and road construction, triggering timely lane-level precise route adjustments, effectively avoiding entering controlled areas or dangerous road sections, and reducing the risk of traffic accidents from the source.

[0114] This embodiment constructs a collaborative fusion and dynamic decision-making mechanism for multi-vehicle perception information based on evidence theory. First, evidence vectors from multiple vehicles regarding the same lane are acquired. The Dempster evidence combination rule is used to fuse the multi-vehicle evidence, and a normalization factor quantifies the consistency between the evidence. When the normalization factor is less than a preset conflict threshold, it indicates a high degree of conflict between multi-vehicle observations. At this point, a weighted average fusion strategy is switched to effectively avoid distortion of the fusion results due to individual abnormal observations, significantly improving the system's robustness and fault tolerance. After fusion, lane status is determined according to a preset decision threshold. When the confidence allocation value for either lane closure or normal traffic conditions reaches the threshold, a deterministic decision is output; otherwise, it is marked as uncertain. The status is not updated temporarily to ensure the reliability of map updates. When a lane is determined to be closed or its status changes, the lane number, geographical coordinates of the closed area, update timestamp, and confidence score are written into the dynamic layer of the high-precision map and pushed to vehicles in the area in real time at a preset frequency. This fusion mechanism makes full use of the wide coverage and high redundancy of crowdsourced data. Through multi-vehicle collaborative verification, it effectively suppresses the missed detections and false detections of single-vehicle perception, improves the accuracy and reliability of lane dynamic status decisions, and ensures the dynamic freshness of the high-precision map through the real-time push mechanism. This provides reliable data support for autonomous vehicles to perceive sudden road conditions such as temporary lane closures and road construction in advance, reducing the risk of traffic accidents from the source and improving road traffic efficiency.

[0115] Secondly, the present invention also provides a high-precision map lane-level dynamic information update system based on crowdsourced data, implemented using a high-precision map lane-level dynamic information update method based on crowdsourced data, including: The crowdsourced data acquisition module is used to collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information. The lane line perception module is used to construct a single-stage detection model of the line anchor using a convolutional neural network. Based on the single-stage detection model of the line anchor, the module performs perception and detection of lane lines in road image data and outputs the start and end coordinates and slope of the lane lines in the image coordinate system as lane line perception information. The traffic cone perception module is used to register road image data of consecutive frames during vehicle travel using an image alignment compensation algorithm, calculate motion difference maps of consecutive frames, and perform multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The data is then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information. The perception information projection module is used to project lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system using camera parameters; The lane-level matching and positioning module is used to determine the road segment where the vehicle is located based on the collected vehicle trajectory data, using a hidden Markov model, filtering the candidate lane set, and combining lane width information and lane slope information to calculate the vehicle's lateral absolute position in the current lane and determine the current lane of the vehicle. The multi-vehicle evidence fusion and update module is used to construct evidence vectors of three probabilistic quality based on evidence theory, which include lane closure status, lane normal traffic status and observation uncertainty status, by accumulating and fusing the evidence vectors of multiple vehicles through evidence combination rules, and then updating the high-precision map after deciding the lane dynamic status according to a preset threshold.

[0116] It should be noted that this system corresponds to the above-mentioned method for updating lane-level dynamic information of high-precision maps based on crowdsourced data. All implementation methods in the above-mentioned method embodiments are applicable to the embodiments of this system and can achieve the same technical effect.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0119] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0122] If the aforementioned functions are implemented as software functional units 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 invention, essentially, or the part that contributes to the prior art, or a portion 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 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 invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0123] Furthermore, it should be noted that in the system and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Moreover, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order; some steps can be executed in parallel or independently of each other. Those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in hardware, firmware, software, or a combination thereof. This is something that those skilled in the art can achieve by using their basic programming skills after reading the description of the present invention.

[0124] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing system. The computing system can be a known general-purpose system. Therefore, the object of the present invention can also be achieved simply by providing a program product containing program code implementing the method or apparatus. That is, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any known storage medium or any storage medium developed in the future. It should also be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent to the present invention. Furthermore, the steps performing the above series of processes can naturally be performed in the order described, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel or independently of each other.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for updating lane-level dynamic information in high-precision maps based on crowdsourced data, characterized in that, Includes the following steps: S1, collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information. S2 uses a convolutional neural network to construct a single-stage detection model for lane anchors. Based on the single-stage detection model for lane anchors, lane lines in road image data are detected and perceived. The start and end coordinates and slope of the lane lines in the image coordinate system are output as lane line perception information. S3 uses an image alignment compensation algorithm to register road image data of consecutive frames during vehicle travel, calculates motion difference maps of consecutive frames, and performs multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The data is then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information. S4 projects lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system through camera parameters; S5. Based on the collected vehicle trajectory data, the hidden Markov model is used to determine the road segment where the vehicle is located, the candidate lane set is filtered, and the lane width information and lane slope information are combined to calculate the vehicle's lateral absolute position in the current lane and determine the current lane of the vehicle. S6, based on evidence theory, constructs the perception information of multiple vehicles on the same lane into three types of evidence vectors with probabilistic quality, including lane closure state, lane normal traffic state, and observation uncertainty state. The evidence vectors of multiple vehicles are accumulated and fused through evidence combination rules, and the high-precision map is updated after deciding the lane dynamic state according to a preset threshold.

2. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 1, characterized in that, Step S2 includes the following sub-steps: S21, a single-stage detection model for line anchors is constructed using a convolutional neural network. The backbone network of the convolutional neural network is used to extract features from the road image. The backbone network of the convolutional neural network includes multiple downsampling convolutional layers connected in sequence, which are used to output high-level feature maps and shallow feature maps respectively. S22, a set of radiating line anchors emanating from the bottom to the top of the image are predefined on the high-level feature map. Each line anchor has a preset starting point coordinate, direction angle and length. The starting point coordinate is fixed at the bottom edge of the image, and the direction angle is selected within a preset angle range. S23, the feature vectors of the areas covered by each anchor are mapped to fixed-length feature vectors through the region of interest alignment operation, and then input to the classification branch and the coarse regression branch respectively. The classification branch outputs the confidence score that the anchor belongs to the lane line, and the coarse regression branch outputs the coordinate offset between the start point of the anchor and the start point of the real lane line and the coordinate offset between the end point of the anchor and the end point of the real lane line. S24. Select the top K line anchors as candidate line anchors based on the confidence scores from high to low, and perform coarse correction on the start and end coordinates of the candidate line anchors based on the output coordinate offset of the coarse regression branch to obtain the rough start and end coordinates of the lane line on the high-level feature map. S25, map the coarse start and end point coordinates of the candidate line anchors from the high-level feature map to the shallow feature map to obtain the corresponding mapping region on the shallow feature map, extract the local feature vector of the mapping region, input it into the fine regression branch, and the fine regression branch outputs the fine offset to correct the coarse start and end point coordinates to obtain the fine start and end point coordinates of the lane line in the image coordinate system. S26, calculate the lane slope based on the fine start and end point coordinates, and output the start coordinates, end coordinates and slope of the left lane line, as well as the start coordinates, end coordinates and slope of the right lane line.

3. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 2, characterized in that, Step S3 includes the following sub-steps: S31, acquire road image data of consecutive frames during vehicle travel, and register adjacent frame images using an image alignment compensation algorithm to obtain the aligned and compensated consecutive frame images; S32, calculate the pixel-by-pixel motion difference based on the aligned and compensated consecutive frame images, and construct the motion difference map of the consecutive frames; S33, the original road image features of the corresponding frame are stitched and fused with the motion difference map in the channel dimension to obtain a multimodal feature map; S34. Input the multimodal feature map into the target detection model for recognition, output the coordinates of the traffic cone center point and the confidence score in the image coordinate system, and filter the traffic cone detection results with a confidence score greater than the confidence threshold as traffic cone perception information.

4. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41, obtain the camera's intrinsic parameter matrix, and use the intrinsic parameter matrix to convert the lane line perception information and traffic cone perception information in the image coordinate system into three-dimensional coordinates in the camera coordinate system; S42, obtain the camera's extrinsic parameter matrix, which includes a rotation matrix and a translation vector. Use the extrinsic parameter matrix to convert the three-dimensional coordinates in the camera coordinate system into three-dimensional coordinates in the vehicle coordinate system. S43, acquire the vehicle's positioning information, which includes the global navigation satellite system position and the vehicle's attitude angle. Use the positioning information to transform the three-dimensional coordinates in the vehicle coordinate system to the geographic coordinate system of the high-precision map, and output the projection results of lane lines and traffic cones in the geographic coordinate system of the high-precision map.

5. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 1, characterized in that, Step S5, which involves using a hidden Markov model to determine the road segment where the vehicle is located based on the collected vehicle trajectory data and filtering the candidate lane set, includes the following sub-steps: S51, the road segments adjacent to the spatial location of the vehicle trajectory in the high-precision map are taken as the set of hidden states of the hidden Markov model, and the trajectory points of the trajectory data collected during the vehicle's driving process are taken as the observation state sequence. S52, for any trajectory point and candidate road segment, calculate the minimum perpendicular distance from the trajectory point to the centerline of the candidate road segment, and calculate the probability of observing each trajectory point under the condition that the vehicle is located in the candidate road segment based on the Gaussian distribution, and obtain the observation probability; S53, calculate the observed displacement distance between two adjacent trajectory points, and calculate the shortest path distance between candidate road segments based on the candidate road segments corresponding to the two adjacent trajectory points. Calculate the state transition probability between the two candidate road segments based on the difference between the observed displacement distance between trajectory points and the shortest path distance between candidate road segments. S54. Based on the spatial distance between the starting trajectory point of the trajectory sequence and each candidate road segment, calculate the probability that the vehicle's initial state is located in each candidate road segment, and obtain the initial state probability. S55. Based on the observation probability, state transition probability and initial state probability, the Viterbi algorithm is used to solve the road segment sequence corresponding to the maximum posterior probability to obtain the optimal matching road segment sequence. S56. Based on the frequency of occurrence of each road segment in the optimal matching road segment sequence, calculate the posterior probability of each road segment, filter road segments with posterior probabilities higher than a preset probability threshold, and construct a candidate road segment set.

6. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 5, characterized in that, Step S5, which combines lane width information and lane slope information to calculate the vehicle's lateral absolute position within the current lane and determine the vehicle's current lane, includes the following sub-steps: S57, obtain the slope of the left lane line and the slope of the right lane line, and calculate the ratio of the distance from the vehicle to the left lane line to the total lane width based on the difference between the slopes of the left lane line and the right lane line and the slope of the right lane line. S58: Obtain the width information of the current lane from the high-precision map, multiply the current lane width by a ratio, and obtain the lateral absolute position of the vehicle relative to the left boundary of the current lane. S59: Obtain all lane information of the road where the vehicle is located in the horizontal direction from the high-precision map, take the leftmost boundary of the road as the starting position in the horizontal direction, arrange them in the horizontal order of the road, and accumulate them according to the width of each lane to construct the cumulative width range of each lane in the horizontal direction of the road. S510: Based on the vehicle's lateral absolute position relative to the left boundary of the current lane and the sequential number of the current lane in the road, the widths of all lanes before the current lane are accumulated, and then added to the vehicle's lateral absolute position in the current lane to obtain the vehicle's cumulative position in the lateral direction of the entire road. S511: Based on the width information of each lane in the road, the lateral space of the road is divided into multiple lane intervals. Each lane interval corresponds to a lane number. When the lateral cumulative position of a vehicle is located in a certain lane interval, the lane number where the vehicle is currently located is determined to be the lane corresponding to that lane interval, and the lane-level matching result is output.

7. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 1, characterized in that, Step S6, based on evidence theory, constructs three types of probability quality evidence vectors from the perception information of multiple vehicles on the same lane, including lane closure state, lane normal traffic state, and observation uncertainty state. This includes the following sub-steps: S61, Define a set of lane state assumptions, which includes lane closure state, lane normal traffic state, and observation uncertainty state; S62, construct an evidence vector based on the probability mass of lane closure state, lane normal traffic state and observation uncertainty state, wherein the sum of the probability mass of lane closure state, lane normal traffic state and observation uncertainty state is 1; S63, Obtain the perception results of a single vehicle on the target lane. The perception results include traffic cone perception information, lane line perception information, and matching error between the lane line and the high-precision map. The matching error includes lane line start and end point position deviation and lane line slope deviation. S64, dynamically assign values ​​to each component of the evidence vector based on the matching quality of the perception results, and output the evidence vector of a single vehicle to the target lane, including: When a traffic cone is detected, and the deviations in the starting and ending points and the slope between the lane line perception information and the high-precision map are both less than the preset deviation threshold, and the confidence level of the traffic cone is greater than the preset confidence threshold, the perception quality is determined to be high confidence. The probability quality of the lane closure state in the evidence vector is set to be no less than the first preset threshold, and the remaining probability quality is allocated to the observation uncertainty state. When no traffic cone is detected, and the deviations in the starting and ending points and the slope between the lane line perception information and the high-precision map are both less than the preset deviation threshold, and the observation is stable for multiple consecutive frames, the lane is determined to be in a normal traffic state. The probability mass of the normal traffic state of the lane in the evidence vector is set to no less than the second preset threshold, and the remaining probability mass is allocated to the observation uncertainty state. When the deviation of the lane line perception information from the start and end point positions or the slope deviation from the high-precision map is greater than the preset deviation threshold, or the lane line detection confidence is lower than the preset lane line confidence threshold, or there is image occlusion or abnormal lighting interference, the perception quality is judged to be low confidence. The probability quality of the uncertain observation state in the evidence vector is set to no less than the third preset threshold, and the remaining probability quality is allocated to the lane closure state and the lane normal traffic state.

8. The method for updating lane-level dynamic information of high-precision maps based on crowdsourced data as described in claim 7, characterized in that, Step S6, which involves accumulating and fusing evidence vectors from multiple vehicles using evidence combination rules and updating the high-precision map after determining the lane dynamic state based on a preset threshold, includes the following sub-steps: S66, obtain the evidence vector corresponding to the observation results of multiple vehicles in the same lane; S67. The Dempster evidence combination rule is used to fuse the evidence vectors of multiple vehicles. The fusion process includes: for any proposition in the lane closure state, lane normal passage state, and observation uncertainty state, the confidence of the corresponding proposition in the evidence vectors of multiple vehicles is multiplied together, and the multiplication results of all intersections equal to the proposition are summed. Then, the normalization is performed by normalization factor, and the confidence assignment value of each state after fusion is calculated. S68. Calculate a normalization factor based on the evidence vectors of multiple vehicles. The normalization factor is obtained by multiplying the confidence scores of all propositions whose intersection is an empty set, summing the results, and then subtracting the sum from 1. The normalization factor is used to measure the degree of conflict between the evidence. When the normalization factor is less than a preset conflict threshold, the evidence is highly conflicting. A weighted average fusion strategy is adopted to arithmetically average the confidence scores of each state in the evidence vectors of multiple vehicles to obtain the fused confidence allocation value. S69, based on the fused confidence assignment value, set a decision threshold and make a lane state decision: When the confidence assignment value for the lane closure status is greater than or equal to the decision threshold, the lane is determined to be in a closed state. When the confidence assignment value of the lane's normal traffic status is greater than or equal to the decision threshold, the lane is determined to be in a normal traffic status. When the confidence assignment value of the observed uncertain state is greater than or equal to the decision threshold, or when the confidence assignment values ​​of both the lane closure state and the lane normal traffic state are less than the decision threshold, the lane state is determined to be uncertain and the map is not updated for the time being. S610, when the lane status is determined to be closed, or changes from closed to normal traffic status, update information is written to the dynamic layer of the high-precision map. The update information includes the closed lane number, the geographical coordinates of the starting point of the closed range, the geographical coordinates of the ending point of the closed range, the status update timestamp, and the confidence score. S611, Set the update frequency of the high-precision map dynamic layer to a preset frequency value, and push the updated map information to vehicles in the area in real time according to the update frequency, so as to provide the latest road condition data for route planning and driving decisions.

9. A high-precision map lane-level dynamic information update system based on crowdsourced data, implemented using the high-precision map lane-level dynamic information update method based on crowdsourced data as described in any one of claims 1-8, characterized in that, include: The crowdsourced data acquisition module is used to collect trajectory information and road image data of crowdsourced vehicles. The crowdsourced vehicles are equipped with positioning and visual perception devices and have built-in high-precision maps. The high-precision maps include lane group information, lane marking information, lane width information, lane centerline information and road boundary information. The lane line perception module is used to construct a single-stage detection model of the line anchor using a convolutional neural network. Based on the single-stage detection model of the line anchor, the module performs perception and detection of lane lines in road image data and outputs the start and end coordinates and slope of the lane lines in the image coordinate system as lane line perception information. The traffic cone perception module is used to register road image data of consecutive frames during vehicle travel using an image alignment compensation algorithm, calculate motion difference maps of consecutive frames, and perform multimodal fusion of the original road image features of the corresponding frames with the motion difference maps. The data is then input into the target detection model for recognition, and the coordinates and confidence of the traffic cone center point in the image coordinate system are output as traffic cone perception information. The perception information projection module is used to project lane line perception information and traffic cone perception information onto the high-precision map geographic coordinate system using camera parameters; The lane-level matching and positioning module is used to determine the road segment where the vehicle is located based on the collected vehicle trajectory data, using a hidden Markov model, filtering the candidate lane set, and combining lane width information and lane slope information to calculate the vehicle's lateral absolute position in the current lane and determine the current lane of the vehicle. The multi-vehicle evidence fusion and update module is used to construct evidence vectors of three probabilistic quality based on evidence theory, which include lane closure status, lane normal traffic status and observation uncertainty status, by accumulating and fusing the evidence vectors of multiple vehicles through evidence combination rules, and then updating the high-precision map after deciding the lane dynamic status according to a preset threshold.

10. A computer-readable storage medium, characterized in that, The storage medium stores a method program for updating lane-level dynamic information of a high-precision map based on crowdsourced data. When the method program is executed, it implements the method for updating lane-level dynamic information of a high-precision map based on crowdsourced data as described in any one of claims 1-8.

Images