Precise navigation method for charging pile based on site guide photo and geographic information

Abstract

The application discloses a charging pile accurate navigation method based on site guide photos and geographic information, relates to the technical field of vehicle-mounted intelligent navigation or computer vision and multi-source positioning fusion, and is used for solving the problem that after a vehicle enters a charging station, the vehicle cannot accurately find a specified charging pile and cannot complete reliable station entry and parking alignment; a photo-geographic structure joint reference data set is constructed and is uniformly calibrated; real-time images and GNSS or IMU information are collected at a vehicle end, a site is locked by searching and geometric verification, a vehicle pose is output by fusion positioning, a target charging pile is recognized by combining images and point clouds after entering the station, a relative pose is estimated, a smooth path is generated and followed, and a pile position anchor point is updated after end alignment.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted intelligent navigation or computer vision and multi-source positioning fusion technology, and more specifically, to a method for precise navigation of charging piles based on site guidance photos and geographic information. Background Technology

[0002] With the increasing popularity of electric vehicles, the number of public and industrial park charging stations is growing rapidly. However, charging piles are often located in complex environments such as underground garages, enclosed parks, or multi-story parking lots. Existing navigation systems mostly rely on station-level POIs and road-level routes, lacking fine-grained guidance such as lane directions, entrance corners, parking space numbers, and the appearance of the charging piles once inside the station. Furthermore, GNSS is prone to drift of several meters under conditions of obstruction and reflection, making it difficult to achieve accurate positioning upon arrival at the station and the charging pile. Some solutions rely solely on image recognition or point cloud positioning, which is easily affected by changes in lighting, obstruction, and interference from similar scenes. Moreover, the passable areas within the station can dynamically change due to construction closures or temporary parking, rendering the planned route unusable. On the other hand, map data updates are lagging, resulting in missing or inaccurate charging pile coordinates, and a lack of traceable correction mechanisms.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for precise navigation of charging piles based on site guidance photos and geographic information, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a preferred embodiment, it includes: Construct a joint reference dataset of site guide photos and geographic information, write the site guide photos and site geographic structure into the joint reference dataset, and create site index entries; Acquire vehicle observation frames, determine candidate sets based on candidate station sorting list, and perform cross-frame geometric consistency competition locking for candidate sets: recursively calculate the cumulative advantage potential value of each candidate station within the observation window, and introduce observation gating coefficient to suppress abnormal frames; drive observation window update based on congestion and clustering degree, and output target stations and control replacement by combining entry boundary, maintenance boundary and inertial potential value. Acquire local environmental observation data to generate local observation frames, identify target charging piles and estimate the relative pose of target charging piles, calculate target parking positions and generate smooth path records; Based on the smooth path record, the system performs path following and precise end-point alignment control to generate vehicle control quantities; it corrects the geographic anchor points of the target charging piles and obtains stable anchor point values, writes the stable anchor points back to the site guidance photos, and records geographic information in conjunction with the reference dataset in a versioned manner.

[0006] In a preferred embodiment, by collecting and uploading site guidance photos, EXIF ​​data or sensor information is read to generate a photo metadata record containing the collection time, collection location, and collection orientation. When a location is missing, the site guidance photo is used as the query image to perform image retrieval in a geographically tagged reference image library to obtain a candidate location range, which is then written into the metadata record. After lens distortion correction and brightness normalization preprocessing of the site guidance photos, local and global features are extracted and written into a joint reference dataset. For local feature matching, a random sampling consistency estimation geometric transformation model is used to eliminate outliers and obtain inlier matching pairs. The model type, threshold, and number of inliers are then associated with the database.

[0007] In a preferred embodiment, a road network, traversable areas, and charging station points of interest are obtained by calling a map data interface or importing map data. The road network map, traversable area boundaries, and point of interest coordinates are then parsed and written into a joint reference dataset. When the geographic information lacks a station-level location, the geographic anchor point of the target charging station is marked as pending confirmation, and the station-level location source is recorded. The joint reference dataset is time-consistent and coordinate-consistent, and a version field, coordinate system identifier, and projection parameter identifier are written into it. Coordinate transformation is performed according to the projection parameters, local coordinates within the station are preserved, and an association with the station identifier is established. Face and license plate area detection is performed on the station guidance photos, and occlusion or blurring is performed according to the bounding box. A station index entry containing a station identifier, photo set identifier, geographic structure version field, visual feature summary, and source identifier is generated for each station, and a retrieval association is established with the dataset.

[0008] In a preferred embodiment, real-time images and positioning information are acquired at the vehicle end, distortion correction and brightness normalization preprocessing are performed on the vehicle images, and local and global features are extracted. The preprocessed images, observation visual features and positioning information are mapped to the same time base according to the timestamp to complete the alignment and generate vehicle observation frames.

[0009] In a preferred embodiment, multiple candidate stations are extracted from the sorted list to form a candidate set, and a variable-length observation window is constructed using the last frame. The window length is driven by and limited within a preset range by the crowding and clustering degree calculated from the candidate score interval sequence and the number of leader switching within the window. For each candidate station and each frame within the window, the local features of the vehicle observation frame are matched with the local features of the station guidance photo associated with the candidate station to obtain a set of candidate matching pairs. The inter-frame geometric transformation model is estimated using random sampling consistency and outliers are eliminated to obtain a set of inliers. Based on this, the inlier size, mean and fluctuation of reprojection error residuals, inlier convex hull coverage, and geometric model drift of adjacent frames are calculated. The observation gating coefficient is calculated using the acceleration and angular velocity output by the inertial measurement unit to attenuate abnormal frames.

[0010] In a preferred embodiment, a convergence index is constructed based on the convergence amount and fluctuation suppression term of the residual as the frame progresses, and the size, coverage, convergence, stability and their coupling are mapped to the injection amount per frame. The cumulative advantage potential value of each candidate site is recursively calculated based on the gated weighted injection amount within the window. The injection rate is pruned in conjunction with the window length to adapt to window expansion and contraction. At the same time, an entry boundary and a retention boundary are constructed based on the candidate set complexity, and an inertial potential value obtained by recursion from the leading indicator is maintained for each candidate. At each time step, the main candidate and the shadow candidate are selected and the advantage interval is calculated. When the entry boundary and the inertial threshold are met, the target site is output. After output, replacement is only performed when the shadow candidate meets the retention boundary and the inertial threshold. At other times, the output is maintained and the potential value, inertial potential value and window length are updated.

[0011] In a preferred embodiment, after obtaining the target station, its geographical structure is read, GNSS geographical coordinates are read from the vehicle observation frame as the vehicle coarse geographical coordinates, and the coarse geographical coordinates are projected onto the nearest roadside in combination with the road network map. Then, an accessibility check is performed based on the connectivity with the boundary of the target station. If the check fails, the next candidate station is switched according to the candidate station sorting list and the check is repeated until the check passes. After passing through, the global position and orientation of the vehicle are solved by fusing GNSS and IMU outputs using extended Kalman filtering in the projected coordinate system, and the position is projected onto the boundary of the passable area for map constraint correction; finally, the location description of the target charging pile is written according to whether the geographical anchor point of the target charging pile exists. When the anchor point is to be confirmed, the appearance category and number visible identifier obtained by the site guidance photo are written as auxiliary description and the target charging pile record is generated.

[0012] In a preferred embodiment, vehicle camera images and LiDAR point clouds are acquired, and timestamp synchronization and coordinate transformation are completed. In the images, a target detection network is used to obtain the bounding box of the target charging pile and extract the key points of the pile. On the point cloud side, candidate pile clusters are selected by neighborhood clustering, and vertical boundary line segments are extracted by random sampling consistency line segment fitting to calculate the position of the pile center point. A two-dimensional and three-dimensional correspondence is established based on the key point pixels and the geometric feature points of the point cloud. The pose is obtained by solving three or more perspective points and transformed to the vehicle coordinate system by extrinsic parameters. The pose is then checked for consistency with the center point of the point cloud. The normal unit vector is calculated from the pile center point and orientation, and the parking position is obtained according to the reserved distance. The current position of the vehicle is mapped as the planning start point and the parking position is mapped as the planning end point. The passable area and road network are discretized to generate a grid map and the obstacle grid obtained by semantic segmentation is updated. A bidirectional A-star search with dynamic weights and pruning is used to generate an initial path point set. A smooth path is generated by cubic Bezier piecewise fitting. After the consistency of the end point and the pose threshold is checked, the relative pose, parking position and smooth path record are output.

[0013] In a preferred embodiment, segmented control points from the smooth path record are read and sampled using a cubic Bézier curve to generate a desired trajectory point sequence. The vehicle's current position, orientation, and speed are obtained as feedback quantities. Lateral and heading deviations are calculated for each control cycle, and steering and speed control quantities are generated and executed accordingly. The speed control quantity is adaptively adjusted based on trajectory curvature and obstacle grid constraints. During the following process, the passable grid map is continuously updated using semantic segmentation and point cloud projection. Forward collision prediction is performed along the desired trajectory. When a trajectory point falls into an obstacle unit, replanning is triggered, and the current path is replaced with a new smooth path record. When the vehicle enters the target area... After marking the neighborhood of the parking position, the frequency of updating the target charging pile center point and normal vector obtained by target charging pile detection, key point extraction and point cloud clustering line segment fitting is increased. The position error and orientation error are calculated and the end alignment control quantity is generated. In the end alignment stage, the estimated spatial center point of the pile is made consistent with the geographic anchor point of the target charging pile. When the anchor point is to be confirmed, the pile center point is transformed from the vehicle coordinate system to the station coordinate system as the anchor point candidate value, and the candidate value is subjected to counting correction based on two-level thresholds to output a stable anchor point value. The stable anchor point value and its source identifier are written into the anchor point field of the joint reference dataset, and the anchor point version and confirmation status are written into the timestamp and version fields.

[0014] The technical effects and advantages of the charging pile precise navigation method based on site guidance photos and geographic information of this invention are as follows: This invention achieves rapid site location by using global feature retrieval and local feature RANSAC verification, suppressing mislocalization in similar scenarios; it improves continuity and robustness in weak signal environments by using GNSS or IMU extended Kalman fusion combined with navigable area projection constraints; during the approach phase, it integrates target detection, key point and point cloud clustering or line segment fitting to obtain reliable pile relative poses, supporting precise docking alignment. Path planning employs dynamic obstacle updates, coupled with bidirectional A... The system is smoothed with Bessel curves, reducing detours and sharp turns. After the terminal results are stabilized through counting, the pile location anchor points and version information are written back to update, forming a traceable self-calibrating closed loop. Attached Figure Description

[0015] Figure 1 This is a curve comparing the positioning error of the charging pile precise navigation method based on site guidance photos and geographic information according to the present invention.

[0016] Figure 2 This is a comparison chart of path search and smoothing for the accurate navigation method for charging piles based on site guidance photos and geographic information according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example This invention discloses a method for precise navigation of charging stations based on site guidance photos and geographic information, such as... Figure 2 As shown, it includes: Step 1: Construct a joint reference dataset of site guide photos and geographic information and complete the standardization and calibration; First, site guidance photos are acquired. These photos are collected and uploaded on-site by the charging station operator or an authorized data acquisition entity. Each photo must contain images reflecting the charging station entrance, lane layout, parking spaces, charging pile appearance, and surrounding fixed environment. For each photo, its metadata is read and a photo metadata record is generated. This metadata comes from the EXIF ​​field of the photo file or sensor information simultaneously recorded by the acquisition terminal. The photo metadata record includes at least the acquisition time, acquisition location, and acquisition orientation, or equivalent information. When a photo file does not contain an EXIF ​​location field or the location field is incomplete, the location field is not directly assigned a value. Instead, the photo is used as a query image in a reference image library to perform image retrieval to obtain candidate location ranges for the photo. The reference image library consists of geotagged images with legitimate sources. The geotags originate from the original acquisition metadata of the reference images or annotation data from the map data provider. Each reference image in the reference image library stores its geotag and source identifier. The retrieved candidate location ranges and retrieval similarity results are written into the photo metadata record to replace the missing EXIF ​​location field.

[0019] Subsequently, image preprocessing is performed on the site guidance photos to generate preprocessed photos. The image preprocessing includes at least lens distortion correction and brightness normalization. Lens distortion correction uses camera intrinsic parameters to correct distortion in the unprocessed photos. These camera intrinsic parameters are derived from the factory calibration parameters of the acquisition terminal or calibration results obtained through a calibration board during acquisition. When the site guidance photos come from different models of acquisition terminals, an intrinsic parameter configuration is established for each model, and the corresponding intrinsic parameter configuration identifier is written into the photo metadata record, enabling distortion correction to complete the processing by calling the corresponding intrinsic parameter configuration according to the image source. Brightness normalization uses histogram equalization, gamma correction, or adaptive exposure compensation to normalize the illumination of the preprocessed photos, reducing brightness differences caused by different acquisition periods.

[0020] After image preprocessing, site visual features are extracted from the preprocessed photos and written into the site guidance photo and geographic information joint reference dataset. The site visual features include at least local and global features. Local features are generated by detecting keypoints in the preprocessed photos using SIFT, ORB, or a learning-based feature point network, producing a set of keypoint coordinates and corresponding descriptor sets. Global features are generated by generating global feature vectors from the preprocessed photos using a global feature extraction network, producing a global representation for image similarity calculation. To enable geometric consistency discrimination of local features, a random sampling consistency method is introduced to perform outlier removal: candidate matching pairs from two photos are used as input, and the random sampling consistency method is used to estimate the geometric transformation model between the two photos, removing matching pairs that do not meet the geometric transformation model, and retaining inlier matching pairs that pass the geometric consistency test. The type of geometric transformation model used, the outlier removal threshold, and the number of inlier matching pairs are written into the joint reference dataset and associated with the site visual features of the corresponding photos.

[0021] Further geographic information is acquired and a site geographic structure is generated. This geographic information is obtained by calling map data interfaces, importing offline map files, or reading open-source map data. The geographic information includes at least road network data, traversable area data, and charging station point of interest (POI) information. Road network data is used to construct road connectivity, traversable area data is used to determine the boundaries of traversable areas inside and outside the station, and charging station POI information is used to identify the charging station's range and center location. The above geographic information is parsed into a site geographic structure and written into a joint reference dataset. The site geographic structure includes at least a road network map, traversable area polygon boundaries, and a set of charging station POI coordinates. When the geographic information contains fine-grained location data for charging piles or parking spaces, it is written as the target charging pile geographic anchor point; when the geographic information only contains site-level locations and not pile-level locations, the target charging pile geographic anchor point field is marked as pending confirmation, and only its upper-level site-level location field and its source identifier are saved.

[0022] After the site guidance photos and geographic information are written into the joint reference dataset, a standardization calibration process is performed on the joint reference dataset. This standardization calibration process includes at least time standardization and coordinate standardization. Time standardization unifies the acquisition time in the photo metadata record with the update timestamp of the geographic information to the same time base, and writes version fields for the site guidance photos and site geographic structures in the joint reference dataset to distinguish between different acquisition batches and different map update batches. Coordinate standardization performs coordinate system annotation and necessary coordinate transformations on the coordinate expressions of data from different sources in the geographic information: it writes the geographic coordinate system identifier for latitude and longitude coordinates and the projection parameter identifier for projected coordinates; when conversion between geographic coordinate expressions and projected coordinate expressions is required, the projection parameters provided by the map data provider or publicly available projection parameter files are used for coordinate transformation, and the source identifier of the projection parameters used is written into the joint reference dataset. If there are local coordinate expressions within a site, their original coordinate expressions are retained and their coordinate system identifiers are written in; they are not directly replaced with latitude and longitude coordinates; at the same time, an association record is established between this local coordinate expression and the site identifier.

[0023] Finally, the site guidance photos undergo compliance processing and are indexed and stored in the database. Compliance processing includes masking or blurring image areas that may contain sensitive personal information, targeting at least face and license plate regions. These face and license plate regions are identified using an object detection network, with the detection results output as sensitive region bounding boxes. Masking or blurring is then applied to the corresponding pixel regions based on these sensitive region bounding boxes. Indexing involves creating a site index entry for each charging station. This entry includes at least a site identifier, a site guidance photo set identifier, a site geographic structure version field, a summary of the site's visual features for each guidance photo, and a source identifier for the photo's metadata. A searchable association is then established between the site index entry and the site guidance photos and site geographic structures in the joint reference dataset.

[0024] Step 2: Solving for candidate site locking and vehicle global localization based on the joint reference dataset; First, real-time vehicle observation data is acquired and vehicle observation frames are generated. The real-time vehicle observation data is collected from the vehicle itself and includes at least real-time vehicle images captured by the vehicle's camera and vehicle positioning information collected by the vehicle's positioning sensor. The vehicle positioning information includes at least the geographic coordinates output by the Global Navigation Satellite System receiver and the attitude angle or angular velocity and acceleration output by the inertial measurement unit. Image preprocessing, consistent with step one, is performed on the real-time vehicle images to obtain preprocessed real-time vehicle images. This image preprocessing includes at least lens distortion correction and brightness normalization. The camera intrinsic parameters used for lens distortion correction are derived from the calibration parameters of the vehicle's camera. Vehicle observation visual features are extracted from the preprocessed real-time vehicle images. These visual features include at least local and global features. Local features are generated by a feature point detection algorithm that detects key points and generates key point descriptors, while global features are generated by a global feature extraction network that produces global feature vectors. The preprocessed real-time vehicle image, vehicle observation visual features, and vehicle positioning information are time-aligned according to the acquisition time to generate a vehicle observation frame. The time alignment is achieved by mapping the timestamp of the real-time vehicle image and the timestamp of the vehicle positioning information to the same time base, and the aligned unified timestamp is written into the vehicle observation frame.

[0025] It's important to note that in urban core areas or large industrial parks, charging stations often exhibit a homogeneous, clustered distribution: the directional photos of multiple stations are highly consistent in overall composition, charging pile shape, and signage style. Furthermore, vehicle observation frames are subject to dynamic interference such as occlusion, specular reflection, and noise from rain, fog, or nighttime conditions. This leads to the global feature similarity assigning similarly high scores to multiple stations simultaneously, resulting in candidate station rankings exhibiting parallel, jump, and short-term reversals across consecutive frames. In this situation, even if a frame temporarily ranks first, it may only be a false advantage caused by momentary changes in lighting or occlusion. Once locked, this can lead subsequent navigation and positioning convergence to incorrect stations, creating a difficult-to-correct chain of errors.

[0026] Therefore, in this embodiment, each candidate site in the candidate site sorting list is further sorted. Perform cross-frame geometric consistency contention locking. First, select the top m sites from the candidate site sorting list to form a candidate set. Where m is stored as a candidate truncation parameter in the parameter table of the site guidance photo and the geographic information joint reference dataset, it is used to balance computational cost and candidate coverage. The vehicle observation frame at time t... Constructing observation windows for the end Window length Subject to the following rules, which are dynamically updated and restricted to: Inside, among which These are all configurable parameters written into the joint reference dataset to control the minimum stable evidence length and the maximum allowable delay.

[0027] In an implementable example, Used to ensure that cross-frame geometric consistency contention lock accumulates to at least the minimum stable evidence length, so that occasional false advantages in a single frame or a few frames do not trigger the target site output; This is used to limit the maximum allowable latency and real-time computing overhead, and to prevent the potential value response from being delayed due to an excessively long window. and It can be set in conjunction with the vehicle camera frame rate and vehicle speed: when the vehicle speed is high or the candidate station ranking list fluctuates greatly, increase the frame rate. To improve stability; reduce [the impact of] lower vehicle speeds or when a faster initial stability lock is required. To improve response. This can be configured in conjunction with the real-time processing budget to ensure that the maximum computational load for RANSAC verification within a window does not exceed the upper limit allowed by the vehicle-side processing cycle. The above example is for illustrating possible implementations and is not limiting. and The specific numerical value.

[0028] For any candidate site With each frame within the window The local features of the frame image are compared with those of the candidate sites. The associated site guides the matching of local features of the photos to obtain a set of candidate matching pairs. ,in For vehicle image pixel coordinates, To indicate the pixel coordinates of the photo. Then... A random sampling consistency estimation model is used to estimate the geometric transformation between two images. Then remove the outliers to obtain the set of interior points. The model type and interior point decision threshold used in RANSAC Maximum number of iterations Both the confidence score p and the model type are provided by the joint reference dataset. It can be linked to site guide photos or site index entries to adapt to different shooting scales and noise levels. p can be used as a global parameter to constrain the real-time computation cost and success probability.

[0029] In an implementable example, the interior point determination threshold Settings related to vehicle camera resolution and lens distortion correction residual level: Higher resolution or more thorough distortion correction... Tinlier can be reduced to improve geometric consistency discrimination accuracy; when the resolution is low or there is some motion blur, Tinlier can be appropriately increased to avoid misclassifying true interior points as exterior points. Maximum number of iterations The confidence level p is used to ensure that the correct model can still be found with sufficient probability when the proportion of outliers is high: it is increased when the candidate matching pair set is larger or the proportion of outliers is higher. Alternatively, increase p to improve the model's estimation success rate; when real-time requirements are stronger, decrease p or limit it. To constrain computational overhead, differentiated configurations for different sites or devices are achieved by fixing versions in a joint reference dataset.

[0030] In obtaining and Then, the observable evidence quantity for that frame is calculated. The in-point scale quantity is defined as: ; For each interior point matching pair Calculate the reprojection error: ; in This is a homogeneous coordinate normalized projection. From this, we obtain the residual mean and residual fluctuation: ; ; To characterize whether the spatial distribution of interior points is sufficiently covered, the area of ​​the convex hull of the interior points on the vehicle image plane is taken as the effective area of ​​the image. Area normalization: ; in The definition originates from the effective imaging mask in vehicle-side image preprocessing, ensuring comparable coverage across different resolutions. To characterize the stability of the geometric model between adjacent frames, the model drift is defined as follows: ; in It is a vectorization operator for expanding matrices in a fixed order, used to transform parameter variations into measurable scalars.

[0031] To suppress the disruption of evidence recursion caused by anomalous frames due to occlusion, reflection, rain, fog, or sharp turns, an observation gating coefficient is introduced. The gating coefficient is calculated from the output of the inertial measurement unit in the vehicle positioning information: the acceleration vector aligned with the timestamp. With angular velocity scalar For input, define: ; in The scaling parameter is used to separate normal driving disturbances from violent motion disturbances. The values ​​can be obtained from offline calibration and written into the parameter table of the joint reference dataset by device or vehicle model, so that the gating strength is consistent and controllable under different IMU noise levels.

[0032] In one feasible example, the offline calibration includes: acquiring the output sequence of the inertial measurement unit under two types of vehicle observation frame conditions: normal driving and sharp turns; statistically analyzing the differences in the distribution of the acceleration vector norm and angular velocity scalar under the two types of conditions; and selecting accordingly. and This allows for observation of the gating coefficient under normal driving conditions. Observation of gating coefficient under conditions close to 1 and intense motion Significant attenuation. Offline calibration can also verify the gating effect on anomalous frame enrichment segments such as those with occlusion, reflection, rain, fog, or nighttime noise: when anomalous frames occur... It should be reduced to decrease its injection contribution to the recursion of the cumulative advantage potential. (Calibrated) and The gating strength is written into the parameter table of the joint reference dataset along with its source identifier and the version is fixed to ensure that the gating strength is consistent and controllable under different IMU noise levels.

[0033] Based on the defined amount of evidence per frame, candidate sites Constructing a recursive injection volume First, define the residual convergence factor: ; The first term indicates whether the error decreases as the frame progresses, and the second term is used for... Suppressing situations where the error decreases but fluctuates drastically; γ is a configurable parameter that can be obtained from offline sequence parameter tuning and written into the joint reference dataset.

[0034] In a feasible example, offline sequence hyperparameter tuning aims to achieve a stable positive accumulation of residual convergence at correct stations as frames progress, while hindering the stable accumulation of residual convergence at incorrect stations. When the residual mean decreases but residual fluctuations are large, γ is increased to suppress the severe fluctuations, thus mitigating the impact of these fluctuations. It is not easily judged as stable convergence; when there are small noise fluctuations in the true convergence sequence, reducing γ avoids excessive penalty that weakens the convergence of correct sites. The hyperparameter-tuned γ and its source identifier are written into the joint reference dataset and the version is fixed.

[0035] Then, the interior point size, coverage, convergence, and drift are collectively mapped to the injection amount: in Used to compress the dynamic range of inlier size and avoid the dominance of maximum inlier number; the first term aggregates four types of information: size, coverage, convergence, and stability; the second term is a coupling term of size and coverage, used to distinguish between inliers concentrated in small areas forming pseudo-consistency and inliers covering the entire map forming true consistency. The parameter set can be obtained and fixed in the database in two ways: first, use grid / Bayes search on the offline validation set to maximize the separation between the injection quantities of correct sites and those of incorrect sites; second, perform dimensional balance settings according to the typical value range of each quantity and then fine-tune it with a small amount of labeled data.

[0036] Maintain the advantage potential for each candidate site. This represents the cumulative advantage at cutoff time t, and is recursively calculated using the gated weighted injection amount within the window: ; in To prevent zero constant; To control the injection rate, the speed at which new evidence overwrites old potential values ​​is controlled, and this is linked to the window length to ensure greater stability during window expansion and greater agility during window contraction. In implementation, the following approaches are recommended: ; in Configurable parameters are added to the database: When the candidate set becomes crowded, causing the window to lengthen, It automatically decreases in size, making it less likely for single-frame perturbations to cause potential value flips; when the candidate gap widens and the window shortens... Increase in size to accelerate convergence.

[0037] In an implementable example, This is used to avoid excessive lag in potential value recursion, which could prevent convergence from occurring in time during the window contraction period. This is to prevent potential value recursion from being overly sensitive to single-frame perturbations, which could lead to potential value flipping during the window expansion period. Used to set near Lt The baseline injection speed at that time. Parameter selection can satisfy the following linkage constraint: when the candidate set is crowded, causing the window to become longer, the clipped... near To improve stability; when the gap between candidates widens, causing the window to shorten, the clipped window... near To improve agility; and through offline sequence parameter tuning, with the joint objective of "lowest false lock rate and shortest first stable lock time", the selected , , Write the joint reference dataset and solidify the version.

[0038] To characterize whether the candidate set is in a highly similar clustered state, the global similarity score sequence is read from the candidate site ranking list. This score, derived from cosine similarity or Euclidean distance, has been used to generate the ranked list. The adjacent interval is defined as follows: ; ; And based on this, the crowding degree and clustering degree are calculated: ; ; This indicates that the smaller the average interval of candidate scores, the more crowded the candidate pool. The less smooth the interval pattern, the more clustered the structure. To directly quantify ranking jumps caused by ties, the leader index is calculated within the window. Number of switching times: ; Then use Drive window scaling: ; in Control the sensitivity of congestion, clustering, and jump to window expansion respectively. The system provides a natural pullback bias to gradually shorten the stabilization window. All four parameters are configured and stored in the database. They can be jointly adjusted and fixed on offline sequences with the goal of minimizing the false lock rate and the shortest first stable lock time.

[0039] In a feasible example, joint regulation , , , At that time, the offline sequence should at least include: a crowded cluster scenario segment where the candidate site ranking list shows ties, jumps, and short-term reversals, and a stable convergence segment where the candidate gap gradually widens. Parameter tuning should be done using the window length. Constrained by the following: within crowded cluster segments, it can expand sufficiently to suppress short-term ranking crossovers; within stable convergent segments, it can gradually shrink back to reduce latency: increase... It can enhance sensitivity to congestion, making windows easier to expand and increase size. This can enhance sensitivity to clustering degree, making cluster structures more susceptible to expansion and increasing... This can enhance sensitivity to the number of leader switching events to suppress frequent jumps and increase... This can enhance the natural shrinkage bias, causing the stabilization window to shorten more quickly. After parameter tuning, the parameters and their source identifiers are written into the joint reference dataset and the version is fixed.

[0040] To avoid frequent switching during candidate set congestion and jitter periods, two distinct potential difference boundary bands are constructed for entry and retention. The dynamic entry boundary is defined as follows: ; in Basic gap requirements, The candidate set complexity is transformed into an additional gap requirement, both of which are input parameters; and the maintenance boundary is defined as follows: ; ; In an implementable example, Used to provide the basic gap requirement, so that the minimum advantage interval must still be reached before the target site is output when the candidate set is not crowded; , , This is used to map the complexity caused by crowding, clustering, and leader switching frequency into an additional gap requirement, so that the more crowded, clustered, or frequently the candidate set changes, the better. The larger the value, the higher the output threshold for the entry stage. κ is used to set the tightness ratio of the holding stage relative to the entry stage: the smaller κ is, the tighter the holding boundary. The lower the value, the less likely the output target site is to be easily overturned; the larger the value, the higher the boundary Bt↓, the more sensitive the system is to new stable leaders, but also the easier it is to switch. κ can be jointly selected through offline sequences under constraints of false lock rate and switching frequency, and the selection results are written into the joint reference dataset and the version is fixed.

[0041] in The control of the tightness ratio between the maintenance phase and the entry phase ensures that the locked target is not easily overturned by short-term fluctuations. To distinguish between short-term and sustained leadership, a maintenance inertial potential value is applied to each candidate. : ; in The inertial memory length is determined by the parameters used for data entry. The smaller the value, the longer it must maintain a leading position for I to reach the threshold.

[0042] In a feasible example, η is used to determine the memory length of the inertial potential: the smaller η is, the less sensitive the inertial potential is to short-term lead, and the longer it needs to maintain a lead to achieve its full potential. Accumulated to a higher level; the larger η is, the more sensitive the inertial potential is to changes in the leading value, which is suitable for situations where the candidate gap widens rapidly and rapid output is required. Used to define the minimum inertia level required for sustained leadership: when the candidate set is in a crowded cluster state, it can be improved. To suppress short-term pseudo-leader-triggered output or replacement; when the candidate set is not crowded and the gap is significant, it can reduce To shorten the initial stable locking time. η and The goal can be jointly set on offline sequences with the lowest false locking rate and the shortest first stable locking time, and written into the solidified version of the joint reference dataset.

[0043] At every moment, the main candidate Shadow Candidate and define the dominance interval. When satisfied and Time output For the target site, among which These are parameters for data entry, used to specify the minimum inertia level required for sustained leadership; once the target site has been output, the shadow candidate only needs to meet these parameters. and If the primary candidate is replaced by a shadow candidate, then the original output is retained and only updated. This dual-track replacement rule prevents output switching from being triggered when candidate sets experience short-term ranking overlaps within crowded clusters. Instead, switching only occurs when the replacement candidate establishes a sustained and sufficiently large stable advantage within the extended window.

[0044] After obtaining the target site, the corresponding site geographic structure is read, and the vehicle positioning information is consistent with the site geographic structure. Specifically, the geographic coordinates output by the Global Navigation Satellite System receiver are read from the vehicle observation frame as the vehicle's coarse geographic coordinates; the set of charging station points of interest coordinates and the road network map are read from the site geographic structure, and the accessibility of the vehicle's coarse geographic coordinates is checked based on the road network map. The accessibility check is achieved by projecting the vehicle's coarse geographic coordinates onto the nearest road edge in the road network map and calculating the connectivity between the projected point and the target site boundary; if the connectivity is not established, the next candidate site in the candidate site ranking list is selected, and the above consistency process is repeated until a target site that meets the accessibility check is obtained.

[0045] After the target station is determined and the vehicle's coarse geographic coordinates pass the reachability check, a global vehicle positioning solution is executed to output the vehicle's global positioning result. The vehicle's global positioning result includes at least the vehicle's current geographic location and current orientation. The vehicle's current geographic location is obtained by fusing the output of the Global Navigation Satellite System receiver from the vehicle's positioning information with the output of the Inertial Measurement Unit; this fusion estimation is implemented using the Extended Kalman Filter algorithm, such as... Figure 1 As shown. The state vector of the extended Kalman filter algorithm is defined as: ; in, and This represents the position components of the vehicle in the preset planar coordinate system. Indicates the vehicle's heading angle. This indicates vehicle speed. The preset planar coordinate system is given by the projected coordinate system determined in step one (coordinate unification), and its projection parameter source identifier has been written into the site guidance photo and geographic information joint reference dataset.

[0046] The extended Kalman filter algorithm uses a vehicle motion model for state prediction: ; in, Let k be the predicted state at time k. Let be the posterior state at time k−1. To control the input vector. (Control input vector) The longitudinal acceleration and angular velocity output from the inertial measurement unit are used, and time alignment is completed under a unified timestamp for the vehicle observation frames. The covariance update for state prediction is as follows: ; in, To predict the covariance matrix, State transition function The Jacobian matrix of the state, The process noise covariance matrix is ​​used to characterize the uncertainty of the motion model. Measurement updates utilize geographic coordinates output from the Global Navigation Satellite System receiver, which are transformed to the preset plane coordinate system using the projection parameters determined in step one to form the measurement vector. ; in, and This represents the plane coordinate measurement components obtained from the coordinate transformation of the output of a Global Navigation Satellite System receiver. The measurement function is defined as: ; in, For measurement function, To measure noise, the Kalman gain is calculated as follows: ; in, Here is the Kalman gain matrix. For measurement function The Jacobian matrix of the state, To measure the noise covariance matrix, which characterizes the uncertainty of global navigation satellite system receiver measurements. The posterior state update is as follows: ; The posterior covariance is updated as follows: ; in, It is an identity matrix.

[0047] In a feasible example, the process noise covariance matrix is ​​used to characterize the uncertainty of the vehicle motion model. It can be set based on the noise level of the inertial measurement unit (IMU), the vehicle speed range, and the control cycle: when the vehicle speed is high or the acceleration fluctuations are large, the process noise covariance matrix is ​​increased to improve the filter's tolerance to model uncertainty; when the vehicle is traveling at a low speed and smoothly, the process noise covariance matrix is ​​decreased to improve state smoothness. The measurement noise covariance matrix is ​​used to characterize the measurement uncertainty of the GNSS receiver. It can be set based on the drift level of the GNSS receiver under obstruction and reflection conditions: when in areas with severe obstruction or reflection, the measurement noise covariance matrix is ​​increased to reduce the impact of measurements on posterior state updates; when the GNSS receiver signal is stable, the measurement noise covariance matrix is ​​decreased to improve position convergence speed. The settings of these matrices can be written into the joint reference dataset as global parameters and fixed to ensure reproducibility in different site environments.

[0048] After outputting the vehicle's current geographic location and current orientation using the extended Kalman filter, the vehicle's global localization result is corrected for map constraints based on the target station's geographic structure. Map constraint correction is achieved by projecting the vehicle's current geographic location onto the boundary of the traversable area within the site's geographic structure: when the vehicle's current geographic location falls outside the traversable area boundary, the nearest point to the boundary is calculated and replaced with the nearest point; when the vehicle's current geographic location falls within the traversable area boundary, the current geographic location remains unchanged. The corrected vehicle's current geographic location, current orientation, and target station identifier are then written into the vehicle's global localization result.

[0049] Finally, the target description of the target charging pile is determined and the binding is completed. If a geographical anchor point for the target charging pile exists in the site's geographical structure, the geographical anchor point is written into the target charging pile record as the target charging pile location description. If the geographical anchor point is in an unconfirmed state, the target charging pile record is limited to a site-level location description, and the available charging pile appearance category, parking space number identifier, or other visible identifiers extracted from the site guidance photos are written into the target charging pile record as auxiliary descriptions. The visible identifiers are identified by the target detection network in the site guidance photos, and the identification results are saved in the form of category labels and bounding boxes and their source photo identifiers are stored in the joint reference dataset. This yields the output result of step two, which includes the vehicle global positioning result, the target site identifier, and the target charging pile record.

[0050] Step 3: Local fine positioning and entry path planning generation for the target charging station; First, local environmental observation data is acquired and local observation frames are generated. This local environmental observation data is collected from the vehicle and includes at least real-time vehicle images captured by the vehicle's camera and point cloud data collected by the lidar. The point cloud data is a set of three-dimensional points obtained by the lidar within a unit of time, measuring distances to the surrounding space. Each point contains at least three-dimensional coordinate information and reflection intensity or equivalent information. Lens distortion correction and brightness normalization are performed on the real-time vehicle images to obtain preprocessed real-time vehicle images. Time synchronization and coordinate transformation are performed on the point cloud data to obtain preprocessed point cloud data with a unified timestamp. Time synchronization is achieved by mapping the timestamps of the real-time vehicle images and the point cloud data to the unified time reference used in step two. Coordinate transformation maps the point cloud coordinates in the lidar coordinate system to the vehicle coordinate system through extrinsic parameter transformation. This extrinsic parameter transformation originates from the vehicle calibration file and is stored in matrix form.

[0051] Subsequently, the target charging pile is identified in the preprocessed real-time vehicle image, and the target charging pile image observations are obtained. Target charging pile identification employs a target detection network to infer from the preprocessed real-time vehicle image, outputting the target charging pile's bounding box and target category label. This target detection network is an end-to-end target detection model based on a convolutional neural network or its equivalent structure, taking an image as input and outputting a set of target bounding boxes and corresponding confidence scores. The target bounding box with the highest confidence score and a target category label matching the target charging pile's auxiliary description is determined as the target charging pile bounding box. To obtain stable features for pose estimation, a set of charging pile key points is further extracted within the area covered by the target charging pile bounding box. This set of key points can be output by a key point detection network or detected by a corner detection algorithm within the target charging pile bounding box area, and the pixel coordinates of each key point are written into the target charging pile image observations.

[0052] In one feasible example, the input to the object detection network is a preprocessed real-time image of the vehicle, and the output includes at least a set of target bounding boxes, a corresponding confidence set, and target category labels. The input to the keypoint detection network is a preprocessed real-time image of the vehicle or the area covered by the target charging pile bounding box, and the output is the pixel coordinates of the charging pile keypoint set. The input to the semantic segmentation is a real-time image of the vehicle, and the output is a pixel-level category map, which is used to project and update the obstacle region to the grid cells. The above deep learning model can be trained by annotating real-time images of vehicles in a charging station scenario. The training data at least covers typical categories such as charging pile appearance, parking lines, road boundaries, and obstacles, and covers dynamic interference conditions such as occlusion, reflection, rain, fog, or nighttime noise, to ensure that the output target bounding boxes, charging pile keypoint sets, and pixel-level category maps remain stable under complex conditions. The above example is for illustrating feasible implementations and does not limit the specific form of the network structure and training strategy.

[0053] When the target charging pile record contains a visible identifier that is a two-dimensional coded identifier, two-dimensional code recognition is performed on the bounding box region of the target charging pile to obtain the two-dimensional coded identifier content. The two-dimensional code recognition uses image binarization, locator detection, and error correction decoding to obtain the two-dimensional coded identifier content, and then performs a consistency check between the two-dimensional coded identifier content and the target charging pile record. The check is achieved by comparing the charging pile number field or its equivalent field in the two-dimensional coded identifier content with the auxiliary description field in the target charging pile record. When the consistency check passes, the pixel region where the two-dimensional coded identifier is located is used as a high-confidence feature region for pose estimation, and its pixel coordinate range is recorded.

[0054] After obtaining the target charging pile image observations, spatial localization of the target charging pile based on point cloud data is performed to obtain the target charging pile spatial observations. The preprocessed point cloud data is clustered according to spatial neighborhood relationships to obtain a point cloud cluster set; the clustering adopts region growing clustering based on Euclidean distance or its equivalent clustering method, and points within a distance threshold are grouped into the same cluster. For each point cloud cluster, its geometric dimensions, point density, and principal direction are calculated, and point cloud clusters that meet the requirements of geometric dimensions within the charging pile's external scale and point density higher than the threshold are selected as candidate charging pile clusters.

[0055] In one feasible example, the distance threshold is used to balance over-splitting and over-merging of clusters. It can be set in relation to the density of the LiDAR point cloud and the distance range of the target charging pile: increasing the distance threshold appropriately when the target distance is greater or the point cloud is sparser avoids splitting the same charging pile into multiple clusters; decreasing the distance threshold appropriately when the target distance is closer or the point cloud is denser avoids merging adjacent objects into the same cluster. The geometric dimensions within the charging pile's outline scale range and the point density threshold can be statistically obtained from offline sampled charging pile point cloud fragments. These ranges and thresholds are then written into a joint reference dataset in a configuration format and fixed in a fixed version, ensuring consistent and controllable filtering rules across different sites and LiDAR configurations.

[0056] For each candidate charging pile cluster, line segment fitting is performed to extract the set of vertical boundary line segments for the charging pile. The line segment fitting uses a random sampling consensus method to fit a straight line model in the point cloud cluster, eliminating outliers and retaining inliers to form line segments. For the set of vertical boundary line segments, two parallel boundary line segment pairs with a spacing satisfying the width range of the charging pile are calculated. The endpoints of these line segment pairs determine the spatial position of the target charging pile in the vehicle coordinate system. To avoid omitting explanations for newly introduced terms, the endpoints refer to the two boundary points of the line segment in three-dimensional space, and their coordinates are calculated from the extreme points of the inliers in the principal direction. The spatial center point of the target charging pile, calculated from the endpoints of the line segment pairs, is denoted as... ,in This represents the three-dimensional coordinate components of the target charging pile's spatial center point in the vehicle coordinate system.

[0057] Subsequently, relative pose estimation of the target charging pile is performed to obtain its relative pose. The relative pose refers to the combination of the vehicle's relative position and relative orientation relative to the target charging pile in the vehicle coordinate system. Relative pose estimation is achieved through image observation and spatial observation unification: a correspondence is established between the pixel coordinates of key points of the charging pile in the image observation and the geometric feature points of the charging pile in the spatial observation, resulting in two-dimensional and three-dimensional corresponding point sets; where the two-dimensional points are pixel coordinates and the three-dimensional points are spatial coordinates in the vehicle coordinate system. Based on the two-dimensional and three-dimensional corresponding point sets, a perspective three-point or higher solution algorithm is used to calculate the pose of the target charging pile in the camera coordinate system; this perspective three-point or higher solution algorithm takes the camera intrinsic parameter matrix and the two-dimensional and three-dimensional corresponding point sets as input and outputs a rotation matrix. With translation vector ,in This indicates the rotational relationship between the target charging station coordinate system and the camera coordinate system. This indicates the position of the origin of the target charging pile's coordinate system in the camera's coordinate system. Then, using an extrinsic parameter transformation from the camera coordinate system to the vehicle coordinate system, the... and Transformed into the relative pose of the target charging pile in the vehicle coordinate system, the following is obtained: and .in, The translation components are the same as those mentioned above. A consistency check is performed; the consistency check is achieved by comparing whether the Euclidean distance between the two in the vehicle coordinate system is less than a preset distance threshold, which is stored in the form of a parameter and read from the joint reference dataset.

[0058] After obtaining the relative pose of the target charging pile, the target parking position is calculated, and a planned start point and a planned end point are generated. The target parking position refers to the location point where the vehicle, after completing its entry into the charging station, forms a specified relative relationship with the target charging pile. The target charging pile's spatial center point is used as the reference point. Given the target charging station's orientation as input, calculate the unit vector of the normal direction directly in front of the target charging station. And calculate the target parking position based on this. The aforementioned can be The orientation component is obtained by extraction, or by cross product of the principal direction and vertical direction of point cloud line segments and normalization. The target docking position is calculated using the following formula: ; Where d is the reserved distance, used to characterize the expected distance between the vehicle's parking point and the target charging station; d is a configurable parameter whose value source is registered in the joint reference dataset. If an example is needed, d can be set to, for example, a number of meters instead of a fixed value. The projection point of the vehicle's current position in the vehicle coordinate system is recorded as the planning starting point. ,Will The corresponding point in the site's geographic coordinate system is recorded as the planning endpoint. The vehicle's current location is derived from the global vehicle positioning result in step two, and then mapped to the site's geographic structure coordinate system through the coordinate transformation determined in step one.

[0059] Subsequently, drivability constraints are constructed and a drivable raster map is generated. The drivable raster map refers to a map representation that discretizes the polygonal boundaries of drivable areas and the road network map in the site's geographic structure into raster cells and assigns a drivable label to each raster cell. The drivable raster map is generated by filling the polygonal boundaries of drivable areas into the raster coordinate system and labeling drivable cells, while labeling impassable areas as obstacle cells. When the site's geographic structure includes a road network map, the road network map is projected onto the raster coordinate system, and raster cells near the road centerline are assigned a lower drivability cost. To introduce real-time obstacle constraints, semantic segmentation is performed on the preprocessed real-time vehicle images to identify drivable and obstacle areas. Semantic segmentation refers to a deep learning inference process that classifies image pixels, outputting a pixel-level category map of the same size as the image. The pixel-level category map is mapped to the raster coordinate system through camera extrinsic parameters, and raster cells labeled as obstacles are updated to obstacle cells, with the update timestamp of the obstacle cells recorded.

[0060] After the traversable grid map is constructed, a path search is performed to obtain the initial path. The path search uses a bidirectional A-path search algorithm to search the traversable grid map from the planned starting point. Searched for the planned destination The collision-free path. The bidirectional A-path search algorithm refers to a search algorithm that simultaneously performs A-path searches from both the planning start point and the planning end point, and merges the two search front edges when they meet to obtain the complete path. The evaluation function of each node n in the A-path search algorithm is defined as: f(n) = g(n) + h(n); Where g(n) represents the cumulative cost from the planning start point to node n, which is obtained by summing the grid movement costs; h(n) represents the heuristic cost from node n to the planning end point, which is calculated from the Euclidean distance or Manhattan distance between node n and the planning end point. (Bidirectional A) In the path search algorithm, evaluation functions are defined for forward search and backward search, respectively. and Dynamic weights are introduced to adjust the search balance on both sides. The dynamic weights are calculated as a combination of the distance from the node to the target and the number of search iterations: ; in, Let α represent the weights at the k-th iteration, and let α represent the weight coefficients of the distance term and the iteration term. , This represents the distance from the current forward search front node to the planned endpoint. This represents the distance from the current leading edge node of the reverse search to the planning starting point, and k represents the current iteration number. This represents the iterative baseline constant used for normalization. Based on The heuristic costs for forward search and backward search are defined as follows: ； ; The evaluation functions of both sides of the search are updated accordingly. To reduce the number of invalid expanded nodes, a greedy pruning strategy is introduced for expanded nodes: when the evaluation function value of a candidate node exceeds the upper bound of the currently known optimal meeting cost, the candidate node is directly discarded and not added to the open set; the open set refers to the set structure in the A-path search algorithm that stores nodes to be expanded, and the discard operation is implemented by not enqueuing. The node sequence output by the bidirectional A-path search algorithm constitutes the initial path, and the node sequence is saved as the initial path point set in raster coordinate order.

[0061] After obtaining the initial set of path points, trajectory smoothing is performed on the initial path to generate a smoothed path. The trajectory smoothing process uses a cubic Bézier curve to piecewise fit the initial path. The cubic Bézier curve is a parametric curve determined by four control points, and its expression is: ; ; in, and These are the start and end control points of the segmented path, respectively. and As an intermediate control point, Here are the coordinates of the point on the curve corresponding to parameter t. Control points are selected by sampling anchor points from the initial path point set according to curvature variation or a fixed step size, and then setting them based on the tangential direction of adjacent anchor points. and The implementation involves writing the control point set and segment range of each segment into the smooth path record, so that the smooth path can be uniquely reproduced from the control point set.

[0062] Finally, the relative pose of the smoothed path and the target charging station is checked for consistency, and the result of step three is output. The consistency check includes endpoint consistency check and attitude consistency check: endpoint consistency check compares the endpoint of the smoothed path with the target docking position. Distance is achieved within the same coordinate system; attitude consistency is verified by comparing the tangential direction at the end of the smooth path with the unit vector of the normal direction of the target charging pile. Whether the included angle meets the preset angle threshold is determined. The relative pose of the vehicle to the target charging station and the target parking position, which have passed the consistency check, are then considered. The smooth path record is also recorded as the output of step three.

[0063] Step 4: Path following and precise alignment of the endpoint, followed by result correction and data write-back update; First, path following control is executed, and vehicle control variables are generated. The segmented control point set from the smooth path record is read, and the desired trajectory point sequence is generated segment by segment according to the cubic Bézier curve expression. The desired trajectory point sequence consists of trajectory point coordinates sampled according to parameter t, with the sampling step size determined by the vehicle control cycle. The current vehicle state is acquired as the following feedback variable. The current vehicle state is jointly determined by the vehicle global positioning result from step two and the output of the inertial measurement unit acquired in real time at the vehicle end, and includes at least the vehicle's current position, current orientation, and speed. For each control cycle, the lateral deviation and heading deviation from the vehicle's current position to the desired trajectory point sequence are calculated. The lateral deviation is the normal distance from the vehicle's current position to the nearest desired trajectory point, and the heading deviation is the angle between the vehicle's current orientation and the tangential direction at the nearest desired trajectory point. Based on the lateral and heading deviations, the vehicle steering control variable and speed control variable are calculated and executed.

[0064] The vehicle control parameters are calculated using a path following control algorithm, which may employ proportional-integral-derivative (PID) control or model predictive control. Taking PID as an example, the lateral deviation is defined as... The heading deviation is Steering control quantity is Then the steering control quantity satisfies: ; in, This is the lateral deviation proportionality coefficient. The differential coefficient for the lateral deviation. This is the heading deviation coefficient. The lateral deviation rate of change over time is calculated from the lateral deviation difference between adjacent control cycles. The speed control quantity is adaptively adjusted based on the path curvature and obstacle constraints: the local curvature is calculated from the desired trajectory point sequence; the speed control quantity is reduced for road segments with high curvature and increased for road segments with low curvature; when the real-time obstacle constraint indicates the presence of obstacle grid cells ahead, the speed control quantity is reduced below the safety threshold and a replanning decision is triggered.

[0065] Online safety constraints and replanning decisions are performed during path following. In each control cycle, local environmental observation data is acquired and the passable grid map is updated. This local environmental observation data includes real-time vehicle images and preprocessed point cloud data. The real-time images are semantically segmented to obtain pixel-level category maps, which are then projected to update grid cells. The point cloud data is spatially projected to map obstacle point sets to the grid coordinate system and update obstacle cells. Collision prediction is performed on the updated passable grid map: Trajectory points within a certain number of prediction steps are checked along the current desired trajectory point sequence to see if they fall into obstacle cells. If any trajectory points fall into obstacle cells, the current smoothed path is determined to be impassable. Step three (path search and trajectory smoothing) is re-executed to generate a new smoothed path record, which replaces the current smoothed path record to continue path following control. The number of prediction steps is a configurable parameter, sourced from station guidance photos and the geographic information joint reference dataset.

[0066] In a feasible example, the prediction step count is set in association with the vehicle control cycle and vehicle speed, ensuring that the forward collision prediction covers the desired trajectory segment length corresponding to at least one safe braking distance: the prediction step count is increased to extend the forward sight distance as vehicle speed increases; the prediction step count is decreased to reduce computational overhead and increase update frequency when vehicle speed is low and the environment is confined. The prediction step count setting and its source identifier are written into a joint reference dataset and fixed to a version, thereby making collision predictions reproducible across different site environments.

[0067] When the vehicle reaches the neighborhood of the target parking spot, it switches to end-point precise alignment control. The neighborhood of the target parking spot is a spatial region centered on the target parking spot with a preset distance threshold as its radius. This preset distance threshold is a configurable parameter and is registered in the site guidance photo and geographic information joint reference dataset. After entering the neighborhood of the target parking spot, real-time vehicle images and preprocessed point cloud data are continuously acquired, and target charging pile identification and relative pose estimation are performed at a higher frequency: in the real-time vehicle images, a target detection network outputs the target charging pile bounding box and target category label, and extracts the charging pile key point set within the bounding box area; in the point cloud data, clustering and line segment fitting are used to obtain the spatial center point of the target charging pile. and the unit vector of the normal direction of the target charging pile The end-effector alignment error is calculated based on the updated relative pose of the target charging station, and an end-effector alignment control quantity is generated based on the alignment error.

[0068] The end-point alignment error includes at least position error and orientation error. Position error is defined as the difference vector between the vehicle's current position and the target parking position. : ; in, This represents the coordinate vector of the vehicle's current location in the site's geographic coordinate system. Let be the coordinate vector of the target parking position in the site's geographic coordinate system. Orientation error is defined as the unit vector of the vehicle's current orientation. Unit vector of the normal direction of the target charging station Angle error between : ; in, This represents the vector dot product. The end-effector alignment control variable is generated based on the position error and orientation error, and the control objective is to make the end-effector alignment control variable generate the vector dot product. Convergence and Convergence. The steering control and speed control quantities during the end-alignment phase are calculated using a proportional-integral-derivative (PID) control algorithm. The inputs to the PLD algorithm are the lateral component of the position error, the longitudinal component of the position error, and the orientation error, respectively. The outputs are the steering control and speed control quantities, which are issued and executed in each control cycle.

[0069] During the precise alignment process at the end point, a correction calculation for the target charging pile location description is performed. If a geographic anchor point for the target charging pile exists in the target charging pile record, this geographic anchor point is aligned with the spatial center point of the target charging pile estimated in the precise alignment stage, and a correction amount is calculated. If the geographic anchor point for the target charging pile is in an unconfirmed state, the spatial center point of the target charging pile estimated in the precise alignment stage is transformed to the site geographic structure coordinate system and used as a candidate value for the geographic anchor point. The transformation is achieved by a coordinate transformation from the vehicle coordinate system to the site geographic structure coordinate system, which is determined jointly by the coordinate alignment in step one and the vehicle global positioning result in step two.

[0070] To suppress the impact of instantaneous measurement noise on the candidate geographical anchor points of the target charging piles, a count correction is performed on the candidate geographical anchor points. The count correction maintains historical anchor point values. With counter c. When the current candidate anchor value Compared with historical anchor values When the Euclidean distance does not exceed the first threshold, Write the historical anchor value and increment the counter; discard the value when the Euclidean distance exceeds the first threshold. Then reset the counter to zero. When the counter reaches the second threshold, output the historical anchor value as the stable anchor value. Both the first and second thresholds mentioned above are configurable parameters, and their sources are registered in the site guidance photos and geographic information joint reference dataset. For example, the first threshold can be described as a distance threshold on the order of several meters, and the second threshold can be described as a counting threshold for several consecutive, consistent frames, without being written with fixed values.

[0071] In one feasible example, a first threshold is used to suppress jumps in candidate anchor point values ​​caused by instantaneous measurement noise. Its magnitude can be matched with the upper bound of the "coordinate transformation error" from the vehicle coordinate system to the site geographic structure coordinate system: when the vehicle global positioning result has greater noise or there are still some viewpoint changes during the end-point precise alignment stage, the first threshold is increased to avoid frequent zeroing; when the vehicle global positioning result is more stable and the end-point precise alignment converges sufficiently, the first threshold is decreased to improve the accuracy of stable anchor point values. A second threshold is used to suppress short-term consistency caused by occasional abnormal frames. Its magnitude can be matched with the control cycle of the end-point precise alignment stage and the target charging pile detection update frequency: the higher the update frequency, the larger the second threshold can be to improve stability; the lower the update frequency, the smaller the second threshold can be to shorten the confirmation time. The above example is only for illustrating the parameter magnitude and feasible implementation methods and does not limit the specific values ​​of the first and second thresholds.

[0072] Finally, perform a data write-back update and complete the versioning record. Set the stable anchor value. Write the site guidance photo, the target charging pile geographic anchor point field of the geographic information joint reference dataset, and the source identifier of the anchor point value. The source identifier includes at least: the site identifier that triggered this write-back update, the unified timestamp of the vehicle observation frame, the real-time image identifier of the vehicle used for estimation, the point cloud data identifier used for estimation, and the identifier of the perspective three-point or higher solution algorithm and counting correction parameters used. If the target charging pile record was originally in a pending confirmation state, its status is updated to confirmed, and a confirmation timestamp and confirmation version field are written; if the target charging pile record already has a target charging pile geographic anchor point, the original anchor point value is saved as a historical version and a version number is written, and then the stable anchor point value is written as the new version anchor point value. After completing the write-back update, the relative pose of the target charging pile in the end precise alignment stage, the target parking position, and the key consistency verification results in the execution process are written into the output record of step four, forming a traceable execution result archive.

[0073] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0074] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0075] Those skilled in the art will recognize that the modules 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 inventive 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 implementation should not be considered beyond the scope of this application.

[0076] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0078] In conclusion, 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 precise navigation of charging piles based on site guidance photos and geographic information, characterized in that, include: Construct a joint reference dataset of site guide photos and geographic information, write the site guide photos and site geographic structure into the joint reference dataset, and create site index entries; Acquire vehicle observation frames, determine candidate sets based on candidate station sorting list, and perform cross-frame geometric consistency competition locking for candidate sets: recursively calculate the cumulative advantage potential value of each candidate station within the observation window, and introduce observation gating coefficient to suppress abnormal frames; drive observation window update based on congestion and clustering degree, and output target stations and control replacement by combining entry boundary, maintenance boundary and inertial potential value. Acquire local environmental observation data to generate local observation frames, identify target charging piles and estimate the relative pose of target charging piles, calculate target parking positions and generate smooth path records; Based on the smooth path record, the system performs path following and precise end-point alignment control to generate vehicle control quantities; it corrects the geographic anchor points of the target charging piles and obtains stable anchor point values, writes the stable anchor points back to the site guidance photos, and records geographic information in conjunction with the reference dataset in a versioned manner.

2. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 1, characterized in that: By collecting and uploading site guidance photos, reading EXIF ​​data or collecting sensor information, a photo metadata record containing the collection time, collection location, and collection orientation is generated. When the location is missing, the site guidance photo is used as the query image to perform image retrieval in a geographically tagged reference image library to obtain the candidate location range and write it into the metadata record. After performing lens distortion correction and brightness normalization preprocessing on the site guidance photos, local and global features are extracted and written into the joint reference dataset. For local feature matching, a random sampling consistency estimation geometric transformation model is used to eliminate outliers and obtain inlier matching pairs. The model type, threshold and number of inliers are then associated with the database.

3. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 2, characterized in that: The system calls map data interfaces or imports map data to obtain road networks, traversable areas, and charging station points of interest (POIs). It then parses and generates road network maps, traversable area boundaries, and POI coordinates, which are written into a joint reference dataset. When the geographic information lacks station-level locations, the geographic anchor point of the target charging station is marked as pending confirmation, and the station-level location source is recorded. The joint reference dataset is then time-consistent and coordinate-consistent, with version fields, coordinate system identifiers, and projection parameter identifiers written into it. Coordinate transformation is performed according to the projection parameters, preserving local coordinates within the station and establishing a connection with the station identifier. After detecting faces and license plate areas in the station guidance photos and performing occlusion or blurring based on bounding boxes, a station index entry is generated for each station, containing a station identifier, photo set identifier, geographic structure version field, visual feature summary, and source identifier. This entry is then linked to the dataset for retrieval.

4. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 3, characterized in that: Real-time images and positioning information are acquired at the vehicle end. Distortion correction and brightness normalization preprocessing are performed on the vehicle images, and local and global features are extracted. The preprocessed images, observation visual features, and positioning information are mapped to the same time base according to the timestamp to complete the alignment and generate vehicle observation frames.

5. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 4, characterized in that: Multiple candidate stations are extracted from the sorted list to form a candidate set. A variable-length observation window is constructed using the last frame. The window length is driven by the crowding and clustering degree calculated from the candidate score interval sequence and the number of leader switching within the window, and is limited to a preset range. For each candidate station and each frame within the window, the local features of the vehicle observation frame are matched with the local features of the station guidance photo associated with the candidate station to obtain a set of candidate matching pairs. The inter-frame geometric transformation model is estimated using random sampling consistency and outliers are removed to obtain a set of inliers. Based on this, the inlier size, mean and fluctuation of reprojection error residuals, inlier convex hull coverage, and geometric model drift of adjacent frames are calculated. The observation gating coefficient is calculated using the acceleration and angular velocity output by the inertial measurement unit to attenuate abnormal frames.

6. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 5, characterized in that: Convergence indices are constructed based on the convergence amount and fluctuation suppression term of the residuals as they advance with each frame. The size, coverage, convergence, stability, and their coupling are mapped to the injection amount per frame. The cumulative advantage potential of each candidate site is recursively calculated based on the gated weighted injection amount within the window. The injection rate is pruned in conjunction with the window length to adapt to window expansion and contraction. At the same time, entry and retention boundaries are constructed based on the complexity of the candidate set, and the inertial potential value obtained by recursion from the leading indicator is maintained for each candidate. At each time step, the primary candidate and shadow candidate are selected and the advantage interval is calculated. When the entry boundary and inertial threshold are met, the target site is output. After output, replacement is only performed when the shadow candidate meets the retention boundary and inertial threshold. At other times, the output is maintained and the potential value, inertial potential value, and window length are continuously updated.

7. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 6, characterized in that: After obtaining the target site, its geographical structure is read. GNSS geographical coordinates are read from the vehicle observation frame as the vehicle's coarse geographical coordinates. The coarse geographical coordinates are then projected onto the nearest roadside in conjunction with the road network map. Based on the connectivity with the boundary of the target site, an accessibility check is performed. If the check fails, the next candidate site is switched according to the candidate site sorting list and the check is repeated until the check passes. After passing through, the global position and orientation of the vehicle are solved by fusing GNSS and IMU outputs using extended Kalman filtering in the projected coordinate system, and the position is projected onto the boundary of the passable area for map constraint correction; finally, the location description of the target charging pile is written according to whether the geographical anchor point of the target charging pile exists. When the anchor point is to be confirmed, the appearance category and number visible identifier obtained by the site guidance photo are written as auxiliary description and the target charging pile record is generated.

8. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 7, characterized in that: The system acquires vehicle camera images and LiDAR point clouds, synchronizes timestamps, and performs coordinate transformation. A target detection network is used to obtain the bounding box of the target charging pile in the image and extracts key points. On the point cloud side, candidate pile clusters are selected by neighborhood clustering, and vertical boundary line segments are extracted using random sampling consistency line segment fitting to calculate the pile center point position. A two-dimensional and three-dimensional correspondence is established based on key point pixels and point cloud geometric feature points. The pose is obtained by solving at least three perspective points and transformed to the vehicle coordinate system using extrinsic parameters, then checked for consistency with the point cloud center point. The normal unit vector is calculated from the pile center point and orientation, and the parking position is obtained according to a reserved distance. The vehicle's current position is mapped to the planning start point, and the parking position to the planning end point. The passable area and road network are discretized to generate a raster map, which is then updated by fusing semantic segmentation to obtain obstacle raster data. A bidirectional A-star search with dynamic weights and pruning is used to generate an initial path point set. A smooth path is further generated using cubic Bezier piecewise fitting, and the relative pose, parking position, and smooth path record are output after consistency verification of the end point and pose threshold.

9. The method for precise navigation of charging piles based on site guidance photos and geographic information according to claim 8, characterized in that: The system reads the segmented control points from the smooth path record and samples them using a cubic Bézier curve to generate the desired trajectory point sequence. It obtains the vehicle's current position, orientation, and speed as feedback parameters. For each control cycle, it calculates the lateral and heading deviations and generates steering and speed control parameters accordingly, which are then executed. The speed control parameters are adaptively adjusted based on the trajectory curvature and obstacle grid constraints. During the following process, it continuously updates the passable grid map using semantic segmentation and point cloud projection. It performs forward collision prediction along the desired trajectory. When a trajectory point falls into an obstacle cell, it triggers replanning and replaces the current path with a new smooth path record. The vehicle enters the neighborhood of the target parking position. The frequency of target charging pile detection, key point extraction, and point cloud clustering line segment fitting is increased to update the pile center point and normal vector. The position error and orientation error are calculated and the end alignment control quantity is generated. In the end alignment stage, the estimated pile spatial center point is made consistent with the geographic anchor point of the target charging pile. When the anchor point is to be confirmed, the pile center point is transformed from the vehicle coordinate system to the site coordinate system as the anchor point candidate value. The candidate value is then subjected to a two-level threshold-based counting correction to output a stable anchor point value. The stable anchor point value and its source identifier are written into the anchor point field of the joint reference dataset. The anchor point version and confirmation status are written into the timestamp and version fields.

Images