A parking method based on semantic topological map

Abstract

The application discloses a parking method based on a semantic topological map, and the method comprises the following steps: constructing a semantic topological map; in response to a trigger event of entering a satellite signal-free area, determining an initial pose of a vehicle in the semantic topological map through multi-modal feature recognition and fusion positioning; screening available parking spaces based on vehicle parameters, and performing physical pre-locking of the parking spaces in response to user selection; planning a semantic navigation path in combination with vehicle parameters and channel attributes, and performing navigation through segmented feature tracking and semantic steering guidance; performing parking confirmation and state resetting when the vehicle reaches a target parking space; and dynamically maintaining the semantic topological map based on crowd-sourced perception data. The application realizes accurate parking guidance in a satellite signal-free environment, avoids the problem of occupied parking spaces, and improves the adaptability and reliability of the parking process.

Description

Technical Field

[0001] This invention relates to the field of parking technology, and more particularly to a parking method based on semantic topology maps. Background Technology

[0002] Existing automatic or semi-automatic parking technologies mainly suffer from the following problems: Location source dependency error: Traditional solutions attempt to extend satellite positioning signals by signal amplification in underground parking lots or use the latitude and longitude of high-precision maps for lane-level navigation. However, the multipath effect is severe in underground environments, and satellite positioning signals are difficult to achieve centimeter-level accuracy. Furthermore, SLAM is essentially based on relative coordinates, and forcibly fusing absolute latitude and longitude will lead to coordinate system drift and computational redundancy. The problem of "what you see is not what you get" is that the existing system only marks "free" on the map and lacks a real-time physical locking mechanism from the time the user selects the space until the vehicle arrives. As a result, when the vehicle arrives at the specified latitude and longitude, it often finds that the parking space has been occupied. The lack of vehicle compatibility and the uniform parking space display did not take into account the actual size of the vehicles, turning radius and autonomous driving capabilities, resulting in the infeasibility of the planned routes. Summary of the Invention

[0003] The purpose of this invention is to disclose a parking method based on semantic topology maps, which solves the technical problems of existing parking solutions that rely on satellite positioning and lack physical locking mechanisms, resulting in inaccurate guidance and easy occupation of parking spaces.

[0004] To achieve the above objectives, this invention discloses a parking method based on a semantic topology map, comprising: In response to a trigger event of entering a region without satellite signal, an initialization process is executed to determine the initial pose of the vehicle in a pre-built semantic topology map. The semantic topology map is pre-built and stored in the cloud. The semantic topology map is a graph structure composed of nodes and edges, and does not contain the absolute latitude and longitude coordinates of the parking space. Nodes represent identifiable semantic feature points in the physical environment. Each feature point is associated with a multimodal feature template for identification and matching. Edges represent drivable channels between adjacent feature points and are associated with channel attributes. Available parking spaces are filtered based on vehicle parameters, a list of candidate parking spaces is generated and presented to the user; In response to the user's selection of a target parking space, the parking space is physically pre-locked; Starting from the current location and ending at the target parking space, a semantic navigation path is planned in the semantic topology map by combining vehicle parameters and channel attributes. During the vehicle's journey, segmented tracking and guidance are performed based on the semantic navigation path until the target parking space is reached. When the vehicle arrives at the target parking space, parking confirmation and status reset are performed.

[0005] Specifically, the initialization process includes: The system acquires trigger events and preloads the entrance layer feature library of the target parking lot based on these events; the trigger events include geofence triggers, visual feature triggers, or manual triggers. During the vehicle's entry process, environmental data is collected in real time through onboard sensors, which include at least cameras and lidar. The real-time collected environmental data is compared in parallel with all pre-loaded entry layer feature libraries to identify the current entry layer and specific feature points. If a match is successful, the local relative pose of the vehicle relative to the identified feature point is calculated, and the global standard pose of the feature point is obtained from the cloud. The vehicle's pose in the global topological coordinate system is calculated through coordinate transformation to complete the initial positioning. If no match is found after a preset scanning time, the downgrade initialization process is executed. The downgrade initialization process includes: extracting regional features based on the surround-view camera to perform regional matching, taking the center of the matched region as the initial position, and providing guidance if accurate positioning cannot be restored within a preset distance.

[0006] Specifically, the parallel comparison employs multimodal feature matching, which includes: A convolutional neural network is used to extract global feature vectors from images, and a point cloud deep learning network is used to extract geometric feature vectors from point clouds. The fusion weight α is dynamically determined based on ambient lighting and texture richness, and the two feature vectors are fused into a joint feature vector. α is calculated based on the normalized average brightness L and the normalized average gradient magnitude G according to the formula α = 0.5 × min(1, max(0, 2L)). Calculate using 0.5))+0.5×min(1,G / 0.3); The cosine similarity between the fused feature vector and the multimodal feature templates of each feature point in the preloaded feature library is calculated. If the similarity exceeds the preset threshold, the recognition is considered successful.

[0007] Specifically, the filtering of available parking spaces based on vehicle parameters includes: Obtain the vehicle's parameters, which include at least the vehicle length, width, height, minimum turning radius, and wheelbase. Read the attribute data of all parking spaces from the semantic map in the cloud, including parking space size and parking space type; The vehicle's parameters are compared with the parking space attributes one by one, and parking spaces whose dimensions do not meet the vehicle's parking requirements are automatically eliminated. Generate a candidate list of the remaining eligible parking spaces and output it.

[0008] Specifically, the execution of the physical pre-locking of the parking space includes: Receive the user's instruction to select a target parking space; The cloud sends an up command to the smart lock associated with the target parking space and waits for the lock to return an up confirmation signal. If confirmation is received, a time-limited temporary locking token is generated, and the parking space is marked as reserved in the cloud. If locking fails, the user is prompted to select again.

[0009] Specifically, the semantic navigation path planning includes: Read the attribute data of all channels in the topology map. The attributes include channel width, channel height, minimum turning radius requirement, and whether it is a one-way street. When using a graph search algorithm for path search, the channel attribute requirements are matched with the vehicle parameters. Specifically, in the cost function, when evaluating a certain edge, the channel width of that edge is compared with the vehicle width plus a safety margin. If it exceeds the limit, the passage cost of that edge is set to infinity to automatically avoid it. In the topological subgraph formed by feasible routes, search for the shortest path from the starting point to the target parking space; The searched path is represented as an ordered sequence of key semantic nodes, each node being associated with a visual feature template, relative pose relationship, and driving action.

[0010] Specifically, the segmented tracking and guidance includes: Guided by the current target node, the following steps are executed repeatedly: identify the current target node, calculate the relative pose, generate local control commands, determine the node arrival, until the entrance of the channel where the target parking space is located is reached; For intersections or forks, the system relies on preset semantic features to trigger turning actions. If no turning guidance features are identified, it triggers local replanning or stops and waits. If the vehicle deviates from the planned path, a new feasible path will be searched and connected from the current location.

[0011] Specifically, the docking confirmation and status reset include: When a vehicle approaches the target parking space, the cloud-controlled indicator light above the parking space switches from the first state to the second state to indicate that the space is reserved for this vehicle. When a vehicle parks, the on-board unit sends a parking signal to the cloud, which then controls the parking lock to descend and turn off the indicator light. If an inconsistency is detected, a fault is reported and the parking space is marked as malfunctioning, while the driver is prompted to confirm safety.

[0012] Specifically, the semantic topology map extracts unique visual feature combinations around each parking space by collecting vehicle scans or crowdsourced data to form a feature fingerprint database. The unique visual feature combinations include at least column numbers, wall billboard textures, fire hydrant locations, and special ground markings. The map is stored in the data format {parking space ID, parking space attribute, associated feature list}, where the associated feature list includes feature point IDs and the relative pose of the parking space with respect to the feature point. The feature vectors stored in the cloud map are processed by dimensionality reduction or hash encoding and an index structure is established.

[0013] Furthermore, it also includes the step of dynamically maintaining and updating feature points in the semantic topology map based on crowdsourced perception data from multiple vehicle trips, specifically including: If, under high-confidence visual relocalization, the observed actual features do not match the expectations and the new features are stably present, a feature change event is triggered. The image data of the new features and their relative anchor point information in the topology map are anonymized and uploaded to the cloud; The cloud system collects new features reported by multiple vehicles in the same topology context. When the number of observations exceeds the threshold and the original features have not been observed for a long time, the new features are entered into the database to replace the old features; if it is a temporary occlusion, only the new features are recorded.

[0014] The beneficial effects of this invention are as follows: By constructing a semantic topological map that does not include the latitude and longitude coordinates of parking spaces, this invention represents the location of parking spaces as relative pose relationships with environmental feature points. Combined with multimodal feature recognition and fusion positioning technology, it achieves accurate initial positioning and navigation in environments without satellite signals. Simultaneously, through a physical pre-locking mechanism that integrates smart parking locks and cloud collaboration, the parking space is locked in real-time from the time the user selects the space until the vehicle arrives, preventing it from being occupied by others. Furthermore, through dynamic maintenance of the semantic map based on crowdsourced perception, automatic updates and optimization of map features are achieved. This invention effectively solves the reliance on satellite positioning in traditional solutions, improving the reliability of parking guidance and user experience. At the same time, through vehicle model adaptation filtering and route planning, it improves the success rate and safety of the parking process. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the parking method based on semantic topology maps of 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] In this invention, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing the invention and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0019] Furthermore, in addition to indicating direction or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain situations to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0020] Furthermore, the terms "installation," "setup," "equipped with," "connection," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0021] Furthermore, the terms "first," "second," etc., are primarily used to distinguish different devices, components, or parts (which may be the same or different in specific type and construction), and are not intended to indicate or imply the relative importance or quantity of the indicated devices, components, or parts. Unless otherwise stated, "a plurality of" means two or more.

[0022] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings.

[0023] like Figure 1 As shown, this embodiment provides a parking method based on a semantic topology map, including the following steps: Step 1: In response to a trigger event that enters a region without satellite signal, execute an initialization process to determine the vehicle's initial pose in a pre-built semantic topology map.

[0024] The semantic topology map is pre-built and stored in the cloud. It does not contain the absolute latitude and longitude coordinates of parking spaces, but is a graph structure composed of nodes and edges. Nodes represent identifiable semantic feature points in the physical environment, and each feature point is associated with a multimodal feature template for identification and matching; edges represent drivable channels between adjacent feature points and are associated with channel attributes.

[0025] The semantic topology map extracts unique visual feature combinations around each parking space by collecting vehicle scanning or crowdsourcing data, forming a feature fingerprint database. The unique visual feature combinations include at least column numbers, wall billboard textures, fire hydrant locations, and special ground markings.

[0026] The multimodal feature template includes at least visual features and geometric features, and the channel attributes include at least channel width, channel height, and turning radius requirements.

[0027] The map is stored in the following data format: {parking space ID, parking space attribute (size / type), associated feature list [{feature point ID: Feature_1, relative pose of parking space relative to feature point (Tx1, Ty1, θ1)}, {feature point ID: Feature_2, relative pose of parking space relative to feature point (Tx2, Ty2, θ2)}, ...]}.

[0028] The feature vectors stored in the cloud map are processed by dimensionality reduction or hash encoding, and an efficient index structure (such as KD-Tree or inverted index) is established to support real-time approximate nearest neighbor retrieval on the vehicle side.

[0029] In this embodiment, the initialization process includes the following sub-steps: Step 101: Obtain the trigger event and preload the entrance layer feature library of the target parking lot based on the trigger event.

[0030] The triggering event includes one of the following types: a. Geofence Trigger: The vehicle positioning module detects that the vehicle has entered the geofence range of the parking lot entrance (e.g., 50 meters around the entrance) and the satellite signal is attenuated. The system records the current GPS coordinates and preloads the semantic topology map metadata of all entrance layers of the parking lot. b. Visual feature trigger: The forward-looking camera identifies visual features of the entrance, including downhill ramps or height restriction bars, in real time and automatically activates the indoor navigation preparation process; c. Manual trigger: Receives the activation command input by the user through clicking the "Enter Indoor Navigation" button on the vehicle's infotainment screen.

[0031] In this step, upon receiving a trigger event, the system downloads the initial feature library for all entrance layers of the parking lot from the cloud based on the recorded location coordinates. The initial feature library is stored independently for each entrance layer, and each feature library contains at least the location anchor point information for the entrance area of ​​that layer. The location anchor points include QR codes on pillars, gate number plates, and wall advertisements, and each anchor point is associated with its standard pose in the global topology map.

[0032] Step 102: During the vehicle's entry process, environmental data is collected in real time through onboard sensors, including at least a camera and a lidar. After determining the entry layer, the system dynamically selects features about to enter the camera's field of view from the feature library of that layer as candidate recognition targets based on the vehicle's current driving direction (estimated by IMU / wheel speed sensor) and field of view.

[0033] Step 103: Perform parallel comparison between the real-time collected environmental data and all pre-loaded entry layer feature libraries to identify the current entry layer and specific feature points. If a match is successful, proceed to step 104; if no match is found after a preset scanning time (e.g., 5 seconds), proceed to the downgrade initialization process (step 107).

[0034] The multimodal feature matching specifically includes: First, a convolutional neural network (such as ResNet-50) is used to extract the global feature vector f from the image. img Point cloud geometric feature vectors f are extracted from point clouds using point cloud deep learning networks (such as PointNet++). pcd ; Secondly, based on the dynamic determination of the current environmental perception quality, the fusion weight α is determined, and the two feature vectors are fused into a joint feature vector: f fused =αf img +(1 α)f pcd The weight α is dynamically adjusted based on ambient lighting and texture richness. Specifically, the system calculates the normalized average brightness L and the mean normalized gradient magnitude G of the current image in real time, and determines α according to the following formula: α = 0.5 × min(1, max(0, 2L) 0.5))+0.5×min(1,G / 0.3) This formula ensures that: when the lighting is good (L>0.75) and the texture is rich (G>0.3), α approaches 1, and image features dominate; when the lighting is low (L<0.25) or the texture is poor (G<0.1), α approaches 0, and point cloud features dominate; the intermediate states transition smoothly, ensuring the robustness of the fusion. Finally, the cosine similarity between the fused feature vector and the multimodal feature templates of each feature point in the preloaded initial feature library in step 101 is calculated:

[0035] If the similarity of a feature point exceeds a preset threshold (e.g., 0.85), the recognition is successful, and the feature point ID and its position in the image are recorded; if no match is found after a preset scanning time (e.g., 5 seconds), the process proceeds to step 107 for exception handling.

[0036] Step 104: Calculate the local relative pose of the vehicle with respect to the identified feature points. If monocular vision is used, the calculation is performed in conjunction with vehicle motion information (wheel speed sensor / IMU); if multi-view vision or depth camera is used, the calculation is performed directly based on 3D information.

[0037] Step 105: Obtain the standard pose of the recognized feature in the global topology map from the cloud semantic topology map. The vehicle's pose in the global topological coordinate system is calculated using the following coordinate transformation formula. Initial positioning complete:

[0038]

[0039] Step 106: The vehicle's infotainment system displays "Indoor navigation activated," and the system enters the preset navigation mode.

[0040] In this embodiment, the preset navigation mode is an indoor navigation mode without satellite positioning signals.

[0041] Step 107: If the floor cannot be identified (e.g., no match in all feature libraries) or if consecutive matching fails, execute the downgrade initialization process: a. The vehicle slows down and activates the surround-view camera to capture images of the area around the entrance; b. Extracting regional features and performing regional matching: The vehicle activates its surround-view cameras to capture multi-view images of the surrounding environment at the entrance. The following regional features are extracted from the images: the main color tone and color distribution of the walls, the average spacing and arrangement of the pillars, the estimated width of the driveway, the ceiling height range (if a depth sensor is available), and special markings (such as "East Zone," "West Zone," etc., if recognizable). These regional features are combined to form a regional feature vector, which is then matched against pre-stored regional feature templates for each entrance area in the cloud map. These regional feature templates are generated statistically during the mapping phase by the vehicle traversing each entrance area, and include typical macroscopic attribute values ​​and allowable fluctuation ranges for that area. The matching algorithm uses weighted nearest neighbor, assigning weights to each regional feature dimension (e.g., lower weight for wall color, higher weight for pillar spacing), calculating the weighted similarity between the current feature vector and each regional template, and selecting the area with the highest similarity exceeding a preset threshold as the matching result to determine the current entrance area of ​​the vehicle (e.g., "East Entrance of B2 Level"). If no area has a similarity exceeding the threshold, the matching fails, and the process proceeds to step d. c. Use the center of the area as the initial position, and gradually correct it through visual repositioning after the vehicle enters the area, while prompting "Lower positioning accuracy, please drive slowly"; d. If, after downgrading and initialization, the vehicle still cannot regain accurate positioning through visual repositioning within a preset distance (e.g., 50 meters), the driver will be prompted with "Insufficient environmental features, please follow the instructions to leave" or guided to a manual service area.

[0042] Step 2: Filter available parking spaces based on vehicle parameters, generate a list of candidate parking spaces, and present it to the user.

[0043] In this embodiment, step 2 includes: Step 201: Obtain the vehicle parameters, which include at least the vehicle length, width, height, minimum turning radius, and wheelbase. These parameters can be automatically read through the vehicle bus or manually entered by the user in the APP.

[0044] Step 202: Read the attribute data of all parking spaces from the cloud semantic map, including parking space size (length, width, height restrictions) and parking space type (normal parking space, mechanical parking space, disabled parking space).

[0045] Step 203: Compare the vehicle parameters with the parking space attributes one by one, and automatically remove parking spaces whose dimensions do not meet the parking requirements of this vehicle.

[0046] For example, if the vehicle width is greater than the mechanical parking space width threshold, all mechanical parking spaces will be invisible; if the vehicle height is greater than the height of a height-restricted parking space, the corresponding parking space will be filtered out. The width threshold for mechanical parking spaces is pre-stored in the cloud map as part of the parking space attributes. The threshold may differ for different types of mechanical parking spaces (for example, the width threshold for lift-and-slide mechanical parking spaces is typically 1.85 meters, while for vertical loop mechanical parking spaces it may be 1.9 meters). The system reads this threshold and compares it with the vehicle's width.

[0047] Step 204: Generate a candidate list of the remaining eligible parking spaces and output it through the vehicle's infotainment system or the app interface.

[0048] Step 3: In response to the user's selection of the target parking space, perform physical pre-locking of the parking space.

[0049] In this embodiment, step 3 includes: Step 301: Receive the user's instruction to select a target parking space.

[0050] Step 302: The cloud sends a raise command to the smart lock associated with the target parking space and waits for the lock to return a confirmation signal that it has been raised. If a confirmation message is received, a validity period of T is generated. lock The system issues a temporary lock token and marks the parking space as "reserved" in the cloud. If the parking lock fails to rise or times out without response, the lock is deemed to have failed, and the user is prompted to select a different location.

[0051] Step 4: Starting from the current location and ending at the target parking space, plan a semantic navigation path in the semantic topology map by combining vehicle parameters and channel attributes.

[0052] In this embodiment, step 4 includes: Step 401: Before performing route planning, read the attribute data of all channels in the topology map. The attributes include channel width, channel height, minimum turning radius requirement, and whether it is a one-way street.

[0053] Step 402: When using a graph search algorithm (such as A* algorithm or Dijkstra algorithm) to perform path search, match the channel attribute requirements with the vehicle parameters.

[0054] Specifically, in the cost function of the path search algorithm, when evaluating whether an edge can be added to the current path, the channel width of that edge is first compared with the vehicle's parameters (vehicle width + safety margin). The safety margin is preset based on the width of the vehicle's rearview mirrors, driver operating habits, and safety redundancy requirements, and is generally 0.3-0.5 meters. The specific value can be manually set by the user on the central control screen. If the vehicle's parameters exceed the channel width (for example, a channel width is less than the vehicle width + 0.4 meters), the passage cost of that edge is set to infinity, causing the search algorithm to automatically avoid that channel.

[0055] Step 403: In the topological subgraph composed of feasible channels, use the A* algorithm or Dijkstra's algorithm to search for the shortest path from the starting point to the target parking space.

[0056] Step 404: Represent the searched path as a series of ordered key semantic nodes.

[0057] For example: Starting point (current feature point A) → Passing through feature point B (red pillar) → Passing through feature point C (right side of fire hydrant) → Passing through feature point D (before speed bump) → Target parking space P-088. Each node stores in the cloud map: the node's visual feature template (for identification), the node's relative pose to adjacent nodes, and the driving action to be performed when reaching the node (such as "go straight", "turn left", "turn right", "decelerate").

[0058] Step 5: During the vehicle's journey, segmented tracking and guidance are performed based on the semantic navigation path until the target parking space is reached.

[0059] In this embodiment, step 5 includes: Step 501: Segmented feature tracking.

[0060] Guided by the current target node, perform the following operations in a loop: a. Current target node identification: Using the same multimodal feature extraction and fusion method as in step 103, the feature vector after fusing the real-time acquired image and point cloud is compared with the multimodal feature template of the current target node sent from the cloud. If the similarity exceeds the threshold, the identification is successful. b. Relative pose calculation: Once the target node is identified, calculate the precise pose of the vehicle relative to that node. ; c. Local control command generation: Based on the current vehicle's pose deviation (including lateral deviation, heading angle deviation and distance) relative to the target node, a pure tracking control algorithm or a model predictive control algorithm is used to calculate the desired steering wheel angle and target vehicle speed, and generate low-level control commands (including steering, throttle, and braking signals) to make the vehicle approach the target node along a smooth trajectory. d. Node arrival determination: When the distance between the vehicle and the target node is less than a preset threshold (e.g., 1 meter) and the vehicle's orientation meets the passage requirements, it is determined that "the current node has been reached", and the target node is updated to the next node in the sequence; e. Loop until the target parking space: Repeat the above process until you reach the entrance of the passage where the target parking space is located.

[0061] Step 502, Semantic redirection guidance.

[0062] For intersections or forks, a "semantic guidance" approach is used instead of a "coordinate guidance" approach. For example, when a left turn is required, the system does not rely on latitude and longitude coordinates to determine the "time to turn left," but instead relies on visual features: "Prepare to turn left when a fire hydrant on the right is detected ahead; complete the left turn and enter the target lane when a speed bump on the left is detected ahead." If the vehicle does not recognize the steering guidance features at the expected location, a local replanning or a stop waiting is triggered.

[0063] Step 503: Real-time deviation correction and local replanning.

[0064] If the vehicle deviates from the planned path during operation (e.g., fails to pass feature point B as expected), the system will execute the following: a. Deviation detection: Visual relocalization is used to detect discrepancies between the current feature point and the expected node in the planned path; b. Local replanning: Starting from the current location and ending at the target parking space, search for feasible paths again in the local topology subgraph; c. Path continuation: Connect the starting point of the new path with the remaining end point of the original path in the topology and continue navigation.

[0065] Step 6: When the vehicle arrives at the target parking space, perform parking confirmation and status reset.

[0066] In this embodiment, step 6 includes: Step 601: When the vehicle approaches the target parking space through visual repositioning, the smart indicator light (red / green light) installed above the parking space switches from red (occupied) to green (exclusive to this vehicle) under cloud control.

[0067] Step 602, Parking Reset: When a vehicle parks, the on-board unit sends a parking signal to the cloud. The cloud controls the parking lock to descend, and the indicator light turns off when the parking lock is fully descended. If an inconsistency is detected (e.g., the parking lock is descended but the light is still green), a fault is reported to the cloud. After receiving the fault report, the cloud marks the parking space as "equipment abnormal" and unavailable. At the same time, if the vehicle has not yet completed parking, the system prompts the driver "parking lock status abnormal, please confirm parking space safety."

[0068] Step 7: Based on the crowdsourced perception data from multiple vehicle trips, dynamically maintain and update the feature points in the semantic topology map.

[0069] In this embodiment, step 7 includes: Step 701, Feature Change Perception: During vehicle operation, based on the current high-confidence visual relocation result (i.e., the vehicle clearly knows its position in the semantic topology map), if the observed visual feature (such as a pillar advertisement) that should appear at the current topological position is inconsistent with the actually observed feature, and the new feature exists stably in multiple consecutive frames, and the vehicle's current visual relocation confidence is higher than a preset threshold (such as the number of matched feature points > 10, and the reprojection error < 0.5 meters), then a "suspected feature change" event is triggered.

[0070] Step 702, Feature Context Binding and Reporting: The vehicle de-identifies the image data of the new feature, along with its anchor point information in the topology map (e.g., located between "B2 pillar" and "fire hydrant", 3 meters to the right of B2 pillar), and uploads it to the cloud. The upload here is the relative positional relationship, not the absolute latitude and longitude.

[0071] Step 703, Cloud Arbitration and Update: The cloud collects new features reported by multiple vehicles in the same topological context (e.g., all passing the right side of pillar B2, 3 meters behind). When the number of observations exceeds a threshold (e.g., observations from 10 different vehicles, or verification by more than 5 vehicles at the same location within a week), and the original feature has not been observed by any vehicle at that location for more than 30 consecutive days, the cloud records the new feature (e.g., the texture of a new billboard) into the database, replacing the old feature, as the new positioning anchor point for that topological node. If a feature is reported as "missing" multiple times in a short period (e.g., within 24 hours), but is subsequently observed again, it is determined to be temporary occlusion, and feature replacement is not performed; only the occlusion period is recorded as historical data.

[0072] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A parking method based on semantic topological maps, characterized in that, Includes the following steps: In response to a trigger event that enters a region with no satellite signal, an initialization process is executed to determine the vehicle's initial pose in a pre-built semantic topology map; The semantic topology map is pre-built and stored in the cloud. The semantic topology map is a graph structure composed of nodes and edges, and does not contain the absolute latitude and longitude coordinates of parking spaces. Nodes represent identifiable semantic feature points in the physical environment. Each feature point is associated with a multimodal feature template for identification and matching. Edges represent drivable channels between adjacent feature points and are associated with channel attributes. Available parking spaces are filtered based on vehicle parameters, a list of candidate parking spaces is generated and presented to the user; In response to the user's selection of a target parking space, the parking space is physically pre-locked; Starting from the current location and ending at the target parking space, a semantic navigation path is planned in the semantic topology map by combining vehicle parameters and channel attributes. During the vehicle's journey, segmented tracking and guidance are performed based on the semantic navigation path until the target parking space is reached. When the vehicle arrives at the target parking space, parking confirmation and status reset are performed.

2. The method according to claim 1, characterized in that, The initialization process includes: The system acquires trigger events and preloads the entrance layer feature library of the target parking lot based on these events; the trigger events include geofence triggers, visual feature triggers, or manual triggers. During the vehicle's entry process, environmental data is collected in real time through onboard sensors, which include at least cameras and lidar. The real-time collected environmental data is compared in parallel with all pre-loaded entry layer feature libraries to identify the current entry layer and specific feature points. If a match is successful, the local relative pose of the vehicle relative to the identified feature point is calculated, and the global standard pose of the feature point is obtained from the cloud. The vehicle's pose in the global topological coordinate system is calculated through coordinate transformation to complete the initial positioning. If no match is found after a preset scanning time, the downgrade initialization process is executed. The downgrade initialization process includes: extracting regional features based on the surround-view camera to perform regional matching, taking the center of the matched region as the initial position, and providing guidance if accurate positioning cannot be restored within a preset distance.

3. The method according to claim 2, characterized in that, The parallel comparison employs multimodal feature matching, specifically including: A convolutional neural network is used to extract global feature vectors from images, and a point cloud deep learning network is used to extract geometric feature vectors from point clouds. The fusion weight α is dynamically determined based on ambient lighting and texture richness, and the two feature vectors are fused into a joint feature vector. α is calculated based on the normalized average brightness L and the normalized average gradient magnitude G according to the formula α = 0.5 × min(1, max(0, 2L)). Calculate using 0.5))+0.5×min(1,G / 0.3); The cosine similarity between the fused feature vector and the multimodal feature templates of each feature point in the preloaded feature library is calculated. If the similarity exceeds the preset threshold, the recognition is considered successful.

4. The method according to claim 1, characterized in that, The filtering of available parking spaces based on vehicle parameters includes: Obtain the vehicle's parameters, which include at least the vehicle length, width, height, minimum turning radius, and wheelbase. Read the attribute data of all parking spaces from the semantic map in the cloud, including parking space size and parking space type; The vehicle's parameters are compared with the parking space attributes one by one, and parking spaces whose dimensions do not meet the vehicle's parking requirements are automatically eliminated. Generate a candidate list of the remaining eligible parking spaces and output it.

5. The method according to claim 1, characterized in that, The execution of physical pre-locking of the parking space includes: Receive the user's instruction to select a target parking space; The cloud sends an up command to the smart lock associated with the target parking space and waits for the lock to return an up confirmation signal. If confirmation is received, a time-limited temporary locking token is generated, and the parking space is marked as reserved in the cloud. If locking fails, the user is prompted to select again.

6. The method according to claim 1, characterized in that, The planned semantic navigation path includes: Read the attribute data of all channels in the topology map. The attributes include channel width, channel height, minimum turning radius requirement, and whether it is a one-way street. When using a graph search algorithm for path search, the channel attribute requirements are matched with the vehicle parameters. Specifically, in the cost function, when evaluating a certain edge, the channel width of that edge is compared with the vehicle width plus a safety margin. If it exceeds the limit, the passage cost of that edge is set to infinity to automatically avoid it. In the topological subgraph formed by feasible routes, search for the shortest path from the starting point to the target parking space; The searched path is represented as an ordered sequence of key semantic nodes, each node being associated with a visual feature template, relative pose relationship, and driving action.

7. The method according to claim 6, characterized in that, The segmented tracking and guidance includes: Guided by the current target node, the following steps are executed repeatedly: identify the current target node, calculate the relative pose, generate local control commands, determine the node arrival, until the entrance of the channel where the target parking space is located is reached; For intersections or forks, the system relies on preset semantic features to trigger turning actions. If no turning guidance features are identified, it triggers local replanning or stops and waits. If the vehicle deviates from the planned path, a new feasible path will be searched and connected from the current location.

8. The method according to claim 1, characterized in that, The docking confirmation and status reset include: When a vehicle approaches the target parking space, the cloud-controlled indicator light above the parking space switches from the first state to the second state to indicate that the space is reserved for this vehicle. When a vehicle parks, the on-board unit sends a parking signal to the cloud, which then controls the parking lock to descend and turn off the indicator light. If an inconsistency is detected, a fault is reported and the parking space is marked as malfunctioning, while the driver is prompted to confirm safety.

9. The method according to claim 1, characterized in that, The semantic topology map extracts unique visual feature combinations around each parking space by collecting vehicle scans or crowdsourced data to form a feature fingerprint database. The unique visual feature combinations include at least column numbers, wall billboard textures, fire hydrant locations, and special ground markings. The map is stored in the data format {parking space ID, parking space attribute, associated feature list}, where the associated feature list contains feature point IDs and the relative pose of the parking space with respect to the feature point. The feature vectors stored in the cloud map are processed by dimensionality reduction or hash encoding and an index structure is established.

10. The method according to claim 1, characterized in that, It also includes the step of dynamically maintaining and updating feature points in the semantic topology map based on crowdsourced perception data from multiple vehicle trips, specifically including: If, under high-confidence visual relocalization, the observed actual features do not match the expectations and the new features are stably present, a feature change event is triggered. The image data of the new features and their relative anchor point information in the topology map are anonymized and uploaded to the cloud; The cloud system collects new features reported by multiple vehicles in the same topology context. When the number of observations exceeds the threshold and the original features have not been observed for a long time, the new features are entered into the database to replace the old features; if it is a temporary occlusion, only the new features are recorded.

Images