A Visual Enhancement Navigation System and Method for Autonomous Vehicles Based on BIM-Digital Twin City Model

Abstract

This invention discloses a vision-enhanced navigation system and method for unmanned vehicles based on a BIM-digital twin city model, including a BIM-digital twin city model database module, an IoT real-time data module, a fusion positioning module, a vision-enhanced recognition module, a cloud scheduling module, and an in-vehicle terminal module. By retrieving local model blocks, structured parameters, and real-time data, and correcting the in-vehicle image recognition results based on joint matching, continuous positioning and stable navigation are achieved in both above-ground and underground scenarios, improving recognition accuracy and decision reliability in complex environments.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, autonomous driving, building information modeling (BIM), digital twins, and vehicle navigation and control technologies. In particular, it relates to a system and method for enhancing and correcting the visual recognition results of unmanned vehicles using a BIM-digital twin city model, and achieving continuous navigation by combining multi-source positioning and real-time data. Background Technology

[0002] This invention relates to the fields of intelligent transportation, autonomous driving, building information modeling (BIM), digital twins, and vehicle navigation and control technologies. In particular, it relates to a system and method for enhancing and correcting the visual recognition results of unmanned vehicles using a BIM-digital twin city model, and achieving continuous navigation by combining multi-source positioning and real-time data.

[0003] Existing autonomous driving solutions typically rely on onboard cameras, LiDAR, satellite positioning equipment, and pre-built map data. In complex urban scenarios, underground spaces, and low-light / occlusive environments, existing technologies suffer from at least the following drawbacks: First, existing maps or high-precision maps are primarily used for path planning, making it difficult to directly apply the structured parameters from BIM-digital twin city models to visual recognition correction, resulting in a lack of geometric and spatial constraints on perception results. Second, the positioning link is discontinuous when switching between above-ground and underground scenes, and relying solely on inertial navigation after satellite positioning failure can easily lead to cumulative drift. Third, visual recognition is prone to failure under conditions such as rain, fog, backlight, and partial occlusion, making it difficult to guarantee the stability of navigation decisions. Fourth, high-precision and static maps are updated slowly, making it difficult to reflect dynamic changes such as construction closures, parking space occupancy, temporary obstacles, and building entrance / exit status in a timely manner.

[0004] Therefore, existing autonomous driving solutions struggle to simultaneously achieve visual correction using structured parameters from BIM-digital twin city models, continuous localization across above-ground and underground scenes, and stable navigation in dynamic environments. To address these issues, it is necessary to provide a technical solution that can directly incorporate structured parameters from BIM-digital twin city models into the visual recognition process and achieve continuous localization and navigation across above-ground, underground, and scene-switching conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a vision-enhanced navigation system and method for unmanned vehicles based on BIM-digital twin city models, in order to solve the problems of insufficient coupling between map data and visual recognition, poor stability of visual recognition in complex environments, and insufficient continuity of positioning in special spaces in the existing technology.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A vision-enhanced navigation system for autonomous vehicles based on a BIM-digital twin city model, such as Figure 1 As shown, it includes a BIM-digital twin city model database module 1, an IoT real-time data module 2, a fusion positioning module 3, a visual enhancement and recognition module 4, a cloud scheduling module 5, and a vehicle terminal module 6.

[0007] The BIM-digital twin city model database module 1, such as Figure 2 As shown, the system includes a model building unit 11, a local model block storage unit 12, a parameter extraction unit 13, and a model updating unit 14. The model building unit 11 is used to integrate BIM models of buildings, roads, transportation facilities, and underground spaces with real-world 3D data to form a BIM-digital twin city model covering both above-ground and underground spaces. The local model block storage unit 12 is used to partition and index the model according to road sections, intersection units, parking areas, underground passage areas, or building entrance areas. The parameter extraction unit 13 is used to extract structured parameters related to vehicle perception and navigation. The model updating unit 14 is used to locally update the corresponding local model blocks based on vehicle-side perception data, real-time IoT data, or manually collected data.

[0008] The IoT real-time data module 2, such as Figure 3 As shown, the system includes a data access unit 21, a data preprocessing unit 22, and a data distribution unit 23. The data access unit 21 is used to access one or more of the following: traffic signal status, road construction information, temporary closure information, parking space occupancy information, and building or parking lot entrance / exit status information. The data preprocessing unit 22 is used to perform time alignment, anomaly removal, and format standardization on the accessed data. The data distribution unit 23 is used, under the coordination of the cloud scheduling module 5, to send real-time data corresponding to the target area to the visual enhancement recognition module 4 and the vehicle terminal module 6 according to the vehicle's current location or future driving segment.

[0009] The fusion positioning module 3, such as Figure 4 As shown, it includes a satellite positioning unit 31, an inertial measurement unit 32, an underground compensation positioning unit 33, and a coordinate registration unit 34. The satellite positioning unit 31 is used to output vehicle position information in ground space. The inertial measurement unit 32 is used to provide continuous pose information when the satellite signal is weakened, interrupted, or the scene is switched. The underground compensation positioning unit 33 is used to output compensation positioning information in underground space. The coordinate registration unit 34 is used to map the ground space positioning results to the city-level model coordinate system and to map the underground space positioning results to the underground local model coordinate system or a preset unified model coordinate system.

[0010] The visual enhancement recognition module 4, such as Figure 5As shown, the system includes a position prediction unit 41, a joint matching unit 42, a local model preloading unit 43, an occlusion evaluation unit 44, and a decision output unit 45. The local model preloading unit 43 is used to receive and cache the local model blocks and structured parameters corresponding to the future driving segment sent by the cloud scheduling module 5. The occlusion evaluation unit 44 is used to perform visibility domain analysis on static and dynamic occlusions ahead based on the local model blocks and real-time data. The decision output unit 45 is used to generate environmental identification results and / or navigation control suggestions for vehicle control based on the matching results of the joint matching unit 42 and the occlusion evaluation results.

[0011] The cloud scheduling module 5, such as Figure 6 As shown, it includes a data interaction unit 51, a resource scheduling unit 52, and a collaborative control unit 53. The data interaction unit 51 is used to realize the transmission of local model blocks, structured parameters, real-time data, and control information between the cloud and the vehicle. The resource scheduling unit 52 is used to allocate model retrieval, caching, and computing resources to different vehicles according to the number of vehicles, vehicle locations, and driving scenarios. The collaborative control unit 53 is used to switch the working parameters of the fusion positioning module 3 and the visual enhancement recognition module 4 according to scenarios such as driving on ground roads, underground parking, or passing through closed parks.

[0012] The vehicle-mounted terminal module 6, such as Figure 6 As shown, it includes a data acquisition unit 61, a model caching unit 62, an instruction execution unit 63, and a status feedback unit 64; the data acquisition unit 61 is used to acquire vehicle images and pose data; the model caching unit 62 is used to cache local model blocks to provide data support in case of communication abnormalities, model missing, or joint matching failure; the instruction execution unit 63 is used to execute navigation control instructions based on the environment recognition results; the status feedback unit 64 is used to feed back vehicle status and incremental data to the cloud scheduling module 5.

[0013] Furthermore, the BIM-digital twin city model database module can index and store the model in blocks according to road sections, intersection units, parking areas, underground passage areas, or building entrance areas; the local model blocks can include geometric boundary information, spatial topological relationships, and structured parameters. The local model preloading unit in the visual enhancement recognition module receives and caches the local model blocks and structured parameters sent by the cloud scheduling module for use by the joint matching unit.

[0014] Furthermore, the fusion positioning module includes a satellite positioning unit, an inertial measurement unit, an underground compensation positioning unit, and a coordinate registration unit. When the above-ground positioning conditions meet the requirements, the satellite positioning unit and the inertial measurement unit work together to output the vehicle's pose; when the underground space or obstructed environment causes the satellite signal to fail to meet the requirements, the underground compensation positioning unit and the inertial measurement unit work together to output the compensated positioning result; the coordinate registration unit maps the positioning result to the city-level model coordinate system, the underground local model coordinate system, or a preset unified model coordinate system according to the vehicle's location.

[0015] Furthermore, the IoT real-time data module can access one or more of the following: traffic signal status, construction closure information, parking space occupancy information, building entrance and exit status information, and temporary obstacle information. After time alignment, anomaly removal, and format standardization, the data is sent to the cloud scheduling module, visual enhancement and recognition module, and vehicle terminal module for environmental recognition correction and navigation decision updates.

[0016] Furthermore, the visual enhancement recognition module extracts edge, corner, contour, texture or landmark features from vehicle images and performs joint matching by combining geometric boundaries, spatial topological relationships and structured parameters in local model blocks, thereby imposing geometric constraints, scale constraints and spatial position constraints on the pure visual recognition results; the occlusion evaluation unit can also perform visible domain analysis on static and dynamic occlusion by combining building boundary relationships, road spatial relationships and real-time data.

[0017] Furthermore, the vehicle terminal module is used to receive navigation control commands and execute corresponding control actions; in the event of communication abnormalities, missing target models, or joint matching failures, the vehicle terminal module can also call cached local model segmentation or degradation control strategies to maintain continuous navigation or safe parking.

[0018] It should be noted that the specific algorithms described in the above embodiments are merely illustrative examples to help those skilled in the art understand and implement the present invention. Other equivalent algorithms or implementation methods can be used to complete the corresponding functions without departing from the core concept of the present invention. All equivalent substitutions or modifications made based on the core concept of the present invention should fall within the protection scope of the present invention.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects: First, by directly introducing local model blocks and their structured parameters into the visual recognition link, and correcting the vehicle image recognition results based on joint matching, geometric constraints, scale constraints, and spatial position constraints can be imposed on the visual recognition results, thereby improving the stability and accuracy of visual recognition in complex environments.

[0020] Second, by integrating the positioning module and the coordinate registration unit, continuous positioning and unified coordinate mapping can be achieved in above-ground space, underground space, and during scene switching, thereby improving navigation continuity and reducing positioning drift.

[0021] Third, through the collaborative mechanism of location prediction, cloud retrieval, and local model block preloading, the data resources corresponding to the target area can be retrieved in advance, thereby reducing model call latency and improving system real-time performance.

[0022] Fourth, by introducing real-time IoT data and combining it with an occlusion assessment mechanism, it can dynamically respond to environmental changes such as construction closures, parking space occupancy, temporary obstacles, and target occlusion, thereby improving the reliability of navigation decisions and the ability to adapt to dynamic scenarios. Attached Figure Description

[0023] Figure 1 is a schematic diagram of the overall structure of the system of the present invention. Figure 2 is a schematic diagram of the structure of the BIM-digital twin city model database module of the present invention; Figure 3 is a schematic diagram of the structure of the IoT real-time data module of the present invention; Figure 4 is a schematic diagram of the fusion positioning module of the present invention; Figure 5 is a schematic diagram of the structure of the visual enhancement recognition module of the present invention; Figure 6 is a schematic diagram of the data interaction between the cloud scheduling module and the vehicle terminal module of the present invention; Figure 7 is a schematic diagram of the application process of the present invention in a ground road scenario; Figure 8 is a schematic diagram of the application process of the present invention in an underground parking lot scenario.

[0024] Figure labeling: 1-BIM-Digital Twin City Model Database Module; 11-Model Building Unit; 12-Local Model Block Storage Unit; 13-Parameter Extraction Unit; 14-Model Update Unit; 2-IoT Real-time Data Module; 21-Data Access Unit; 22-Data Preprocessing Unit; 23-Data Distribution Unit; 3-Fusion Positioning Module; 31-Satellite Positioning Unit; 32-Inertial Measurement Unit; 33-Underground Compensation Positioning Unit; 34-Coordinate Registration Unit; 4-Visual Enhancement Recognition Module; 41-Location Prediction Unit; 42-Joint Matching Unit; 43-Local Model Preloading Unit; 44-Occlusion Assessment Unit; 45-Decision Output Unit; 5-Cloud Scheduling Module; 51-Data Interaction Unit; 52-Resource Scheduling Unit; 53-Collaborative Control Unit; 6-Vehicle Terminal Module; 61-Data Acquisition Unit; 62-Model Cache Unit; 63-Instruction Execution Unit; 64-Status Feedback Unit. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments describe in detail the specific implementation methods of each module to enable those skilled in the art to implement the present invention. Those skilled in the art can make modifications or substitutions to the specific implementation methods without departing from the concept of the present invention.

[0026] Example 1: Ground Road Navigation Scenario In this embodiment, as Figure 7 As shown, the vehicle is driving in an urban road environment.

[0027] (i) Integrated positioning and coordinate registration The satellite positioning unit in the fusion positioning module receives GPS / BeiDou satellite signals with an update frequency of 10Hz; the inertial measurement unit uses a six-axis IMU with a sampling frequency of 100Hz; data fusion is performed through extended Kalman filtering (EKF) to output the vehicle's current three-dimensional position (x, y, z) and attitude (roll angle, pitch angle, yaw angle). The coordinate registration unit uses a pre-calibrated seven-parameter Bursa model to convert the positioning results in the geodetic coordinate system into the city-level BIM-digital twin city model coordinate system. The coordinate system type is a left-handed coordinate system, and the unit is m. The conversion formula is Equation (1). Where ΔX, ΔY, and ΔZ are translation parameters, m is the scale factor, and εx, εy, and εz are rotation parameters.

[0028]

[0029] Where ΔX, ΔY, ΔZ are translation parameters, m is the scale factor, and εx, εy, εz are rotation parameters.

[0030] (II) Determination of future driving sections The position prediction unit in the visual enhancement recognition module predicts the trajectory within a future time window Δt based on the vehicle's current pose, velocity v, heading angle θ, and planned path using a uniform motion model; in this embodiment, Δt is taken as 3s. Specifically, the vehicle's position sequence within the prediction time domain is calculated, as shown in equation (2);

[0031] Where Ts is the control period, which is 0.1s in this embodiment. Then, the predicted trajectory is spatially overlaid with the local model block index map to extract all local model block IDs traversed by the predicted trajectory, including the upcoming intersection, adjacent road segments, and expected turning areas. The location prediction unit ultimately outputs a list L of future driving segments, along with the priority of each local model block, with closer blocks having higher priority.

[0032] (III) Cloud-based scheduling and data distribution The cloud-based scheduling module receives a list L of future driving sections and requests corresponding local model blocks from the BIM-digital twin city model database module (each block size can be adjusted according to the scenario; for example, above-ground roads can use 200m×200m blocks, and underground parking lots can use 50m×50m blocks, including triangular meshes, boundary line sets, and semantic labels). Simultaneously, it requests structured parameters from the parameter extraction unit (such as lane width 3.5m±0.2m, minimum turning radius 12m, and lane line position relationship with a left-right lane line spacing of 3.75m, etc.). Furthermore, the cloud-based scheduling module also obtains real-time data of the target area from the IoT real-time data module (traffic light phases, polygonal boundaries of construction closure areas, coordinates and dimensions of temporary obstacles). This data is compressed (using the Draco compression algorithm) and then transmitted via a 5G V2X link to the vehicle terminal module and the visual enhancement and recognition module, with transmission latency typically controlled within 50ms (depending on network conditions).

[0033] (iv) Joint matching and visual correction The data acquisition unit in the vehicle-mounted terminal module uses a forward-facing monocular camera (1920×1080 pixels resolution, 30fps) to acquire images of the front. The joint matching unit performs the following steps: Step A: Image Feature Extraction. For example, the ORB feature extraction algorithm can be used, where the FAST corner threshold can be set to 20 and the number of pyramid layers can be set to 8 (SIFT, SURF, or deep learning-based feature extraction methods can also be used) to extract key points and descriptors in the image. At the same time, Canny edge detection (low threshold 50, high threshold 150) is used to extract lane line edges and traffic sign contours.

[0034] Step B: Model Projection. Based on the vehicle's pose in the model coordinate system obtained from the current fusion localization, and combined with the camera intrinsic matrix K (where focal lengths fx and fy can be set to 1200 pixels, and optical centers cx and cy can be set to 960 pixels and 540 pixels respectively, with specific values ​​determined based on the actual camera calibration results) and the pre-calibrated extrinsic matrix [R|t], the three-dimensional boundary points within the local model block, including lane center lines, stop lines, and curb boundaries, are projected onto the image plane, as shown in Equation (3):

[0035] Obtain the two-dimensional projection lines of the model on the image.

[0036] Step C: Joint Matching. The extracted image edges are registered with the model projection lines using ICP (Iterative Closest Point) matching. The matching error function is defined as shown in equation (4):

[0037] in For image edge points, The nearest point on the model projection line is used. The optimal pose correction is solved using nonlinear least squares (Levenberg-Marquardt algorithm) to correct the visual recognition results. For example, when the lane line detected by vision deviates, the joint matching unit uses the lane boundary geometric constraints in the model (known lane line equations) to force the detection results back to the range defined by the model.

[0038] Step D: Compensation and correction for missing features. When rain, fog, backlight, or partial occlusion cause some lane line features to be missing, the joint matching unit initiates a model-based structured compensation mechanism. For example, if the left lane line is obscured by water stains and cannot be detected, but the right lane line is clear, the position of the left lane line is calculated by adding the normal vector to the coordinates of the right lane line and multiplying it by the width, using the lane width structured parameter (fixed value 3.5m) and the lane line position relationship (parallel left and right, normal distance 3.5m) in the model. The specific calculation formula is shown in equation (5):

[0039] Where w is the lane width This is a unit vector perpendicular to the lane direction. For traffic sign recognition, if the sign in the image is blurry, the joint matching unit reads the precise location, size, and text content of the sign from the model blocks and overlays them into the recognition result.

[0040] (v) Obstruction assessment and decision output The occlusion assessment unit analyzes visibility within a 120° field of view in front of the vehicle using a ray casting algorithm, based on building boundaries (3D polygons) and road spatial relationships within local model blocks, combined with the vehicle's current position and real-time data (such as coordinates of temporary obstacles). For each potential target of interest (such as traffic lights or pedestrian crossings), it calculates whether the line connecting the center of the vehicle's camera to the target surface intersects with any static occlusion (buildings, pillars) or dynamic occlusion (other vehicles, construction fences). If an intersection occurs, it is considered an occlusion, and an occlusion mask is output.

[0041] The decision output unit receives the environmental recognition results (including corrected lane lines, traffic signs, and obstacle distances) and occlusion assessment results after joint matching, and generates control suggestions. For example, when joint matching shows that the lane lines ahead are complete and unobstructed, it outputs control suggestions such as "go straight, recommended speed 30km / h"; when occlusion assessment shows that the traffic light ahead is obstructed by a large vehicle, it combines the real-time traffic light status from IoT data (currently red light with 15 seconds remaining) and outputs "decelerate, prepare to stop". The control suggestions are sent to the command execution unit of the on-board terminal module via the CAN bus, and the vehicle's drive-by-wire system executes acceleration, braking, or steering.

[0042] Example 2: Automated parking scenario in an underground parking lot

[0043] In this embodiment, as Figure 8 As shown, the vehicle drives from the ground level into the underground parking lot.

[0044] (a) Positioning mode switching and underground compensation positioning When the fusion positioning module detects that the satellite signal strength is below the threshold CN0 < 30dB-Hz and the duration exceeds 3 seconds, it automatically switches from the "satellite + IMU" mode to the "underground compensation positioning + IMU" mode. The underground compensation positioning unit adopts a multi-source fusion scheme: the wheel speed odometer outputs 4096 pulses per revolution to provide relative displacement increments; the geomagnetic matching module performs dynamic time warping (DTW) matching between the real-time geomagnetic sequence and the pre-acquired geomagnetic reference map to obtain absolute position correction; in the underground space where UWB base stations have been deployed, time difference of arrival (TDOA) positioning is used, with a positioning accuracy of ±10cm. The above data is fused through unscented Kalman filtering (UKF) to output the local coordinates of the vehicle in the underground parking lot; the origin of the coordinates is set at the parking lot entrance. The coordinate registration unit maps this local coordinate to the coordinate system of the underground parking lot BIM-digital twin city model.

[0045] (II) Prediction of Future Driving Sections and Model Loading The location prediction unit predicts the vehicle's trajectory within the next 2 seconds using an Ackerman kinematic model, based on the planned path within the underground parking lot (generated by the A* algorithm on a drivable area grid map) and the current vehicle heading angle and steering wheel angle. The future driving segment includes: the area ahead of the lane, the area where the target parking space is located, and possible meeting areas. The cloud-based scheduling module retrieves the corresponding local model blocks, which include: lane boundaries (3D wireframes), column grid structure (cylinders or cuboids, radius 0.3m, height 2.5m), parking space borders (rectangles, length 5m, width 2.5m), entrance / exit locations (ramp start and end coordinates), and height restriction information (2.0m). Structured parameters include: lane width (6m), parking space size (5m × 2.5m), column grid spacing (8m × 8m), height restriction (2.0m), and maximum ramp gradient (15%).

[0046] (III) Application of joint matching in underground environments Due to the dim lighting and repetitive textures in underground parking lots, the joint matching unit, in addition to using ORB features, also incorporates the line feature extraction algorithm LSD and structured light projection. The extracted column edges and parking space corners are matched with the column grid positions and parking space borders in the local model blocks. The parking space size of 5m × 2.5m in the structured parameters is used as a scale constraint; when the visually detected parking space width deviation exceeds 10%, it is automatically corrected to the model's standard value. Specifically, the cost function is constructed, as shown in equation (5).

[0047]

[0048] in These are the coordinates of the corner points detected by vision. For the model projection corner points, Geometric dimensions measured visually (such as parking space width). Let be the standard size in the structured parameters, where the constraint weight λ is 0.7. The visual recognition result is optimized by minimizing this cost function.

[0049] (iv) Dynamic real-time data response The IoT real-time data module accesses data from parking space geomagnetic sensors (each parking space's occupancy status is updated at 1Hz), entrance / exit gate status (open / closed), and information on temporary obstacles within the passageway (detected by site cameras). When the target parking space status changes from "vacant" to "occupied," the IoT module pushes the change information to the cloud scheduling module within 100 milliseconds (depending on the reporting frequency of the IoT sensors). The cloud scheduling module then triggers path replanning, switching the target parking space to an adjacent vacant parking space (automatically selected based on the parking space layout in the local model block). The decision output unit regenerates the parking path (a fifth-order polynomial curve) based on the new target parking space and controls the vehicle to travel at a low speed (<5km / h) to the new parking space.

[0050] (v) Communication degradation processing When the vehicle enters deep underground, causing a 5G signal interruption, the model cache unit of the onboard terminal module has pre-stored local model blocks of the entire underground parking lot (the cache capacity can be determined according to the onboard terminal hardware configuration; for example, 1GB can cover an underground parking lot model of approximately 5000m²). The joint matching unit continues to use the locally cached model and structured parameters for visual enhancement. If joint matching fails (e.g., the model deviates significantly from the current scene), the decision output unit automatically switches to a degraded control strategy: relying solely on the onboard ultrasonic radar (12 probes, detection range 0.2~5m) for obstacle avoidance, slowly driving at a low speed (e.g., ≤3km / h) to the exit or open area according to the last effective path memory, and illuminating the hazard warning lights.

[0051] It should be noted that the frequencies, thresholds, sizes, weights, and algorithms described in the above embodiments are preferred embodiments of the present invention and are used to illustrate the present invention rather than to limit the scope of protection of the present invention. Without departing from the core concept of the present invention, those skilled in the art can adjust the relevant parameters according to specific scenarios, hardware configurations, and performance requirements, or use equivalent algorithms to achieve the same function, all of which should fall within the scope of protection of the present invention.

[0052] Those skilled in the art will understand that any equivalent substitutions, simplifications, or extensions made to this invention without departing from its spirit and essence should fall within the protection scope of this invention.

Claims

1. A BIM-Digital Twin City Model based visual augmented navigation system for self-driving vehicles, characterized in that, include: The system includes a BIM-digital twin city model database module, an IoT real-time data module, a fusion positioning module, a visual enhancement and recognition module, a cloud dispatch module, and a vehicle terminal module. The cloud-based scheduling module is communicatively connected to the BIM-digital twin city model database module, the IoT real-time data module, the fusion positioning module, the visual enhancement recognition module, and the vehicle-mounted terminal module. The BIM-digital twin city model database module is used to store local model blocks divided by road sections, intersection units or parking areas, as well as structured parameters extracted from the model. The structured parameters include at least one or more of lane width, turning radius, parking space size, and height restriction information, which are used to provide geometric constraints, scale constraints or spatial location constraints. The IoT real-time data module is used to access, preprocess and distribute dynamic real-time data of the target area, providing dynamic environment support for visual enhancement recognition and navigation decision-making. The fusion positioning module is used to output the vehicle's current position and / or pose information, and to register the positioning results to the model coordinate system corresponding to the BIM-digital twin city model; The visual enhancement recognition module includes a position prediction unit and a joint matching unit. The position prediction unit is used to determine the future driving segment based on the vehicle's current position, speed, heading, and planned path. The joint matching unit is used to extract visual features from the vehicle image and perform joint matching by combining the geometric boundaries, spatial topological relationships, and structured parameters in the corresponding local model blocks to correct the visual recognition results and generate environmental recognition results. The cloud scheduling module is used to retrieve the corresponding local model blocks, structured parameters and IoT real-time data according to the future driving section, and send them to the visual enhancement recognition module and the vehicle terminal module. The vehicle-mounted terminal module is used to collect vehicle-mounted images and pose data, and execute navigation control commands based on the environment recognition results.

2. The system according to claim 1, characterized in that, The BIM-digital twin city model database module includes a model building unit, a local model block storage unit, a parameter extraction unit, and a model updating unit; The model building unit is used to integrate BIM models of buildings, roads, transportation facilities and underground spaces with real-world 3D data to form a BIM-digital twin city model covering above-ground and underground spaces. The local model block storage unit is used to partition and index the model according to road sections, intersection units, parking areas, underground passage areas or building entrance and exit areas. The parameter extraction unit is used to extract structured parameters related to vehicle perception and navigation. The structured parameters include at least one or more of the following: lane boundary, channel width, turning radius, ramp information, parking space size, height restriction information, column grid boundary, and entrance / exit location. The model update unit is used to perform local updates on corresponding local model blocks based on vehicle-side perception data, real-time IoT data, or manually collected data.

3. The system according to claim 1, characterized in that, The IoT real-time data module includes a data access unit, a data preprocessing unit, and a data distribution unit. The data access unit is used to access one or more of the following: traffic signal status, road construction information, temporary closure information, parking space occupancy information, and building or parking lot entrance / exit status information. The data preprocessing unit is used to perform time alignment, anomaly removal, and format standardization on the accessed data. The data distribution unit is used, under the coordination of the cloud scheduling module, to send real-time data corresponding to the target area to the visual enhancement recognition module and the vehicle terminal module according to the vehicle's current location or future travel segment.

4. The system according to claim 1, characterized in that, The fusion positioning module includes a satellite positioning unit, an inertial measurement unit, an underground compensation positioning unit, and a coordinate registration unit; The satellite positioning unit is used to output vehicle location information in terrestrial space; The inertial measurement unit is used to provide continuous pose information when satellite signals are weakened, interrupted, or the scene is switched. The underground compensation positioning unit is used to output compensation positioning information in underground space. The compensation positioning information is obtained through one or more of the following methods: wheel speed odometer, geomagnetic matching, or pre-deployed UWB base station. The coordinate registration unit is used to map the above-ground spatial positioning results to the city-level model coordinate system, and to map the underground spatial positioning results to the underground local model coordinate system or the preset unified model coordinate system.

5. The system according to claim 1, characterized in that, The visual enhancement recognition module also includes a local model preloading unit, an occlusion evaluation unit, and a decision output unit; The local model preloading unit is used to receive and cache the local model blocks and structured parameters corresponding to the future driving segment sent by the cloud scheduling module; The occlusion evaluation unit is used to perform visibility domain analysis on static and dynamic occlusion in front of the camera based on local model blocks and real-time data. The decision output unit is used to generate environmental identification results and / or navigation control suggestions for vehicle control based on the joint matching results and occlusion evaluation results.

6. The system according to claim 1, characterized in that, The cloud-based scheduling module includes a data interaction unit, a resource scheduling unit, and a collaborative control unit; The data interaction unit is used to realize the transmission of local model blocks, structured parameters, real-time data and control information between the cloud and the vehicle. The resource scheduling unit is used to allocate model retrieval, caching, and computing resources to different vehicles based on the number of vehicles, vehicle location, and driving scenario. The collaborative control unit is used to switch the operating parameters of the fusion positioning module and the visual enhancement recognition module according to scenarios such as driving on surface roads, underground parking, or passing through closed parks.

7. A vision-enhanced navigation method for unmanned vehicles based on a BIM-digital twin city model, applied to the system according to any one of claims 1 to 6, characterized in that, Includes the following steps: S1. Obtain the vehicle's current position and / or pose information through the fusion positioning module, and register the positioning results to the model coordinate system corresponding to the target scene; S2. The position prediction unit in the visual enhancement recognition module determines the future driving segment based on the vehicle's current position, speed, heading, and planned path. S3. The cloud scheduling module retrieves the local model blocks, structured parameters, and real-time IoT data of the target area corresponding to the determined future driving section, and sends them to the visual enhancement recognition module and the vehicle terminal module via wireless communication. S4. The vehicle-mounted image is acquired by the camera in the vehicle-mounted terminal module. The joint matching unit combines the local model block, structured parameters and vehicle-mounted image to perform joint matching, correct the visual recognition result and obtain the environment recognition result. S5. Based on the environmental identification results, real-time data of the target area, and vehicle positioning results, a navigation control command is generated and executed by the vehicle terminal module through the drive-by-wire system.

8. The method according to claim 7, characterized in that, Step S1 includes: when the ground space satellite positioning signal meets the requirements, the satellite positioning result is fused with the inertial measurement result; when the underground space, building obstruction area or tunnel environment causes the satellite positioning signal to fail to meet the requirements, the underground compensation positioning result is fused with the inertial measurement result; and according to the scene where the vehicle is located, the corresponding coordinate transformation relationship is called to map the fused positioning result to the target model coordinate system.

9. The method according to claim 7, characterized in that, The joint matching in step S4 includes: extracting one or more of the edge, corner, landmark outline, or texture features from the vehicle image; extracting one or more of the lane boundary, channel boundary, column boundary, parking space border, traffic facility outline, or entrance / exit location corresponding to the vehicle image from the local model block corresponding to the future driving section; and using the structured parameters to apply geometric constraints, scale constraints, or spatial location constraints to the visual feature matching results to output the corrected environment recognition results.

10. The method according to claim 7, characterized in that, Step S5 and subsequent steps include: incrementally updating the local model blocks based on real-time data of the target area, vehicle-side perception data, or manually collected data; and in the event of communication anomalies, model missing, or joint matching failure, invoking the local model block cached in the vehicle terminal module or the degradation control strategy to maintain continuous navigation or safe parking of the vehicle.

Images