An AI image processing empowered vehicle-mounted visual inertial navigation method

Abstract

The application relates to the technical field of vehicle navigation, and discloses an AI image processing enabled vehicle visual inertial navigation method; the method acquires road scene images through a vehicle visual acquisition module, extracts road features after semantic analysis by an AI image processing module, and autonomously constructs a real-time local road network topology model which can be dynamically expanded; vehicle pose data calculated by an inertial measurement unit is combined to correct inertial navigation drift through visual road features, and accurate registration of the vehicle and the road network is completed; a double-layer iterative route search is performed through a route search engine module, real-time road conditions are dynamically corrected in combination, and finally navigation guidance instructions are output; the application does not need to rely on a pre-stored electronic map, can autonomously complete the whole-process navigation in a no-map or map failure scene, and greatly improves the scene adaptability and real-time performance of vehicle navigation.

Description

Technical Field

[0001] This invention belongs to the field of vehicle navigation technology, specifically an AI image processing-enabled vehicle visual inertial navigation method. Background Technology

[0002] Current mainstream vehicle navigation systems rely on Global Navigation Satellite System (GNSS) and pre-stored electronic maps to complete route search and guidance. In scenarios where GNSS signals are blocked, such as urban canyons, tunnels, and underground parking lots, Visual Inertial Navigation System (VINS) is usually introduced to achieve continuous positioning, but the route search process still relies entirely on pre-stored electronic maps.

[0003] Existing technologies have the following inherent defects: First, they rely heavily on pre-stored maps. In scenarios where there are no maps or the maps are outdated, such as newly built roads, non-standard rural roads, and temporary road detours due to construction, route search will fail directly, resulting in poor scenario adaptability. Second, route search is based on static map data and cannot be adjusted in conjunction with real-time dynamic road conditions, resulting in insufficient real-time navigation. Third, the cost of collecting, updating, and maintaining high-precision electronic maps is high, making it difficult to cover all non-standard road scenarios.

[0004] Therefore, there is an urgent need for an in-vehicle navigation method that can break free from the reliance on pre-stored electronic maps and autonomously complete road network construction and route search based on real-time visual perception, in order to solve the aforementioned problems in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an AI-powered visual inertial navigation method for vehicles. This method constructs a local road network topology in real time using AI vision, combines visual-inertial navigation fusion to complete pose registration, and achieves two-layer iterative route search based on the self-constructed road network. This completely eliminates the reliance on pre-stored electronic maps and significantly improves the scene adaptability, real-time performance, and economy of vehicle navigation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an AI image processing-enabled vehicle visual inertial navigation method, comprising the following steps: S1. Sensor data synchronous acquisition: The vehicle-mounted vision acquisition module acquires visual images of the road scene in front of the vehicle, and the inertial measurement unit synchronously acquires six-axis inertial motion data. The two types of data are synchronized at the microsecond level through a hardware clock.

[0007] S2. Inertial Navigation Pose Calculation: The inertial navigation data calculation module performs strapdown inertial navigation calculation on the six-axis inertial data, and outputs real-time vehicle pose data including three-dimensional position, three-axis attitude, and three-dimensional velocity.

[0008] S3. Real-time local road network topology construction: The AI ​​image processing module performs semantic segmentation and feature extraction on visual images to construct a local road network topology model with lane units as nodes. The model continues to expand and complete splicing correction as vehicles travel.

[0009] S4. Road Network Pose Fusion Registration: Using visually extracted lane line features as the observation benchmark, combined with the extended Kalman filter algorithm to correct the cumulative drift error of the inertial navigation system, the vehicle pose and the local road network are accurately registered in space.

[0010] S5. Two-layer iterative route search: Taking the navigation target point with absolute geographic coordinates as the endpoint, the target point is iteratively mapped to the continuously expanding local road network by combining the inertial navigation displacement and the vehicle's absolute position corrected by vision. The weighted cost A* algorithm is used to perform global route search, while local dynamic adjustments are performed in combination with real-time traffic conditions to generate the final navigation route.

[0011] S6. Dynamic navigation guidance output: Based on the vehicle's real-time speed, the system calculates the timing of navigation operation prompts, generates navigation guidance commands that combine voice and visual input, and outputs them to the driver through the in-vehicle interactive system.

[0012] Furthermore, in step S3, the local road network topology construction specifically includes three sub-steps: semantic segmentation, feature extraction, and topology construction and stitching correction. The deep learning semantic segmentation model outputs pixel-level road element classification results, extracts multi-dimensional features such as lanes, signs, and obstacles, and constructs a topological graph structure road network with attributes. Combined with inertial navigation attitude correction for monocular scale drift, the road network stitching is completed through inter-frame feature matching, and the accumulated topology error is eliminated through loop closure detection.

[0013] Furthermore, the coverage distance of the local road network topology model is dynamically adjusted according to vehicle speed levels, and a vehicle speed threshold hysteresis interval is set to avoid frequent switching caused by vehicle speed fluctuations, thus balancing the road network precision in low-speed scenarios with the route lead time in high-speed scenarios.

[0014] Furthermore, the global route search in step S5 adopts the weighted cost A* algorithm. The cost function integrates three factors: driving distance, number of lane changes, and road traffic priority. The road traffic priority is automatically determined and assigned a corresponding weight based on the number of lanes and road type identified by visual recognition, ensuring that the searched route takes into account both traffic efficiency and driving safety.

[0015] Furthermore, the local dynamic adjustment in step S5 sets four distinct triggering conditions, corresponding to different traffic restriction scenarios. After being triggered, the local road change route is quickly planned and connected to the global route, with a response time of less than 50ms, ensuring the real-time performance and continuity of navigation.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention autonomously constructs a local road network topology in real time using AI vision, and uses it as the sole basic data for route search. It does not rely on any pre-stored electronic maps and can effectively adapt to scenarios where traditional navigation fails, such as newly built roads, rural non-standard roads, and temporary construction detours. This greatly expands the applicability of in-vehicle navigation and completely eliminates the dependence on pre-stored maps, making it highly adaptable to different scenarios. This invention constructs a two-layer route search architecture of "global optimal search + local dynamic adjustment". The global route is continuously iterated and optimized as the road network expands to ensure the global optimal route. When local road conditions change, the route can be corrected within 50ms. The response speed to sudden road conditions is far superior to that of traditional static map navigation. Its layer iterative search mechanism makes navigation both accurate and real-time. This invention uses visually stable lane line features as the observation benchmark and corrects the cumulative drift error of the inertial navigation system through extended Kalman filtering. The vehicle pose positioning accuracy is better than 0.5m and the heading angle accuracy is better than 0.5°. At the same time, it combines inertial navigation attitude correction to correct monocular visual scale drift, ensuring the geometric accuracy of the autonomously constructed road network and providing a reliable foundation for high-precision route search. Its deep fusion of vision and inertial navigation results in high positioning and registration accuracy. This invention eliminates the costs of collecting, updating, and maintaining high-precision electronic maps in traditional navigation systems. It can achieve full-process navigation using only vehicle-mounted cameras and inertial measurement units. It has low hardware costs, wide scenario coverage, and extremely high economic efficiency and practical value, significantly reducing the cost of navigation systems. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] In the attached diagram: Figure 1 This is a block diagram of the overall system architecture of an embodiment of the present invention; Figure 2 This is a block diagram showing the internal sub-module connections of the AI ​​image processing module in an embodiment of the present invention; Figure 3 This is a block diagram showing the internal sub-module connections of the route search engine module in an embodiment of the present invention; In the diagram: 1. Vehicle-mounted visual acquisition module; 2. Inertial measurement unit; 3. AI image processing module; 4. Semantic segmentation submodule; 5. Road feature extraction submodule; 6. Local road network topology construction submodule; 7. Inertial navigation data calculation module; 8. Pose fusion submodule; 9. Route search engine module; 10. Global path planning submodule; 11. Local dynamic adjustment submodule; 12. Navigation guidance generation module; 13. Human-computer interaction output module. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This embodiment discloses an AI image processing-enabled in-vehicle visual inertial navigation method, applied to civilian vehicles equipped with an in-vehicle perception computing unit. It can autonomously complete the entire navigation process—road network construction, pose registration, route search, and navigation guidance—in scenarios without pre-stored electronic maps or with GNSS signal failure. The overall system architecture upon which this method relies is as follows: Figure 1 As shown, the core functional modules include vehicle-mounted visual acquisition module 1, inertial measurement unit 2, AI image processing module 3, inertial navigation data calculation module 7, pose fusion submodule 8, route search engine module 9, navigation guidance generation module 12, and human-computer interaction output module 13.

[0021] The specific implementation steps of this embodiment are as follows: Step 1: Synchronous Data Acquisition from Multiple Sensors The vehicle-mounted visual acquisition module 1, installed on the upper inner side of the vehicle's windshield, continuously acquires visual images of the road scene directly in front of the vehicle. In this embodiment, the vehicle-mounted visual acquisition module 1 uses a 2-megapixel monocular vehicle-mounted camera with a frame rate of 30fps, an image resolution of 1920×1080, and a field of view covering 120° horizontally and 60° vertically, capable of completely acquiring multi-lane road and traffic sign information ahead. Simultaneously, the inertial measurement unit 2, installed at the vehicle's center of gravity, synchronously acquires the vehicle's six-axis inertial motion data, including three-axis acceleration and three-axis angular velocity data. The inertial measurement unit 2 uses industrial-grade MEMS inertial navigation devices with a sampling frequency of 100Hz, an accelerometer range of ±16g, and a gyroscope range of ±2000° / s, which can accurately capture changes in the motion state of the vehicle during driving. The vehicle-mounted vision acquisition module 1 and the inertial measurement unit 2 achieve time synchronization through the hardware clock module of the vehicle-mounted computing unit, with a synchronization accuracy of no less than 1ms. Each frame of visual image corresponds to a unique hardware timestamp, which is matched one-to-one with the inertial data at the same time, ensuring that the time base of subsequent visual and inertial navigation fusion processing is consistent and avoiding fusion errors caused by time deviation.

[0022] Step 2: Strapdown calculation of inertial navigation data The inertial navigation data processing module 7 receives the six-axis inertial motion data collected by the inertial measurement unit 2 and uses the strapdown inertial navigation processing algorithm to sequentially complete attitude update, velocity update, and position update: First, the three-axis attitude angles of the vehicle are updated using gyroscope data to obtain the real-time attitude matrix; then, the acceleration in the carrier coordinate system is converted into the acceleration in the navigation coordinate system by combining the accelerometer data and the attitude matrix, and the three-dimensional velocity of the vehicle is obtained through integration; the velocity is then integrated to obtain the three-dimensional position coordinates of the vehicle, and finally, the real-time vehicle pose data containing position, attitude, and velocity is output; since pure inertial navigation processing has drift errors that accumulate over time, high-precision positioning cannot be achieved by relying solely on inertial navigation, this embodiment subsequently uses visual road feature observation to continuously correct the accumulated errors of inertial navigation.

[0023] Step 3: AI Image Processing and Real-time Local Road Network Topology Construction The acquired road scene visual images are input into the AI ​​image processing module 3 for layered processing. The internal sub-module connection structure of the AI ​​image processing module 3 is as follows: Figure 2 As shown, it includes a semantic segmentation submodule 4, a road feature extraction submodule 5, and a local road network topology construction submodule 6. The specific processing procedure is as follows: 3.1 Pixel-level road semantic segmentation The semantic segmentation submodule 4 employs an improved DeepLabv3+ deep learning model optimized for in-vehicle road scenes to perform pixel-level semantic segmentation on the input single-frame visual image and output a road semantic segmentation map. In this embodiment, the semantic segmentation is set with 12 recognition categories, including: straight-ahead lane, dedicated left-turn lane, dedicated right-turn lane, reverse lane, pedestrian crossing, road median, traffic lights, prohibitory traffic signs, directional traffic signs, static obstacles, dynamic obstacles, and background area. Model inference is implemented through an in-vehicle embedded AI acceleration unit, and the inference time for a single frame image is less than 30ms, which meets the frame rate requirements for real-time processing.

[0024] 3.2 Multi-dimensional road feature extraction Road feature extraction submodule 5 performs post-processing and multi-dimensional feature extraction on the output road semantic segmentation map, specifically extracting four types of features: Lane geometry features: The boundary lines of each lane and the overall road edge are extracted using the Canny edge detection algorithm. The results are then converted to a bird's-eye view through perspective transformation, and geometric parameters such as lane width, number of lanes, total road width, road curvature, and extension direction are calculated. Traffic rule features: Extract the location, category and meaning of various traffic signs, including speed limit signs, turning restriction signs, no entry signs, etc., and convert them into traffic rule constraint attributes of the corresponding road network nodes; Obstacle characteristics: Extract the category, location, occupation range and movement trend of static obstacles (disabled vehicles, construction roadblocks, etc.) and dynamic obstacles (pedestrians, non-motorized vehicles, etc.), and mark the occupied areas as impassable nodes in the road network; (4) Topological connection features: Based on the ORB feature point matching results of continuous frame images, calculate the relative pose transformation of vehicles between frames, and infer the topological connection relationships of the road ahead, such as bifurcation, intersection, and turning, to provide a connection basis for road network splicing and topology construction.

[0025] 3.3 Local Road Network Topology Construction and Splicing Correction Based on the various road features extracted above, the local road network topology construction submodule 6 constructs a real-time local road network topology model with a directed topology graph structure, using the vehicle's current position as the origin and the vehicle's driving direction as the positive Y-axis. This model uses segmented lane units as topology nodes and the connection relationships between lanes (straight, left turn, right turn, lane change) as topology edges. Each topology node is accompanied by attribute information such as lane width, speed limit, steering restriction, and obstacle occupancy status.

[0026] The coverage distance of the local road network topology model is dynamically adjusted according to vehicle speed: when the real-time vehicle speed is below 60km / h, the road network coverage distance is set to 200m to ensure the road network precision in low-speed scenarios; when the vehicle speed is between 60-100km / h, the coverage distance is set to 350m to balance accuracy and lead time; when the vehicle speed is above 100km / h, the coverage distance is set to 500m to ensure the lead time for route search in high-speed scenarios; at the same time, a hysteresis interval of 5km / h is set for the vehicle speed threshold to avoid frequent switching of coverage distance caused by small fluctuations in vehicle speed.

[0027] During vehicle travel, the local road network topology model is dynamically updated in real time: for each frame of road image acquired by the vehicle-mounted vision acquisition module 1, the local road network topology construction submodule 6 completes a front-end expansion and back-end cropping, with the front end extending the newly added road network nodes in the direction of vehicle travel, and the back end cropping historical nodes that exceed the coverage area.

[0028] The specific mechanism for inter-frame road network stitching and error correction is as follows: combining the attitude data calculated by inertial navigation to correct the scale drift of monocular vision, solving the problem that monocular cameras cannot obtain absolute depth; calculating the relative pose of vehicles through inter-frame feature point matching to complete the coordinate transformation and stitching of road network nodes in adjacent frames; and introducing a closed-loop detection mechanism, when a vehicle travels to a previously identified road area and detects the same road features, performing global correction on the accumulated road network topology error to eliminate topology shift after long-distance travel and ensure the continuity and accuracy of the road network.

[0029] Step 4: Visual-Inertial Fusion and Road Network Pose Registration The pose fusion submodule 8 receives real-time vehicle pose data output by the inertial navigation data calculation module 7, and simultaneously receives road features and local road network topology models output by the AI ​​image processing module 3. It completes the fusion correction of vision and inertial navigation, as well as the spatial registration of the vehicle and the road network. The specific process is as follows: using the lane line features extracted by visual semantic segmentation as the external stable observation benchmark, the vehicle heading angle calculated by the inertial navigation is matched with the extension direction of the lane line to calculate the heading angle deviation; at the same time, the vehicle lateral position calculated by the inertial navigation is matched with the position of the lane centerline to calculate the lateral position deviation.

[0030] An extended Kalman filter algorithm is used to establish the state equation of the inertial navigation system (INS) pose and the observation equation of the visual observation. The heading deviation and lateral position deviation obtained from the visual observation are used as observation inputs to perform optimal estimation and real-time correction of the position and attitude errors calculated by the INS. Finally, the accurate pose mapping result of the vehicle in the local road network topology coordinate system is obtained. In this embodiment, the vehicle pose positioning accuracy after fusion correction is better than 0.5m and the heading angle accuracy is better than 0.5°, which can meet the pose reference requirements of high-precision route search.

[0031] Step 5: Two-layer iterative route search The route search engine module 9 receives the vehicle pose mapping results output by the pose fusion submodule 8, and simultaneously receives the real-time local road network topology model output by the AI ​​image processing module 3, performing a two-layer iterative route search operation of "global optimal search + local dynamic adjustment"; the internal submodule connection structure of the route search engine module 9 is as follows: Figure 3 As shown, it includes a global path planning submodule 10 and a local dynamic adjustment submodule 11. The specific search process is as follows: 5.1 Global Optimal Route Search The global route planning submodule 10 receives the user-preset navigation target point, which is in absolute geographic coordinates. Since the local road network topology model is a relative coordinate system with the vehicle as the origin and only covers a limited area in front of the vehicle, it cannot complete a long-distance global route search in one go. Therefore, an iterative expansion search mechanism is adopted: As the vehicle moves forward, the local road network topology model continues to expand in the direction of travel. The pose fusion submodule 8 calculates the vehicle's absolute geographic coordinates in real time by combining the cumulative displacement of the inertial navigation system with the visual error correction results. At the same time, a visual feature correction mechanism is introduced. When road nodes with matching absolute coordinates, such as intersections and traffic signs, are identified, the absolute position error accumulated by the inertial navigation system is corrected to avoid inaccurate target point mapping after long-distance travel. Finally, the navigation target point is iteratively mapped to the current local road network topology.

[0032] The global path planning submodule 10 uses the weighted cost A search algorithm to perform a heuristic search in the feasible node space of the local road network topology to generate a global navigation route from the vehicle's current position to the navigation target point. This route includes a sequence of lane units passed through in sequence and the corresponding turning and lane changing operation sequences.

[0033] In this embodiment, the cost function of the weighted cost A algorithm combines three weighting factors. The total cost = travel distance cost × 1.0 + lane change frequency cost × 20 + road traffic priority cost × travel distance. The specific rules are as follows: (1) Travel distance cost: The total estimated travel distance of the route; the shorter the distance, the lower the cost. (2) Lane change cost: The number of lane changes in the path. The fewer the number of lane changes, the lower the cost, reducing the complexity of driving operations. A single lane change corresponds to an equivalent distance cost of 20m. (3) Road traffic priority cost: Traffic priority is determined based on the road features identified by visual recognition. The more lanes in the same direction and the higher the road grade, the higher the priority and the lower the corresponding weight coefficient. The priorities from high to low are: main roads with ≥3 lanes in the same direction (weight 0.8), secondary arterial roads with 2 lanes in the same direction (weight 1.0), single-lane branch roads in one direction (weight 1.2), and non-motorized vehicle lanes (weight 1.5).

[0034] During vehicle operation, the global path planning submodule 10 completes a route search iteration every 100ms, and optimizes and updates the global navigation route in real time by combining the latest local road network topology data to ensure that the route always conforms to the latest road conditions.

[0035] 5.2 Local Dynamic Adjustment and Correction The local dynamic adjustment submodule 11 monitors the dynamic road condition changes identified by vision in real time. When any traffic restriction trigger condition is detected, local route replanning is immediately initiated. In this embodiment, the trigger conditions for local dynamic adjustment specifically include four categories: (1) In the lane corresponding to the global navigation route, static obstacles occupy more than 80% of the lane, making the lane completely impassable; (2) Dynamic obstacles continue to encroach on the driving lane, and it is expected that the traffic efficiency of vehicles passing through the area will decrease by more than 50%; (3) Visual recognition of temporary construction areas, lane closure signs, temporary traffic control, etc., resulting in the interruption of the original route; (4) If traffic accidents or congestion occur on the road ahead, causing the estimated travel time of the original route to increase by more than 100%, after triggering a local adjustment, the local dynamic adjustment submodule 11 takes the current position of the vehicle as the starting point and the next topological node of the global navigation route as the temporary endpoint. Based on the remaining feasible lanes and feasible domain, it uses the dynamic window method to quickly plan and generate a local obstacle avoidance and rerouting route. At the same time, it reconnects the endpoint of the local route with the global navigation route to complete the update of the overall global navigation route.

[0036] In this embodiment, the response time of local dynamic adjustment is less than 50ms, which can realize rapid response and route correction to sudden road conditions, ensuring navigation continuity and driving safety.

[0037] Step 6: Dynamic navigation guidance generation and output The navigation guidance generation module 12 receives the final updated global navigation route and local adjustment information, and calculates the timing of advance prompts for various navigation operations based on the vehicle's real-time driving speed: for turning operations, a voice prompt is issued 100-300m in advance based on the current vehicle speed, with the prompt distance increasing as the vehicle speed increases; for lane changing operations, a prompt is issued 50-150m in advance; and for deceleration and obstacle avoidance operations, prompts are provided in stages based on the distance to the obstacle and the current vehicle speed.

[0038] The dynamic navigation guidance instructions generated by the navigation guidance generation module 12 include two parts: first, voice broadcast content, which clearly informs the driver of the next operation instructions and precautions; second, visual guidance information, including lane-level guide arrows, obstacle location markers, route diagrams, etc. Finally, the dynamic navigation guidance instructions are transmitted to the human-machine interaction output module 13. In this embodiment, the human-machine interaction output module 13 includes an in-vehicle central control display screen and an in-vehicle voice system. The in-vehicle central control display screen displays a visual navigation interface, and the in-vehicle voice system broadcasts navigation guidance information, providing the driver with clear and intuitive navigation guidance.

[0039] Throughout the vehicle's journey, the above steps are executed cyclically, sensors continuously collect data, local road networks are dynamically updated in real time, route search is continuously iterated and optimized, and navigation guidance is dynamically adjusted, ultimately achieving continuous and accurate in-vehicle navigation in scenarios without pre-stored electronic maps.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI image processing-enabled vehicle-mounted visual inertial navigation method, comprising the steps of acquiring visual images of the road scene in front of the vehicle through an onboard visual acquisition module (1), synchronously acquiring six-axis inertial motion data of the vehicle through an inertial measurement unit (2), and performing strapdown calculation on the six-axis inertial motion data through an inertial navigation data processing module (7) to obtain real-time vehicle pose data, characterized in that, The method further includes the following steps: S1. Real-time local road network topology construction: The AI ​​image processing module (3) performs semantic analysis and road feature extraction on the visual image of the road scene. The real-time local road network topology model covering the driving range in front of the vehicle is constructed with the current position of the vehicle as the origin. The local road network topology model continues to expand forward as the vehicle moves and completes inter-frame stitching correction. S2, Road Network Pose Fusion Registration: The pose fusion submodule (8) spatially registers the real-time pose data of the vehicle with the real-time local road network topology model, and corrects the inertial navigation cumulative error by using the visually extracted lane line features as the observation benchmark, so as to obtain the accurate pose mapping result of the vehicle in the local road network coordinate system. S3, Double-layer iterative route search: The route search engine module (9) takes the navigation target point in absolute geographic coordinates as the endpoint, combines the inertial navigation cumulative displacement and the vehicle absolute position corrected by visual features, iteratively maps the navigation target point to the continuously expanding local road network topology, and performs global optimal search and local dynamic adjustment based on the pose mapping result and the local road network topology model to generate a complete vehicle navigation route. S4. Dynamic navigation guidance output: The navigation guidance generation module (12) calculates the timing of prompts by combining the vehicle's real-time speed and navigation route, generates dynamic navigation guidance instructions, and outputs them to the driver through the human-machine interaction output module (13).

2. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 1, characterized in that, Step S1 specifically includes: S11, Pixel-level semantic segmentation: The road scene visual image is segmented into pixels by the semantic segmentation submodule (4) in the AI ​​image processing module (3), and a road semantic segmentation map containing 12 types of road elements is output. S12, Multi-dimensional road feature extraction: The road feature extraction submodule (5) extracts lane geometric features, traffic rule features, obstacle features and topological connectivity features from the road semantic segmentation map; S13. Local road network topology construction and splicing: The local road network topology model is constructed based on the extracted road features by the local road network topology construction sub-module (6). The monocular visual scale drift is corrected by combining inertial navigation attitude data. The road network splicing of adjacent frames is completed by inter-frame feature matching. The accumulated topology error is corrected by closed-loop detection.

3. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 2, characterized in that, The real-time local road network topology model uses lane units as topology nodes and lane connections as topology edges. Each node is equipped with attributes such as lane width, speed limit, steering restriction, and obstacle occupancy status. The model coverage distance is dynamically adjusted according to vehicle speed: 200m when the vehicle speed is below 60km / h, 350m when the vehicle speed is between 60-100km / h, and 500m when the vehicle speed is above 100km / h. A hysteresis interval of 5km / h is set for the vehicle speed threshold.

4. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 1, characterized in that, The specific steps of step S2 are as follows: the pose fusion submodule (8) uses the lane line features extracted by semantic segmentation as the external observation benchmark, matches the vehicle heading angle calculated by the inertial navigation with the lane line extension direction to obtain the heading deviation, matches the vehicle lateral position with the lane centerline position to obtain the lateral position deviation, and uses the extended Kalman filter algorithm to perform optimal estimation and correction of the position and attitude error of the inertial navigation to obtain the pose mapping result of the vehicle in the local road network coordinate system.

5. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 1, characterized in that, The route search engine module (9) includes a global path planning submodule (10) and a local dynamic adjustment submodule (11). Step S3 specifically includes: S31, Global Optimal Route Search: The global route planning submodule (10) uses the weighted cost A* search algorithm to search for a global navigation route from the vehicle's current location to the navigation target point in the feasible node space of the local road network topology; as the local road network topology continues to expand, a global route iteration optimization is completed every 100ms; S32, Local dynamic adjustment and correction: The local dynamic adjustment submodule (11) monitors the dynamic road conditions in real time. When the global navigation route is restricted, the local obstacle avoidance and road change route is replanned with the current position of the vehicle as the starting point and the next node of the global route as the temporary end point, and the global navigation route is updated.

6. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 5, characterized in that, The cost function of the weighted cost A* search algorithm includes three weighting factors: travel distance cost, lane change frequency cost, and road traffic priority cost; The road traffic priority is determined based on the road features identified by visual recognition. The more lanes in the same direction, the higher the priority and the lower the corresponding weight coefficient. The priorities from high to low are: main roads with ≥3 lanes in the same direction, secondary arterial roads with 2 lanes in the same direction, one-way single-lane branch roads, and non-motorized vehicle lanes, with corresponding weight coefficients of 0.8, 1.0, 1.2, and 1.5, respectively.

7. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 5, characterized in that, The triggering conditions for the local dynamic adjustment include: (1) The proportion of static obstacles in the lanes of the global route exceeds 80%, making the lanes completely impassable; (2) Dynamic obstacles continuously encroach on the driving lane and are expected to reduce traffic efficiency by more than 50%; (3) Identify that temporary construction, lane closures, or traffic control have caused the original route to be interrupted; (4) The road ahead is congested, which will increase the estimated travel time of the original route by more than double.

8. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 1, characterized in that, The specific steps of step S4 are as follows: the navigation guidance generation module (12) calculates the advance warning distance for steering, lane changing and deceleration operations based on the real-time driving speed of the vehicle, generates dynamic navigation guidance instructions containing voice broadcast content and lane-level visual guidance information, and outputs them synchronously with the voice system through the vehicle display screen of the human-machine interaction output module (13).

9. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 1, characterized in that, The vehicle-mounted visual acquisition module (1) and the inertial measurement unit (2) achieve time synchronization through the vehicle-mounted hardware clock, with a synchronization accuracy of not less than 1ms. Each frame of visual image corresponds to a unique timestamp and is matched with the inertial data at the same moment.

10. The AI ​​image processing-enabled vehicle vision-inertial navigation method according to claim 2, characterized in that, The update mechanism of the local road network topology model is as follows: For each frame of road image acquired by the vehicle-mounted vision acquisition module (1), the local road network topology construction submodule (6) completes a front-end road network expansion and back-end node clipping that exceeds the range, and simultaneously completes inter-frame stitching and error correction to realize continuous dynamic update of the road network.

Images