Navigation guidance information intelligent generation method, device, equipment and storage medium

Abstract

The application discloses a kind of navigation induction information intelligent generation method, device, equipment and storage medium, the method is by obtaining the map data containing lane level information, according to map data construction structured semantic graph, structured semantic graph includes key decision point information, lane action sequence and the alignment relationship of landmark and action;Structured semantic graph is input to natural language generation model, generates initial induction text, and executes safety check to initial induction text to determine safe induction text;Obtain user feature data, and according to user feature data, the safe induction text is personalized adjusted, generates final broadcast text and controls navigation equipment to broadcast output, can provide accurate structured context basis, eliminate the safety risk caused by model illusion, realize navigation broadcast on the basis of ensuring safety and accuracy, provide personalized experience adapted to different driving habits.

Description

Technical Field

[0001] This invention relates to the field of intelligent conference control technology, and in particular to a method, apparatus, device and storage medium for intelligent generation of navigation and guidance information. Background Technology

[0002] As navigation electronic maps evolve from traditional road-level maps to high-precision lane-level maps, the way navigation guidance information is generated faces fundamental challenges.

[0003] Traditional navigation systems use a "preset template + parameter filling" broadcast generation mode, such as "{distance} ahead {turning action}", and their guidance information has the following three structural defects: (1) Single information dimension: It only outputs distance-action pairs and lacks the fusion expression of multi-dimensional spatial information such as lane topology, landmark reference, and traffic situation; (2) Fixed expression pattern: All drivers receive the exact same broadcast text at the same location, and it is impossible to dynamically adjust the information density and expression style according to the individual cognitive level of the driver and the real-time scene; (3) Lack of safety mechanism: When artificial intelligence (AI) is used to generate models, model illusion may lead to deviations in key safety information such as incorrect steering direction and incorrect lane number, and existing solutions lack a systematic verification mechanism.

[0004] The aforementioned defects are particularly prominent in complex intersection scenarios, which may cause drivers to miss intersections, make emergency lane changes, or violate traffic rules, thus affecting driving safety.

[0005] Existing technologies analyze user queries, combine points of interest (POIs) with road network data, and use language models to generate navigation text containing landmarks, replacing the traditional mechanical broadcasts without landmarks.

[0006] Its technical defects are as follows: (a) It only addresses the issue of “enriching the broadcast content in the landmark dimension”, without involving a deep understanding of the intersection topology; the solution cannot parse the complete lane action sequence between “current lane → target exit lane”, and is not suitable for complex intersections that require continuous lane changes; (b) The generated content is still template-oriented: the landmark information is inserted in a fixed sentence pattern and fails to generate a guide that conforms to the natural language expression habits and has contextual coherence; (c) Lack of adaptability: The broadcast content is not tailored to all drivers based on their driving skill level and historical behavior data.

[0007] Existing technologies can utilize deep learning models to directly map map features to navigation text in an end-to-end manner, generating natural language instructions.

[0008] Its technical defects: (a) The model is uninterpretable and has the risk of factual errors: the end-to-end black box model may output turning instructions that are inconsistent with the real road topology (such as generating a left turn as a right turn), which is unacceptable in safety-critical scenarios; (b) Lack of verifiable mechanism: The generated results are output directly without any verification process to ensure the consistency between the generated content and the map's true value; (c) Uniform output: Only one text is generated, which cannot be differentiated according to the cognitive level and preferences of different drivers. Summary of the Invention

[0009] The main objective of this invention is to provide an intelligent method, apparatus, device, and storage medium for generating navigation guidance information, aiming to solve the technical problems of existing navigation guidance information generation methods, such as single information dimension, fixed expression mode, lack of security verification mechanism, and lack of personalized adaptation.

[0010] In a first aspect, the present invention provides a method for intelligently generating navigation guidance information, the method comprising the following steps: Acquire map data containing lane-level information, and construct a structured semantic graph based on the map data. The structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial lead text, and a security check is performed on the initial lead text to determine a secure lead text. The system acquires user characteristic data, personalizes the safety guidance text based on the user characteristic data, generates the final broadcast text, and controls the navigation device to broadcast it.

[0011] Optionally, the step of acquiring map data containing lane-level information and constructing a structured semantic graph based on the map data, wherein the structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions, including: Traverse the lane connection relationships on the vehicle's driving path, determine the node type based on the number of subsequent lanes, the number of front drive lanes, and the vehicle heading angle deviation, and use nodes with branching points, merging points, or turning points as the key decision point information. The system searches forward along the lane connections to the target exit lane, and constructs the lane action sequence by comparing the lateral offset of the center lines of adjacent lanes. The lane action sequence includes the lane change direction and the number of lane changes. For each lane change or turning action node, surrounding points of interest are retrieved with a preset search radius. Based on the street-facing attributes and building outline data of the POI, points of interest located in back streets, obscured by buildings, or not directly observable from the road are filtered out. Retain street-facing points of interest with a popularity rating higher than a preset rating threshold as target points of interest; Calculate the spatial orientation of the target point of interest relative to the action point, convert the spatial orientation into natural language spatial relation words, and form the alignment relationship between the landmark and the action; A structured semantic graph is constructed based on the key decision point information, the lane action sequence, and the alignment relationship.

[0012] Optionally, the process of traversing the lane connections along the vehicle's travel path, determining the node type based on the number of subsequent lanes, the number of front-wheel drive lanes, and the heading angle deviation, and using nodes with branching points, merging points, or turning points as the key decision point information includes: Iterate through the lane connections along the vehicle's travel path; When the number of subsequent lanes is greater than the number of preceding lanes, the node type is determined as a branch point; When the number of subsequent lanes is less than the number of preceding lanes, the node type is determined as a merging point; When the deviation of the vehicle heading angle is greater than a preset angle threshold, the node type is determined as a turning point; Nodes with branching points, merging points, or turning points are used as key decision point information.

[0013] Optionally, the step of inputting the structured semantic graph into a natural language generation model to generate initial lead text, and performing security checks on the initial lead text to determine secure lead text, includes: The structured semantic graph is organized into structured prompt words, and the structured prompt words are input into the natural language generation model to guide the natural language generation model to generate initial inductive text and corresponding structured action annotations through thought chain reasoning. The turning action words, lane number descriptions, and distance values ​​in the initial guidance text are extracted and compared deterministically with the original map data to perform the first level of verification. When the first-level verification fails, the backup broadcast engine is triggered to generate backup guidance information, and the backup guidance information is identified as the security guidance text. When the first-level verification passes, the existence and location rationality of the landmarks referenced in the initial inducement text are verified to perform the second-level verification, and the initial inducement text that passes the second-level verification is determined as the security inducement text.

[0014] Optionally, when the first-level verification passes, verifying the existence and location rationality of the referenced landmarks in the initial inducement text to perform a second-level verification, and determining the initial inducement text that passes the second-level verification as the security inducement text, includes: If the first-level verification passes, the map database is queried to verify whether the referenced landmarks in the initial guidance text exist. If the referenced landmark does not exist, or if the referenced landmark exists but the calculated deviation between the location described in the initial guiding text and the actual location in the map database exceeds a preset distance threshold, the second-level verification is deemed to have failed, and the description of the referenced landmark is deleted or the description of the referenced landmark is corrected to a distance description. When the referenced landmark exists and the deviation distance does not exceed the preset distance threshold, the second-level verification is determined to be successful, and the initial inducement text that has passed the second-level verification is determined as the security inducement text.

[0015] Optionally, the step of acquiring user feature data, personalizing the safety guidance text based on the user feature data, generating the final broadcast text, and controlling the navigation device to broadcast the text includes: Acquire user characteristic data and parse the user characteristic data to determine the user's driving ability level and broadcast style preference; Adjust the information density of the safety guidance text according to the driving ability level; When the driving ability level is the preset novice level, the predictive reminders and step-by-step action breakdown are retained, and emotional reassurance statements are added. When the driving ability level is the preset proficient driving level, the prediction reminders and supplementary explanations are removed and simplified to the core action instructions; Based on the broadcast style preference, the safety guidance text after adjusting the information density is stylized and rewritten to generate the final broadcast text; The final broadcast text is sent to the in-vehicle voice synthesis module for voice broadcast, and the navigation device is controlled to display lane-level turning visual indication signals synchronized with the final broadcast text.

[0016] Optionally, after acquiring user feature data, personalizing the safety guidance text based on the user feature data, generating the final broadcast text, and controlling the navigation device to broadcast the text, the intelligent generation method for navigation guidance information further includes: After the navigation device outputs a broadcast message, the consistency between the vehicle's driving trajectory and the guidance instructions in the final broadcast text is monitored. When a negative feedback event is detected, the current sample is marked as a negative feedback sample. The negative feedback event includes sudden braking or the vehicle entering the wrong lane and then correcting itself within a preset time after the broadcast. The negative feedback samples are used to adjust the generation strategy of the natural language generation model or update the user feature data.

[0017] Secondly, to achieve the above objectives, the present invention also proposes an intelligent navigation guidance information generation device, the intelligent navigation guidance information generation device comprising: The map construction module is used to acquire map data containing lane-level information and construct a structured semantic map based on the map data. The structured semantic map includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The security verification module is used to input the structured semantic graph into the natural language generation model, generate initial induction text, and perform security verification on the initial induction text to determine the secure induction text; The broadcast control module is used to acquire user characteristic data, personalize the safety guidance text based on the user characteristic data, generate the final broadcast text, and control the navigation device to broadcast it.

[0018] Thirdly, to achieve the above objectives, the present invention also proposes an intelligent navigation guidance information generation device, which includes: a memory, a processor, and an intelligent navigation guidance information generation program stored in the memory and executable on the processor. The intelligent navigation guidance information generation program is configured to implement the steps of the intelligent navigation guidance information generation method described above.

[0019] Fourthly, to achieve the above objectives, the present invention also proposes a storage medium storing a navigation guidance information intelligent generation program, wherein the navigation guidance information intelligent generation program, when executed by a processor, implements the steps of the navigation guidance information intelligent generation method described above.

[0020] The intelligent generation method for navigation guidance information proposed in this invention acquires map data containing lane-level information, constructs a structured semantic graph based on the map data, and the structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial guidance text, and a safety check is performed on the initial guidance text to determine safe guidance text. User feature data is acquired, and the safe guidance text is personalized based on the user feature data to generate the final broadcast text, which is then broadcast by the navigation device. This method provides accurate structured context for the natural language generation model by constructing a structured semantic graph containing key decision points, lane action sequences, and alignment relationships between landmarks and actions, thereby generating guidance text that is information-rich and naturally expressed. Simultaneously, by performing safety checks on the initial guidance text, the safety risks caused by model illusions are eliminated, and personalized adjustments based on user feature data ensure that navigation broadcasts provide a personalized experience adapted to different driving habits while ensuring safety and accuracy. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the intelligent generation method for navigation guidance information of the present invention; Figure 3 This is a flowchart illustrating the second embodiment of the intelligent generation method for navigation guidance information of the present invention; Figure 4 This is a flowchart illustrating the third embodiment of the intelligent generation method for navigation guidance information of the present invention; Figure 5 This is a flowchart illustrating the fourth embodiment of the intelligent generation method for navigation guidance information of the present invention; Figure 6 This is a flowchart illustrating the fifth embodiment of the intelligent generation method for navigation guidance information of the present invention; Figure 7 This is a functional block diagram of the first embodiment of the intelligent navigation guidance information generation device of the present invention.

[0022] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] The solution of this invention mainly involves: acquiring map data containing lane-level information; constructing a structured semantic graph based on the map data, wherein the structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions; inputting the structured semantic graph into a natural language generation model to generate initial guidance text, and performing security checks on the initial guidance text to determine safe guidance text; acquiring user feature data, and making personalized adjustments to the safe guidance text based on the user feature data to generate the final broadcast text and control the navigation device to broadcast it. This solution provides a precise structured context for the natural language generation model by constructing a structured semantic graph containing key decision points, lane action sequences, and alignment relationships between landmarks and actions, thereby generating guidance text with rich information dimensions and natural expression; simultaneously, by performing security checks on the initial guidance text, the safety risks caused by model illusions are eliminated, and personalized adjustments are made based on user feature data, enabling navigation broadcasts to provide a personalized experience adapted to different driving habits while ensuring safety and accuracy. This solves the technical problems of existing navigation guidance information generation methods, such as single information dimensions, fixed expression patterns, lack of security checks, and lack of personalized adaptation.

[0025] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0026] like Figure 1 As shown, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0027] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0028] like Figure 1As shown, the memory 1005, which serves as a storage medium, may include an operating device, a network communication module, a user interface module, and a navigation guidance information intelligent generation program.

[0029] The device of the present invention calls the intelligent generation program for navigation guidance information stored in the memory 1005 through the processor 1001, and performs the following operations: Acquire map data containing lane-level information, and construct a structured semantic graph based on the map data. The structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial lead text, and a security check is performed on the initial lead text to determine a secure lead text. The system acquires user characteristic data, personalizes the safety guidance text based on the user characteristic data, generates the final broadcast text, and controls the navigation device to broadcast it.

[0030] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: Traverse the lane connection relationships on the vehicle's driving path, determine the node type based on the number of subsequent lanes, the number of front drive lanes, and the vehicle heading angle deviation, and use nodes with branching points, merging points, or turning points as the key decision point information. The system searches forward along the lane connections to the target exit lane, and constructs the lane action sequence by comparing the lateral offset of the center lines of adjacent lanes. The lane action sequence includes the lane change direction and the number of lane changes. For each lane change or turning action node, surrounding points of interest are retrieved with a preset search radius. Based on the street-facing attributes and building outline data of the POI, points of interest located in back streets, obscured by buildings, or not directly observable from the road are filtered out. Retain street-facing points of interest with a popularity rating higher than a preset rating threshold as target points of interest; Calculate the spatial orientation of the target point of interest relative to the action point, convert the spatial orientation into natural language spatial relation words, and form the alignment relationship between the landmark and the action; A structured semantic graph is constructed based on the key decision point information, the lane action sequence, and the alignment relationship.

[0031] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: Iterate through the lane connections along the vehicle's travel path; When the number of subsequent lanes is greater than the number of preceding lanes, the node type is determined as a branch point; When the number of subsequent lanes is less than the number of preceding lanes, the node type is determined as a merging point; When the deviation of the vehicle heading angle is greater than a preset angle threshold, the node type is determined as a turning point; Nodes with branching points, merging points, or turning points are used as key decision point information.

[0032] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: The structured semantic graph is organized into structured prompt words, and the structured prompt words are input into the natural language generation model to guide the natural language generation model to generate initial inductive text and corresponding structured action annotations through thought chain reasoning. The turning action words, lane number descriptions, and distance values ​​in the initial guidance text are extracted and compared deterministically with the original map data to perform the first level of verification. When the first-level verification fails, the backup broadcast engine is triggered to generate backup guidance information, and the backup guidance information is identified as the security guidance text. When the first-level verification passes, the existence and location rationality of the landmarks referenced in the initial inducement text are verified to perform the second-level verification, and the initial inducement text that passes the second-level verification is determined as the security inducement text.

[0033] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: If the first-level verification passes, the map database is queried to verify whether the referenced landmarks in the initial guidance text exist. If the referenced landmark does not exist, or if the referenced landmark exists but the calculated deviation between the location described in the initial guiding text and the actual location in the map database exceeds a preset distance threshold, the second-level verification is deemed to have failed, and the description of the referenced landmark is deleted or the description of the referenced landmark is corrected to a distance description. When the referenced landmark exists and the deviation distance does not exceed the preset distance threshold, the second-level verification is determined to be successful, and the initial inducement text that has passed the second-level verification is determined as the security inducement text.

[0034] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: Acquire user characteristic data and parse the user characteristic data to determine the user's driving ability level and broadcast style preference; Adjust the information density of the safety guidance text according to the driving ability level; When the driving ability level is the preset novice level, the predictive reminders and step-by-step action breakdown are retained, and emotional reassurance statements are added. When the driving ability level is the preset proficient driving level, the prediction reminders and supplementary explanations are removed and simplified to the core action instructions; Based on the broadcast style preference, the safety guidance text after adjusting the information density is stylized and rewritten to generate the final broadcast text; The final broadcast text is sent to the in-vehicle voice synthesis module for voice broadcast, and the navigation device is controlled to display lane-level turning visual indication signals synchronized with the final broadcast text.

[0035] The device of the present invention, through processor 1001 calling the intelligent generation program for navigation guidance information stored in memory 1005, also performs the following operations: After the navigation device outputs a broadcast message, the consistency between the vehicle's driving trajectory and the guidance instructions in the final broadcast text is monitored. When a negative feedback event is detected, the current sample is marked as a negative feedback sample. The negative feedback event includes sudden braking or the vehicle entering the wrong lane and then correcting itself within a preset time after the broadcast. The negative feedback samples are used to adjust the generation strategy of the natural language generation model or update the user feature data.

[0036] This embodiment, through the above-described scheme, acquires map data containing lane-level information, constructs a structured semantic graph based on the map data, and the structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial guidance text, and a safety check is performed on the initial guidance text to determine safe guidance text. User feature data is acquired, and the safe guidance text is personalized based on the user feature data to generate the final broadcast text, which is then controlled by the navigation device for broadcast output. By constructing a structured semantic graph containing key decision points, lane action sequences, and alignment relationships between landmarks and actions, accurate structured contextual basis is provided for the natural language generation model, thereby generating guidance text with rich information dimensions and natural expression. Simultaneously, by performing safety checks on the initial guidance text, safety risks caused by model illusions are eliminated, and personalized adjustments are made based on user feature data, enabling navigation broadcasts to provide a personalized experience adapted to different driving habits while ensuring safety and accuracy.

[0037] Based on the above hardware structure, an embodiment of the intelligent generation method for navigation guidance information of the present invention is proposed.

[0038] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the intelligent generation method for navigation guidance information according to the present invention.

[0039] In the first embodiment, the intelligent generation method for navigation guidance information includes the following steps: Step S10: Obtain map data containing lane-level information, and construct a structured semantic graph based on the map data. The structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions.

[0040] It should be noted that road connectivity and attribute status are obtained using map data containing lane-level information, and a semantic graph is generated using a structured construction method. Based on the key decision point information, lane action sequences, and alignment relationships between landmarks and actions contained in the structured semantic graph, the map data is transformed from raw geometric information into understandable semantic information in a structured manner.

[0041] Step S20: Input the structured semantic graph into the natural language generation model to generate initial inductive text, and perform security verification on the initial inductive text to determine secure inductive text.

[0042] It should be understood that by using structured semantic graphs as input data to provide structured context, the reasoning ability of natural language generation models is used to generate initial guidance text, and the security verification of key action information is performed on the initial guidance text to determine the safe guidance text, thereby achieving the flexibility of natural language generation of navigation guidance information and the security and reliability of key instructions.

[0043] Step S30: Obtain user feature data, personalize the safety guidance text based on the user feature data, generate the final broadcast text, and control the navigation device to broadcast it.

[0044] Understandably, by using user characteristic data to characterize users' driving habits and preferences, and employing personalized adjustment strategies to adapt safety guidance text, the final broadcast text is generated based on the adjustment results and the navigation device is controlled to broadcast the text, thereby achieving personalized and accurate broadcasting of navigation guidance information for different driving users.

[0045] This embodiment, through the above-described scheme, acquires map data containing lane-level information, constructs a structured semantic graph based on the map data, and the structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial guidance text, and a safety check is performed on the initial guidance text to determine safe guidance text. User feature data is acquired, and the safe guidance text is personalized based on the user feature data to generate the final broadcast text, which is then controlled by the navigation device for broadcast output. By constructing a structured semantic graph containing key decision points, lane action sequences, and alignment relationships between landmarks and actions, accurate structured contextual basis is provided for the natural language generation model, thereby generating guidance text with rich information dimensions and natural expression. Simultaneously, by performing safety checks on the initial guidance text, safety risks caused by model illusions are eliminated, and personalized adjustments are made based on user feature data, enabling navigation broadcasts to provide a personalized experience adapted to different driving habits while ensuring safety and accuracy.

[0046] Furthermore, Figure 3 This is a flowchart illustrating the second embodiment of the intelligent generation method for navigation guidance information of the present invention, as shown below. Figure 3 As shown, based on the first embodiment, a second embodiment of the intelligent generation method for navigation guidance information of the present invention is proposed. In this embodiment, step S10 specifically includes the following steps: Step S11: Traverse the lane connection relationships on the vehicle's driving path, determine the node type based on the number of subsequent lanes, the number of front drive lanes, and the vehicle heading angle deviation, and use nodes with branching points, merging points, or turning points as the key decision point information.

[0047] It should be noted that the road topology is obtained by utilizing the lane connection relationship on the vehicle's driving path, and the lane change characteristics are analyzed by using the node type determination logic. The node type is determined based on the number of subsequent lanes, the number of front drive lanes, and the vehicle heading angle deviation. Nodes with branching points, merging points, or turning points are used as key decision point information to achieve accurate identification and structured marking of key decision point information.

[0048] Furthermore, step S11 specifically includes the following steps: Iterate through the lane connections along the vehicle's travel path; When the number of subsequent lanes is greater than the number of preceding lanes, the node type is determined as a branch point; When the number of subsequent lanes is less than the number of preceding lanes, the node type is determined as a merging point; When the deviation of the vehicle heading angle is greater than a preset angle threshold, the node type is determined as a turning point; Nodes with branching points, merging points, or turning points are used as key decision point information.

[0049] It should be understood that the road topology is obtained by utilizing the lane connection relationship on the vehicle's driving path, and the changes in the number of lanes and directional deviations are analyzed using node type determination logic. When the number of subsequent lanes is greater than the number of front-wheel drive lanes, it is determined as a branch point; when the number of subsequent lanes is less than the number of front-wheel drive lanes, it is determined as a merging point; and when the vehicle heading angle deviation is greater than a preset angle threshold, it is determined as a turning point. Nodes with branch point, merging point, or turning point types are used as the key decision point information to achieve quantitative determination and accurate identification of key decision point types.

[0050] Step S12: Search forward along the lane connection relationship to the target exit lane, and construct the lane action sequence by comparing the lateral offset of the center lines of adjacent lanes. The lane action sequence includes the lane change direction and the number of lane changes.

[0051] It is understandable that the search path is determined by using lane connection relationships and target exit lane information, the lane geometric features are analyzed by using the lateral offset comparison method, and the lane action sequence is constructed based on the lateral offset of the center lines of adjacent lanes. The lane action sequence includes the lane change direction and the number of lane changes, thereby realizing the quantitative description and structured construction of the lane action sequence.

[0052] Step S13: For each lane change or turning action node, retrieve surrounding points of interest with a preset search radius. Based on the street-facing attributes and building outline data of the POI, filter out points of interest located on back streets, obscured by buildings, or not directly observable from the road.

[0053] It should be understood that by using lane change or turning action nodes as the retrieval center to obtain surrounding points of interest data, and using a preset search radius for spatial retrieval, points of interest located on back streets, obscured by buildings, or not directly observable from the road are filtered out based on the street-facing attributes and building outline data of the POI, thus achieving accurate filtering of the visibility of points of interest in the alignment relationship between landmarks and actions.

[0054] Step S14: Retain the points of interest that are located on the street and have a popularity score higher than the preset score threshold as target points of interest.

[0055] Understandably, the street-facing attributes and popularity scores of surrounding points of interest are used to obtain characteristic information of points of interest. The point of interest selection logic is used to evaluate the quality of the points of interest. Points of interest that are street-facing and have a popularity score higher than a preset score threshold are retained as target points of interest, thus ensuring the salience and recognizability of target points of interest in the alignment relationship between landmarks and actions.

[0056] Step S15: Calculate the spatial orientation of the target point of interest relative to the action point, convert the spatial orientation into natural language spatial relation words, and form the alignment relationship between the landmark and the action.

[0057] It should be understood that by utilizing the location information of the target point of interest and the action point to obtain the spatial relative state, and using spatial orientation calculation and semantic conversion methods to process geospatial data, the spatial orientation is converted into natural language spatial relation words according to the spatial orientation of the target point of interest relative to the action point, forming the alignment relationship between the landmark and the action, and realizing the accurate spatial semantic association between the landmark description and the driving action.

[0058] Step S16: Construct a structured semantic graph based on the key decision point information, the lane action sequence, and the alignment relationship.

[0059] It is understandable that by using key decision point information, lane action sequences, and alignment relationships to obtain core elements of the map, and employing a structured integration method to organize multi-source semantic data, a structured semantic map is constructed based on the key decision point information, lane action sequences, and alignment relationships, thereby achieving complete expression and standardized encapsulation of navigation path semantic information.

[0060] This embodiment, through the above-described scheme, traverses the lane connections along the vehicle's travel path, determines node types based on the number of subsequent lanes, the number of forward-drive lanes, and the vehicle heading angle deviation, and uses nodes with branching points, merging points, or turning points as key decision point information; it searches forward along the lane connections to the target exit lane, constructing the lane action sequence by comparing the lateral offset of the centerlines of adjacent lanes, the lane action sequence including lane change direction and number of lane changes; for each lane change or turning action node, it retrieves surrounding points of interest with a preset search radius, filtering out points of interest located on back streets, obscured by buildings, or not directly observable from the road based on the street-facing attributes and building outline data of the POIs; it retains points of interest that are street-facing and have a visibility score higher than a preset score threshold. Points are used as target points of interest; the spatial orientation of the target points of interest relative to action points is calculated, and the spatial orientation is converted into natural language spatial relation words to form the alignment relationship between the landmark and the action; a structured semantic graph is constructed based on the key decision point information, the lane action sequence, and the alignment relationship; key decision points and lane action sequences can be accurately identified through lane topology analysis and geometric feature calculation, and the reliability and recognizability of landmark information are ensured by combining a point of interest filtering mechanism based on visibility and popularity, transforming the original map data into a structured semantic graph containing spatial semantic relationships, providing accurate, safe, and structured contextual basis for subsequent natural language generation models, thereby reducing the difficulty of model understanding and reducing the risk of illusion caused by data noise.

[0061] Furthermore, Figure 4 This is a flowchart illustrating the third embodiment of the intelligent generation method for navigation guidance information of the present invention, as shown below. Figure 4 As shown, based on the first embodiment, a third embodiment of the intelligent generation method for navigation guidance information of the present invention is proposed. In this embodiment, step S20 specifically includes the following steps: Step S21: Organize the structured semantic graph into structured prompt words, input the structured prompt words into the natural language generation model, and guide the natural language generation model to generate initial inductive text and corresponding structured action annotations through thought chain reasoning.

[0062] It should be noted that the structured context information of the navigation path is obtained by using a structured semantic graph, and structured prompt words are constructed using a prompt word organization method. The structured prompt words are then input into a natural language generation model, which is then guided to generate initial guidance text and corresponding structured action annotations through thought chain reasoning. This achieves flexible natural language generation of navigation guidance information and structured synchronous annotation of key actions.

[0063] Step S22: Extract the turning action words, lane number descriptions, and distance values ​​from the initial guidance text, and perform a deterministic comparison with the original map data to perform the first level of verification.

[0064] Understandably, the initial guidance text is used to obtain the text content to be verified, and the navigation instruction elements are processed using key information extraction and deterministic comparison methods. Based on the turn action words, lane number descriptions and distance values ​​extracted from the initial guidance text, a deterministic comparison is performed with the original map data to perform the first level of verification, thereby realizing the accuracy verification of key navigation instruction information and the avoidance of safety risks.

[0065] Step S23: When the first-level verification fails, the backup broadcast engine is triggered to generate backup guidance information, and the backup guidance information is determined as the security guidance text.

[0066] It should be understood that the verification status of the initial guidance text is obtained by using the result of the first-level verification, and the verification failure is handled by the backup broadcast engine triggering mechanism. When the first-level verification fails, the backup broadcast engine is triggered to generate backup guidance information, and the backup guidance information is determined as the safety guidance text. This achieves a safety risk protection and reliable output of guidance information when navigation guidance information generation fails.

[0067] Step S24: When the first-level verification passes, verify the existence and location rationality of the referenced landmarks in the initial inducement text to perform the second-level verification, and determine the initial inducement text that passes the second-level verification as the security inducement text.

[0068] It is understandable that by using the initial guidance text that has passed the first level of verification to obtain the landmark description information to be verified, the semantic verification process is processed using the landmark existence and location rationality verification logic. Based on the verification of the existence and location rationality of the landmarks referenced in the initial guidance text when the first level of verification is passed, the second level of verification is performed. The initial guidance text that has passed the second level of verification is determined as the safe guidance text, thereby realizing the verification of the authenticity of the landmark references in the navigation guidance information and the guarantee of the spatial logic rationality.

[0069] Furthermore, step S24 specifically includes the following steps: If the first-level verification passes, the map database is queried to verify whether the referenced landmarks in the initial guidance text exist. If the referenced landmark does not exist, or if the referenced landmark exists but the calculated deviation between the location described in the initial guiding text and the actual location in the map database exceeds a preset distance threshold, the second-level verification is deemed to have failed, and the description of the referenced landmark is deleted or the description of the referenced landmark is corrected to a distance description. When the referenced landmark exists and the deviation distance does not exceed the preset distance threshold, the second-level verification is determined to be successful, and the initial inducement text that has passed the second-level verification is determined as the security inducement text.

[0070] It should be noted that the initial guidance text and map database obtained after the first-level verification are used to obtain landmark verification data. The verification process is handled by the logic of landmark existence query and location deviation calculation. If the referenced landmark does not exist or if the referenced landmark exists but the calculated deviation distance between the location described in the initial guidance text and the actual location in the map database exceeds a preset distance threshold, the second-level verification is determined to be unsuccessful, and the description of the referenced landmark is deleted or corrected to a distance description. If the referenced landmark exists and the deviation distance does not exceed the preset distance threshold, the second-level verification is determined to be successful, and the initial guidance text that passes the second-level verification is determined to be the safe guidance text. This achieves accurate determination of the second-level verification result and ensures the reliability of the landmark description in the safe guidance text.

[0071] This embodiment, through the above-described scheme, organizes the structured semantic graph into structured prompt words, inputs these structured prompt words into a natural language generation model, and guides the natural language generation model to generate initial guidance text and corresponding structured action annotations through thought chain reasoning; it extracts turning action words, lane number descriptions, and distance values ​​from the initial guidance text, and performs a deterministic comparison with the original map data to perform a first-level verification; if the first-level verification fails, a backup broadcast engine is triggered to generate backup guidance information, which is then identified as the safety guidance text; if the first-level verification passes, the existence of landmarks referenced in the initial guidance text is verified. The system performs a second-level check to verify the location's rationality. The initial guidance text that passes the second-level check is identified as the safe guidance text. It can guide the natural language generation model to produce natural, fluent initial guidance text with structured annotations through structured prompts and thought chain reasoning mechanisms. Combined with the first-level check based on deterministic comparison of the original map data and the second-level check based on the existence and rationality of landmarks, the accuracy of key navigation instructions and semantic information is double-verified. If the check fails, a backup broadcast engine is triggered as a fallback. This achieves an efficient balance between the flexibility of natural language and the safety and reliability of key instructions in the process of generating navigation guidance information, effectively avoiding the navigation safety risks caused by model illusions.

[0072] Furthermore, Figure 5 This is a flowchart illustrating the fourth embodiment of the intelligent generation method for navigation guidance information of the present invention, as shown below. Figure 5 As shown, based on the first embodiment, a fourth embodiment of the intelligent generation method for navigation guidance information of the present invention is proposed. In this embodiment, step S30 specifically includes the following steps: Step S31: Obtain user feature data and parse the user feature data to determine the user's driving ability level and broadcast style preference.

[0073] It should be noted that by using user feature data to obtain users' driving behavior characteristics and preference information, and by using data analysis and feature mapping methods to process the user profile construction process, it is possible to accurately construct and determine the dimensions of the user cognitive profile required for personalized adjustment of navigation guidance information.

[0074] Step S32: Adjust the information density of the safety guidance text according to the driving ability level.

[0075] Understandably, by using driving ability levels to obtain users' driving skill assessment results, and employing a dynamic information density adjustment strategy to handle the level of detail in safety guidance texts, the complexity of navigation guidance information can be accurately matched and personalized to users' cognitive acceptance capabilities.

[0076] Step S33: When the driving ability level is the preset novice level, retain the prediction reminder and step-by-step action breakdown and add emotional reassurance statements.

[0077] It should be understood that by using the preset novice level identifier in the driving ability level to obtain user type information, and employing information retention and emotion enhancement strategies to process the guiding text content, the system maintains forward-looking guidance and detailed operation instructions to meet the high dependence of novice drivers on route prediction and action details. At the same time, it introduces emotional support content to achieve a personalized broadcast effect that reduces the anxiety of novice drivers and provides sufficient decision-making information.

[0078] Step S34: When the driving ability level is the preset proficient driving level, remove the prediction reminders and supplementary explanations and simplify them into core action instructions.

[0079] Understandably, by using preset proficiency level indicators to obtain the user's driving skill status, and employing information noise reduction and core command focusing strategies to process the guidance text, redundant forward-looking prompts and background explanations are omitted for experienced drivers, retaining only key execution commands. This achieves efficient and concise navigation broadcasts while reducing interference with the driver's attention. Step S35: Stylize the safety guidance text after adjusting the information density according to the broadcast style preference to generate the final broadcast text.

[0080] It should be understood that by leveraging broadcast style preferences to obtain users' language expression habits and emotional tendencies, and using natural language style adjustment strategies to process the wording, tone, and expression of safety guidance texts, personalized style conversion of the text is performed while maintaining semantic accuracy, and the final broadcast text is generated.

[0081] Step S36: Send the final broadcast text to the vehicle voice synthesis module for voice broadcast, and control the navigation device to display lane-level turning visual indication signals synchronized with the final broadcast text.

[0082] It should be noted that a multimodal interactive channel is constructed by using an in-vehicle voice synthesis module and navigation display equipment. The audio-visual synchronization strategy is used to process the presentation of guidance information. Based on the final broadcast text, voice broadcast is executed and lane-level turning visual indication signals are displayed in conjunction, thereby realizing the synchronous enhancement of navigation instructions in the auditory and visual dimensions.

[0083] This embodiment, through the above-described scheme, acquires user feature data, analyzes the user feature data to determine the user's driving ability level and broadcast style preference; adjusts the information density of the safety guidance text according to the driving ability level; when the driving ability level is a preset novice level, it retains the predictive reminders and step-by-step action breakdowns and adds emotional reassurance statements; when the driving ability level is a preset experienced driving level, it removes the predictive reminders and supplementary explanations and simplifies them to core action instructions; it stylizes the safety guidance text after adjusting the information density according to the broadcast style preference to generate the final broadcast text; and sends the final broadcast text. The system uses an in-vehicle voice synthesis module to provide voice prompts and controls the navigation device to display lane-level visual steering indicators synchronized with the final spoken text. By analyzing user feature data, it can accurately identify driving ability levels and speaking style preferences, and adjust the density of guidance text information based on level differences. For novice drivers, it retains predictive reminders and emotional reassurance to reduce driving anxiety, while for experienced drivers, it simplifies the core instructions to avoid information overload. By combining stylized rewriting with a synchronized output mechanism of voice prompts and lane-level visual indicators, it achieves a precise match between the complexity of navigation guidance information and the user's cognitive ability, thereby improving driving safety and personalized interactive experience.

[0084] Furthermore, Figure 6 This is a flowchart illustrating the fifth embodiment of the intelligent generation method for navigation guidance information of the present invention, as shown below. Figure 6 As shown, based on the first embodiment, a fifth embodiment of the intelligent generation method for navigation guidance information of the present invention is proposed. In this embodiment, after step S30, the intelligent generation method for navigation guidance information further includes the following steps: Step S40: After the navigation device outputs the broadcast, monitor the consistency between the vehicle's driving trajectory and the guidance instructions in the final broadcast text.

[0085] It should be noted that the actual driving status data of the vehicle is obtained by using the driving time period after the broadcast output, and the driving behavior verification process is processed by using trajectory tracking and instruction matching degree analysis strategies. Based on the consistency monitoring results between the vehicle driving trajectory and the guidance instructions in the final broadcast text, the quantitative evaluation of the compliance of the guidance instruction execution and the accurate acquisition of closed-loop optimization data are achieved.

[0086] Step S50: When a negative feedback event is detected, the current sample is marked as a negative feedback sample. The negative feedback event includes sudden braking or the vehicle entering the wrong lane and then correcting itself within a preset time after the broadcast.

[0087] Understandably, by using vehicle driving behavior monitoring data to obtain driving response status, and using negative feedback event recognition logic to process the abnormal behavior judgment process, negative feedback events are judged based on the behavior characteristics of sudden braking or vehicle correction after entering the wrong lane within a preset time after the broadcast, and the current sample is marked as a negative feedback sample, so as to accurately capture navigation guidance error scenarios and effectively accumulate failure data required for subsequent model optimization.

[0088] Step S60: Adjust the generation strategy of the natural language generation model or update the user feature data using the negative feedback samples.

[0089] It should be understood that by using labeled negative feedback samples to obtain the basis for model defect identification and user behavior deviation, and using a closed-loop learning mechanism to process the generation strategy adjustment and feature data update, the generation strategy of the natural language generation model can be optimized or the user feature data can be corrected based on the negative feedback samples, so as to realize the continuous self-evolution and dynamic improvement of the navigation guidance information generation system and the personalized adaptation capability.

[0090] In its specific implementation, this embodiment proposes an intelligent generation system for navigation guidance information based on a large model; its core architecture is a six-level pipeline of "lane-level environmental perception → semantic graph construction → large model thinking chain reasoning → safety-rule dual-layer verification → user cognition adaptive matching → closed-loop feedback optimization".

[0091] This system comprises four core modules, all of which run within the vehicle navigation domain controller: Module 1: Lane-level Environmental Perception and Structured Module Input: Current location coordinates, high-precision map (HD Map) lane-level data of the target intersection, and real-time traffic flow data (congestion level, traffic density).

[0092] Function: Parse the lane centerline geometry, lane attributes (type / direction / width), lane connection relationships (front drive lane ID / follower lane ID list), intersection polygon geometry and surrounding POI data from HD Map data, and output a structured intersection feature vector.

[0093] Output: A structured feature vector containing the lane sequence of the target intersection / exit, the distance to adjacent intersections, the coordinates and categories of surrounding POIs, and the annotations of high-frequency lane change / deceleration areas in the traffic flow.

[0094] Module 2: Large Model Inference and Generation Module Input: intersection feature vector (output of module 1), user profile label, current scene context (vehicle speed, current lane ID, distance from intersection).

[0095] Processing: The intersection feature vectors are transformed into a structured semantic graph (Model 1); the semantic graph is used as the Prompt input to guide the large model to reason step by step according to the thought chain of "scene analysis → timing selection → language generation", generating natural language guidance text containing landmark references, timing predictions and action instructions, and simultaneously outputting action annotation JSON for the verification module to use.

[0096] Output: Natural language-guided information to be verified + structured action annotations (extracted_actions).

[0097] Module 3: Security Verification and Rule Engine Module (Core Security Mechanism) Input: Guidance information and action annotations to be verified (output of module 2), original high-precision map data (as a truth reference), and traffic rule base.

[0098] Processing: Two levels of verification are performed: The first level of hard verification extracts turning action words, lane numbers, and distance values ​​from the generated text using regular expressions and NER, and performs a deterministic comparison with the original map data. If any key field does not match, it is marked as FAIL and a backup template is triggered; The second level of soft verification verifies the existence and location rationality of landmarks, and corrects or deletes inaccurate descriptions.

[0099] Output: Passed security guidance information (PASS), or error type code + backup template identifier (FAIL).

[0100] Module Four: Personalized Broadcasting and Feedback Learning Module Inputs: Safety guidance information (Module 3 PASS output), user profile (driving style, landmark preference, cognitive habits, historical feedback data).

[0101] Processing: Determine user level based on user driving ability score (Score_drive); adjust information density and broadcast timing according to level; match landmark references according to preferred landmark types; stylize broadcasts according to preferred broadcast styles; collect deviation data between vehicle trajectory and guidance instructions after broadcast.

[0102] Output: Personalized final broadcast text + broadcast timing offset + broadcast repetition count + user feedback tags (for closed-loop optimization).

[0103] In its specific implementation, the core control method of this implementation includes the following steps: S1. Obtain intersection scenario: When the vehicle travels to the predetermined trigger distance (configurable, default 800m) from the target intersection, the vehicle positioning module obtains the current location coordinates and loads the lane-level data of the intersection from the high-precision map database, including lane center lines, lane connection relationships and surrounding POI data.

[0104] S2. Structured Intersection Understanding: The algorithm for constructing a complex intersection semantic graph (Model 1) is invoked to parse the intersection data. Three sub-steps are executed: key decision point extraction → lane topology chain construction → landmark-action alignment, outputting a structured semantic graph S in JSON format.

[0105] S3. Large Model Generates Initial Prompt Information: The structured semantic graph S, user profile labels U, and current scene context C are organized into a structured prompt and input into the large language model. The large model reasones step by step according to the thought chain, outputting natural language prompt text and structured action annotations.

[0106] S4. Safety-Rule Dual-Layer Validation: First Layer Hard Validation: Extracts turning action words, lane numbers, and distance values ​​from the generated text and performs a deterministic comparison with the map's true values. If a mismatch occurs, the text is discarded and a backup template is triggered. Second Layer Soft Validation: Verifies the existence and location rationality of referenced landmarks. If inaccurate, the landmark description is corrected or deleted.

[0107] S5. Personalized Packaging: After verification, the user cognitive adaptive matching algorithm (Model 3) is invoked to adjust the information density, replace landmarks, and stylize the safety text according to the user's driving ability level, landmark preferences, and broadcast style preferences.

[0108] S6. Multimodal output of guidance information: Personalized broadcast text is broadcast through the speech synthesis module and simultaneously displayed on the in-vehicle screen as a dynamic lane-level turn arrow overlay (AR-HUD or central control screen) to show the turning direction and target lane. The voice and visual information are time-aligned and content is consistent.

[0109] S7. Effect Evaluation and Closed-Loop Optimization: After the broadcast, continuously monitor the vehicle's driving trajectory and compare the actual lane-changing / steering behavior with the guidance instructions. If any of the following negative feedback events are detected: sudden braking (deceleration > 4 m / s²) occurs within T seconds after the broadcast (T=5s) or the vehicle enters the wrong lane and then corrects itself, the sample is marked as negative feedback, stored as part of the fine-tuning dataset, and periodically used to optimize the generation strategy of the large model.

[0110] Core Algorithm Model Model 1: Semantic Graph Construction Algorithm for Complex Intersections This model transforms vector road network data from high-precision maps into a "structured semantic graph" that large models can directly understand and reason about. It is a key intermediate representation layer that connects "cold map data" with "semantic reasoning".

[0111] enter The target intersection's high-resolution (HD) map lane-level data specifically includes: Lane centerline geometry: a 3D curve described by an ordered sequence of latitude and longitude coordinates; Lane attributes: Lane type (regular / bus / HOV / emergency), direction of travel, lane width; Lane connectivity: a list of preceding lane IDs and a list of succeeding lane IDs for each lane; Intersection geometry: The closed polygonal boundary of the intersection area; Surrounding POI data: POI unique identifier, name, category label, latitude and longitude coordinates, popularity score (popularity value from map data).

[0112] Algorithm steps Step 1: Extraction of Key Decision Points It iterates through the connections between all lanes on the vehicle's current path and identifies key locations where the driver needs to make lane change or steering decisions.

[0113] The judgment rules are as follows: Branching point determination: If the number of subsequent lanes of the current lane is greater than 1, that is, the downstream of the lane branches into multiple selectable lanes, it is marked as a "branching point" (the driver needs to choose which subsequent lane to enter). Merging point determination: If the number of front-wheel drive lanes in the current lane is greater than 1, it is marked as a "merging point" (attention should be paid to merging vehicles); Turning point determination: Calculate the angle θ between the direction vector of the target exit lane and the direction vector of the current lane. If θ > the preset turning threshold (default 20°), mark it as a "turning point".

[0114] For each key decision point, record its location coordinates (latitude and longitude), path distance from the vehicle's current location (accumulated along the lane centerline), and the lane segment ID to which the point belongs.

[0115] 2. Step 2: Lane Topology Chain Construction Starting from the vehicle's current lane, the system searches forward along lane connections (following lanes) until it reaches the target exit lane, thus constructing a complete lane action sequence. L_seq=[(lane_1,action_1),(lane_2,action_2),…,(lane_n,action_n)].

[0116] The calculation method for action_i is as follows: compare the lateral offset Δd of the geometric centerline of lane_i and lane_{i+1} at the intersection cross-section. If Δd < -1.5m (leftward deviation exceeds one standard lane width): action = "Change lanes to the left"; If Δd > 1.5m: action = "Change lanes to the right"; If |Δd|≤1.5m: action="Keep current lane".

[0117] Note: The lane width threshold (default 1.5m) can be configured according to the actual road standard (3.0-3.5m for urban roads and 3.5-3.75m for highways). Here, half the width is taken as an approximate threshold.

[0118] 3. Step 3: Landmark-Motion Alignment For each "lane change" or "turn" action node in the lane action sequence L_seq, search for surrounding POIs with the node's coordinates as the center and a search radius R (configurable, default 200m); then filter and sort the searched POIs. (a) Filtering: Filter out POIs located on back streets, obscured by buildings, or not directly observable from the road (judged by the "street frontage" attribute of the POI and OSM building outline data), and prioritize retaining POIs that are street frontage and have a visibility score higher than the threshold; (b) Sorting: Sort in descending order by weighted score sorting_score=α×(1-d_i / D_max)+β×popularity_i, where d_i is the straight-line distance from POI to the action point, D_max is the search radius, popularity_i is the normalized score of POI popularity, α=0.6 and β=0.4 are the default weights, and the top 1 is selected; (c) Binding: For the selected POI, generate a triple (action_i, POI_name, relative_position), where relative_position is mapped to a natural language spatial relation word by calculating the azimuth and distance between the POI coordinates and the action point coordinates. The mapping rules are shown in Table 1 below: Table 1. Example table of mapping rules:

[0119] Output The structured semantic graph S = {"decision_points": list of key decision points, "lane_action_sequence": lane action sequence L_seq, "poi_action_triples": list of landmark-action triples} is output in JSON format and can be directly used as the structured data part of the large model Prompt.

[0120] Model 2: Generate-Verify Dual-Model Inference Model This model is the core security mechanism of the system, and it consists of a "generation sub-model" (based on LLM-based thought chain reasoning generation) and a "verification sub-model" (based on rule engine-based deterministic verification).

[0121] The generation sub-model is responsible for naturalness and richness, while the verification sub-model is responsible for security and accuracy. The two are functionally decoupled but sequential in process, ensuring that security verification does not depend on the reliability of the generation model itself.

[0122] Generating sub-models (LLM-based mind chain generation) enter Semantic graph S (Model 1 output format) User profile tag U = {driving experience (years), historical emergency braking frequency (times / 100 km), landmark preference type (enumerated value), historical error rate at this type of intersection} Current scenario context C = {Current vehicle speed (km / h), Current lane ID, Path distance to target intersection (m), Current traffic congestion level (1-5)} Algorithm logic (thought chain reasoning generation) Step 1: Construct a structured Prompt The semantic graph S, user profile U, and scene context C are assembled into a structured input according to a preset Prompt template.

[0123] The Prompt template contains the following semantic blocks: role setting ("You are an experienced navigator"), scene data area (S, U, C data in JSON format), reasoning guidelines (step-by-step reasoning requirements), output format constraints (JSON schema), and security constraints ("It is forbidden to fabricate non-existent landmark names").

[0124] Step 2: Large Model Thinking Chain Reasoning After receiving the Prompt, the large model proceeds with step-by-step reasoning according to the following thought chain (the Prompt guides the model to explicitly output the intermediate reasoning process): Step 1 - Scenario Analysis: Analyze the spatial relationship between the current lane and the target exit to identify driving difficulties. The reasoning should include: the number of lane changes required, the target lane number for each lane change, the presence of areas with heavy traffic, and any merging points requiring attention. Output format: "[Scenario Analysis] Currently in lane {m}, target lane {n}, requiring {k} left / right lane changes. Driving difficulties: {difficulty description}". Step 2 - Timing Selection: Based on landmark locations and current vehicle speed, calculate the optimal time to change lanes / turn. Calculation logic: t=remaining_distance / current_speed; If t < preset safe time threshold T_safe (default 8s) → it is judged as "urgent" and an immediate reminder is required. The broadcast timing offset is set to -Δt_urgent (early trigger). Otherwise → select the landmark closest to the action point that meets the condition of "still within the line of sight when the vehicle arrives" as the reference trigger point.

[0125] Output format: "[Timing Selection] Current distance {m} meters, estimated arrival time {t} seconds. Recommended action trigger point: {POI_name} (distance from action point {d} meters). Urgency level: {urgent / normal}." Step 3 - Language Generation: Organize the above reasoning results into natural language broadcast text, following a four-part structure of "predictive reminder → landmark reference → action instruction → supplementary explanation": - Anticipatory reminders: Inform drivers in advance about the characteristics of the intersection (e.g., "You need to change lanes continuously at the intersection ahead, so prepare in advance"); - Landmark reference: Indicates the timing of an action based on a landmark (e.g., "after passing the bus stop ahead"). - Action instructions: Give clear lane-level operations (such as "change one lane to the left to the middle lane"); - Additional notes: Safety tips or reassurance information may be added as needed.

[0126] Language style requirements: Use colloquial Chinese expressions and avoid mechanical phrases like "XX meters ahead".

[0127] Step 3: Structured Output The large model outputs a JSON object containing the following fields: { "reasoning_chain": "[Scene Analysis]... [Timing Selection]... [Language Generation]...", "generated_text": "The final triggering text (plain text string)", "extracted_actions": [ {"type": "lane change", "direction": "left", "lane_count": 2}, {"type": "Turning", "direction": "Left Turn"} ], "reference_poi": "bus stop", "timing_offset_s": -2.0 } 2.2 Validation Sub-model (Rule Engine + Deterministic Validation) enter The generated text and extracted actions output by the sub-model Raw high-precision map data (as a ground truth reference) Traffic rule database (regulations such as turning restrictions and speed limits) Algorithm logic (two-level verification) Level 1: Hard verification (consistency verification of critical actions) (a) Information extraction: The following structured fields are extracted from generated_text using regular expressions and the NER model: Set of turning action words: {turn left, turn right, go straight, make a U-turn, keep left, keep right} Lane number description: {Lane 1, Leftmost Lane, Middle Lane, Rightmost Lane, Nth Lane on the Left, ...} → Mapped to a unified lane index number Distance values ​​and units: Extract all combinations of numbers and units (e.g., "300 meters" or "800 meters") and convert them to values ​​in meters. (b) Deterministic comparison process: Step A: Compare turning actions: Perform a precise string match between the extracted turning action words and the preset turning directions of the target intersection in the original map data. If they do not match, the verification result will be FAIL, triggering a backup template announcement with the error code: ERR_TURN_MISMATCH. Step B: Compare Lane Numbers: After mapping the lane descriptions in the text to lane indices, compare them with the lane numbers corresponding to the target exit in the original map data. If they do not match, the verification result will be FAIL, triggering a backup template broadcast with the error code: ERR_LANE_MISMATCH. Step C: Distance Comparison: Compare the extracted distance value with the actual path distance from the current location to the target intersection on the map. If the deviation exceeds ±30%, FAIL will not be triggered, but it will be recorded as WARNING (distance ambiguity deviation does not affect the safe direction), and soft verification correction will be performed. Step D: All the above comparisons pass → Verification result = PASS, proceed to the second level of soft verification.

[0128] Level 2: Soft verification (verification of the existence and location rationality of landmarks) (a) Existence verification: Perform exact and fuzzy matching (edit distance ≤ 2) in the POI database using the POI names appearing in the generated text as keywords.

[0129] If the match fails (POI does not exist or has been changed) → delete the landmark description and replace it with a distance description (e.g., change "after passing the bus stop" to "after driving about 200 meters"). (b) Location rationality verification: If the POI exists, calculate the straight-line distance d_real between the POI coordinates and the reference action point.

[0130] If d_real > the preset upper limit (default 300m), it is judged as "inaccurate location description" → the precise distance in the generated text is changed to a vague description (such as changing "gas station 150 meters ahead" to "gas station not far ahead"); if d_real ≤ 300m, the original description is retained.

[0131] Output Verification result label: PASS or FAIL If PASS: Output the final security guidance text T_safe that passes both levels of verification. If FAIL: Output error type code + trigger backup template identifier, and the system will automatically switch to the rule-based backup broadcast engine.

[0132] Model 3: User Cognitive Adaptive Matching Algorithm Based on the safety text T_safe that passed the validation of Model 2, this model is personalized according to the driver's cognitive level and preferences. While ensuring that the safety information is not distorted, the expression is adjusted to suit different users.

[0133] enter Safety guidance text T_safe (Model 2 PASS output) User profile U={ Driving experience (years), Historical frequency of emergency braking (times / 100 km / h) Historical frequency of getting lost (times / 100km, grouped by intersection type) Preferred landmark types (enumerated values: chain brands, iconic buildings, public facilities, natural landscapes, no preference), Preferred broadcasting style (enumeration values: concise, detailed).

[0134] Algorithm steps Step 1: Determining the User's Driving Ability Level The driving ability score Score_driver is calculated using a weighted formula based on three dimensions: driving experience, frequency of emergency braking, and frequency of wrong turns. Score_driver = w1 × (1 - Normalized value of driving experience) + w2 × Normalized value of emergency braking frequency + w3 × Normalized value of wrong turn frequency The normalization process uses Min-Max normalization (with preset reference ranges for each dimension), with default weights w1=0.30, w2=0.35, w3=0.35 (the weights can be dynamically adjusted based on user feedback data).

[0135] Determine user level based on Score_driver: Score_driver>0.6→Level L1 (Beginner / Low Proficiency): Needs more predictive reminders and reassuring statements; 0.3≤Score_driver≤0.6→Level L2 (General Proficiency): Provides standard information density; Score_driver < 0.3 → Level L3 (High Proficiency): Simplify information and remove redundant reassurance.

[0136] Step 2: Adaptive adjustment of information density Based on the user level, the information density of T_safe is adjusted as shown in Table 2 below: Table 2. Example of information density adjustment:

[0137] Step 3: Landmark Preference Matching Read preferred landmark types from user profiles.

[0138] If the landmark category of the POI referenced in T_safe happens to match the user's preference type, add a "familiarity marker" prefix when referencing it (e.g., "It's the XX intersection you often pass by"); if it does not match, search the POI list in the semantic graph S for a candidate POI with the same preference type. If it exists and meets the location rationality condition (d_real≤300m), replace the original POI reference and re-pass the soft validation.

[0139] Step 4: Presentation Style Packaging The text is rewritten in a stylized manner based on the user's preferred broadcast style.

[0140] This step is completed by calling the large model using a preset style Prompt template. The Prompt explicitly constrains "not to change the content and direction of the action instructions" to maintain safety.

[0141] The example comparison table is shown in Table 3 below: Table 3. Style Example Table:

[0142] Output Personalized final broadcast text T_final Broadcast timing offset Δt (positive value = delayed broadcast, negative value = early broadcast, unit: seconds) The number of times the announcement is repeated is n_repeat (L1 users repeat the announcement 1-2 times for key actions, while L3 users only repeat it once). Example 1: Complex Continuous Lane Change Scenario Scenario Description: The vehicle is currently in the second lane (out of four lanes traveling in the same direction) next to the rightmost bus lane on an eight-lane main road in the city. The navigation route indicates a left-turn ramp is required at an interchange approximately 800 meters ahead. It is currently rush hour, and traffic density is high in the two left lanes. Traditional navigation would typically announce: "Turn left in 800 meters."

[0143] Execution process of this invention: (1) Map perception and understanding (S1-S2): The system loads the intersection data from the high-precision map. The semantic map construction algorithm analysis found that: the current lane (index 2) → target exit (left turn ramp, which needs to pass through index 3 → index 4) requires two consecutive left lane changes; there is a bus stop (POI_001) between index 3 lane and index 4 lane in the middle; the current left lane (index 3 and index 4) has a congestion probability of >70% in the historical traffic flow data during this period.

[0144] (2) Large Model Reasoning and Generation (S3): Output of Large Model Thinking Chain: Scene Analysis → "Need to change lanes from lane 2 to lane 4 (left turn lane) continuously, requiring a total of 2 lane changes. Driving difficulty: The left lane is congested during the evening rush hour, and it is necessary to find a lane change window in advance." Timing selection → "Current distance 800m, speed 40km / h, estimated arrival in 72 seconds."

[0145] POI_001 (bus stop) is 120m from the first lane change point; it is recommended to use this POI as the trigger reference. Language generation → Generation of preliminary guiding text.

[0146] (3) Safety verification (S4): Hard verification compares the “change lanes to the left” in the generated text with the direction of the left turn ramp on the map → PASS; verify that the “bus stop” POI actually exists in the map database and is 120m away from the first lane change point (<300m threshold) → PASS.

[0147] (4) Personalized packaging (S5, taking L1 novice user as an example): The user profile shows 0.8 years of driving experience, 3.2 times / 100 km of emergency braking frequency, 1.5 times / 100 km of wrong turn rate, Score_driver=0.72→L1 novice level.

[0148] The information density has been adjusted to retain the complete four-segment structure, with added step-by-step lane change prompts and emotional reassurance.

[0149] Final announcement: "The intersection ahead is a bit complicated. We need to change direction continuously from the right to the far left. Please prepare in advance."

[0150] Look to your right. After you pass the bus stop on your right, make sure it's safe on your left. First, change lanes to the left, hold your position, and then change lanes to the left again. You'll then be in the left turn lane.

[0151] No rush, there's plenty of time. (5) Broadcast output (S6): The above text (detailed style) is broadcast in voice, and the target lane trajectory of the two lane changes is displayed on the AR-HUD with dynamic arrow animation.

[0152] (6) Comparison of effects: After the traditional method was announced, the wrong-way rate at the intersection was 18.3% (based on historical statistics). After the new method was announced, the wrong-way rate dropped to 4.7%, and the number of emergency braking incidents decreased by 62%.

[0153] Example 2: Benchmark guidance for a simple standard intersection Scenario description: A standard intersection on a secondary arterial road in the city. The vehicle is currently traveling straight in the middle lane. The navigation system indicates that a right turn is required to enter the auxiliary road 300 meters ahead.

[0154] The road conditions are simple: there are no complex interchanges at the intersection, the right-turn lane is independent, and there are no obvious landmarks in the vicinity. Traditional navigation will announce: "Please turn right in 300 meters."

[0155] Execution process of this invention: (1) Semantic graph construction (S1-S2): Semantic graph algorithm analysis found that: the current lane is straight to the right turn lane, only one lane change (to the right) is needed; the key decision point is 280 meters away from the intersection as the "branch point" (straight and right turn are separated); there are no POIs that meet the popularity threshold within a search radius of 200m → the landmark-action triplet is empty.

[0156] (2) Large model generation and verification (S3-S4): The large model recognizes that the scene is simple and has no landmarks, and generates preliminary guidance: "Turn right at the intersection ahead, you can move to the right now." Hard verification confirms that the "turn right" direction is correct → PASS.

[0157] The soft check detected no landmark reference → no verification required, pass directly.

[0158] (3) Personalized Packaging (S5, taking L3 proficient user as an example): User Score_driver=0.12→L3 high proficiency. The information density is adjusted to remove the predictive reminders and supplementary explanations, and only retain the concise instructions. The final broadcast is: "Turn right at the intersection ahead, keep to the right." The broadcast timing is offset by -1.0s (1 second in advance).

[0159] (4) Effect: In simple scenarios, skilled users can obtain extremely simple instructions, information noise is reduced, and the broadcast time is shortened from 3.2 seconds in the traditional solution to 1.1 seconds.

[0160] Example 3: Degradation processing for scenes with missing landmarks Scenario Description: A newly built development zone road has complete high-precision map data, but sparse POI data for the surrounding area. Vehicles need to turn left 500 meters ahead to enter the internal road of the park.

[0161] The POI database within a 200m search radius only contains "Construction Sites" (with a popularity score of 0, and is filtered out) and "Temporary Fences" (a non-standard POI category).

[0162] Traditional navigation systems announce: "Please turn left in 500 meters."

[0163] Execution process of this invention: (1) Semantic graph construction (S1-S2): Lane topology chain construction successful (straight-through → left-turn lane, 1 lane change).

[0164] In the landmark-action alignment step, if there are no valid POIs within a 200m search radius, the landmark-action triplet list will be empty. The poi_action_triples field of the semantic graph S will be marked as null.

[0165] (2) Large Model Generation (S3): The landmark data area in the Prompt is empty. In the "Language Generation" step of the thought chain, the large model is guided by the Prompt to a degraded mode—no landmark reference sentences are generated, and distance is used as the reference instead. Initial prompting: "You need to turn left into the park 500 meters ahead, you can change lanes to the left now." (3) Security check (S4): Hard check confirms "left turn" is correct → PASS.

[0166] If no POI reference is detected by soft validation, the existence / location validity validation is skipped, and the application passes directly.

[0167] The verification log records a downgrade flag: FALLBACK_NO_POI.

[0168] (4) Personalized Packaging (S5): Add reassurance and compensation to the final broadcast for L1 new users (as the lack of landmark reference may increase anxiety): "Turn left into the park 500 meters ahead, slowly move to the left, and I will remind you again when you reach the intersection." A secondary reminder broadcast will be added 150 meters from the intersection.

[0169] (5) Effect: In scenarios with sparse POI data, the system automatically downgrades to the "distance reference + secondary reminder" mode to avoid mentioning non-existent or inaccurate landmarks, thus maintaining the executability of instructions and the driver's trust.

[0170] Example 4: Comparison of Differentiated Output for Different User Levels in the Same Scenario Scenario description: The exit ramp of an urban expressway requires two consecutive right lane changes from the leftmost fast lane to enter the right exit ramp. There is a chain gas station 800 meters before the exit (POI_002, popularity rating 0.85). The section before the exit is a solid line (lane changing is prohibited) 200 meters before the exit.

[0171] The output for Level L1 users (new driver, 0.5 years of driving experience, Score_driver=0.81) is as follows: T = 800m from the exit: "We're about to exit the expressway. We need to change from the far left to the far right, which will require two changes."

[0172] See the gas station on your right? After passing the gas station, start changing lanes to the right. First, change lanes to the middle to maintain a safe distance, then change lanes to the far right. Be prepared in advance, so you don't get stuck when you reach the solid line.

[0173] T = 400m from the exit (Second reminder): Gas station is approaching, prepare to change lanes.

[0174] Information characteristics: Complete four-part format + split lane change + emotional reassurance + solid line warning + secondary reminder. Total broadcast word count: approximately 120 words.

[0175] The broadcast output for Level L3 users (skilled, 8 years of driving experience, Score_driver=0.08) is as follows: T = 800m from the exit: After passing the gas station, turn right onto two lanes, with a solid line ahead.

[0176] Information characteristics: concise instructions + key constraints (solid lines) + no secondary reminders; total number of words broadcast: approximately 16 words.

[0177] Comparison Explanation: L1 and L3 users receive completely different broadcast messages, but the core safety information (lane change direction, number of lane changes, landmark references, solid line constraints) remains consistent. The differences are only reflected in information redundancy, sentence complexity, and broadcast frequency, without changing the content of safety-critical information.

[0178] This embodiment, through the above-described scheme, monitors the consistency between the vehicle's trajectory and the guidance instructions in the final broadcast text after the navigation device broadcasts the information. When a negative feedback event is detected, the current sample is marked as a negative feedback sample. The negative feedback event includes sudden braking or the vehicle entering the wrong lane and then correcting itself within a preset time after the broadcast. The generation strategy of the natural language generation model is adjusted or the user feature data is updated based on the negative feedback sample. This allows for the accurate identification of negative feedback events by monitoring the consistency between the vehicle's trajectory and the guidance instructions after the broadcast, marking negative feedback data with behavioral samples such as sudden braking or lane correction, and dynamically adjusting the generation strategy of the natural language generation model or updating user feature data based on the negative feedback sample. This achieves closed-loop self-optimization and continuous evolution of the personalized adaptation capabilities of the navigation guidance information generation system, effectively reducing driving risks and improving the user interaction experience.

[0179] Accordingly, the present invention further provides an intelligent navigation guidance information generation device.

[0180] Reference Figure 7 , Figure 7 This is a functional block diagram of the first embodiment of the intelligent navigation guidance information generation device of the present invention.

[0181] In a first embodiment of the intelligent navigation guidance information generation device of the present invention, the intelligent navigation guidance information generation device includes: The map construction module 10 is used to acquire map data containing lane-level information and construct a structured semantic map based on the map data. The structured semantic map includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions.

[0182] The security verification module 20 is used to input the structured semantic graph into the natural language generation model, generate initial induction text, and perform security verification on the initial induction text to determine the secure induction text.

[0183] The broadcast control module 30 is used to acquire user feature data, personalize the safety guidance text according to the user feature data, generate the final broadcast text, and control the navigation device to broadcast the text.

[0184] The steps for implementing each functional module of the intelligent navigation guidance information generation device can be referred to in the various embodiments of the intelligent navigation guidance information generation method of the present invention, and will not be repeated here.

[0185] Furthermore, this embodiment of the invention also proposes a storage medium storing a navigation guidance information intelligent generation program. When the navigation guidance information intelligent generation program is executed by a processor, it implements the operations described in the above-described navigation guidance information intelligent generation method embodiment.

[0186] Those skilled in the art will understand that all or part of the steps in the methods described above can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium is a computer-readable storage medium, including: USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program code.

[0187] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0188] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0189] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligently generating navigation guidance information, characterized in that, The intelligent generation method for navigation guidance information includes: Acquire map data containing lane-level information, and construct a structured semantic graph based on the map data. The structured semantic graph includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The structured semantic graph is input into a natural language generation model to generate initial lead text, and a security check is performed on the initial lead text to determine a secure lead text. The system acquires user characteristic data, personalizes the safety guidance text based on the user characteristic data, generates the final broadcast text, and controls the navigation device to broadcast it.

2. The intelligent generation method for navigation guidance information as described in claim 1, characterized in that, The process involves acquiring map data containing lane-level information, constructing a structured semantic graph based on the map data, and the structured semantic graph including key decision point information, lane action sequences, and alignment relationships between landmarks and actions, including: Traverse the lane connection relationships on the vehicle's driving path, determine the node type based on the number of subsequent lanes, the number of front drive lanes, and the vehicle heading angle deviation, and use nodes with branching points, merging points, or turning points as the key decision point information. The system searches forward along the lane connections to the target exit lane, and constructs the lane action sequence by comparing the lateral offset of the center lines of adjacent lanes. The lane action sequence includes the lane change direction and the number of lane changes. For each lane change or turning action node, surrounding points of interest are retrieved with a preset search radius. Based on the street-facing attributes and building outline data of the POI, points of interest located in back streets, obscured by buildings, or not directly observable from the road are filtered out. Retain street-facing points of interest with a popularity rating higher than a preset rating threshold as target points of interest; Calculate the spatial orientation of the target point of interest relative to the action point, convert the spatial orientation into natural language spatial relation words, and form the alignment relationship between the landmark and the action; A structured semantic graph is constructed based on the key decision point information, the lane action sequence, and the alignment relationship.

3. The intelligent generation method for navigation guidance information as described in claim 2, characterized in that, The process of traversing the lane connections along the vehicle's travel path determines the node type based on the number of subsequent lanes, the number of forward-drive lanes, and the heading angle deviation. Nodes with branching points, merging points, or turning points are designated as key decision point information, including: Iterate through the lane connections along the vehicle's travel path; When the number of subsequent lanes is greater than the number of preceding lanes, the node type is determined as a branch point; When the number of subsequent lanes is less than the number of preceding lanes, the node type is determined as a merging point; When the deviation of the vehicle heading angle is greater than a preset angle threshold, the node type is determined as a turning point; Nodes with branching points, merging points, or turning points are used as key decision point information.

4. The intelligent generation method for navigation guidance information as described in claim 1, characterized in that, The step of inputting the structured semantic graph into a natural language generation model to generate initial lead text, and performing security checks on the initial lead text to determine secure lead text, includes: The structured semantic graph is organized into structured prompt words, and the structured prompt words are input into the natural language generation model to guide the natural language generation model to generate initial inductive text and corresponding structured action annotations through thought chain reasoning. The turning action words, lane number descriptions, and distance values ​​in the initial guidance text are extracted and compared deterministically with the original map data to perform the first level of verification. When the first-level verification fails, the backup broadcast engine is triggered to generate backup guidance information, and the backup guidance information is identified as the security guidance text. When the first-level verification passes, the existence and location rationality of the referenced landmarks in the initial inducement text are verified to perform the second-level verification, and the initial inducement text that passes the second-level verification is determined as the security inducement text.

5. The intelligent generation method for navigation guidance information as described in claim 4, characterized in that, Upon passing the first-level verification, the existence and location rationality of the referenced landmarks in the initial inducement text are verified to perform a second-level verification. The initial inducement text that passes the second-level verification is determined as the safe inducement text, including: If the first-level verification passes, the map database is queried to verify whether the referenced landmarks in the initial guidance text exist. If the referenced landmark does not exist, or if the referenced landmark exists but the calculated deviation between the location described in the initial guiding text and the actual location in the map database exceeds a preset distance threshold, the second-level verification is deemed to have failed, and the description of the referenced landmark is deleted or the description of the referenced landmark is corrected to a distance description. When the referenced landmark exists and the deviation distance does not exceed the preset distance threshold, the second-level verification is determined to be successful, and the initial inducement text that has passed the second-level verification is determined to be the security inducement text.

6. The intelligent generation method for navigation guidance information as described in claim 1, characterized in that, The steps of acquiring user feature data, personalizing the safety guidance text based on the user feature data, generating the final broadcast text, and controlling the navigation device to broadcast the text include: Acquire user characteristic data and parse the user characteristic data to determine the user's driving ability level and broadcast style preference; Adjust the information density of the safety guidance text according to the driving ability level; When the driving ability level is the preset novice level, the predictive reminders and step-by-step action breakdown are retained, and emotional reassurance statements are added. When the driving ability level is the preset proficient driving level, the prediction reminders and supplementary explanations are removed and simplified to the core action instructions; Based on the broadcast style preference, the safety guidance text after adjusting the information density is stylized and rewritten to generate the final broadcast text; The final broadcast text is sent to the in-vehicle voice synthesis module for voice broadcast, and the navigation device is controlled to display lane-level turning visual indication signals synchronized with the final broadcast text.

7. The intelligent generation method for navigation guidance information as described in claim 1, characterized in that, After acquiring user feature data, personalizing the safety guidance text based on the user feature data, generating the final broadcast text, and controlling the navigation device to broadcast the text, the intelligent generation method for navigation guidance information further includes: After the navigation device outputs a broadcast message, the consistency between the vehicle's driving trajectory and the guidance instructions in the final broadcast text is monitored. When a negative feedback event is detected, the current sample is marked as a negative feedback sample. The negative feedback event includes sudden braking or the vehicle entering the wrong lane and then correcting itself within a preset time after the broadcast. The negative feedback samples are used to adjust the generation strategy of the natural language generation model or update the user feature data.

8. A navigation guidance information intelligent generation device, characterized in that, The intelligent navigation guidance information generation device includes: The map construction module is used to acquire map data containing lane-level information and construct a structured semantic map based on the map data. The structured semantic map includes key decision point information, lane action sequences, and alignment relationships between landmarks and actions. The security verification module is used to input the structured semantic graph into the natural language generation model, generate initial induction text, and perform security verification on the initial induction text to determine the secure induction text; The broadcast control module is used to acquire user characteristic data, personalize the safety guidance text based on the user characteristic data, generate the final broadcast text, and control the navigation device to broadcast it.

9. A navigation guidance information intelligent generation device, characterized in that, The intelligent navigation guidance information generation device includes: a memory, a processor, and an intelligent navigation guidance information generation program stored in the memory and executable on the processor, wherein the intelligent navigation guidance information generation program is configured to implement the steps of the intelligent navigation guidance information generation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a navigation guidance information intelligent generation program, which, when executed by a processor, implements the steps of the navigation guidance information intelligent generation method as described in any one of claims 1 to 7.

Images