Active navigation method for automobile

Abstract

The invention discloses an active navigation method for an automobile, and relates to the technical field of intelligent traffic and Internet of Vehicles, and the active navigation method comprises the following steps: acquiring intention data and environment data, and extracting feature vectors; calculating an intention probability, a prior preference probability based on historical records and a navigation resistance cost; and calculating comprehensive recommendation scores of the candidate destinations in combination with the results. And when the score meets an activation threshold value dynamically adjusted based on the vehicle speed and a preset safety marginal condition, generating an active navigation instruction. Then, a recommended path is planned based on the time, distance and cost sensitivity parameters of the user, and refined guidance for the facility node is performed at the end of the travel. And finally, updating the user model in a closed-loop manner by utilizing dominant interaction and implicit track feedback data. According to the method, high-confidence-coefficient destination active recommendation is realized by fusing multi-dimensional features, manual operation is reduced, a personalized navigation strategy is continuously optimized through adaptive learning, and driving experience and safety are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation and Internet of Vehicles, and in particular to an active navigation method for a vehicle. BACKGROUND

[0002] With the continuous improvement of the intelligent level of vehicles, the vehicle navigation system has become a standard configuration of vehicles. The existing vehicle navigation technology mainly adopts a passive response type of interaction mode, that is, the driver needs to input text or issue specific voice instructions through a touch screen to set a destination, and then the system will calculate a path and start guiding. This interaction mode requires the driver to perform explicit operations before or during driving, which is not only cumbersome, but also easy to distract the driver's attention during vehicle driving, increasing the risk of driving safety.

[0003] Although some advanced navigation systems begin to try to provide a destination recommendation function based on historical records, the prediction logic is usually simple, and is often based on only a single frequency statistics or a fixed time rule (for example, only recommending home or company during morning and evening peak hours), lacking deep fusion and analysis of multi-dimensional situational information such as real-time schedule of users, temporary voice intent, and current external environmental resistance. This results in insufficient accuracy and timeliness of the recommended results, and often pops up interfering information when the user does not need navigation, or the recommended destination does not meet the actual needs of the user at the moment.

[0004] In addition, the existing path planning algorithm mostly adopts a general shortest time or shortest distance strategy, and fails to fully quantify the individualized sensitivity of users to toll fees, driving mileage, and time cost, and cannot dynamically adjust the planning weight value according to the long-term driving habits of users. At the same time, the guidance of traditional navigation usually terminates at the POI center coordinate or roadside of the destination, lacking the ability of fine identification and guidance of specific facility nodes (such as underground parking lot entrances and drop-off areas) around large destinations, resulting in the need for the driver to spend effort to find the specific entrance after reaching the vicinity of the destination, and making it difficult to achieve truly door-to-door intelligent service. SUMMARY

[0005] The purpose of the present application is to provide an active navigation method for a vehicle, which at least solves the technical problems of the existing vehicle navigation system that relies on passive input instructions of the driver, cannot achieve high-confidence destination prediction based on multi-dimensional scene features, and lacks individualized path adaptive guidance.

[0006] The present application provides the following solutions:

[0007] An active navigation method for a vehicle, comprising the steps of multi-source data collection, feature processing, comprehensive decision-making, path planning and model iterative updating. First, multi-source input data is collected, which includes intent data obtained through voice and device interconnection channels, and environmental data obtained through vehicle bus and cloud services. The system preprocesses the multi-source input data, extracts the intent feature vector through beamforming and semantic slot analysis, simultaneously analyzes the vehicle position, traffic flow and weather information to construct the environmental state vector, and performs spatio-temporal alignment and normalization processing based on a sliding time window.

[0008] In the decision-making calculation stage, the method calculates the intent probability representing the user's travel demand confidence based on the intent feature vector, using a logistic regression model to process the time urgency feature and the voice semantic confidence. At the same time, based on the environmental state vector, the system traverses the historical trip records in the user model, and calculates the prior preference probability in the current spatio-temporal scenario through spatio-temporal clustering screening and exponential time decay weighting. In addition, the navigation resistance cost is calculated according to the path traffic congestion coefficient and the weather influence factor.

[0009] The system further calculates the comprehensive recommendation score of the candidate destination, which is composed of three components linearly weighted: the first component is the intent-driven component, determined by the product of the intent probability and the intent matching degree; the second component is the habit-driven component, corresponding to the prior preference probability; the third component is the environmental resistance component, corresponding to the navigation resistance cost.

[0010] In the trigger judgment stage, the system performs a winner-takes-all logic to lock the candidate destination with the highest score, and introduces a double threshold judgment mechanism: on the one hand, it judges whether the highest score exceeds the activation threshold dynamically associated with the vehicle speed, and on the other hand, it judges whether the difference between the highest score and the second highest score meets the preset safety margin. Only when the above conditions are met and the cooling period is not reached, the system generates an active navigation instruction.

[0011] In the path planning stage, the system plans the recommended path based on the road network topology graph. The comprehensive travel cost of a road segment is composed of time dimension cost, distance dimension cost and economic dimension cost, and the weight coefficients of each dimension correspond to the time sensitivity, distance sensitivity and cost sensitivity stored in the user model. When the vehicle travels to the vicinity of the destination, the system retrieves the associated facility node set and modifies the navigation endpoint to a specific parking lot entrance or drop-off point according to the end behavior preference record, realizing fine guidance.

[0012] In the model optimization phase, the system dynamically adjusts the route according to the real-time environmental changes during vehicle driving, and constructs a closed-loop feedback mechanism. The system collects explicit interaction feedback and implicit behavior feedback of the user, wherein the implicit feedback is determined by comparing the distance between the actual driving endpoint of the vehicle and the recommended destination. Based on the feedback data, the system adjusts the destination confidence weight using the incremental learning rule, and analyzes the attribute residual between the recommended path and the actual path using the least squares method or the gradient descent algorithm to calibrate the sensitivity weight of each item in the user preference parameter.

[0013] Through the above scheme, the following beneficial technical effects are obtained:

[0014] The present application fuses the explicit intention features of voice semantics and device interconnection data, the implicit preference probability mined based on historical travel records, and the environmental impedance component reflecting real-time road conditions, and constructs a multi-dimensional comprehensive recommendation score model. This mechanism enables the navigation system to change from passive instruction reception to active demand perception, accurately predicts the travel destination without manual input by the user, reduces the frequency of human-computer interaction during driving, and reduces the risk of distraction of the driver.

[0015] The present application adopts a double-determination logic of an activation threshold based on vehicle speed dynamic adjustment and a relative safety margin based on score difference. The logic can automatically adjust the sensitivity of the system according to the driving state of the vehicle, suppress the recommendation of low confidence in high-speed driving or ambiguous decision (such as two candidate destinations with similar scores), avoid the interference of invalid information on the driver, and ensure that the timing of active navigation triggering is accurate and meets the safety driving requirements.

[0016] The present application plans the path using personalized weight parameters including time, distance and cost dimensions, and realizes fine guidance to specific parking lots or drop-off points in combination with endpoint facility node data, and updates the user model in a closed loop through collection of explicit interaction and implicit trajectory feedback. This design not only solves the problem of traditional navigation guidance ambiguity, but also enables the system to continuously adapt to the personalized driving habits of the user (such as avoiding tolls or preferring high speed) as the use time increases, realizing the continuous evolution of the navigation strategy. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a method flowchart of an active navigation method for a vehicle provided by one or more embodiments of the present application.

[0018] Figure 2 is a comprehensive recommendation score calculation and decision logic diagram of an active navigation method for a vehicle provided by one or more embodiments of the present application.

[0019] Figure 3Fig. 1 is a schematic diagram of a model parameter updating mechanism of an active navigation method for a vehicle according to one or more embodiments of the present application. DETAILED DESCRIPTION

[0020] The technical solutions of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0021] With reference to the drawings Figure 1 , the present application provides an active navigation system 100, comprising an intention perception module 101, an environment perception module 102, a user model construction module 103, an active decision module 104 and an interaction feedback module 105. The intention perception module 101 is configured to collect and process real-time behavior data of a user to generate an intention feature vector. The intention perception module 101 is connected with a vehicle-mounted microphone array and a mobile terminal data interface. The intention perception module 101 extracts a user input feature set at time t. The feature set is defined as: ;

[0022] wherein, is a semantic feature vector of a voice instruction, which is obtained by analyzing a collected voice stream through a natural language processing algorithm, and the vector represents a keyword feature related to travel; is a smart device data feature vector, which is obtained by reading calendar events and email content of associated devices, and the vector represents structured data containing a timestamp and a location coordinate.

[0023] The intention perception module 101 calculates an intention probability of a potential travel demand existing at the current time based on the feature set .

[0024] The intention perception module 101 outputs the calculated and to the active decision module 104. The environment perception module 102 is configured to collect navigation environment state data. The environment perception module 102 obtains external data through a vehicle networking communication interface and a vehicle-mounted sensor interface. The environment perception module 102 generates an environment state vector at time t , which is defined as follows:

[0025] ;

[0026] wherein, ​traffic congestion coefficient representing the target area or the current path, being a normalized value; weather impact factor representing the degree of blockage of meteorological conditions on vehicle passing; density and distribution feature vector of surrounding points of interest. The environment perception module 102 outputs the environment state vector to the active decision module 104 and the user model construction module 103. The user model construction module 103 is configured to store historical travel data and calculate user preference probabilities. The user model construction module 103 includes a database unit that stores a candidate destination set and corresponding historical travel records.

[0027] The user model construction module 103 calculates the prior preference probability for any candidate destination in combination with the current environment state . This probability represents the historical tendency of the user to choose the destination under the current environment state, and the calculation formula is as follows:

[0028] ;

[0029] wherein, represents the total number of historical travel records; is an indicator function, taking a value of 1 when the destination of the historical travel is and when the similarity between the environment state and the current environment state vector satisfies a preset condition, and 0 otherwise; represents a time decay factor, taking a value in the range of , used to adjust the influence of historical time distance on the weight; is the timestamp of the historical record; is a normalization constant.

[0030] The user model construction module 103 also stores a user preference parameter set defined as:

[0031] ;

[0032] wherein, , , correspond to the weight parameters of the user for passing time, travel distance, and passing cost, respectively. The user model construction module 103 outputs and to the active decision module 104.

[0033] The proactive decision-making module 104 is configured to calculate a recommendation score based on the input data and generate a trigger command. The proactive decision-making module 104 is connected to the intent perception module 101, the environment perception module 102, and the user model construction module 103, respectively.

[0034] The proactive decision-making module 104 targets each candidate destination. Calculate the overall recommendation score .

[0035] The active decision-making module 104 executes the threshold determination logic. The activation threshold is set to... The proactive decision-making module 104 filters out those that meet the requirements. optimal destination And send it to the interactive feedback module 105.

[0036] The interactive feedback module 105 is configured to perform human-computer interaction operations and collect feedback data. The interactive feedback module 105 is connected to the vehicle display screen and the vehicle audio output device.

[0037] After receiving the active navigation command, the interactive feedback module 105 renders and displays an active navigation card on the in-vehicle display screen. The active navigation card contains the optimal destination. The system displays the name and estimated arrival time of the navigation card. The interactive feedback module 105 responds to the user's click on the active navigation card, generating a navigation start signal and transmitting it to the navigation engine. The interactive feedback module 105 collects the user's interactive feedback behavior data. The data is then transmitted to the user model building module 103 for parameter updates. The user model building module 103 updates the optimal destination based on the feedback data. Historical weight The updated formula is as follows:

[0038] ;

[0039] in, Values These correspond to accepting recommendations, no action taken, and rejecting recommendations, respectively. This represents the learning rate parameter; This represents the correlation coefficient between environmental states. The updated weights will affect the prior preference probability at the next time step. The calculation results.

[0040] The specific execution process of each step in the active navigation method provided by this invention is described below:

[0041] Multi-source data parallel collection and initialization: After the vehicle-mounted system is started or a user gets on the vehicle, the system initializes the intention perception process and the environment perception process. The two processes run in parallel in independent threads. The intention perception process establishes an audio stream connection with the vehicle-mounted microphone array, and attempts to establish an encrypted data channel with the bound mobile terminal through Bluetooth or WiFi protocol, requests synchronization of calendar and email application data access permissions. The environment perception process simultaneously reads vehicle sensor data through the vehicle bus (CAN / LIN), and initiates a request for road condition data and geographic information data to the cloud server through the vehicle networking module (T-BOX).

[0042] Feature extraction and model parameter retrieval: The system preprocesses the acquired raw data. For voice data, noise reduction processing and keyword extraction are performed to generate semantic feature vectors For calendar and email data, time entities and place entities in the text are parsed to generate device data feature vectors . At the same time, the system accesses the local or cloud user model database, retrieves the current user's ID, and retrieves the user's corresponding historical preference parameter set and historical trip record table. The system filters out the historical environmental state data related to the current scene according to the current timestamp and GPS positioning coordinates.

[0043] Comprehensive recommendation score calculation and threshold determination: The system inputs the generated feature vectors and retrieved model parameters into the preset decision algorithm. The system traverses each destination in the candidate destination set , calculates the comprehensive recommendation score using the aforementioned formula. After calculation, the system performs comparison logic to compare the highest score calculated with the preset activation threshold . If , it is determined that the active navigation condition is met, and the system generates a trigger signal containing the destination coordinates and the recommendation reason (such as detecting a meeting schedule); if , the system remains in a silent monitoring state and returns to continue sampling in the next cycle.

[0044] Rendering and response of active interaction interface: In response to the generated trigger signal, the system graphics processing unit (GPU) renders an active navigation card on the top layer of the current vehicle machine display interface (HMI). The card contains the destination name, estimated time of arrival (ETA), and one-key navigation touch area. The system starts an interaction timer to monitor the user's operation behavior within the preset time window. If a click signal of the one-key navigation touch area is detected, the system immediately locks the destination Send path planning instructions to the underlying navigation engine and switch to the navigation guidance interface; if the timer times out and no operation is detected, the card is automatically collapsed and the behavior is marked as implicit rejection feedback.

[0045] Dynamic navigation execution and environmental monitoring: During navigation guidance, the system polls the environmental perception module at a preset frequency (e.g., every 30 seconds). The system continuously calculates the real-time travel cost function of the currently planned route. If the traffic congestion coefficient of the road ahead is detected... A mutation occurred, causing the current route to... If the preset tolerance threshold is exceeded, the system automatically triggers replanning logic based on the user's preference parameter set. Weights in (e.g.) The system recalculates the optimal route and provides voice prompts to the user regarding route changes.

[0046] Feedback Data Recording and Model Updates: The system records user interaction behaviors (clicks, ignores, closes) and route selection behaviors (accepting replanning, manually switching routes), generating feedback data. The system will With the current environment state vector The associated data is stored in the user model database. The system executes a model update process in the background, correcting the destination based on feedback data. Historical weights and user preference parameter sets The values ​​in the data complete the self-learning loop.

[0047] Detailed destination guidance: When the vehicle's location coordinates are close to the destination The straight-line distance is less than the threshold. When the distance is 200 meters, the system retrieves the destination. Distribution characteristics of surrounding points of interest The system combines the user's recorded destination behavior preferences (e.g., preference for underground parking lot entrances) with the recommended parking location or specific entrance coordinates on the map, and plays corresponding voice guidance information until the vehicle reaches the specific coordinates.

[0048] The present stage specifically includes an intention data acquisition sub-process, an environment data acquisition sub-process, and a multi-source data fusion preprocessing sub-process. In the intention data acquisition sub-process, the system acquires original input through two independent data channels. The first channel is a speech signal processing channel. A vehicle-mounted microphone array acquires in real time an in-vehicle sound field signal, and an audio processing unit (DSP) performs beamforming and echo cancellation (AEC) processing on the original sound wave signal to suppress environmental background noise. The noise-reduced audio stream is input to a voice activity detection (VAD) unit, and when a valid speech frame is detected, the system triggers an automatic speech recognition (ASR) engine to convert the audio signal into a text string. Subsequently, a natural language understanding (NLU) unit performs entity recognition (NER) on the text string to extract semantic slot data representing locations, times, and event types. The second channel is a device interconnection channel. The vehicle-mounted terminal establishes a communication session with the mobile terminal through a Bluetooth protocol stack (such as a PBAP or MAP profile) or a Wi-Fi point-to-point connection. After security authentication, the system calls the calendar and mail application interface (API) of the mobile terminal to capture iCalendar format schedule objects and mail header information. The system uses a preconfigured regular expression rule library to parse the captured data, extract structured fields containing timestamps and geographic locations, and perform privacy desensitization operations to remove sensitive characters related to names and contact information.

[0049] In the environment data acquisition sub-process, the system synchronously acquires vehicle state and external environment information. Vehicle state data is read through a controller area network (CAN) bus, including the current GPS latitude and longitude coordinates, heading angle, and instantaneous speed of the vehicle. External environment information is acquired by sending an HTTP / HTTPS request to a cloud traffic service platform through a vehicle-mounted T-BOX module. The request message contains the current vehicle location coordinates and a query radius. The system receives the JSON format response message returned by the cloud, parses the road traffic flow data (such as average speed, congestion index) and weather data (such as precipitation probability, visibility value) contained therein. At the same time, the system loads local high-precision map data blocks to retrieve the attribute data of points of interest (POIs) around the current coordinates.

[0050] In the multi-source data fusion preprocessing sub-process, the system performs spatio-temporal alignment and standardization processing on the above-mentioned asynchronously acquired data. Due to different sampling frequencies and network delays of various sensors, the system sets a sliding time window (For example, 500 ms). The system aligns all data falling within the same time window Data frames within the same time instance are considered as sampled data. For data with low sampling frequency (e.g. cloud traffic), the system adopts zero-order hold (ZOH) or linear interpolation method to up-sample the data to match the main frequency of the system. Subsequently, the system performs normalization on the heterogeneous data to map physical quantities with different dimensions to the numerical interval of [0, 1] to construct a standardized feature vector. For the normalization of traffic congestion coefficient , the following formula is used:

[0051]

[0052] wherein, is the real-time average speed of the current road segment, is the highest speed limit of the road segment. If the result is less than 0, it is truncated to 0; if it is greater than 1, it is truncated to 1. For the weather influence factor , the system assigns values based on the pre-set resistance coefficient table according to the weather code. After the above processing, the system writes the generated feature set and the environmental state vector into the shared memory area for subsequent decision modules to call.

[0053] This part elaborates on the specific implementation process of how the active decision module calculates the recommended score, performs threshold decision, and manages trigger instructions based on the input data. The active decision process first performs the candidate destination screening step. Since the historical location set stored in the user model database contains a large amount of data, in order to reduce the delay of real-time calculation, the active decision module first performs spatial filtering based on the GPS coordinates of the current vehicle. The system sets a dynamic search radius (e.g. 50 kilometers) and only keeps the historical locations within the radius as the effective candidate set . In addition, if the device data feature vector extracted by the intent perception module contains explicit location text information (e.g. location field in the calendar), the system will convert the text to coordinates through geocoding, and if the location corresponding to the coordinates is not in the historical set, it will be temporarily added to .

[0054] For each destination in the effective candidate set , the active decision module calculates the comprehensive recommended score . This calculation process involves three-dimensional quantitative evaluation:

[0055] The first dimension is the explicit intent matching evaluation. The module calculates the product of the intent probability and the intent matching degree . Among them, the intent matching degree function ​Adopting feature vector similarity calculation logic. If the speech semantics or calendar location text in the feature set contains the character or has a cosine similarity higher than a preset value with the name or label of the destination (such as a company or home), the value is 1, otherwise the value is 0. This item reflects the explicit travel demand expressed by the user at present.

[0056] The second dimension is the implicit habit preference evaluation. The module calls the priori preference probability output by the user model construction module . This value reflects the conditional probability of the user going to the destination based on historical statistical rules under the current time, weather and geographical location.

[0057] The third dimension is the navigation resistance cost evaluation. The module calculates the navigation resistance cost function , which is defined as:

[0058] ;

[0059] wherein, represents the average traffic congestion coefficient of the planned path from the current location to the destination ; and are resistance weight coefficients. This item is used for weight reduction processing of destinations with poor road conditions or bad weather. The comprehensive recommendation score is obtained by the weighted sum formula:

[0060]

[0061] wherein, wherein, represents the intent matching degree function, which is used to calculate the correlation value of the candidate destination and the feature set ; represents the navigation resistance cost function, which is calculated based on and ; is a preset adjustment coefficient, which is used to balance the weight relationship between explicit intent, implicit habit and environmental resistance. For example, the system is set to ensure that when the user has a clear schedule, the priority of the explicit intent is higher than the daily habit.

[0062] After the calculation is completed, the system performs the winner-takes-all and threshold limit logic. The system first selects the optimal destination with the highest score from all candidate results:

[0063] ;

[0064] ​​Subsequently, the system compares the score with a preset activation threshold . Only when , the system enters the trigger decision stage. In the trigger decision stage, in order to prevent frequent pop-ups from disturbing the user, the system introduces a cooling time mechanism. The system maintains a trigger timestamp register . When the score threshold condition is met, the system checks the difference between the current time and . If ( is a preset cooling period, for example, 10 minutes), the system intercepts this trigger instruction; if , the system allows triggering and updates to the current time.

[0065] Once the trigger instruction is generated, the active decision module sends a data packet containing the latitude and longitude of the destination , the POI name and the recommended confidence to the interactive feedback module. At the same time, the decision module pre-sends a pre-planning instruction to the underlying navigation engine, requesting the background thread to pre-calculate the route data to , but does not immediately render the route on the map, thereby ensuring zero-delay navigation startup when the user clicks on the card.

[0066] This part elaborates in detail how the system monitors environmental changes, evaluates route advantages and disadvantages, and implements path switching during the navigation guidance stage. In the process of vehicle driving according to the current planned route, the active decision module runs the path monitoring subprogram in the background. This subprogram implements a dual monitoring mechanism: periodic polling monitoring and event-driven monitoring. Periodic polling monitoring refers to the system actively requesting the latest traffic congestion coefficients of each road segment along the current route from the environmental perception module at a preset time interval (for example, every 60 seconds); event-driven monitoring refers to the system listening to the vehicle-to-everything (V2X) broadcast messages or cloud traffic event push interface, and triggering the evaluation logic as soon as a sudden event message (such as a traffic accident or a temporary road closure construction) occurs on the front path is received.

[0067] When the monitoring mechanism detects changes in environmental state data, the system starts the route re-evaluation process. The active decision module first recalculates the passing cost of the current remaining route based on the latest environmental state vector. This calculation process is weighted according to the preference parameter set stored in the user model. Specifically, the system multiplies the predicted passing time of the remaining road segment, the remaining distance, and the passing fee of the remaining road segment by the user's corresponding time weight , distance weight and fee weight , and sums the weighted results to obtain the real-time total cost of the current route .

[0068] Subsequently, the system initiates a path search algorithm (e.g. improved A* or Dijkstra) in the background to search for alternative routes from the current location to the destination in the road network topology graph. During the search process, the algorithm also utilizes the aforementioned set of user preference parameters as the basis for calculating the edge weights of the road network. For example, if the user model shows a high value (indicating that the user is sensitive to cost), the algorithm will significantly increase the cost value of toll roads when expanding nodes, thereby tending to search for alternative routes that avoid toll stations. The algorithm outputs the candidate alternative routes with the lowest cost and their corresponding total cost. The system executes the switching decision logic to calculate the path switching benefit . In order to avoid frequent route jumps due to small benefit differences, the system sets a dynamic switching threshold . This threshold is not a fixed value, but is positively correlated with the remaining journey length. Only when (e.g. expected time saving exceeds 5 minutes or total cost reduction is more than 15%) does the system determine the necessity of route switching.

[0069] Once the need for switching is determined, the interactive feedback module performs the active prompting operation. The system broadcasts the reason for the change in road conditions and the advantages of the new route (e.g. congestion ahead, new route planned, expected time saving of 8 minutes) through text-to-speech (TTS), and displays the trajectory difference between the original route and the new route in split-screen or pop-up form on the center screen. The system sets a confirmation countdown window (e.g. 10 seconds), and if the user clicks confirm or does not operate (default acceptance) within the countdown, the system sends instructions to the navigation engine to guide the trajectory to smoothly switch to the new route; if the user clicks cancel or rejects through voice instructions, the system retains the original route and suspends the re-planning prompt for the same congestion event for a preset silent period (e.g. 15 minutes).

[0070] This section elaborates on how the intent perception module converts unstructured speech and device data into quantifiable intent probability values. The core of intent calculation lies in the construction of feature vectors and probability mapping. First, for the speech feature vector , the system does not directly use the original audio waveform, but constructs a high-dimensional vector based on natural language processing results. The system has a pre-set navigation intent semantic library, which classifies words into strong intent categories (such as navigate to, route, go to) and weak intent categories (such as nearby, where). The system performs word segmentation on the recognized text and calculates the term frequency (TF) of each category of words in the sentence and the confidence score output by the automatic speech recognition engine. ​​Each dimension of the feature vector corresponds to a weighted score of a different semantic category. For example, if the speech contains the phrase "going to the company", the values corresponding to the strong intent and the location dimension of the company are set to high.

[0071] For the device data feature vector , the system performs a time window based validity mapping. The system iterates through all the resolved schedule events, and filters out the future events whose start time is later than the current system time . For each future event, the system calculates the time difference . The system constructs a time urgency feature using an inverse proportion decay function.

[0072] ;

[0073] wherein is a normalization coefficient, is a tiny amount to prevent the denominator from being zero, is a preset associated time window (e.g. 2 hours). If exceeds the time window, the feature value is 0. The feature vector is composed of and a binary flag (0 or 1) representing the completeness of the location information. Based on the constructed feature vector , the system calculates the normalized intent probability using a logistic regression model. The specific linear weighting and nonlinear activation process is described by the following formula:

[0074] ;

[0075] wherein, wherein represents the weight vector of the speech data, represents the weight vector of the device data, represents the bias parameter, represents the activation function. These weight coefficients are obtained through supervised learning on a large number of historical samples (containing feature input and user's actual start navigation label), reflecting the contribution of different signal sources to the judgment of user's travel intent. For example, after training, the weight of the speech feature containing strong intent keywords is usually significantly higher than that of the device feature containing only schedule reminders. is the bias parameter, used to set the basic sensitivity threshold of the system. is the Sigmoid activation function, used to nonlinearly compress the result of the weighted sum to the interval (0, 1), and the output value represents the confidence of the user's travel demand at the current time.

[0076] This section elaborates on how the user model construction module mines the user's location preferences in context based on historical data. The core of the user preference prediction is to compute the conditional probability , which is essentially a similarity-weighted Bayesian inference process. The algorithm iterates through all the historical trip records in the database, and for each record , evaluates its contextual similarity to the current time and the current environment .

[0077] First, the algorithm defines the specific decision logic of the indicator function . This function is used to filter whether a historical record is a valid sample. The decision contains two layers of matching in time dimension and space dimension: in time dimension, the system introduces the matching mechanism of TimeSlot and DateType. The system divides a day into several equal-length time slots (e.g., every 30 minutes for a slot) and labels the date as weekday or non-weekday. Only when the departure time of the historical record falls within the time slot tolerance range (e.g., 1 hour before and after) of the current time and the date type is consistent, the time dimension matching is successful.

[0078] In space dimension, the system computes the Euclidean distance between the departure point coordinates of the historical record and the current vehicle positioning coordinates. Only when the distance is less than the preset starting clustering radius (e.g., 500 meters), the space dimension matching is successful. The indicator function performs logical and operation of the above two dimensions. Only when both time and space dimensions are matched, and the destination of the historical record is indeed , the function outputs 1, otherwise 0. This means that the system only refers to those historical behaviors that occur in similar time and place.

[0079] Second, the algorithm introduces a time decay mechanism to weight the valid samples. The system uses an exponential decay model to handle the timestamp of the historical record. The decay weight calculation term is , where is a forgetting factor between 0 and 1. This mechanism gives higher weight to recent data. For example, the contribution of the trip that occurred last week to the current prediction is significantly higher than the trip that occurred a year ago. This enables the model to adapt to recent changes in the user's living habits (such as changing jobs or moving), and the weight of old habits will automatically decrease over time.

[0080] Finally, the system performs probability normalization processing. For each candidate destination , the system accumulates the decay weights of all valid historical records that satisfy the indicator function, and obtains the non-normalized score of the destination. Then, the system calculates the normalization constant , i.e., the sum of scores of all candidate destinations. By dividing the score of each destination by , the system outputs the prior preference probability that satisfies the probability axiom. This probability value directly reflects the statistical likelihood of the user going to the destination according to the historical habits in the current spatio-temporal environment.

[0081] With reference to the accompanying Figure 2 , this part elaborates on how the active decision module fuses multi-dimensional features to generate the final trigger instruction, and focuses on the weighted fusion strategy and dual threshold judgment mechanism adopted in the decision-making process. The active decision module adopts a linear weighted fusion model to calculate the comprehensive recommendation score of each candidate destination . This calculation logic aims to unify heterogeneous input signals into a scalar dimension for comparison, specifically involving the coupling processing of three core components:

[0082] The first component is the intention-driven component, whose value is determined by the product of the intention probability and the intention matching degree , and is weighted by the adjustment coefficient . This component represents the strength of the "current explicit demand" in a physical sense. When the system detects an explicit travel instruction through voice semantic analysis or schedule parsing, and the candidate destination highly matches the semantic features of the instruction (such as place names, event-related locations), the value of this component will jump up. At this time, the coefficient ensures that this component dominates the total score, thereby ensuring that the system can override historical habits and prioritize responding to the user's real-time needs when there is an explicit instruction input.

[0083] The second component is the habit-driven component, whose value directly comes from the prior preference probability output by the user model, and is weighted by the adjustment coefficient . This component represents the historical behavior inertia. In the silent start-up scenario where there is no explicit voice or schedule input (for example, the user starts the vehicle in the morning on weekdays without speaking and without a schedule), the intention-driven component tends to zero, and the comprehensive score is mainly supported by this habit-driven component. By presetting the baseline value, the system ensures that in the absence of instruction input, it can still accurately push high-frequency destinations (such as the company or home) based on time rules and geofencing algorithms.

[0084] The third component is the environmental impedance component, whose value is derived from the navigation resistance cost , and is negatively weighted by the coefficient (i.e. as a penalty term). This component introduces the constraint effect of the physical environment on the trip. Even if a destination perfectly matches the user's historical habits (e.g. a restaurant frequently visited on Fridays), if the environmental perception module detects that the road segment leading to this destination currently has severe congestion or extreme weather (resulting in a significant increase in travel time), this component will substantially lower the overall score of this destination by a large negative weight. This mechanism endows the decision logic with a dynamic risk-aversion feature, which can automatically suppress the recommendation to go to areas with poor traffic conditions, or sink the recommendation order.

[0085] After the score calculation of all candidate destinations is completed, the active decision module executes a final decision logic based on a double-threshold to ensure the high confidence of the recommendation result:

[0086] The first threshold is an absolute threshold determination. The system filters out the candidate destination with the highest current score , and compares its score with a pre-set activation threshold . In this embodiment, is designed as a dynamic variable associated with the driving state of the vehicle, rather than a fixed constant. When the vehicle CAN bus data indicates that the vehicle speed exceeds a pre-set safety limit (e.g. 80 km / h), the system automatically adjusts the value of upwards. This dynamic adjustment strategy aims to increase the triggering threshold in high-speed driving state, and reduce the distraction of unnecessary information to the driver's attention, only when the recommendation confidence is extremely high, the system will disturb the driver.

[0087] The second threshold is a relative threshold determination (confidence margin check). The system calculates the difference between the highest score and the second highest score . Only when and at the same time (safety margin) is satisfied, the system finally locks as the recommended target. If is less than the safety margin, it indicates that the system's judgment on the two candidate destinations (e.g. a company office building and a coffee shop next door) is ambiguous. In this case, the system executes a conservative disambiguation strategy, postpones the triggering of the active recommendation, and instead waits for more feature inputs in the next data collection period, or only displays a weak prompt in the interface sidebar rather than a pop-up strong interference card.

[0088] This section details how the system utilizes user preference parameters to construct a dynamic cost function to generate personalized routes, and how it achieves precise guidance based on behavioral habits at the end of the journey. During the route planning phase, the system's underlying navigation engine does not use a single shortest path or fastest path algorithm, but instead employs a directed graph search strategy based on multi-objective weighting. The system first abstracts the road network data into a topological graph containing nodes and edges, where each edge... This represents a section of actual road. The system assigns each edge... Dynamically calculate a comprehensive passage cost All paths belonging to the planning path The sum of the costs of each segment constitutes the total cost function of the path. Comprehensive toll costs The computational logic deeply integrates real-time environmental data with personalized user parameters. Specifically, the cost is a weighted composite of three independent sub-dimensions:

[0089] First, there's the time dimension cost, corresponding to the formula in... The system obtains road segment information in real time. Traffic congestion coefficient (Or real-time average vehicle speed), combined with the physical length of the road segment, to calculate the estimated travel time. For severely congested road segments, the travel time increases significantly, causing this value to rise. The time sensitivity weight is extracted from the user model. If the user's historical behavior shows that they tend to choose the shortest route (e.g., they often take the highway), the weight value is larger, which makes the algorithm tend to avoid congested road sections when searching.

[0090] Secondly, there is the distance dimension cost, corresponding to the formula in... Item. Among them. This refers to the physical mileage of the road segment. This reflects the user's sensitivity to driving mileage. If a user prefers to take shortcuts to save fuel or mileage, this weight increases the total cost of long-distance detours, prompting the algorithm to search for paths with shorter physical distances. Finally, there's the economic dimension cost, corresponding to the formula... Item. Among them. This refers to the toll fees for a road segment (such as highway tolls and bridge tolls). This reflects users' sensitivity to costs. For toll roads, It is a positive value; for toll-free sections, The value is zero. If the user model indicates that the user is price-sensitive (e.g., frequently avoids toll booths in their past travel history), Setting it to a high value, thus imposing a severe penalty on toll roads, makes the planned... Automatically avoid toll-charging paths.

[0091] The system traverses the road network by Dijkstra or A* graph search algorithm to find the path that minimizes the total cost function as the recommended route. This mechanism ensures that the navigation solution is not a one-size-fits-all, but a personalized optimal solution dynamically generated according to the user's time-saving, fuel-saving or money-saving preference.

[0092] At the end of the trip, the system performs the final 100-meter fine-tuning logic. When the vehicle positioning coordinates and the destination Euclidean distance is less than the preset arrival threshold (eg 200 meters), the system automatically activates the end-point guidance mode.

[0093] At this time, the system no longer regards the destination as a single coordinate point, but retrieves high-precision map data or POI depth information to retrieve a set of facility nodes associated with the destination. This set includes but is not limited to: front door entrance, side door entrance, underground parking entrance, ground parking area and temporary drop-off point.

[0094] The system calls the end-point behavior preference record in the user model . This record stores the user's final parking location information when going to this destination or similar destinations (such as shopping malls) in the past. The system matches the set of facility nodes with the preference record. If the match is successful (for example, the user always habitually parks at the underground second floor garage entrance of a certain shopping mall), the system automatically corrects the navigation end-point from the POI center coordinates to the coordinates of the entrance.

[0095] At the human-computer interaction level, the system generates guidance instructions according to the corrected end-point coordinates. The map display interface is automatically zoomed in to the maximum level, and a guidance line is rendered to the specific entrance (such as the parking lot gate), and the specific guidance information is broadcasted (such as "Arriving at the destination, guiding you to the underground parking lot entrance"), thus solving the problem of traditional navigation ending at the roadside, and realizing the complete closed loop from door to door.

[0096] This part elaborates how the interactive feedback module captures the user's explicit instructions and implicit behaviors by monitoring the state of the human-computer interaction interface and the actual driving trajectory of the vehicle, and converts them into standardized feedback signals for subsequent model optimization.

[0097] ​The collection process of user feedback data is divided into explicit interaction collection channel and implicit behavior verification channel. The system monitors the input state of the vehicle-mounted human-machine interface (HMI) in real time through an event listener. In the explicit interaction collection channel, the system defines a clear user operation response logic. When the active navigation card is rendered on the screen, the system monitors the coordinate signal of the display screen touch layer. If it is detected that the touch point coordinate falls within the pixel area of the start navigation or confirmation control, the system generates a positive explicit feedback signal (marked as ). If it is detected that the touch point falls within the close, no longer prompt control area, or the user clicks the modify destination button on the card and inputs a new location, the system generates a negative explicit feedback signal (marked as ). At the same time, the system keeps listening to the voice channel, and if the automatic speech recognition (ASR) engine parses negative semantic instructions such as cancel, no, go somewhere else, the system also generates a negative feedback signal, and records the real destination input by the user as a correction sample.

[0098] In the implicit behavior verification channel, the system processes feedback data in non-interactive scenarios. The system starts a background timer when the navigation card is popped up, setting a response time window (e.g., 15 seconds). If the timer expires and no touch or voice input signal is detected, the card is automatically closed, and the system marks this behavior as ignore or no operation feedback (marked as .

[0099] In order to further distinguish the true meaning of ignore behavior (i.e., whether the user thinks the recommendation is wrong or correct but does not need navigation), the system continuously tracks the subsequent driving trajectory of the vehicle in the background. The system records the driving end coordinates of the vehicle after the card disappears. The system performs a geofencing comparison logic:

[0100] If the calculation finds that the distance between and the recommended destination is less than the determination threshold, it indicates that the system prediction is accurate, but the user is familiar with the route and does not need guidance. At this time, the system will correct the previous label, marking it as a weak positive sample, to strengthen the accuracy of the intent prediction model, but at the same time, reduce the active disturbance weight of this route in the user preference parameters (i.e., consider that the user prefers silent driving on this path).

[0101] If the distance between and is greater than the determination threshold, it indicates that the system prediction of the destination is wrong. At this time, the system marks the feedback as a substantial negative sample, and the real location corresponding to correct labels are stored in the training dataset to correct the prediction model.

[0102] All feedback data is structured as a five-tuple log , where is the timestamp, is the recommended destination. This log file is encrypted and stored in the non-volatile partition of the on-board memory, waiting for the system to trigger the model training and parameter updating process during idle time (e.g. when the vehicle is turned off and charging, or parked at night and connected to Wi-Fi). Through this multi-dimensional collection method, the system can distinguish between "prediction errors" and "interaction redundancies", thereby achieving more accurate self-evolution.

[0103] Referring to the accompanying Figure 3 , this part elaborates on how the system uses the collected feedback data and real driving trajectory data to correct the weight coefficients in the core algorithm and the user preference parameters, in order to achieve the convergence of prediction accuracy. The model updating mechanism mainly consists of two parallel updating loops: the destination prediction model updating loop and the path planning parameter updating loop. Both loops are triggered and executed by the background computing unit during the vehicle's idle period after being turned off or when the system is idle.

[0104] In the destination prediction model updating loop, the system aims to correct the calculation weight of the prior preference probability . The system maintains a historical destination weight table, which stores the confidence weights of each candidate destination under different environmental conditions . After receiving a complete feedback event , the system performs weight correction based on the error gradient. If the feedback is positive , the system increases the weight of the recommended destination . The update logic follows the incremental learning rule:

[0105] ;

[0106] where is the preset learning rate (e.g. 0.01), which determines the correction amplitude of the model by a single feedback; is the environmental similarity coefficient, which limits the weight update range to only similar scenarios to the current environment (e.g. only updating the weight corresponding to Monday morning rush hour, without affecting the weight on weekends).

[0107] If the feedback is negative , the system performs a two-way correction: first, it reduces the weight of the incorrect recommended destination , and at the same time, it increases the weight of the destination actually visited by the user the weights of the attributes. Through this penalty-reward mechanism, the system can gradually suppress the recommendation probability of low-frequency or false locations, and promote the priority of real frequently-visited locations. In the path planning parameter updating loop, the system aims to calibrate the user preference parameter set . The updating process is based on the analysis of the difference between the recommendation and the actuality. The system retrieves the attribute vector of the recommended path (containing the predicted time , distance , cost ), and the attribute vector of the actual traveled path (containing the actual time , distance , cost ).

[0108] The system calculates the attribute residual vector . Subsequently, the system adjusts the coefficients in each item of using the least square method or gradient descent algorithm, so that the path cost function calculated under this set of parameters can determine to be better than .

[0109] Specifically, if the data shows that the user frequently chooses a path that is longer in distance but costs 0 (i.e. and ), the system detects that the current setting of is not sufficient to reflect the user's sensitivity to cost. The system then increases the value of and correspondingly decreases the value of or . After multiple iterations, will converge to the parameter combination that best fits the user's driving behavior characteristics.

[0110] In addition, the system also contains adaptive adjustment logic for the decision threshold . The system counts the proportion of negative feedback (false positive rate ) in the last trigger records. If exceeds the preset tolerance upper limit (e.g. 15%), the system automatically increases the activation threshold by a step (e.g. 0.05), thereby reducing the sensitivity of the system and reducing the invalid disturbance to the user; on the contrary, if the user's adoption rate is very high, the system can appropriately lower the threshold to cover more potential travel needs.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An active navigation method for automobiles, characterized in that, Includes the following steps: Collect multi-source input data, including intent data obtained through voice and device interconnection channels, and environmental data obtained through vehicle bus and cloud services; The multi-source input data is preprocessed to extract intent feature vectors and environment state vectors, and spatiotemporal alignment processing is performed. The intent probability is calculated based on the intent feature vector, the prior preference probability is calculated based on the environmental state vector and the historical travel records in the user model, and the navigation resistance cost is calculated based on the environmental state vector. A comprehensive recommendation score for candidate destinations is calculated based on the intent probability, the prior preference probability, and the navigation resistance cost. Determine whether the comprehensive recommendation score meets the preset triggering conditions. If so, generate an active navigation instruction and plan a recommended path based on user preference parameters. As the vehicle travels along the recommended route, the recommended route is dynamically adjusted based on real-time environmental changes, and the user model is updated based on user feedback data regarding the active navigation commands.

2. The active navigation method for automobiles according to claim 1, characterized in that, The steps of preprocessing the multi-source input data to extract the intent feature vector and the environment state vector specifically include: Beamforming and echo cancellation are performed on the speech signal in the intent data, and semantic slot data is extracted through natural language understanding; Simultaneously, calendar objects and email information are captured through mobile terminal interfaces, and timestamps and geolocation fields are parsed and extracted. The vehicle position coordinates, heading angles, road traffic flow data, and meteorological data obtained from the cloud are analyzed in the environmental data. A sliding time window is set, data frames falling within the same sliding time window are aligned, and physical quantities of different dimensions are normalized and mapped to a standard numerical range to construct the intention feature vector and the environment state vector.

3. The active navigation method for automobiles according to claim 1, characterized in that, The steps for calculating the intent probability based on the intent feature vector specifically include: Based on the difference between the start time of the scheduled event and the current system time, a time urgency feature is constructed using an inverse proportional decay function; The input vector is constructed by combining the time urgency feature with the confidence score of the speech semantics; The input vector is fed into a pre-trained logistic regression model, and the intent probability, which represents the confidence level of the user's travel demand, is calculated through linear weighting and a non-linear activation function.

4. The active navigation method for automobiles according to claim 1, characterized in that, The steps for calculating prior preference probabilities based on the environmental state vector and historical travel records in the user model specifically include: Iterate through the historical travel records and use indicator functions to filter out valid historical records that match the current time slice and date type and whose starting point spatial distance is within the preset clustering radius; A time decay mechanism is introduced, and the weight of the effective historical records is processed using an exponential decay model, so that the weight of recent records is higher than that of distant records. The effective historical record weights of each candidate destination are accumulated and normalized to obtain the prior preference probability under the current spatiotemporal environment.

5. The active navigation method for automobiles according to claim 1, characterized in that, The specific calculation logic for the comprehensive recommendation score of the candidate destination based on the intent probability, the prior preference probability, and the navigation resistance cost is as follows: The product of the intent probability and the intent matching degree is calculated as the intent-driven component, wherein the intent matching degree characterizes the similarity between the candidate destination and the speech or schedule text. The prior preference probability is used as a habit-driven component; The navigation resistance cost is calculated based on the average traffic congestion coefficient of the route and weather impact factors, and the navigation resistance cost is used as an environmental impedance component. The comprehensive recommendation score is obtained by linearly weighting and summing the intention-driven component, the habit-driven component, and the environmental impedance component.

6. The active navigation method for automobiles according to claim 1, characterized in that, The steps for determining whether the comprehensive recommendation score meets the preset triggering conditions specifically include: Implement a winner-takes-all logic to select the single candidate destination with the highest score; Determine whether the comprehensive recommendation score of the unique candidate destination is greater than the activation threshold, which is dynamically adjusted according to the vehicle speed; Determine whether the difference between the comprehensive recommendation score of the only candidate destination and the second highest score is greater than a preset safety margin; If all of the above judgments are true, and the difference between the current time and the previous trigger time is greater than the preset cooldown period, then the preset trigger condition is satisfied.

7. The active navigation method for automobiles according to claim 1, characterized in that, The steps for planning recommendation paths based on user preference parameters specifically include: In the road network topology map, the comprehensive traffic cost is calculated for each road segment. The comprehensive traffic cost is a weighted composite of time dimension cost, distance dimension cost and economic dimension cost. The weight coefficients of the time dimension cost, distance dimension cost, and economic dimension cost correspond to the time weight, distance weight, and cost weight in the user model, respectively. The recommended path is the route that minimizes the total cost function of the path, which is then searched using a graph search algorithm.

8. The active navigation method for automobiles according to claim 7, characterized in that, It also includes refined endpoint guidance steps: When the distance between the vehicle and the destination is less than a preset arrival threshold, retrieve the set of facility nodes associated with the destination, which includes parking lot entrances and drop-off points; The system retrieves the endpoint behavior preference records from the user model and matches the facility node set with these records. The navigation endpoint is corrected to the coordinates of the successfully matched facility node, and a guidance instruction extending to the coordinates of the facility node is generated.

9. The active navigation method for automobiles according to claim 1, characterized in that, The user model is updated based on user feedback data regarding the active navigation commands, wherein the step of collecting the feedback data includes: Monitor touch operations and voice commands on the interactive interface to obtain explicit feedback; If no explicit feedback is detected within the preset response time window, record the actual destination coordinates of the vehicle's subsequent journey. The distance between the actual driving destination coordinates and the recommended destination is compared. If the distance is greater than the judgment threshold, it is marked as negative implicit feedback, and the actual driving destination coordinates are stored as the real label in the training dataset.

10. The active navigation method for automobiles according to claim 1, characterized in that, The steps of updating the user model based on user feedback data regarding the active navigation instructions specifically include: Based on positive or negative feedback, the confidence weights of candidate destinations under environmental conditions are adjusted using incremental learning rules. Calculate the residual vectors between the recommended route and the user's actual driving route in terms of time, distance, and cost attributes; The time weight, distance weight, and cost weight in the user preference parameters are calibrated using the least squares method or gradient descent algorithm, so that the computational cost of the actual driving path under the user preference parameters is better than that of the recommended path.

Images