A travel method and system based on multi-scene fusion perception

Abstract

This invention belongs to the field of travel method technology, and particularly relates to a travel method and system based on multi-scenario fusion perception. It collects ETC gantry data, LBS data, static road network features, dynamic traffic status, and spatiotemporal context data. A spatiotemporal consistency verification algorithm is used to achieve data spatiotemporal alignment. This algorithm includes timestamp correction based on vehicle clock offset and spatial coordinate matching based on 100-meter marker unit mapping. After fusing and aligning the multi-source data, the system identifies individual travel origin and destination points, service area information, and travel time, extracting full travel features. User behavior is labeled and weighted based on the 4W model, and users are categorized into commuter, commercial, and long-distance categories using a hierarchical clustering algorithm, mining features such as travel frequency, frequently used origin / destination (OD), and frequently used routes.

Description

Technical Field

[0001] This invention belongs to the field of travel method technology, and in particular relates to a travel method and system based on multi-scenario fusion perception. Background Technology

[0002] With the increasing density of highway networks and the surge in cross-regional travel demand, individual travel behaviors are becoming more diversified and dynamic, posing multiple technical bottlenecks to existing travel service systems. On the one hand, while ETC gantry data possesses high-precision spatiotemporal attributes, its distribution is sparse; while LBS data is continuous, it suffers from location drift and latency issues. Both types of data lack effective spatiotemporal consistency verification mechanisms, resulting in insufficient accuracy in identifying the start and end points of individual travel activity chains, capturing stop information, and reconstructing travel trajectories, making it difficult to support accurate travel feature mining and demand prediction. On the other hand, traditional travel information dissemination relies on a single channel, often employing a broadcast-style push model, failing to incorporate user travel profiles for personalized customization. This leads to low information compliance rates and an inability to provide differentiated services for different user types such as commuters, commercial users, and long-distance travelers.

[0003] Existing technologies lack adaptability to extreme traffic scenarios (such as high traffic volume during holidays, multiple concurrent accidents, and special weather conditions) and special road network structures (mountain highways, cross-sea bridges, and dense road networks in urban clusters). They often rely on static rules to formulate guidance strategies, making it difficult to dynamically respond to changes in road network conditions and real-time user needs. Furthermore, the accompanying service system is incomplete, lacking full coverage of pre-trip planning, in-trip guidance, and post-trip feedback. Information dissemination suffers from high response delays and content homogenization, failing to meet users' demands for precise and efficient travel services and hindering proactive traffic management and operational efficiency improvements on highway networks. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a travel method and system based on multi-scenario fusion perception.

[0005] In view of this, the present invention provides a travel method based on multi-scenario fusion perception, comprising the following steps: Step 1: Collect ETC gantry data, LBS data, static road network features, dynamic traffic status and spatiotemporal context data, and achieve spatiotemporal alignment of data through a spatiotemporal consistency verification algorithm. The spatiotemporal consistency verification algorithm includes timestamp correction based on vehicle clock offset and spatial coordinate matching of 100-meter pile unit mapping. Step 2: Merge and align the multi-source data to identify the origin and destination of individual trips, information on service areas stopped at, and trip time, and extract full travel features; Step 3: Tag and assign weights to user behavior based on the 4W model, and divide users into categories such as commuting, operation, and long distance through hierarchical clustering algorithm to mine travel frequency, frequently used OD, and frequently used route features; Step 4: A hybrid teaching LSTM encoder and decoder model is adopted to fuse historical trajectory and road environment information. The model includes a gantry position attention mechanism and a historical trajectory and road environment fusion layer to predict the origin and destination of the trip, the driving route and the information of the service areas to be stopped. Step 5: Based on user profiles and travel stages, provide personalized services throughout the entire travel process—before, during, and after the trip—through a multi-channel information dissemination system comprised of roadside electromechanical equipment and vehicle-mounted terminals.

[0006] Preferably, in the spatiotemporal consistency verification algorithm, the timestamp correction formula is: ; in, N is the number of successfully matched samples in the vehicle's history and N≥5; During spatial coordinate matching, LBS points are projected onto the nearest 100-meter pile unit. LBS points with a lateral offset of ≤50 meters and a distance of >200 meters from the nearest highway are marked as non-highway status.

[0007] Preferably, the gantry position attention mechanism includes: Map the gantry ID to a d-dimensional vector space. Construct attention score ; Normalization yields attention weights Generate context vector ; The gantry embedding dimension d=64 and the attention hidden layer dimension da=128.

[0008] Preferably, the historical trajectory and road environment fusion layer concatenates static road network features, dynamic traffic conditions, and spatiotemporal context into an environment vector. The decoder input at each step The data is fed into the LSTM unit to obtain... The output layer adds a road network constraint mask to ensure path topology connectivity.

[0009] Preferably, the hybrid teaching strategy employs a linearly decreasing decay function. =max(0.3, 0.7 - 0.005) In the early stages of training (epochs 1-30), the probability of teacher forcing is 70%, and it gradually decreases to 30% in the later stages (epochs 31-80). The model convergence condition is that the change in the edit distance of the validation set is less than 0.001 for 5 consecutive rounds or the decrease in training loss is less than 1e-4 / epoch.

[0010] Preferably, the user profile construction also incorporates association rule methods to mine travel-related characteristics, identify features such as whether an individual frequently uses service areas and travel distance, and support differentiated guidance; travel information prediction also integrates spatial data such as gantry location and road network topology to improve the accuracy of path prediction.

[0011] Preferably, the multi-channel information release system includes roadside variable message signs, lane indicators, variable speed limit signs, as well as vehicle-mounted navigation systems, apps, and in-vehicle T-BOX. The in-vehicle T-BOX supports the 3GPPR16 version C-V2X protocol, and the average communication latency from the roadside to the vehicle is <100ms. Personalized services are pushed according to the importance of information, and emergency events trigger audible and visual alarms through the C-V2XPC5 interface.

[0012] Preferred strategies are adopted for special road network scenarios such as mountain highways, cross-sea bridges, and dense road networks in urban agglomerations, including tunnel entry and exit time difference compensation, cross-sea travel profiles, and cross-city commuting heat maps. For special users such as commercial vehicles, commuters, and emergency vehicles, we provide customized services such as fatigue driving warning, lane-level priority navigation, and green wave escort.

[0013] A travel system based on multi-scenario fusion perception includes: Multi-source data fusion processing module: executes the spatiotemporal consistency verification algorithm as described in claim 1 or 2 to achieve accurate alignment between ETC gantry data and LBS data; Individual travel activity chain reconstruction module: After merging and aligning the data, the origin and destination points, stopover information and travel time are reconstructed; User profile building and mining module: Executes the profile building method described in claim 1 or 6 to generate user classification and travel feature tags; Travel information prediction module: Deploy the hybrid teaching LSTM encoder-decoder model as described in claims 1, 3, 4, and 5 to predict travel-related information; Accompanying travel service module: Executes the full-process personalized service as described in claim 1 or 7 through a multi-channel information dissemination system; The system architecture adopts a four-layer architecture of end-edge-cloud, including vehicle terminal, edge computing node, cloud computing center and application service layer.

[0014] The beneficial effects of this invention are as follows: Through multi-source data fusion and innovative algorithm architecture, it achieves a breakthrough in the accuracy of individual travel activity chain reconstruction and travel information prediction. The spatiotemporal consistency verification algorithm effectively solves the spatiotemporal deviation problem between ETC gantry data and LBS data. The 100-meter pile unit mapping and anomaly filtering mechanism control the positioning error within a reasonable range, providing a reliable data foundation for trajectory reconstruction. The gantry position attention mechanism and historical trajectory-environment fusion layer introduced by the hybrid teaching LSTM model highlight the weight of key nodes and deeply integrate the physical characteristics of the road network, significantly improving the accuracy of path prediction. The continuity of long-distance travel trajectories and adaptability to special scenarios are significantly better than traditional models, and the response delay of guidance information in extreme traffic scenarios is greatly shortened.

[0015] Meanwhile, the end-to-end accompanying service system and personalized adaptation mechanism have significantly improved the quality of travel services and the efficiency of road network operation. The multi-channel information release system achieves collaborative linkage between the roadside and the vehicle through C-V2X technology, keeping communication latency at an extremely low level. Combined with user profiling and hierarchical push strategies based on the 4W model, it provides customized services for different user types such as commuters, operators, and emergency responders, resulting in a significant improvement in user information compliance. In special scenarios such as mountainous highways, cross-sea bridges, and dense road networks in urban clusters, the scenario-based adaptation algorithm has achieved significant results, including a significant reduction in accident rates, a marked improvement in cross-city commuting efficiency, and a substantial increase in emergency vehicle passage efficiency. This not only meets users' needs for precise and safe travel but also provides efficient support for proactive traffic management on highways. Attached Figure Description Figure 1 This is the overall flowchart of the present invention; Figure 2 It is a dot plot; Figure 3 This is a workflow diagram. Detailed Implementation

[0016] The following will refer to the appendices in the embodiments of this application. Figure 3 The technical solutions in the embodiments of this application are clearly described. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this application are within the scope of protection of this application.

[0017] The purpose of this invention is to provide a system and method for reconstructing individual travel activity chains and providing accompanying travel services based on highway gantry data and mobile internet data. This solves the problems of inaccurate individual travel activity chain reconstruction, limited information dissemination channels, and lack of accompanying services in the prior art, thereby achieving precise travel services and proactive traffic control on the highway network.

[0018] The present invention achieves the above objectives through the following technical solutions: Individual travel activity chain reconstruction module; Based on highway gantry data, and by integrating ETC data and LBS data, we can accurately reconstruct the individual travel activity chain. The origin, destination, service area information, and travel time of individual trips are identified through gantry data. By leveraging historical data mining, individual travel profiles can be constructed to achieve accurate prediction of travel information; Accompanying travel service module; Construct a multi-channel information dissemination system, including roadside electromechanical equipment (variable message boards, lane indicators, variable speed limit signs) and vehicle terminals (navigation systems, apps, vehicle information processing units). To enable personalized customization of information release content, methods, and frequency, thereby improving information compliance rates; Provides comprehensive information services throughout the entire process, from pre-trip to during-trip to post-trip. Multi-source data fusion processing module; A spatiotemporal consistency verification algorithm is used to perform spatiotemporal alignment on ETC gantry data and LBS data. Individual travel chain identification algorithm; By combining association rule methods to mine the association characteristics of residents' travel, stability can be identified; User profile building and mining module; User behaviors are labeled and weighted based on the 4W model (When, Where, Who, What); Cluster analysis was performed using the model to categorize ETC users into commuter, commercial, and long-distance users. Passengers are classified using hierarchical clustering algorithms to identify individual travel characteristics, such as travel frequency, frequently used origin and destination (OD), frequently used routes, travel distance, and whether they frequently use service areas. Travel information prediction algorithm; A hybrid teaching LSTM encoder-decoder model is adopted, which integrates historical trajectory and road environment information. By combining attention mechanisms with spatial data such as gantry location and road network topology, the spatial accuracy of path prediction can be improved. Based on user profiles and travel stages, personalized information is disseminated through multiple channels to provide accompanying travel services.

[0019] To address the aforementioned issues, this model introduces two core innovative modules based on the standard Seq2Seq architecture: a gantry position attention mechanism; Different weights are assigned to different gantry nodes to highlight key nodes that play a decisive role in the route direction (such as merging and diverging points, provincial border gantries, and large hubs).

[0020] In the encoder output hidden state (H=[h1,h2,...,h...), t After that, a learnable spatial embedding layer is introduced to map each gantry ID to a d-dimensional spatial vector; ; Where: H=[h1,h2,...,h t [ ] represents the hidden state sequence output by the encoder, with a length of t; pᵢ is the embedding vector of the i-th gantry ID; d is the dimension of the embedding vector (which is also the dimension of the hidden state); It is a d-dimensional real space; Attention score for gantry location construction: ; in These can be training parameters; Normalized attention weights: ; Generate context vectors: ; Gantry embedding dimension (d=64); Attention hidden layer dimension (da=128); During training, a teacherforcing + scheduled sampling strategy was used. Automatically identify route decision points (such as the Luogang Interchange where Guangzhou North Ring Road turns into Guangzhou-Shenzhen Expressway). Suppress the effects of noise gantry (such as short-term false triggering); Enhance the ability to model the path coherence of long-distance travel.

[0021] Historical Trajectory and Road Environment Fusion Layer By deeply integrating static / dynamic road environment features with dynamic trajectory features, the model's ability to identify physically feasible paths is enhanced.

[0022] The following is a summary of the input features as shown in Table 1;

[0023] Fusion structure; The above features are concatenated into an environment vector. ; ; In each step of the decoder input, the environment vector of the previous predicted gantry is included. With encoder context vector splicing: ; Feed into the decoder LSTM unit: ; The output layer adds a road network constraint mask: the softmax probability is calculated only for downstream gantries reachable from the current gantry, forcing the path to conform to topological connectivity.

[0024] The specific implementation of individual travel activity chain reconstruction; Individual travel chain reconstruction; The individual travel chain reconstruction technology integrates ETC and LBS data to extract full travel features from the highway network and identify the starting point, destination, and activity information along the way for individual highway trips, such as whether they stop at service areas and the duration of their stay in service areas, thus realizing the identification and reconstruction of the entire travel chain on the highway network.

[0025] Individual travel profile mining; Individual travel profile mining technology uses the unique identifier ID of ETC vehicles to conduct historical data mining and identify individual highway travel characteristics, such as travel frequency, frequently used origin-destination (OD) routes, travel distance, and whether service areas are frequently used, to form an individual travel profile and support precise differentiated guidance for the highway network in the Greater Bay Area.

[0026] Individual travel information prediction; Individual travel information prediction technology predicts the origin and destination, route, and whether service areas will be stopped for an individual's next trip, providing support for the overall operation prediction of the Greater Bay Area road network and individualized guidance and push notifications.

[0027] I. Multi-source data fusion processing module: Detailed description of the spatiotemporal consistency verification algorithm; Target; Achieving precise alignment between ETC gantry data (high-precision time / location but sparse) and LBS data (continuous but with drift / delay) under a unified spatiotemporal benchmark provides reliable input for reconstructing individual travel chains.

[0028] Timestamp alignment rules: based on the deviation correction between GPS time and gantry equipment clock; Background issue: ETC gantry equipment uses local NTP synchronization, which results in a clock drift of ±500ms; LBS data (such as mobile signaling) reporting delays can be as high as 1–3 seconds; Matching directly with the original timestamp can lead to broken tracks or misjudged stopover behavior.

[0029] Deviation correction formula: Suppose the original recorded time of a certain vehicle at gantry gi is... If the most recent time of the corresponding LBS point is tLBS, then: ; Where (N) is the number of historically successful matching samples for the vehicle (≥5 times), used to estimate the vehicle's clock offset.

[0030] Correct the new record: ; (3) Spatial coordinate matching: Error control of mapping between 100-meter stake coordinates and LBS latitude and longitude; Unified road network representation: The highway is divided into line segment units (each segment is 100 meters long) according to the 100-meter markers. Each cell is assigned a unique ID and its starting / ending latitude and longitude (WGS84) and direction vector are stored.

[0031] Mapping error control methods: LBS point projection to the nearest road segment: The Haversine distance is used to calculate the vertical distance from the LBS point to each 100-meter pile unit; The unit with the smallest distance and the same direction is selected as the affiliation.

[0032] Error tolerance threshold: Lateral offset ≤50 meters (excluding interference from underpasses, etc.); The longitudinal offset is obtained by interpolation to obtain the virtual station number: ; in, To extend the projected distance of the road section, ; Anomaly filtering: If an LBS point is more than 200 meters away from the nearest highway, it is marked as a non-highway and will not participate in the fusion process. If the speed changes at three consecutive LBS points exceed 80km / h, it is considered a location jump and the location will be removed.

[0033] Hybrid teaching LSTM model: complete training parameters and reproducibility verification; The selection criteria for key parameters of the model structure are shown in Table 2 below;

[0034] Hybrid teaching strategy (Teacher Forcing + Scheduled Sampling); Objective: To mitigate exposure bias and improve the robustness of the model in actual inference.

[0035] Strategy: Early training phase (epochs 1–30): Use real-historical gantry (teacher forcing) with 70% probability, and model prediction with 30% probability; Later in training (epochs 31–80): Gradually reduce the teacher forcing ratio, eventually reaching 30% : 70%; Decay function: linearly decreasing; ; (3) Convergence criteria and training stability; Convergence condition: The edit distance (ED) of the validation set changed by less than 0.001 for five consecutive rounds. Or the training loss decreases by less than 1e-4 / epoch.

[0036] Training loss curve characteristics (based on 5 independent runs): The training loss steadily decreased from an initial 2.85 to 0.42 ± 0.03; The validation ED decreased from 0.21 to 0.058 ± 0.004; A standard deviation of <0.005 indicates that the algorithm is highly reproducible.

[0037] Early termination mechanism: If the ED is not improved for 8 consecutive rounds, it will be terminated early.

[0038] The specific implementation of accompanying travel services Construction of a multi-channel information dissemination system Roadside electromechanical equipment: Install variable message signs, lane indicators, variable speed limit signs, etc. at key nodes to achieve roadside information coverage. Vehicle-mounted terminal: Integrated with navigation systems or apps to enable vehicle-mounted information push; the onboard information processing unit (T-BOX) supports the C-V2X communication protocol. Information channel coordination: Enables coordinated dissemination of information between roadside and vehicle terminals, ensuring continuity of information delivery. V2X communication technology: Employing C-V2X technology, it achieves vehicle-road-cloud collaboration through the PC5 interface (direct communication) and the Uu interface (cellular network). Building a "multi-channel information dissemination system" involves the selection of various hardware devices and details of communication protocols. The following is a detailed description of roadside variable message signs, vehicle-mounted T-BOXs, and edge computing nodes, including hardware models, communication protocol details, and data from data transmission latency tests between devices.

[0039] Roadside variable message signs Interface Standard: RS485 / Ethernet: To ensure compatibility and flexibility, a variable message sign (VMS) supporting both RS485 and Ethernet interfaces is selected. For example, a certain brand of VMS supports both interface standards, allowing for flexible deployment across different network architectures.

[0040] Communication protocol: HTTP / HTTPS or MQTT protocols are used for data exchange to ensure the real-time nature and security of information dissemination.

[0041] In-vehicle T-BOX (Telematics Box) C-V2X protocol version: 3GPP R16: The in-vehicle T-BOX supports the latest C-V2X communication technology based on the 3GPP Release 16 standard, which provides enhanced support for direct communication between vehicles and infrastructure, such as low-latency and high-reliability vehicle-to-everything (V2X) services.

[0042] Edge computing nodes Hardware configuration: CPU model: Intel Xeon E-2288G or equivalent processor, capable of meeting high-speed data processing requirements.

[0043] Computing power metrics: Provides at least 200 GFLOPS (billion floating-point operations per second) of computing power, suitable for real-time data analysis and decision support.

[0044] Data transmission delay test data; Average communication delay from roadside equipment to vehicle terminal: In a real-world testing environment, the average communication latency from the roadside equipment sending information to the vehicle-side terminal receiving and displaying it is less than 100ms. This result was obtained after simulating real-world traffic scenarios and taking into account factors such as signal propagation time and network processing time, demonstrating that the system has a rapid response capability and is suitable for applications with high real-time requirements, such as highways.

[0045] Personalized information dissemination mechanism; Tiered and categorized release: Information is divided into three levels, A, B, and C, based on its importance (e.g., priority for national highways), with emergency events pushed to the entire network at the ministerial level and ordinary congestion-level pushes to specific areas. Real-time dynamic adjustment: Adjust the content and method of information release dynamically based on real-time traffic conditions and user feedback; User profile adaptation: Commuters receive lane-level navigation during the morning rush hour; Long-distance users receive consumption discounts when stopping at service areas; Message priority: Emergency events (accidents, construction) are triggered directly by the C-V2XPC5 interface, triggering audible and visual alarms; general information (service area recommendations) is pushed via 5G messages or the App. Full-process implementation of accompanying services; Pre-trip services: Based on user profiles, push personalized route planning, service area recommendations, weather warnings and other information; Services during travel: Based on real-time traffic conditions, push lane-level navigation, congestion warnings, service area recommendations, and other information; Post-trip services: providing information such as trip summary, satisfaction survey, and personalized service recommendations; Dynamic update mechanism: Gradient boosting decision trees (such as LightGBM) and neural networks are used to optimize profile updates, combined with real-time ETC data streams to achieve incremental learning; System integration and operation; System architecture design; Construct a four-layer architecture of "device-edge-cloud-application": vehicle terminal (device), edge computing node (edge), cloud computing center (cloud), and application service (application); Deploy gantry systems, integrate ETC data, and build a platform to reconstruct individual travel activity chains; A hybrid teaching LSTM encoder-decoder model is adopted, which integrates historical trajectory and road environment information; Unified representation of basic road network: Expressway sections are divided according to 100-meter markers, and a unified representation of 100-meter markers is achieved based on a unified coordinate system; Service triggering and execution; Based on the type of traffic incident and user travel characteristics, the corresponding service content is automatically triggered; Emergency event triggering: When an emergency event such as an accident or construction is detected, an audible and visual alarm will be triggered directly through the C-V2XPC5 interface; Normal event trigger: When congestion, full service areas, or other similar situations are detected, detour suggestions will be pushed via 5G messages or the App; Scheduled trigger: Automatically push personalized travel suggestions during peak travel times (such as 7-9 am and 5-7 pm); Dynamic adjustment and optimization; We will dynamically adjust the content and method of information dissemination based on real-time traffic conditions and user feedback. Real-time traffic integration: Road segment speed data is updated every 5 minutes, and the optimal route is recalculated based on the user's current driving location; User feedback loop: Collect user satisfaction ratings through the vehicle-side APP and dynamically adjust the push strategy in combination with real-time traffic data; Incremental learning: An improved method similar to the LRU algorithm is used to update the user model and record the freshness of the service type; Effectiveness evaluation and continuous optimization; By analyzing service effectiveness through statistical reports, we can continuously optimize the travel activity chain reconstruction algorithm and the accompanying service strategy. Evaluation indicators include: information delivery time, information compliance rate, user satisfaction, and road network operating speed difference; The accuracy of travel time prediction is assessed using the root mean square error. The system depends on the following data sources and their update frequencies, as shown in Table 3 below:

[0046] Extreme traffic scenario testing; (1) High traffic volume during holidays (average daily traffic volume ≥ 100,000 vehicles); Test section: Dongguan section of Guangzhou-Shenzhen Expressway (peak daily traffic volume of 123,000 vehicles during the 2024 Spring Festival holiday). Comparison models: conventional LSTM, Transformer trajectory prediction, and Gaode Navigation route planning; This invention is characterized by: Path prediction accuracy: 86.7% (conventional LSTM: 61.2%; Gaode: 72.5%) Induced information response delay: <3 minutes (traditional systems average 8–15 minutes); Thanks to accurate predictions of diversion demand, the average speed on the main line increased by 12.4 km / h. Key advantages: By leveraging historical holiday profiles and real-time gantry flow density, attention weights are dynamically adjusted to effectively identify "tidal travel" OD pairs and avoid homogenization of guidance strategies.

[0047] (2) Multiple concurrent incidents (3 or more incidents occurring simultaneously); Test scenario: During the May Day holiday in 2024, rear-end collisions occurred simultaneously in three locations: Humen Bridge, Nansha, and Huangpu. Invention capabilities: Generate a layered guidance scheme within 5 minutes: Layer 1 (3km upstream of the accident): Lane-level speed limit + variable message signs to guide lane changes; Layer 2 (adjacent interchange): Detour suggestions will be issued in advance at the Nanlang Interchange of the Guang'ao Expressway; Layer 3 (City Entrance Ramp): Alternative routes are pushed through the toll station's app.

[0048] The user compliance rate reached 78.3% (compared to only 41.6% for traditional broadcast-style inducements). Technical support: Multi-event parallel processing engine + dynamic road network impedance update mechanism to ensure that the induction strategies do not conflict with each other.

[0049] (3) Special weather (heavy rain / dense fog); Test conditions: Persistent heavy fog in the Pearl River Estuary in June 2024 (visibility <200m); Enhancement measures of this invention: Automatically activate "safe mode": lower the service area stop prediction threshold by 30% and push the nearest service area in advance; In conjunction with the meteorological API, when LBS data is missing, anomaly detection of gantry passage time is enabled to determine slow passage. Lane-level navigation switches to "low visibility guidance mode" to highlight emergency lanes and escape routes.

[0050] Results: The success rate of advance guidance at service areas increased to 92.1%; The accident rate under severe weather conditions decreased by 27% year-on-year (compared to the same period in 2023).

[0051] Special road network scenario testing; (1) Mountain expressway (multiple tunnels / long slopes); Test section: Yunfu section of Shantou-Zhanjiang Expressway (including 12 tunnels, with a maximum longitudinal slope of 4.8%). Challenges: LBS signal loss, large gantry spacing (up to 8.2km), and sudden changes in vehicle speed; The solution proposed in this invention is to introduce a tunnel entry and exit time difference compensation algorithm, combined with slope correction and travel time prediction. In areas without signal coverage, inertial simulation combined with road network topology constraints is used to maintain trajectory continuity. For heavy-duty trucks, a "braking performance degradation" factor is added to the environmental fusion layer.

[0052] Results: The path prediction error within the tunnel group was 0.071 (the conventional model had an error of 0.25+ due to signal loss). The advance warning time for abnormal truck parking is 4.2 minutes.

[0053] (2) Cross-sea bridge (long distance with no service area); Test section: Shenzhen-Zhongshan Bridge (24km long, no intermediate service areas); This invention is innovative in that it establishes a "cross-sea travel profile" to identify whether a user has cross-sea experience (historical travel frequency ≥ 3 times). For first-time users, a "fuel / battery check reminder" + "emergency parking lane location map" will be forcibly pushed to them; If the vehicle speed is detected to be consistently below 60km / h, the C-V2X emergency broadcast will be automatically triggered.

[0054] Results: First-time users' anxiety complaints decreased by 63%; Emergency response time has been reduced to within 90 seconds.

[0055] (3) Dense road network in urban clusters (cross-city expressways in the Greater Bay Area); Test area: Guangzhou-Foshan-Zhaoqing + Shenzhen-Dongguan-Huizhou metropolitan area (average daily cross-city traffic of 450,000 vehicles). Challenges: Highly mixed OD (Original Direction) patterns, frequent short-path intersections, and easily conflicting guiding information; This invention is groundbreaking in that it constructs a heat map of cross-city commuting to identify high-frequency cross-city origin-destination (OD) routes (such as Foshan-Guangzhou Science City and Dongguan-Shenzhen Nanshan). Establish virtual commuter corridors for commuters and provide lane-priority navigation for fixed time periods; Supports coordinated guidance across multiple administrative regions: Information released in Guangzhou can be synchronized to the Shenzhen VMS screen.

[0056] Metric: Average travel time for cross-city commuters decreased by 11.3 minutes; The conflict rate in the ramp weaving area decreased by 34%.

[0057] Special user scenario testing; (1) Forecast of long-distance stops for operating vehicles (freight trucks); Sample: Cross-border freight vehicles between Guangdong and Hong Kong (average daily mileage > 500km); This invention's capabilities include: integrating ETC consumption records (gasoline / food expenses), dwell time, and historical stopover habits; Introducing a "fatigue driving risk index": If driving continuously for more than 4 hours without stopping at a service area, the system will automatically push the nearest rest stop to you; The accuracy rate of predicting service area selection reached 89.4% (compared to only 67.2% for the baseline model).

[0058] Business Value: By collaborating with merchants in the service area, targeted promotional offers were delivered, increasing user dwell time by 22%.

[0059] (2) Lane-level navigation accuracy for commuter users; Test subjects: Commuters on the Guangzhou-Shenzhen Riverside Expressway during the morning rush hour (7:00–9:00); This invention achieves: Based on historical lane selection preferences (such as frequently using the leftmost fast lane) and combined with real-time congestion, the optimal lane is dynamically recommended. Lane change suggestions (accurate to 100-meter markers) are sent to the vehicle terminal via the C-V2XUu interface. Supports integration with high-precision maps, providing a notification such as "Exit on the right 2km ahead".

[0060] Accuracy indicators: Lane recommendation accuracy: 94.6%; User-initiated adoption rate: 81.7% (traditional navigation lane suggestion adoption rate <50%).

[0061] (3) Priority passage guidance for emergency vehicles; Test scenario: An ambulance travels from Zhongshan People's Hospital to a top-tier hospital in Guangzhou; Mechanism of this invention: It connects to the 120 dispatch system and automatically identifies the license plates of emergency vehicles. "Green Wave Escort Mode" activated: Upstream variable message signs display "Emergency Lane, please give way"; Simultaneously push avoidance prompts to other vehicles within a 5km radius via their mobile apps; Service area entrances are dynamically closed to prevent emergency traffic from being cut off.

[0062] Effect: The average speed throughout the journey has been increased to 85 km / h (compared to approximately 55 km / h under normal conditions). Arrival time was reduced by 38%, and the company received real-name authentication from the health department.

[0063] Conclusion: Technological advancement beyond simple extension; The test results above show that the present invention exhibits significantly better adaptability and robustness than the prior art in three dimensions: extreme traffic, complex road networks, and diverse users, as shown in Table 4 below;

[0064] Therefore, this invention is not a simple improvement on LSTM or general trajectory prediction models, but an integrated intelligent travel system of "perception-cognition-decision-service" built for the complex operating environment of Chinese highways, which has significant originality and engineering application value.

[0065] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A travel method based on multi-scenario fusion perception, characterized in that: Includes the following steps: Step 1: Collect ETC gantry data, LBS data, static road network features, dynamic traffic status and spatiotemporal context data, and achieve spatiotemporal alignment of data through a spatiotemporal consistency verification algorithm. The spatiotemporal consistency verification algorithm includes timestamp correction based on vehicle clock offset and spatial coordinate matching of 100-meter pile unit mapping. Step 2: Merge and align the multi-source data to identify the origin and destination of individual trips, information on service areas stopped at, and trip time, and extract full travel features; Step 3: Tag and assign weights to user behavior based on the 4W model, and divide users into categories such as commuting, operation, and long distance through hierarchical clustering algorithm to mine travel frequency, frequently used OD, and frequently used route features; Step 4: A hybrid teaching LSTM encoder and decoder model is adopted to fuse historical trajectory and road environment information. The model includes a gantry position attention mechanism and a historical trajectory and road environment fusion layer to predict the origin and destination of the trip, the driving route and the information of the service areas to be stopped. Step 5: Based on user profiles and travel stages, provide personalized services throughout the entire travel process—before, during, and after the trip—through a multi-channel information dissemination system comprised of roadside electromechanical equipment and vehicle-mounted terminals. 2.The travel method based on multi-scene fusion perception according to claim 1, characterized in that: In the spatiotemporal consistency verification algorithm, the timestamp correction formula is: ； wherein, N is the number of samples for which the vehicle history matches successfully and N > 5; During spatial coordinate matching, LBS points are projected onto the nearest 100-meter pile unit. LBS points with a lateral offset of ≤50 meters and a distance of >200 meters from the nearest highway are marked as non-highway status.

3. The travel method based on multi-scenario fusion perception according to claim 1, characterized in that: The gantry position attention mechanism includes: Mapping gantry id to d-dimensional space vector , constructing attention score ; normalizing the attention weights generating the context vector ; The gantry embedding dimension d=64 and the attention hidden layer dimension da=128. 4.The travel method based on multi-scene fusion perception of claim 1, wherein: The historical trajectory and road environment fusion layer concatenates static road network features, dynamic traffic conditions, and spatiotemporal context into an environment vector. The decoder input at each step The data is fed into the LSTM unit to obtain... The output layer adds a road network constraint mask to ensure path topology connectivity.

5. The travel method based on multi-scene fusion perception according to claim 1, characterized in that: The hybrid teaching strategy employs a linearly decreasing decay function. In the early stages of training (epochs 1-30), the probability of teacher forcing was 70%, which gradually decreased to 30% in the later stages (epochs 31-80). The model convergence condition is that the change in the edit distance of the validation set is less than 0.001 for 5 consecutive rounds or the decrease in training loss is less than 1e-4 / epoch.

6. The travel method based on multi-scenario fusion perception according to claim 1, characterized in that: The user profile construction also incorporates association rule methods to mine travel-related characteristics, identifying features such as whether an individual frequently uses service areas and travel distance, supporting differentiated guidance; travel information prediction also integrates spatial data such as gantry location and road network topology to improve the accuracy of path prediction.

7. A travel method based on multi-scenario fusion perception according to claim 1, characterized in that: The multi-channel information release system includes roadside variable message signs, lane indicators, variable speed limit signs, as well as vehicle-mounted navigation systems, apps, and in-vehicle T-BOX. The in-vehicle T-BOX supports the 3GPPR16 version C-V2X protocol, and the average communication latency from the roadside to the vehicle is <100ms. Personalized services are pushed according to the importance of information, and emergency events trigger audible and visual alarms through the C-V2XPC5 interface.

8. A travel method based on multi-scenario fusion perception according to claim 1, characterized in that: For special road network scenarios such as mountain highways, cross-sea bridges, and dense road networks in urban agglomerations, adaptation strategies such as tunnel entry and exit time difference compensation, cross-sea travel profiles, and cross-city commuting heat maps are adopted respectively. For special users such as commercial vehicles, commuters, and emergency vehicles, we provide customized services such as fatigue driving warning, lane-level priority navigation, and green wave escort.

9. A travel system based on multi-scene fusion perception, based on the travel method based on multi-scene fusion perception as described in any one of claims 1-8, characterized in that: include: Multi-source data fusion processing module: executes the spatiotemporal consistency verification algorithm as described in claim 1 or 2 to achieve accurate alignment between ETC gantry data and LBS data; Individual travel activity chain reconstruction module: After merging and aligning the data, the origin and destination points, stopover information and travel time are reconstructed; User profile building and mining module: Executes the profile building method described in claim 1 or 6 to generate user classification and travel feature tags; Travel information prediction module: Deploy the hybrid teaching LSTM encoder-decoder model as described in claims 1, 3, 4, and 5 to predict travel-related information; Accompanying travel service module: Executes the full-process personalized service as described in claim 1 or 7 through a multi-channel information dissemination system; The system architecture adopts a four-layer architecture of end-edge-cloud, including vehicle terminal, edge computing node, cloud computing center and application service layer.

Images