A boarding point intelligent optimization method and device and a computer readable storage medium

Abstract

The application discloses a kind of intelligent optimization methods, device and computer readable storage medium of boarding point, belong to intelligent transportation technical field;The method aims at solving the problem that existing boarding point recommendation is ignored actual driving behavior additional cost due to dependence on theoretical path planning;The present application includes: offline stage, obtain historical vehicle trip data, calculate the difference between benchmark path and actual path, identify scheduled deviation mode such as U-turn and quantify its additional cost value, then aggregate cost with geographic spatial information as index to build boarding point cost feature knowledge base;Online stage, generate candidate boarding point after receiving user request, for each candidate point, respectively calculate the first type of cost based on real-time traffic and the second type of cost representing historical additional cost obtained from knowledge base query, then calculate comprehensive score based on the two types of cost to sort and recommend.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and computer technology, specifically to a method, apparatus, and computer-readable storage medium for intelligent optimization of boarding points. Background Technology

[0002] In ride-hailing services, recommending a convenient pick-up point for both passengers and drivers is crucial for improving user experience and platform operational efficiency. Existing technologies typically employ recommendation strategies based on real-time traffic conditions and estimated arrival times. This strategy first generates a set of candidate pick-up points that comply with traffic rules based on the passenger's initial location. Then, using a route planning engine and real-time traffic data, it calculates the estimated travel time from the driver's current location to each candidate pick-up point. Finally, the candidate point with the shortest estimated travel time is prioritized for recommendation.

[0003] However, the aforementioned mainstream technical solutions mainly rely on idealized route planning models, which are based on road network topology and macroscopic traffic flow speeds, failing to fully consider the actual behavioral patterns and constraints of drivers at the microscopic driving level. For example, a pick-up point with a theoretically shortest estimated arrival time may require the driver to perform U-turns, sharp turns, or complex detours in actual driving. These operations will incur significant time and mileage costs in real road environments (such as the presence of physical barriers, no-U-turn signs, or temporary traffic control), but these costs are often underestimated or ignored in the initial route planning.

[0004] This discrepancy between theoretical planning and actual execution leads to a "cost gap." Some existing technologies attempt to address this issue through post-event analysis, analyzing the actual driving trajectory after an order is completed. If detours or other irregularities are found, the corresponding pick-up point is marked as a low-quality point and its priority is reduced in subsequent recommendations. However, this approach is essentially a passive feedback and delayed correction; it cannot anticipate and mitigate potential costs at the time of recommendation generation, and it struggles to provide effective initial judgments for new areas or pick-up points with sparse data.

[0005] Therefore, existing pick-up point recommendation technologies generally suffer from the following shortcomings: First, they lack effective learning from historical driving experience, failing to systematically mine and quantify the additional costs associated with predetermined deviation patterns caused by specific pick-up points from massive amounts of historical vehicle trip data; second, their cost measurement dimensions are singular, mainly focusing on estimated time or distance while ignoring the comprehensive economic costs incurred by passengers, drivers, and the platform due to ineffective driving. How to transform the collective driving experience contained in historical trajectories into predictable cost characteristics and integrate them into real-time recommendation decisions to achieve a paradigm shift from "post-event correction" to "pre-event avoidance" is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0006] To address the problem in the background art that existing pick-up point recommendation methods fail to effectively utilize historical vehicle travel data to predict and avoid additional costs caused by predetermined deviation patterns, this invention provides a pick-up point intelligent optimization method, apparatus, and computer-readable storage medium.

[0007] This invention provides an intelligent optimization method for pick-up points, comprising: acquiring historical vehicle trip data, the historical vehicle trip data including the start position, end position, and trajectory point sequence of the trip; for each historical vehicle trip, calculating a baseline path with the start position as the starting point and the end position as the end point, and determining a baseline cost metric for the baseline path; analyzing the actual driving path of the historical vehicle trip to identify a predetermined deviation pattern from the baseline path, and determining an actual cost metric for the actual driving path; when the predetermined deviation pattern is identified, calculating an additional cost value based on the difference between the actual cost metric and the baseline cost metric; aggregating the additional cost values ​​of the historical vehicle trips using the geospatial information and road attributes of the trip end position as indexes to calculate the cost associated with the index. Cost statistical features are aggregated and stored in a pick-up point cost feature knowledge base for online querying. Upon receiving a passenger's ride request, a set of candidate pick-up points is generated based on the passenger's location and preset rules. For each candidate pick-up point, a first type of cost from the driver's current location to that pick-up point is calculated, based on real-time traffic conditions and path planning. For each candidate pick-up point, the pick-up point cost feature knowledge base is queried based on its geospatial information and road attributes to obtain a second type of cost associated with that pick-up point, representing additional costs incurred due to predetermined deviation patterns during historical trips. For each candidate pick-up point, a comprehensive score is calculated based on the first and second types of costs. Candidate pick-up points are ranked according to the comprehensive score, and the ranking results are pushed to the user.

[0008] In one implementation, the pick-up point cost feature knowledge base adopts a two-layer structure of "spatial precision + attribute generalization," where the spatial precision layer aggregates based on geospatial information, and the attribute generalization layer aggregates based on road attributes. When a query fails or the result has low confidence based on the geospatial information of a candidate pick-up point in the spatial precision layer, a query is performed in the attribute generalization layer based on the road attributes corresponding to the candidate pick-up point. This two-layer structure, through synergy and complementarity, ensures recommendation accuracy in areas with sufficient data while solving the generalization coverage problem in scenarios with sparse data or cold starts.

[0009] Furthermore, the step of analyzing the actual driving path of the historical vehicle journey involves using a combination of at least two algorithms from the following: direction angle sequence abrupt change detection, trajectory curvature cumulative analysis, and path shape template matching, to identify the predetermined deviation pattern. By fusing multiple algorithms for decision-making, the accuracy and robustness of the predetermined deviation pattern identification can be improved.

[0010] Optionally, when obtaining the second type of cost, the cost statistical features obtained from the boarding point cost feature knowledge base are modified in conjunction with the current real-time traffic congestion situation to obtain the second type of cost. This allows historical statistical costs to better adapt to the current dynamic traffic environment and improves the real-time performance of predictions.

[0011] In one implementation, the baseline path is a theoretically shortest travel time path or a theoretically shortest travel distance path calculated by a path planning service; the predetermined deviation mode includes at least one of U-turn, sharp turn, or unnecessary detour.

[0012] Furthermore, the first type of cost includes estimated travel time and estimated travel distance, while the second type of cost is the expected value of predicted additional costs. The step of calculating the comprehensive score includes: obtaining the walking distance from the passenger to the candidate pick-up point, and weighting the estimated travel time, the walking distance, and the expected value of predicted additional costs. This comprehensive scoring mechanism can balance driver pick-up efficiency, passenger walking convenience, and potential additional travel costs.

[0013] Optionally, in the ranking results pushed to the user, at least some of the pick-up points include recommendation reasons characterizing their accessibility, which are generated based on the second type of cost. This helps improve the user's understanding and trust in the recommendation results, assisting them in making better decisions.

[0014] This invention also provides an intelligent pick-up point optimization device, comprising: a data acquisition module for acquiring historical vehicle trip data; a baseline route calculation module for calculating a baseline route and determining its baseline cost metric; a deviation pattern recognition module for identifying predetermined deviation patterns in the actual driving path; an additional cost quantification module for calculating additional cost values; a cost feature aggregation module for aggregating additional cost values ​​to calculate cost statistical features; a knowledge base construction module for storing cost statistical features into a pick-up point cost feature knowledge base; a candidate point generation module for generating a set of candidate pick-up points; a first type of cost calculation module for calculating a first type of cost; a second type of cost prediction module for querying the pick-up point cost feature knowledge base to obtain a second type of cost; a comprehensive scoring module for calculating a comprehensive score; and a recommendation module for sorting candidate pick-up points and pushing results.

[0015] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the intelligent optimization method for boarding points as described in any of the preceding claims.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieved precise cost savings and efficiency improvements. By constructing and utilizing a knowledge base of pick-up point cost characteristics, this invention can quantify and predict the additional cost value generated by predetermined deviation patterns, and proactively avoid pick-up points with high "behavioral cost" risks during recommendations. This directly reduces the vehicle's invalid mileage and time, lowers the actual cost for passengers and the operating costs for drivers, and improves the order-taking efficiency per unit time.

[0017] 2. Enhanced the intelligence and predictability of the recommendation system. This invention transforms the collective driving experience contained in massive historical vehicle travel data into quantifiable and queryable knowledge. This upgrades recommendation decisions from path planning that relies solely on instantaneous, theoretical data to predictive decisions that integrate historical experience. The system can "anticipate" and avoid locations that have historically proven to be "troublesome," achieving an intelligent leap from "computation" to "cognition."

[0018] 3. A data-driven, self-evolving optimization closed loop has been constructed. The travel data for each completed order can be used to update the pick-up point cost characteristic knowledge base. This means the system can continuously learn and automatically adapt to changes in the road network structure (such as new traffic controls, road construction, etc.), dynamically adjusting the cost assessment of each pick-up point, forming a sustainable, self-optimizing closed-loop system that ensures the long-term effectiveness of the recommendation strategy. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the intelligent optimization method for boarding points according to the present invention.

[0020] Figure 2 This is a schematic diagram of the data flow of each module in the offline learning stage in one embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the data flow of each module in the online recommendation stage according to one embodiment of the present invention.

[0022] Figure 4 This is a structural block diagram of the intelligent optimization device for the loading point of the present invention.

[0023] Figure 5 A schematic diagram of direction angle sequence mutation detection in the identification of predetermined deviation patterns (such as U-turns).

[0024] Explanation of reference numerals in the attached figures S1: Obtain historical vehicle trip data; S2: Calculate baseline route; S3: Identify pre-booking deviation patterns; S4: Calculate additional cost values; S5: Aggregate cost statistical features; S6: Generate candidate pick-up points; S7: Calculate overall score; S8: Sort and push results. Detailed Implementation To address the technical problem in existing pick-up point recommendation methods that over-rely on theoretical route planning while ignoring actual driver behavior, leading to discrepancies between planning and execution, this invention proposes an intelligent pick-up point optimization method, device, and computer-readable storage medium. The core idea of ​​this method is to construct a pick-up point cost feature knowledge base by offline learning of predetermined deviation patterns, such as U-turns or detours, from historical vehicle travel data and quantifying them into comparable additional cost values. During online recommendation, the system considers not only the first type of cost calculated based on real-time traffic conditions, such as estimated arrival time, but also queries the knowledge base to obtain the second type of cost associated with candidate pick-up points, i.e., predicted behavioral additional costs. By comprehensively evaluating these two types of costs, this invention can recommend a pick-up point with the lowest overall cost in actual execution, thereby effectively avoiding unnecessary time and money waste caused by inefficient driving behavior.

[0025] The technical solution of this invention involves several core terms. Historical vehicle trip data refers to historical data containing the start and end positions of the trip, as well as the sequence of trajectory points. A baseline path is a reference path calculated by a route planning service with a specific optimization objective (such as shortest time or distance). Correspondingly, the baseline cost metric is the quantified cost of the baseline path, such as the theoretical optimal distance and theoretical optimal time. A predetermined deviation pattern refers to a predefined inefficient driving pattern relative to the baseline path, such as a U-turn or detour. When a predetermined deviation pattern occurs, an additional cost value is generated; this value is the difference between the actual cost metric and the baseline cost metric caused by the occurrence of the predetermined deviation pattern. Geospatial information and road attributes are used as indexes to aggregate the additional cost values. This information includes geographical locations (such as road segments and grids) and road features (such as grade and number of lanes). The result of statistical calculations on the aggregated additional cost values ​​is called cost statistical features, such as average value and frequency of occurrence. These statistical features are stored in a pick-up point cost feature knowledge base. When making online recommendations, the first type of cost is the cost obtained based on real-time traffic route planning, such as estimated travel time or distance; while the second type of cost is the predicted additional cost obtained from the knowledge base, representing the deviation patterns caused by historical trips.

[0026] like Figure 1As shown, the present invention provides an intelligent optimization method for pick-up points, characterized by comprising two main parts: an offline learning stage and an online recommendation stage. In the offline learning stage, step S1 is first executed to acquire historical vehicle trip data, which includes the start and end positions of the trips and the sequence of trajectory points. Next, step S2 is executed to calculate a baseline path and determine its baseline cost metric for each historical vehicle trip, using its start position as the starting point and its end position as the end point. Then, step S3 is executed to identify a predetermined deviation pattern from the baseline path by analyzing the actual driving path. When a predetermined deviation pattern is identified, step S4 is executed to calculate an additional cost value based on the difference between the actual cost metric and the baseline cost metric. Finally, step S5 is executed to aggregate these additional cost values ​​using the geospatial information of the trip's end position and road attributes as indexes, calculate cost statistical features, and store them in a pick-up point cost feature knowledge base. In the online recommendation stage, upon receiving a passenger's ride request, the system executes step S6 to generate a set of candidate pick-up points. For each candidate boarding point, the first type of cost and the second type of cost are calculated separately. Based on these two types of costs, step S7 is executed to calculate a comprehensive score. Finally, step S8 is executed to sort the results according to the scores and push them out. This process realizes a complete closed loop from data to knowledge to decision-making by turning historical experience into knowledge, which significantly improves the predictability and accuracy of recommendations.

[0027] Furthermore, to address the issues of data sparsity and cold start, the boarding point cost feature knowledge base can adopt a two-layer structure of "spatial precision + attribute generalization". The spatial precision layer aggregates data based on geospatial information (such as road segment ID or Geohash encoding) to provide high-precision cost estimates for data-rich areas. The attribute generalization layer aggregates data based on road attributes (such as road grade, number of lanes, presence of a median strip, etc.) to provide references for boarding points with sparse or entirely new data. When a query based on the geospatial information of a candidate boarding point fails in the spatial precision layer, or when the confidence level of the query result (e.g., based on the number of historical orders) is low, the system automatically falls back to the attribute generalization layer to query based on the road attributes corresponding to that boarding point. This complementary two-layer structure ensures both accuracy in hotspot areas and broad coverage of new areas, greatly enhancing the robustness and applicability of the method.

[0028] As a preferred implementation, to improve the accuracy of predetermined deviation pattern recognition, the step of analyzing the actual driving path of historical vehicle trips can employ a combination of various algorithms. For example... Figure 5As shown, abrupt change detection of the azimuth sequence is an effective method. It calculates the change in azimuth angle within a sliding window on the trajectory point sequence. For normal driving trajectories (e.g., trajectory points P0 to P8, corresponding to times T1 to T10), the azimuth angle changes smoothly. However, for trajectories exhibiting U-turn behavior (e.g., a trajectory with a starting point P1, an ending point P7, and an inflection point P4), if the azimuth angle difference Δθ falls within a preset U-turn judgment range (e.g., 150° to 210°), a preliminary U-turn can be identified. Furthermore, trajectory curvature accumulation analysis can be combined to calculate the cumulative turning angle of the path over a certain length; if it exceeds a threshold (e.g., 160°), it is considered a deviation. The path shape template matching algorithm identifies deviations by comparing the actual path with preset "U-shaped" or "P-shaped" templates based on similarity. Combining at least two of these three algorithms, for example using a majority voting strategy (i.e., only confirming a deviation pattern when at least two algorithms determine its existence), can effectively filter out false positives caused by factors such as GPS signal drift, thus significantly improving the accuracy and reliability of deviation pattern recognition.

[0029] To better adapt historical statistical data to current traffic conditions, the process of obtaining Category II costs can include a correction step. Cost statistics (such as historical average additional costs) obtained from the pick-up point cost feature knowledge base are calculated based on historical average traffic conditions. However, actual additional costs are significantly affected by real-time traffic congestion; for example, the time required to complete a U-turn during severe congestion is much longer than during smooth traffic. Therefore, after obtaining historical cost statistics, they can be corrected by incorporating the real-time traffic congestion index of the current road segment. For example, the historical additional cost can be multiplied by a coefficient related to the current congestion index; if the current congestion level is higher than the historical average, the predicted additional cost is adjusted upwards accordingly. This real-time correction allows Category II costs to more accurately reflect the potential risks under current traffic conditions, thus making the final comprehensive score more timely and accurate.

[0030] In a specific implementation of this invention, the benchmark path can be the theoretical shortest travel time path or the theoretical shortest travel distance path calculated by a route planning service, providing a clear and consistent benchmark for cost comparison. The predetermined deviation pattern can be specified as various typical inefficient driving behaviors, including but not limited to U-turns, sharp turns, or unnecessary detours. These patterns are the main reasons why actual driving costs exceed theoretically planned costs. By identifying and quantifying these specific patterns, this invention can accurately capture the hidden costs ignored by existing technologies.

[0031] When calculating the overall score, the first type of cost may include the estimated travel time (ETA) and the estimated travel distance, while the second type of cost is the expected value of the predicted additional costs. The overall score can be calculated using a weighted combination approach, for example, by weighting the estimated travel time, the passenger's walking distance to the candidate boarding point, and the expected value of the predicted additional costs. A typical scoring function could be: The weighting coefficients here. , , This reflects a trade-off between the importance of driver efficiency, passenger convenience, and cost avoidance. The scoring function is designed so that candidate points with shorter ETAs, closer walking distances, and lower prediction costs receive higher scores, thus achieving a balance among multiple optimization objectives.

[0032] To enhance user experience and increase user trust in the recommendation results, this invention can also include recommendation reasons representing the accessibility of at least some pick-up points in the ranking results pushed to users. These recommendation reasons are generated based on the second type of cost. For example, when the second type of cost of a candidate pick-up point is significantly lower than other points, the system can generate prompts such as "Recommending this point can save you about X yuan in detour costs" or "The driver does not need to turn around, making pick-up smoother." This approach translates complex model decisions into language that users can easily understand, not only helping users make more informed choices but also allowing users to intuitively perceive the technical advantages of this invention.

[0033] like Figure 4 As shown, this invention also provides an intelligent pick-up point optimization device, the internal modules of which correspond to the aforementioned method steps. The device includes: a data acquisition module for acquiring historical vehicle trip data; a baseline route calculation module for calculating the baseline route and baseline cost metric; a deviation pattern recognition module for identifying predetermined deviation patterns; an additional cost quantification module for calculating additional cost values; a cost feature aggregation module for aggregating additional cost values ​​and calculating cost statistical features; a knowledge base construction module for storing statistical features in a pick-up point cost feature knowledge base; a candidate point generation module for generating candidate pick-up points; a first-type cost calculation module for calculating first-type costs; a second-type cost prediction module for acquiring second-type costs; a comprehensive scoring module for calculating a comprehensive score; and a recommendation module for sorting and pushing results. These modules work together to realize the intelligent pick-up point optimization function described in this invention.

[0034] Furthermore, this invention also provides a computer-readable storage medium storing computer program instructions. When these instructions are executed by a processor, the aforementioned intelligent optimization method for boarding points can be implemented. This indicates that the technical solution of this invention can be implemented and deployed on general-purpose computing devices (such as servers and vehicle terminals) through software programming.

[0035] To more clearly illustrate the technical solution of this invention, the offline phase first acquires historical vehicle trip data containing trajectory point sequences. For each trip, a route planning service is invoked to calculate a baseline path with the shortest time as the objective, obtaining a baseline cost metric (theoretically optimal time). Next, a single direction angle sequence mutation detection algorithm is used to identify predetermined deviation patterns, i.e., U-turn behavior. When a U-turn is detected, the difference between the actual travel time and the baseline path time is calculated and multiplied by a time cost coefficient to obtain an additional cost value. Then, Geohash encoding is used to grid the trip termination position, aggregating the additional cost values ​​into the corresponding Geohash grid cells, and calculating the average additional cost of each grid as a cost statistical feature, stored in a single-layer boarding point cost feature knowledge base. In the online phase, upon receiving a request, candidate boarding points are generated. For each point, the estimated travel time based on real-time traffic conditions is calculated as a first-class cost, and the average additional cost of its corresponding Geohash grid is retrieved from the knowledge base as a second-class cost. Finally, a simplified scoring function (e.g., the reciprocal of the first-class cost minus the second-class cost) is used for scoring and ranking. This basic implementation fully embodies the core idea of ​​learning behavioral costs from historical data and using them for online decision-making.

[0036] Building upon the aforementioned basic implementation, a superior implementation further defines the construction method of the pick-up point cost feature knowledge base. This knowledge base employs a two-layer structure, comprising a geographic dimension statistics table and an attribute dimension statistics table. The geographic dimension statistics table uses "segment ID + driving direction side" or Geohash encoding as the primary key (geo_key), storing precise statistical values ​​such as avg_extra_cost (historical average extra cost) and u_turn_freq (frequency of U-turns). The attribute dimension statistics table uses road attribute combinations (such as "L3_No U-turn_One-way Street") as the primary key (attr_key), storing the average cost features of roads with similar attributes. When a new pick-up point appears, and it is not recorded in the geographic dimension table or the number of recorded orders (order_count) is less than a preset threshold (e.g., 30), the system extracts the road attributes of that point and matches them with the attr_key in the attribute dimension table, thereby obtaining an initial cost estimate based on attribute similarity. Through this synergy of "spatial precision + attribute generalization," the cold start and data sparsity problems are effectively solved, ensuring that the system can provide evidence-based cost predictions at any location.

[0037] In terms of deviation pattern recognition, another embodiment refines the implementation of the deviation pattern recognition module. This module integrates three algorithms: direction angle sequence abrupt change detection, trajectory curvature cumulative analysis, and path shape template matching. The direction angle detection algorithm slides a 50-meter window along the trajectory, detecting whether the difference in direction angle between the starting and ending points is between 150° and 210°. The curvature analysis algorithm calculates whether the cumulative curvature exceeds 160° within a 100-meter path length. The shape matching algorithm uses Dynamic Time Warping (DTW) technology to compare the actual path with a "U-shaped" template. To improve accuracy, a majority voting strategy is adopted, meaning that a U-turn is only confirmed when at least two algorithms determine that a U-turn has occurred. Furthermore, map matching results are used for verification to confirm that the vehicle has indeed passed through the opposite lane of the same road segment. This multi-algorithm fusion and verification mechanism greatly improves the robustness of the recognition results and effectively eliminates various interferences.

[0038] When acquiring the second type of cost, a specific implementation adds a real-time correction step. After the comprehensive scoring module retrieves the expected historical additional cost value C_history from the knowledge base, it does not use it directly, but instead uses a formula... Corrections are needed. Specifically, This is the ratio of the current real-time congestion index to the historical average congestion index for a given road segment. For example, if the knowledge base records a historical additional cost of 2 yuan for a certain point, and the current congestion index for that road segment is 1.5 times the historical average, then the prediction cost C_predict will be adjusted to 2 × 1.5 = 3 yuan. This design allows the prediction cost to dynamically reflect real-time traffic changes, as congestion significantly increases the time cost of complex operations such as U-turns, thus making the recommendation decision more realistic.

[0039] In one embodiment that includes a complete scoring function, the comprehensive scoring module uses a formula. The candidate points are scored. Here, This is the estimated travel time, and Distance(P, passenger) is the passenger's walking distance. This refers to the predicted behavioral cost. The weighting coefficients w1, w2, and w3 can be determined based on operational data, for example, set as w1=0.4, w2=0.3, and w3=0.3. A numerical example illustrates this: Candidate point P1 has an ETA of 2 minutes, a walking distance of 50 meters, and a predicted cost of 0.3 yuan; candidate point P2 has an ETA of 1.5 minutes, a walking distance of 80 meters, and a predicted cost of 2.5 yuan. Traditional methods would recommend P2 with the shortest ETA. However, according to this invention, Score(P1) = 0.4×(1 / 2)+0.3×(1 / 50)-0.3×0.30 = 0.116, while Score(P2) = 0.4×(1 / 1.5)+0.3×(1 / 80)-0.3×2.50 = -0.480. Ultimately, P1, with its higher score, was prioritized and successfully avoided the high behavioral cost of P2 (2.5 yuan), saving the user approximately 2.2 yuan.

[0040] Finally, when pushing the ranking results to the user, the recommendation module can attach context-aware reasons for the recommendation. For P1 recommended in the above example, the system can display on the interface: "This pick-up point is recommended, and it is estimated to save approximately 2.2 yuan in detour costs." This reason for the recommendation comes directly from the calculation results of the second type of cost, conveying the cost considerations within the model to the user in an intuitive and quantitative way. This transparent communication method not only guides users to choose the better pick-up point overall, but also significantly enhances users' perception and trust in the intelligence of the recommendation system, thereby improving user satisfaction and platform efficiency.

[0041] The following will be combined with the appendix Figures 1 to 5 This document provides a detailed description of a preferred and complete embodiment of the present invention. This embodiment integrates all the foregoing technical features and demonstrates a specific application scheme. First, as... Figure 4 As shown, the system architecture of the intelligent boarding point optimization device includes a data acquisition module, a baseline path calculation module, a deviation pattern recognition module, an additional cost quantification module, a cost feature aggregation module, a knowledge base construction module, a candidate point generation module, a first-type cost calculation module, a second-type cost prediction module, a comprehensive scoring module, and a recommendation module. These modules collaborate with each other through data interfaces to complete the entire process from offline learning to online recommendation.

[0042] During the offline learning phase, the data interaction and processing of each module are as follows: Figure 2 As shown. The data acquisition module collects trip data containing complete trajectory point sequences (timestamps, latitude and longitude, and orientation angles) from the historical order database. For each trip, the baseline path calculation module calls the path planning engine (such as A...). The algorithm, aiming to minimize travel time, calculates the baseline path and its theoretical optimal time T_opt and theoretical optimal distance D_opt. Simultaneously, the deviation pattern recognition module performs map matching on the actual trajectory and executes the deviation pattern recognition step using a voting method combining the aforementioned three algorithms: orientation angle detection, curvature analysis, and template matching. For example... Figure 5 As shown, the orientation angle detection identifies turning behavior by recognizing abrupt changes in the orientation angle sequence of trajectory points within the range of 150° to 210°. For trips where deviations are identified, the additional cost quantification module performs an additional cost quantification step, using the formula... Calculate the additional cost value. Among them, the mileage cost weighting coefficient α can be taken as 1.5 to 3.0 yuan / km, and the time cost weighting coefficient β can be taken as 0.5 to 1.0 yuan / minute.

[0043] Subsequently, the knowledge base construction module aggregates the calculated additional cost values. The knowledge base adopts a two-tier structure of "spatial precision + attribute generalization," containing geographic dimension statistics tables and attribute dimension statistics tables. The geographic dimension table uses Geohash encoding as the key, while the attribute dimension table uses combinations of L1-L3 level attributes such as road grade, number of lanes, and proximity to the central median as the key. Each key is associated with the historical average additional cost. Cost statistics features include the frequency of U-turns (u_turn_freq) and the total number of orders (order_count). The knowledge base is updated in batches every morning at midnight, and also supports real-time incremental writing after each order is completed, forming a sustainable self-evolving closed loop. For newly opened areas, the system adopts a cold start strategy. When the number of orders in a geographical unit is less than 30, it migrates cost features from the attribute dimension table by matching road attributes, or uses the weighted average of the spatial nearest neighbor units as an initial estimate, ensuring the universality of the service.

[0044] During the online recommendation phase, the data interaction and processing of each module are as follows: Figure 3 As shown. Upon receiving a ride request, the candidate point generation module generates a set of candidate pick-up points {P1, P2, ..., Pn} within a 200-meter radius of the passenger. For each candidate point P, the comprehensive scoring module performs a comprehensive scoring step. First, it calls the real-time traffic route planning service to calculate the first type of cost from the driver's current location to point P, i.e., the estimated travel time. Then, using the geographical location of point P, the historical expected value of additional costs is obtained by querying the knowledge base of boarding point cost characteristics. And combined with the current congestion coefficient Correcting the predicted behavioral costs This is the second type of cost. Finally, the complete scoring function is used. Calculate the overall score, where the weights w1, w2, and w3 can be set to 0.4, 0.3, and 0.3, respectively.

[0045] Finally, the recommendation module sorts all candidate points from highest to lowest Score(P), performs a sorting and push step, and pushes the sorted list to the user. As shown in the previous numerical example, a point with the shortest theoretical ETA but an extremely high C_predict (such as P2, scoring -0.480 points) will be surpassed by a point with a better overall cost (such as P1, scoring 0.116 points). The recommendation results will be accompanied by explanatory reasons, such as "saving you approximately 2.2 yuan in detour costs," thus intuitively conveying the technical advantages of the invention to the end user. Through this entire process, the invention achieves a logical leap from "shortest estimated time" to "lowest actual overall cost," accurately saving costs for both drivers and passengers and improving the operational efficiency of the entire transportation system.

Claims

1. A method for intelligent optimization of boarding points, characterized in that, include: Obtain historical vehicle trip data, which includes the start and end positions of the trip and the sequence of trajectory points. For each historical vehicle journey, a baseline path is calculated with the starting position of the journey as the starting point and the ending position as the ending point, and the baseline cost metric of the baseline path is determined. Analyze the actual driving paths of the historical vehicle trips to identify predetermined deviation patterns from the baseline path and determine the actual cost metric of the actual driving paths; When the predetermined deviation pattern is identified, an additional cost value is calculated based on the difference between the actual cost measure and the baseline cost measure; The additional cost values ​​of historical vehicle trips are aggregated using geospatial information of the trip termination location and road attributes as indexes to calculate cost statistics associated with the indexes. The cost statistics features obtained from the aggregated statistics are stored in the boarding point cost feature knowledge base for online querying; After receiving a passenger's ride request, a set of candidate pick-up points is generated based on the passenger's location and preset rules; For each candidate pick-up point, calculate the first type of cost from the driver's current location to that candidate pick-up point. The first type of cost is obtained based on path planning according to real-time traffic conditions. For each candidate pick-up point, based on its geospatial information and road attributes, the pick-up point cost feature knowledge base is queried to obtain a second type of cost associated with the candidate pick-up point. The second type of cost represents the additional cost value caused by the predetermined deviation pattern in the historical trip. For each candidate boarding point, a comprehensive score is calculated based on the first type of cost and the second type of cost; Candidate pick-up points are ranked based on their overall scores, and the ranking results are pushed to users.

2. The method according to claim 1, characterized in that, The knowledge base for the cost features of the pick-up point adopts a two-layer structure of "spatial precision + attribute generalization". The spatial precision layer aggregates according to geospatial information, and the attribute generalization layer aggregates according to road attributes. When the query fails or the confidence of the result is low in the spatial precision layer based on the geospatial information of the candidate pick-up point, the query is performed in the attribute generalization layer based on the road attribute corresponding to the candidate pick-up point.

3. The method according to claim 1, characterized in that, The step of analyzing the actual driving path of the historical vehicle journey involves using a combination of at least two algorithms from the following: direction angle sequence mutation detection, trajectory curvature cumulative analysis, and path shape template matching, to identify the predetermined deviation pattern.

4. The method according to claim 1, characterized in that, When obtaining the second type of cost, the cost statistics features obtained from the boarding point cost feature knowledge base are modified in conjunction with the current real-time traffic congestion situation to obtain the second type of cost.

5. The method according to claim 1, characterized in that, The baseline path is the theoretical shortest travel time path or the theoretical shortest travel distance path calculated by the path planning service; the predetermined deviation mode includes at least one of U-turn, sharp turn, or unnecessary detour.

6. The method according to claim 1, characterized in that, The first type of cost includes estimated travel time and estimated travel distance, while the second type of cost is the expected value of the predicted additional costs. The steps for calculating the comprehensive score include: obtaining the walking distance from the passenger to the candidate boarding point, and weighting the estimated travel time, the walking distance, and the expected value of the predicted additional cost.

7. The method according to claim 1, characterized in that, In the ranking results pushed to users, at least some of the boarding points are accompanied by recommendation reasons that characterize their accessibility, and these recommendation reasons are generated based on the second type of cost.

8. A smart optimization device for boarding points, characterized in that, include: The data acquisition module is used to acquire historical vehicle trip data, which includes the start position, end position, and trajectory point sequence of the trip. The baseline route calculation module is used to calculate a baseline route for each historical vehicle trip, with the starting position of the trip as the starting point and the ending position as the ending point, and to determine the baseline cost metric of the baseline route. A deviation pattern recognition module is used to analyze the actual driving path of the historical vehicle trip to identify a predetermined deviation pattern between the reference path and the actual driving path, and to determine the actual cost metric of the actual driving path. An additional cost quantification module is used to calculate an additional cost value based on the difference between the actual cost metric and the baseline cost metric when the predetermined deviation pattern is identified. The cost feature aggregation module is used to aggregate the additional cost values ​​of historical vehicle trips, using the geospatial information of the trip termination location and road attributes as indexes, in order to calculate the cost statistical features associated with the indexes. The knowledge base construction module is used to store the cost statistical features obtained from aggregate statistics into the boarding point cost feature knowledge base for online querying; The candidate pick-up point generation module is used to receive a passenger's ride request and generate a set of candidate pick-up points based on the passenger's location and preset rules. The first type of cost calculation module is used to calculate the first type of cost from the driver's current location to each candidate pick-up point. The first type of cost is obtained based on path planning according to real-time traffic conditions. The second type of cost prediction module is used to query the cost feature knowledge base of the pick-up point for each candidate pick-up point based on its geospatial information and road attributes, so as to obtain the second type of cost associated with the candidate pick-up point. The second type of cost represents the additional cost value caused by the predetermined deviation pattern in the historical trip. The comprehensive scoring module is used to calculate a comprehensive score for each candidate boarding point based on the first type of cost and the second type of cost; The recommendation module is used to sort candidate boarding points based on comprehensive scores and push the sorting results to users.

9. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the intelligent optimization method for boarding points as described in any one of claims 1-7.

Images