A method for real-time prediction of a destination or a loading / unloading behavior based on vehicle trajectories and transportation routes of road logistics vehicles

Abstract

The application provides a method for predicting a destination or a loading and unloading behavior in real time based on a vehicle trajectory and a transportation line of a highway logistics vehicle, and is an automatic, high-precision and expandable method, which does not need to preset a fence, can dynamically find a new logistics node, significantly improves the identification accuracy of the loading and unloading behavior and the prediction accuracy of an estimated time of arrival (ETA) of long-distance transportation, and is suitable for intelligent scheduling, in-transit monitoring and logistics network optimization.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent transportation and logistics information integration, and is specifically applicable to a method for real-time prediction of destination or loading and unloading behavior based on the vehicle trajectory and transportation route of highway logistics vehicles. Background Technology

[0002] The current mainstream technical approaches for vehicle stop point identification and logistics behavior analysis include: (1) Rule-based detection: This method sets fixed speed and duration thresholds (e.g., speed below 5 km / h and lasting for more than 10 minutes) to determine the stop point. Its logic is simple, but it cannot distinguish the inherent semantics of the stop behavior (e.g., loading and unloading, resting, congestion, refueling, entering and exiting highways, etc.). (2) Geo-fencing-based tagging: This method relies on pre-marked polygonal geofences such as warehouses and logistics parks. When a vehicle's trajectory enters the fenced area and stops, it is marked as loading and unloading. This method relies heavily on the completeness and accuracy of the fence data, cannot discover new and unrecorded loading and unloading locations, and is difficult to adapt to the dynamic expansion and changes of the logistics network. (3) Classification methods based on simple trajectory features: Some studies have attempted to extract simple features around the stop point (such as stop time, whether it is during working hours) for classification, but the feature system is simple and has failed to explore the continuous trajectory mode transformation pattern before and after loading and unloading behavior (such as the typical pattern of high speed driving - low speed crawling - long time stop - low speed crawling - high speed driving).

[0003] However, the main defects of the existing mainstream technologies are as follows: (1) Coarse identification granularity and lack of business semantics: Existing methods are stuck in the binary division of "staying" and "moving", failing to delve into the business level of "why stay", resulting in a large number of loading and unloading events with business value being submerged in noise such as rest and refueling. (2) Poor scalability and high maintenance cost: They rely heavily on preset rules or manually labeled geofences, and the algorithm's generalization ability is weak. Logistics nodes are changing rapidly, and manual maintenance costs are high and lagging. (3) Prediction model is divorced from actual business context: In route planning and ETA prediction, only the road network topology and real-time road conditions are usually considered, ignoring key business contexts such as "what kind of transportation task the vehicle is performing" and "whether the driver has a fixed route preference", resulting in large prediction deviations in long-distance and complex scenarios. (4) Architecture design fails to distinguish data completeness: It does not clearly distinguish between offline analysis based on complete historical data and online prediction based on fragmented real-time data, and often uses the same complex model, resulting in high computational load on the real-time system or insufficient depth of offline analysis. Summary of the Invention

[0004] Technical solution: In view of the shortcomings of existing methods, the present invention provides a method for real-time prediction of destination or loading and unloading behavior based on the vehicle trajectory and transportation route of highway logistics vehicles. The specific steps include using a deep neural network structure to construct a behavior recognition model for predicting the next loading and unloading POI or city / district based on the input vehicle trajectory points. A complete transportation task in the vehicle trajectory is defined as loading stop segment A - driving segment - unloading stop segment B. By mapping the start and end points in each transportation task to the features of loading and unloading POIs, a route profile table for each vehicle or fleet is generated, and a loading and unloading POI knowledge base is constructed. Based on the loading and unloading POI knowledge base and route profile table, the historical trajectory of the vehicle over a period of time is collected, and the behavior recognition model is trained and iterated until the model outputs stably, and the POI knowledge base and vehicle route profile are updated. By inputting real-time trajectory segments of vehicles, the predicted destination or loading / unloading behavior of vehicles can be obtained.

[0005] As an improvement, the real-time trajectory segments of the vehicle also include preprocessing, specifically: obtaining the sequence of vehicle trajectory points. , where each point This includes timestamps, longitude, latitude, speed, orientation, and address information; Multiple key feature points are extracted from the vehicle trajectory points and segmented. Non-loading stop points are filtered out and the output trajectory or segment with behavior labels is generated.

[0006] As an improvement, key feature points include different modes such as high-speed segment, low-speed segment, micro-dwelling point, and dwell segment, where the micro-dwelling point is defined as the vehicle trajectory point. speed For N consecutive points below the threshold And the total duration exceeds High-speed sections are defined as vehicle trajectory points. speed Greater than The low-speed segment of the driving route is defined as the vehicle trajectory point. speed exist A driving section is defined as a road segment where multiple micro-stop points converge.

[0007] As an improvement, the methods for filtering non-loading stop points include: (1) using map data to identify infrastructure stop sections of highway service areas, highway entrances, highway exits, toll stations, gas stations, and maintenance stations, excluding loading and unloading activities of special vehicles; (2) analyzing the speed curves of the trajectories before and after the stop point and comparing them with historical average travel time and changes in driving direction, and identifying continuous same-direction stop periods, excluding low-speed stop periods due to traffic light waiting or long congestion periods; (3) feature points where the straight-line distance between the stop point and the main road, including national highways and provincial highways, is less than a threshold, i.e., feature points deviating from the road distance.

[0008] As an improvement, the behavior recognition model adopts a multi-dimensional spatiotemporal feature system, which includes temporal pattern features, spatiotemporal distribution features, and trajectory morphology features. Temporal pattern features are obtained by extracting the preceding / following motion patterns in the trajectory segments, i.e., the average velocity of the trajectory M kilometers prior to the segment, and the micro-motions during the dwell period to obtain the velocity variance, position standard deviation, and total change in orientation angle for all points. ; The spatiotemporal distribution characteristics include the duration of stay and whether the time of day is a weekday, i.e., the time of occurrence; The trajectory morphology features include the terminal branch depth index and the local turning degree. The terminal branch depth index is the ratio of the shortest path length from the stop point to the nearest main road to the straight-line distance. The local turning degree is the standard deviation of the direction angle within the trajectory range of a set threshold distance before and after the stop point.

[0009] As an improvement, the spatiotemporal features also include textual semantic features. The descriptive text of the location of the stop point is obtained through inverse geocoding, and the text feature vector is extracted using a pre-trained word embedding model. The descriptive text includes the addresses of highway entrances and exits, rest areas, gas stations, charging stations, and vehicle maintenance and inspection stations.

[0010] As an improvement, when training and iterating the behavior recognition model, the specific training method using the autoregressive training mechanism is as follows: (1) Cold start: Manually label a high-quality dwell time sample. (2) Perform model training: using Training the initial model (3) Automatic annotation: using For massive amounts of unlabeled data For prediction, select high confidence levels, i.e., prediction probabilities higher than a threshold. The samples were added to the labeling pool; (4) Manual review and correction: The model prediction results were reviewed by a small-scale sampling, and erroneous samples were corrected to form a labeling pool. (5) Iterative optimization: As a new training set, the model is trained. Repeat steps 3-4-5, iterating continuously: ; The specific iterative process uses existing models Label the samples that do not meet the criteria, when the predicted probability is higher than the target. That is, if the labels are considered correct, they are added to a new training set, and then the next model is trained. This process is repeated until there are no more samples with high prediction probabilities, thus obtaining the final optimized set.

[0011] As an improvement, the construction of the loading point POI knowledge base also includes the generation of electronic fences. For each loading and unloading location, the satellite map and POI layer of the map provider are called to identify its convex hull or minimum bounding rectangle as the electronic fence of that POI.

[0012] As an improvement, the loading point POI knowledge base includes attribute statistical calculation parameters, specifically including loading and unloading frequency and average dwell time. Service vehicle type / cargo type This refers to the total number of valid vehicle stops observed at these loading point points (POIs).

[0013] As an improvement, an offline analysis engine and a real-time prediction engine are also included; the offline analysis engine is used to input a full historical trajectory over a period of time into the recognition model, train and iterate the model, output the constructed vehicle route image, and update the loading and unloading POI knowledge base, model parameters and historical profiles. The real-time prediction engine is used to input real-time vehicle trajectory segments into the recognition model, perform recognition through the model, and output the next possible unloading POI or city, district, and county based on the loading / unloading POI knowledge base and historical route graphics. AND Calculation, where ;

[0014] The average speed is determined by both current road conditions and historical speeds on the same route. AND This refers to the "estimated time of arrival of the vehicle," short for Estimated Time of Arrival.

[0015] As a specific embodiment of the present invention, the present invention also provides a computer medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the above-described method.

[0016] Beneficial effects: This invention proposes a method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles. It is an automated, high-precision, and scalable method that does not require pre-set fences, can dynamically discover new logistics nodes, and significantly improves the accuracy of loading / unloading behavior identification and the accuracy of ETA prediction for long-distance transportation. It is suitable for intelligent scheduling, in-transit monitoring, and logistics network optimization.

[0017] Compared with conventional methods, the specific advantages of the present invention are: (1) to achieve fine-grained and high-precision identification of loading and unloading behavior: it does not rely on fixed fences, but learns the spatiotemporal pattern fingerprint of loading, unloading, getting on and off highways, refueling and other behaviors from the trajectory data itself.

[0018] (2) Establish a data-driven self-iterative annotation system: Design an autoregressive closed loop of “small sample initiation -> model recognition -> result feedback -> model optimization” to greatly reduce the dependence on manual annotation.

[0019] (3) Construct a logistics geographic knowledge graph: automatically mine, deduplicate, and count loading and unloading POIs from massive trajectories to form a dynamically updated logistics facility knowledge base.

[0020] (4) Empowering accurate logistics forecasting: The identified loading and unloading behaviors and route attribution results are used as key features and injected into the downstream transportation task forecasting and ETA calculation model, which significantly improves the forecasting accuracy.

[0021] (5) Design a system architecture with optimized resources: Clearly separate the offline deep mining model from the online lightweight prediction model to achieve accurate matching of computing resources and business needs. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the logic technology of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] This invention proposes a method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles. This method can achieve prediction using the original sequence of vehicle trajectory points as input. , where each point It includes timestamps, longitude, latitude, speed, direction angle, and address information. The output includes: (1) behavioral labels for each trajectory point or segment (e.g., high-speed driving, low-speed urban driving, loading / unloading stops, entering / exiting highways, entering / exiting gas / charging stations, rest stops, and even vehicle annual inspections and maintenance). (2) A list of mining loading / unloading POIs. Each POI includes attributes such as geographical range, historical loading and unloading frequency, and average dwell time. (3) A historical route attribution table for each vehicle, recording the origin and destination POIs, route frequency, and mileage statistics for each transport. (4) Based on the loading and unloading POIs where the vehicle last stopped and the current real-time trajectory, output the predicted next unloading POI and dynamic ETA. Example

[0025] This embodiment specifically provides a method for identifying the stopping behavior of highway logistics vehicles and assigning transportation routes based on vehicle trajectories. The specific steps include: Step 1: Multi-granularity trajectory segmentation, using a hierarchical segmentation strategy to aggregate continuous trajectory points into semantic segments.

[0026] Micro-level dwell points and macro-level trajectory segments were defined. The specific content is as follows: (1) Micro-level dwell point detection: A sliding window algorithm based on speed-duration was adopted. For trajectory points If its speed N consecutive points below the threshold (e.g., 5 km / h), and the total duration exceeds If the time interval is 5 minutes, then the area is determined to be a micro-dwelling point.

[0027] (2) Macroscopic trajectory segment division: Combining the context of speed and road type, three types of macroscopic segments are divided: (2.1) High-speed section: The speed is consistently greater than the set speed, such as 45 km / h, and the vehicle is located on a highway or national road. It can be basically ruled out that the vehicle is completing the loading and unloading behavior.

[0028] (2.2) Low speed range: speed between For example, at speeds of 5-45 km / h, the rate of change in steering angle is large, which usually corresponds to driving in urban areas, and can also basically rule out the possibility that the vehicle is completing loading and unloading operations.

[0029] (2.3) Dwelling section: It is composed of micro dwelling points and is the key area for identifying loading and unloading behavior. Refueling and charging behavior, vehicle repair and annual inspection, etc. usually occur here, but are usually related to specific facilities and institutions.

[0030] Step 2: Contextual geographic filtering based on multi-source data. To reduce the interference of non-loading and unloading stops on the model, a multi-level geographic filter is constructed.

[0031] First layer: Static infrastructure filtering: Using high-precision map data, this layer excludes trajectory points falling on preset types of stopping points such as highway entrances / exits, highway service areas, toll stations, gas stations, and repair shops. Unless it's a special vehicle, such as a fuel truck, the likelihood of loading or unloading at these stopping points is extremely low. In the machine learning feature list, labels such as "highway exit | highway entrance | highway service area | toll station | gas station | repair shop" are added.

[0032] The second layer: Dynamic context filtering. This layer combines real-time traffic information to filter out prolonged low-speed stops caused by traffic light waiting or severe congestion. This can be achieved by analyzing the speed curves of the trajectory before and after the stop point and comparing them with historical average travel time and changes in travel direction. In machine learning, features such as "continuous same-direction stops" are added to the feature list.

[0033] Third layer: Spatial morphology filtering: Rule: Straight-line distance from the stopping point to the main road (such as national highway, provincial highway) < (For example, a distance of 10 meters) is generally unlikely to be a loading or unloading operation. This rule is based on the fact that vehicles involved in general loading, unloading, refueling, maintenance, and annual inspections tend to deviate from the road and enter surrounding areas, especially since loading and unloading usually occur inside factories or warehouses, rather than close to the roadside. In machine learning, a "distance from road" feature should be added to the feature list.

[0034] Step 3: Multi-dimensional feature engineering of loading and unloading behavior. This process constructs a feature system that comprehensively describes loading and unloading behavior. Furthermore, the features include temporal pattern features, spatiotemporal distribution features, and trajectory morphology features. 3.1 Temporal pattern characteristics: including preceding / following motion patterns and micro-motions during the dwell period.

[0035] (1) Preceding / Postceding motion pattern: Extracting the preceding pause segment The average speed of the kilometer trajectory, and whether it is mainly high-speed travel. For example, long-distance high-speed travel is often carried out before unloading.

[0036] (2) Micro-motion during the dwell period: Calculate the velocity variance, position standard deviation (whether stationary), and total change in direction angle for all points within the dwell segment. This represents operations such as "moving the vehicle" and "reversing the load." In this invention, unloading is considered to occur after a long transportation distance; if there is entry / exit on a highway, unloading usually occurs shortly after exiting the highway. Conversely, loading occurs not far from the previous unloading location; if there is entry / exit on a highway, loading generally occurs shortly after re-entering the highway.

[0037] 3.2 Spatiotemporal distribution characteristics: including duration of stay: Time of occurrence: time of day (converted to sine-cosine encoding), whether it is a weekday.

[0038] 3.3 Trajectory morphology characteristics: including terminal branch depth index and local turning degree.

[0039] In this invention, the "end-point branch road depth index" measures the depth to which a vehicle enters the branch road where the stopping point is located. It is calculated as the ratio of the shortest path length from the stopping point to the nearest main road to the straight-line distance. Local turning angle: Within a certain range, such as 500 meters, before and after the stopping point, the standard deviation of the direction angle is calculated.

[0040] 3.4 Text Semantic Features: Descriptive text of the stop location (e.g., "XX Logistics Park", "XX Factory", "XX Gas Station", "XX Charging Station", "XX Service Area", "XX Expressway Exit") is obtained through reverse geocoding. A pre-trained word embedding model is used to extract text feature vectors. Additionally, if the stop occurs at a previously identified loading / unloading POI, the probability of loading / unloading is higher.

[0041] Step 4: Autoregressive training of the behavior recognition model. Model selection: Initially, Gradient Boosting Decision Tree (GBDT), such as LightGBM, can be used because it is friendly to heterogeneous features and missing values, and has strong interpretability. Later, it can be upgraded to a deep neural network (such as a Transformer-based model) to capture more complex temporal dependencies.

[0042] The autoregressive training process described in this invention is as follows: (4.1) Cold start: Manually label a certain number of high-quality dwell segment samples, for example, 1000. .

[0043] (4.2) Model training: Use the initial model for training.

[0044] (4.3) Automatic annotation: using For massive amounts of unlabeled data To make a prediction, select high confidence levels (prediction probability > 100%). The samples were added to the labeling pool.

[0045] (4.4) Manual review and correction: A small-scale sampling review of the model prediction results is conducted to correct erroneous samples and form a complete set of results. .

[0046] (4.5) Iterative optimization: As a new training set, the model is trained. Repeat steps 4.3-4.4-4.4 iteratively: ;

[0047] This process significantly reduces the cost of manual annotation and enables the model to adapt to changes in the logistics network.

[0048] Step 5: POI mining and knowledge base construction for loading and unloading, including electronic fence generation and attribute statistical calculation.

[0049] 5.1 Electronic Fence Generation: For each loading and unloading location, the satellite map and POI layer from the map provider are retrieved, and its convex hull or minimum bounding rectangle is identified as the electronic fence for that POI. The map provider's POI is not mandatory, but merely a means to improve accuracy. If it is not available, it can be replaced by a rectangle from existing technologies such as Geohash L9. Any technology that can achieve the same function is within the scope of protection of this patent.

[0050] 5.2 Attribute Statistics Calculation: This includes loading and unloading frequency statistics, average dwell time, and knowledge base updates. Loading and unloading frequency statistics are based on the historical total number of times recorded from massive vehicle trajectories. Average dwell time: The 10th, 50th, and 90th quantiles of dwell time are specifically defined as follows: 10% of dwell times are less than or equal to the 10th quantile, 50% are less than or equal to the 50th quantile, and 90% are less than or equal to the 90th quantile. These statistics can be used to estimate the loading and unloading time of a POI (Point of Interest), which is a cost factor considered by drivers. Knowledge base update: After new trajectory data flows in, incremental updates to POIs are triggered (addition, merging, frequency updates).

[0051] Step 6: Vehicle transportation route attribution and profiling, including single transportation identification, route abstraction, and statistical profiling.

[0052] 6.1 Single Transport Recognition: In the vehicle trajectory, a continuous sequence of "loading stop segment A" -> "driving segment" -> "unloading stop segment B" is identified and defined as a complete transport task. .

[0053] 6.2 Route Abstraction: The origin and destination points of each transport are mapped to the POIs generated in step 5. Thus, a specific transport is abstracted as a route. The granularity of the route can also be city-to-city, i.e. .

[0054] 6.3 Statistical profiling: Construct a route profile table for each vehicle or each fleet. See Table 1 for an example image in this embodiment.

[0055] Table 1 Examples of statistical images obtained in Example 1

[0056] Step 7: Implementation of offline and real-time strategy separation, including offline analysis engine and real-time prediction engine.

[0057] 7.1 Offline Analysis Engine Input: Complete historical trajectories of vehicles over several months or years. Task: Execute all steps 1-6 above to train a high-precision behavior recognition model, mine the full POI knowledge base, and construct vehicle route profiles. The model is complex and has a long running cycle (e.g., once a day). Output: Updated model parameters, POI database, and historical profiles.

[0058] 7.2 Real-time prediction engine: Input: Real-time trajectory segments of the vehicle in the current and recent period (e.g., the last 6 hours). Task: (1) Load the offline trained lightweight model, such as a pruned model, to perform real-time behavior recognition. (2) Combine the real-time recognized behavior, current location, and vehicle historical route profile (e.g., the driver often runs the Shanghai->Nanjing route and has just finished loading in the Shanghai warehouse) to predict the next possible unloading POI or city / district.

[0059] Dynamic ETA calculation can also be obtained in this step: ;

[0060] The average speed is determined by both current road conditions and historical speeds on the same route. AND This parameter is used to estimate when the driver will complete the transport and unload the goods; it is the abbreviation for Estimated Time of Arrival.

[0061] The following data, obtained from testing on a real trajectory dataset of over 100,000 freight vehicles and approximately 1 billion kilometers, demonstrates that the method described in this invention exhibits significant advantages over traditional baseline methods (based on fixed duration + geofencing), as shown in Table 2.

[0062] Table 2. Performance Comparison Examples of the Invention and Traditional Baseline Methods

[0063] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles, characterized in that, The specific steps include using a deep neural network structure to construct a behavior recognition model that predicts the next loading / unloading POI or city / district based on the input vehicle trajectory points; A complete transportation task in the vehicle trajectory is defined as loading stop segment A - driving segment - unloading stop segment B. By mapping the start and end points in each transportation task to the features of loading and unloading POIs, a route profile table for each vehicle or fleet is generated, and a loading and unloading POI knowledge base is constructed. Based on the loading and unloading POI knowledge base and route profile table, the historical trajectory of the vehicle over a period of time is collected, and the behavior recognition model is trained and iterated until the model outputs stably, and the POI knowledge base and vehicle route profile are updated. By inputting real-time trajectory segments of vehicles, the predicted destination or loading / unloading behavior of vehicles can be obtained.

2. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, The real-time trajectory segments of the vehicle also include preprocessing, specifically: obtaining the sequence of vehicle trajectory points. , where each point This includes timestamps, longitude, latitude, speed, orientation, and address information; Multiple key feature points are extracted from the vehicle trajectory points and segmented. Non-loading stop points are filtered out and the output trajectory or segment with behavior labels is generated.

3. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 2, characterized in that, Key feature points include different modes such as high-speed segment, low-speed segment, micro-dwelling point, and dwell segment, where the micro-dwelling point is defined as the vehicle trajectory point. speed For N consecutive points below the threshold And the total duration exceeds High-speed sections are defined as vehicle trajectory points. speed Greater than The low-speed segment of the driving route is defined as the vehicle trajectory point. speed exist A driving section is defined as a road segment where multiple micro-stop points converge.

4. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 2, characterized in that, The methods for filtering non-loading stop points include: (1) using map data to identify infrastructure stop sections of highway service areas, highway entrances, highway exits, toll stations, gas stations, and maintenance stations, excluding loading and unloading activities of special vehicles; (2) analyzing the speed curves of the trajectory before and after the stop point and comparing them with historical average travel time and changes in driving direction, and identifying continuous same-direction stop periods, excluding low-speed stop periods due to traffic light waiting or long congestion periods; (3) the stop point is a feature point whose straight-line distance from the main road, including national highways and provincial highways, is less than a threshold, i.e., a feature point whose distance from the road is less than a threshold.

5. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, The behavior recognition model adopts a multi-dimensional spatiotemporal feature system, which includes temporal pattern features, spatiotemporal distribution features, and trajectory morphology features. Temporal pattern features are obtained by extracting the preceding / following motion patterns in the trajectory segments, i.e., the average velocity of the trajectory M kilometers prior to the segment, and the micro-motions during the dwell period to obtain the velocity variance, position standard deviation, and total change in orientation angle for all points. ; The spatiotemporal distribution characteristics include the duration of stay and whether the time of day is a weekday, i.e., the time of occurrence; The trajectory morphology features include the terminal branch depth index and the local turning degree. The terminal branch depth index is the ratio of the shortest path length from the stop point to the nearest main road to the straight-line distance. The local turning degree is the standard deviation of the direction angle within the trajectory range of a set threshold distance before and after the stop point.

6. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectory and transportation route of highway logistics vehicles according to claim 5, characterized in that, Spatiotemporal features also include textual semantic features. The descriptive text of the location of the stop point is obtained through inverse geocoding, and the text feature vector is extracted using a pre-trained word embedding model. The descriptive text includes the addresses of highway entrances and exits, rest areas, gas stations, charging stations, and vehicle maintenance and inspection stations.

7. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, When training and iterating the behavior recognition model, the specific training method using the autoregressive training mechanism is as follows: (1) Cold start: Manually label a high-quality dwell time sample. (2) Perform model training: using Training the initial model (3) Automatic annotation: using For massive amounts of unlabeled data For prediction, select high confidence levels, i.e., prediction probabilities higher than a threshold. The samples were added to the labeling pool; (4) Manual review and correction: The model prediction results were reviewed by a small-scale sampling, and erroneous samples were corrected to form a labeling pool. (5) Iterative optimization: As a new training set, the model is trained. Repeat steps 3-4-5, iterating continuously: ； The specific iterative process involves using the existing model... Label the samples that do not meet the criteria, when the predicted probability is higher than the target. That is, if the labels are considered correct, they are added to a new training set, and then the next model is trained. This process is repeated until there are no more samples with high prediction probabilities, and finally the optimized set is obtained.

8. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, The construction of the loading point POI knowledge base also includes generating electronic fences. For each loading and unloading location, the satellite map and POI layer of the map provider are called to identify its convex hull or minimum bounding rectangle as the electronic fence of that POI.

9. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, The loading point POI knowledge base includes attribute statistics and calculation parameters, specifically including loading and unloading frequency and average dwell time. Service vehicle type / cargo type This refers to the total number of valid vehicle stops observed at the POI (Point of Interest) at the loading location.

10. The method for real-time prediction of destination or loading / unloading behavior based on vehicle trajectories and transportation routes of highway logistics vehicles according to claim 1, characterized in that, It also includes an offline analysis engine and a real-time prediction engine; The offline analysis engine is used to input a full historical trajectory over a period of time into the recognition model, train and iterate the model, output the constructed vehicle route image, and update the loading and unloading POI knowledge base, model parameters, and historical profiles. The real-time prediction engine is used to input real-time vehicle trajectory segments into the recognition model, identify them, and output the next possible unloading POI or city, district, and county based on the loading / unloading POI knowledge base and historical route graphs. ETA Calculation, where The average speed is determined by both current road conditions and historical speeds on the same route. ETA This refers to the vehicle's estimated arrival time.

Images