Large truck right turning behavior intelligent detection method and system

Abstract

The invention relates to an intelligent detection method and system for the right-turning behavior of a large truck, and the method comprises the steps: extracting the positioning data of three key track points before, during and after turning distributed according to a time sequence in the right-turning process of the truck, and combining the fence information of a geographic information system; multi-dimensional feature vectors including speed, time, distance, direction and space area are constructed; spatial discrimination characteristics based on a distance threshold value and a region position are innovatively proposed, and four-level classification discrimination is performed according to the distance and the region attribution by calculating the distance and the relative position relation between a vehicle track point and a geo-fence. Through feature optimization processing and classification model training, three-classification prediction of vehicle parking behaviors is realized: direct passing, slow driving and parking. According to the invention, the right-turning parking behavior of the large truck can be accurately identified in real time, and an effective technical means is provided for traffic safety supervision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent traffic and vehicle monitoring, and particularly relates to a large truck right-turn behavior intelligent detection method and system. BACKGROUND

[0002] With the rapid development of the transportation industry, large trucks play an important role in road transportation. However, due to the characteristics of large body, large turning radius, and inner wheel difference, large trucks are prone to traffic accidents during right-turning, especially causing serious threats to pedestrians and non-motor vehicles. In order to improve traffic safety, many areas have begun to implement the regulation that large trucks must stop before right-turning, such as Shandong Province which has clearly stipulated by legislation that large passenger cars and heavy load trucks should stop and observe before right-turning.

[0003] The large truck described in the present application refers to a truck with a total mass greater than or equal to 12000 kg.

[0004] The existing vehicle monitoring technology mainly has the following problems:

[0005] 1. Insufficient information amount of single point judgment method: The traditional stop behavior judgment method is usually based on the speed information of a single or two trajectory points, which cannot fully reflect the dynamic behavior change of the vehicle during turning, resulting in low judgment accuracy.

[0006] 2. Limited accuracy of geofencing judgment: The existing technology mainly uses simple binary judgment (inside or outside the fence) to judge the position relationship between the vehicle and the geofence, which cannot accurately reflect the spatial relationship between the vehicle and the stop reference position.

[0007] 3. Identification accuracy and real-time performance need to be improved: The identification accuracy of the existing method in complex traffic environment is not high, and the processing speed is slow, which is difficult to meet the demand of real-time and all-round monitoring.

[0008] 4. Single feature dimension: The existing technology mainly relies on single-dimensional features (such as speed) for judgment, and lacks comprehensive analysis of multi-dimensional features.

[0009] In recent years, the patent technology in the related field is mainly concentrated in video recognition and single-dimensional judgment method. For example, CN119274148A discloses a large truck right turn must stop detection method based on DeepSort algorithm, which mainly detects illegal behaviors based on video stream and target tracking technology, but the performance is limited in bad weather conditions and camera missing conditions. CN116895148B discloses a vehicle intersection right turn non-stop monitoring method, which takes three consecutive positioning points on the vehicle driving track line for judgment, but lacks fine analysis of spatial position relationship. CN119541246A discloses a vehicle monitoring and early warning system and method based on satellite positioning technology, which mainly focuses on driving safety warning of vehicles, but does not optimize the right turn parking behavior.

[0010] Therefore, there is an urgent need for an intelligent identification method capable of comprehensively analyzing the time sequence characteristics and spatial position relationship of vehicle trajectory, accurately and real-timely identifying the right turn parking behavior of large trucks. SUMMARY

[0011] In view of the problems of low accuracy, poor real-time performance and insufficient automation of the parking behavior judgment in the prior art, an intelligent detection method and system for right turn behavior of large trucks are proposed, which can real-timely and accurately identify the right turn parking behavior of large trucks, and provide an effective technical means for traffic safety supervision.

[0012] The technical scheme of the present application is as follows: an intelligent detection method for right turn behavior of large trucks, specifically comprising the following steps:

[0013] S1, data acquisition and trajectory point extraction: collecting high-frequency continuous positioning data stream in the right turn process of the vehicle, abstracting the large truck as a moving particle with the installation position of the vehicle-mounted positioning terminal as the center, and extracting three key trajectory points before, during and after the turn as a sample according to the time sequence distribution, each trajectory point containing latitude and longitude coordinates, speed, time stamp and direction angle information;

[0014] S2, spatial relationship calculation: obtaining geographic information system fence data, and calculating the distance and relative position relationship of the three trajectory points to the geographic fence;

[0015] S3, multi-dimensional feature construction: constructing a multi-dimensional feature vector based on the three key trajectory points, including speed dimension feature, time dimension feature, distance dimension feature, direction dimension feature and spatial region dimension feature;

[0016] S4, feature optimization processing: after the data collected in step S1 is abstractly extracted from the trajectory points, the multi-dimensional feature vector corresponding to each group of three trajectory point data is calculated as a sample data, the feature vectors in all sample data are sequentially standardized, the existing feature vectors are subjected to related feature expansion and importance evaluation method for feature selection, derivative features are generated from the original features, and the optimal feature vector subset is selected;

[0017] S5, intelligent classification prediction: the sample data corresponding to the optimal feature vector subset selected in step S4 is input into a classification model for prediction, the classification model is obtained by training through a multi-model integration strategy, and the training data label is obtained by manual annotation of video monitoring data, and one of three parking states is output: direct passing, slow driving and parking;

[0018] S6, result output and storage: store the prediction result in the database, generate a classification label with description, and record the prediction timestamp.

[0019] Further, the three key trajectory points in step S1 are the right turn event locked by recognizing the continuous change of the vehicle heading angle, and three key trajectory points representing the timing state during the turning process are extracted therefrom:

[0020] 1) Pre-turn trajectory point: the trajectory point at which the vehicle straight driving ends and enters the turning area;

[0021] 2) Mid-turn trajectory point: the first trajectory point entering the core area of the intersection, that is, the positioning data point recorded when the vehicle first enters the preset monitoring range based on the virtual electronic fence;

[0022] 3) Post-turn trajectory point: the trajectory point at which the vehicle completes the turning action and resumes straight driving;

[0023] The data of each trajectory point includes smoothed latitude and longitude coordinates, instantaneous speed, timestamp and heading angle information.

[0024] Further, the speed dimension feature in step S3 is the instantaneous speed of the three trajectory points; the time dimension feature is the time interval between adjacent trajectory points; the distance dimension feature is the moving distance between adjacent trajectory points, and the distance from each trajectory point to the fence; the direction dimension feature is the change of the direction angle between adjacent trajectory points; and the space region dimension feature is the comprehensive discrimination feature based on distance threshold and region position.

[0025] Further, the space region dimension feature in step S3 is determined according to the distance from the current key trajectory point to the geographic fence and the region position, and four-level classification is performed according to the distance and region attribution, the four-level classification including:

[0026] Level 0: the key trajectory point is greater than the preset distance threshold from the geofence and is located outside the preset area of the geofence;

[0027] Level 1: the key trajectory point is greater than the preset distance threshold from the geofence and is located inside the preset area of the geofence;

[0028] Level 2: the key trajectory point is less than or equal to the preset distance threshold from the geofence;

[0029] Level 3: data anomaly or unable to determine;

[0030] The preset distance threshold is determined according to the vehicle volume and road conditions.

[0031] Further, the feature expansion in step S4: generate derived features through mathematical transformation and combination, further expand the feature space from the original dimension features in step S3.

[0032] Further, the importance evaluation method in step S4 is to comprehensively use variance threshold filtering, correlation analysis and tree model based three indicators for feature selection, adopt a weighted comprehensive score mechanism, normalize the above three indicators and sum them up according to the preset weight to obtain a comprehensive quality score, and sort and filter out a stable feature subset with sufficient information, strong correlation with the target and significant contribution to classification according to the score.

[0033] A large truck right turn behavior intelligent detection system, comprising:

[0034] A data acquisition module for real-time acquisition of vehicle positioning trajectory data and geofence data of a geographic information system;

[0035] A data preprocessing module for parsing trajectory data and extracting key trajectory point information;

[0036] A feature engineering module for calculating speed, time, distance, direction and spatial area features;

[0037] A classification model module for loading a trained model and making predictions;

[0038] A real-time prediction module for real-time classification of new data;

[0039] A result output module for storing prediction results in a database and generating reports;

[0040] Each module performs information transmission through the large truck right turn behavior intelligent detection method of any one of claims 1 to 6.

[0041] Preferably, the feature engineering module comprises:

[0042] A distance calculation unit is configured to calculate the moving distance between the key trajectory points and the distance from the key trajectory points to the geofence.

[0043] A region discrimination unit is configured to determine the spatial region feature according to the distance threshold and the region position.

[0044] A feature generation unit is configured to generate the derived features.

[0045] A feature selection unit is configured to screen the optimal feature subset.

[0046] Preferably, the classification label generated by the result output module is in the format of "number - Chinese description", including "0 - directly pass", "1 - slow driving" and "2 - stop".

[0047] The application of the intelligent detection method for the right-turn behavior of the large truck can be applied to the traffic supervision system to monitor the stopping behavior of the large truck when right-turning in real time, identify the illegal behavior and generate an alarm; applied to the vehicle safety evaluation system to evaluate the driving behavior and safety awareness of the driver; used to identify the high-risk intersection with safety hazards; and used to statistically analyze the vehicle passing situation at the intersection and optimize the traffic signal timing.

[0048] The intelligent detection method and system for the right-turn behavior of the large truck have the following beneficial effects: the three key trajectory points before, during and after the turning are extracted and jointly analyzed, the dynamic behavior features of the vehicle during the turning process can be comprehensively captured, the motion state change of the vehicle can be more accurately reflected, and the identification accuracy is significantly improved; the spatial discrimination features based on the distance threshold and the region position are innovatively proposed, the distance from the vehicle trajectory point to the geofence and the relative position relationship are calculated, the four-level classification discrimination is performed according to the distance and the region attribution, compared with the simple binary judgment of the traditional geofence, the spatial relationship between the vehicle and the parking reference position can be more finely reflected, and the accuracy of the spatial position judgment is improved; the multi-dimensional feature vector including the speed, time, distance, direction and spatial region is constructed, the behavior features of the vehicle are comprehensively reflected through the multi-dimensional feature fusion, compared with the single-dimensional judgment method, the accuracy and robustness of the identification are improved; the multi-model integration strategy and feature optimization processing are adopted, the processing speed is significantly improved while ensuring the identification accuracy, and the real-time monitoring requirement is met; based on the satellite positioning and the geographic information system data, the method is not affected by the adverse weather conditions, has stronger environmental adaptability than the video recognition method, and has less construction investment. BRIEF DESCRIPTION OF DRAWINGS

[0049] FIG. 1 is a schematic diagram of the overall process of the method of the present application;

[0050] FIG. 2 is a schematic diagram of three-point trajectory analysis of the present application;

[0051] Fig. 3 is a schematic diagram of the spatial region discrimination of the present application;

[0052] Fig. 4 is a flow chart of the feature engineering process of the present application;

[0053] Fig. 5 is a schematic diagram of the system architecture of the present application. DETAILED DESCRIPTION

[0054] The present application will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments are implemented on the premise of the technical solutions of the present application, and detailed implementation modes and specific operation processes are given, but the protection scope of the present application is not limited to the following embodiments.

[0055] The intelligent detection method and system for large truck right-turn behavior of the present application, aiming at the insufficient information amount of single-point or two-point judgment method, constructs time-series associated trajectory features by extracting three key trajectory points before, during and after turning, and comprehensively captures the dynamic behavior changes of the vehicle during turning; aiming at the insufficient precision of simple binary judgment of geographic fence, realizes fine judgment of the spatial position of the vehicle by calculating the distance and relative position relationship of the vehicle trajectory point to the geographic fence, combining distance threshold and regional position for four-level classification discrimination; aiming at the low recognition accuracy of single-dimensional feature, constructs a multi-dimensional feature vector containing speed, time, distance, direction and spatial region, and improves the recognition accuracy through multi-dimensional feature fusion. After feature optimization processing and classification model training, the present application realizes three-classification prediction of the subsequent right-turn behavior of the vehicle parking: direct passing, slow driving and parking, which assists to improve the correctness of behavior recognition and improves the environmental adaptability of the large truck right-turn behavior.

[0056] An intelligent detection method for large truck right-turn behavior based on satellite positioning and geographic information system data, comprising the following steps:

[0057] S1, data acquisition and trajectory point abstract extraction: using a vehicle-mounted satellite positioning terminal to collect high-frequency continuous positioning data stream of the vehicle in real time, and abstracting the large truck as a moving mass point with the installation position of the vehicle-mounted positioning terminal as the center. The original positioning data is filtered and denoised to correct satellite positioning errors and signal drift. The right-turn event is locked by recognizing the continuous change characteristics of the vehicle heading angle, and three key trajectory points representing the time-series state of the turning process are extracted therefrom, such as Figure 2 a three-point trajectory analysis schematic diagram,

[0058] 1) Pre-turn trajectory point: the time point when the vehicle straight driving ends and enters the turning area;

[0059] 2) Now: the first trajectory point when the vehicle enters the core area of the intersection, i.e., the positioning data point recorded when the vehicle first enters the preset monitoring range based on the virtual electronic fence;

[0060] 3) Next: the point in time when the vehicle completes the turning action and resumes straight-line driving.

[0061] The data of each trajectory point includes smoothed latitude and longitude coordinates, instantaneous speed, timestamp, and heading angle information.

[0062] S2, spatial topology relationship calculation: obtain the preset right-turn parking virtual electronic fence data in the geographic information system (GIS). The core area of the intersection is defined as a spatial area extending a certain range based on the preset boundary. Based on the coordinates of the moving particle, the spatial geometric relationship between the three key trajectory points and the virtual electronic fence is calculated. For the large size of large trucks, a spatial distance tolerance threshold is set, and the perpendicular distance from the trajectory point to the electronic fence boundary and the relative orientation (inside the fence, outside the fence, or edge area) are calculated to represent the overall spatial state of the vehicle.

[0063] S3, multi-dimensional feature construction: based on the three key trajectory points, a multi-dimensional feature vector is constructed, including:

[0064] Speed dimension feature: instantaneous speed of the three trajectory points;

[0065] Time dimension feature: time interval between adjacent trajectory points;

[0066] Distance dimension feature: moving distance between adjacent trajectory points, and distance from each trajectory point to the fence;

[0067] Direction dimension feature: change in direction angle between adjacent trajectory points;

[0068] Spatial region dimension feature: comprehensive discriminant feature based on distance threshold and region position.

[0069] S4, feature optimization processing: after abstracting and extracting the data collected in step S1, the multi-dimensional feature vector corresponding to each group of three trajectory point data is calculated as a sample data. The feature vectors in all sample data are sequentially standardized, the existing feature vectors are expanded and evaluated for importance, and the most optimal feature vector subset is selected from the original features and derived features.

[0070] S5, intelligent classification prediction: input the sample data corresponding to the optimal feature vector subset selected in step S4 into the classification model for prediction. The classification model is obtained by training through a multi-model integration strategy, and the training data labels are obtained by manual annotation of video monitoring data. Cross-validation and hyperparameter optimization are used to output one of three parking states: direct passing, slow driving, and parking. Direct passing corresponds to the behavior of the vehicle passing through the turning area directly without obvious deceleration. Slow driving corresponds to the behavior of the vehicle decelerating during turning but not completely stopping. Parking corresponds to the behavior of the vehicle completely stopping at least once during turning.

[0071] S6, result output and storage: store the prediction result in the database, generate a classification label with description, and record the prediction timestamp.

[0072] Example 1: Intelligent detection method for large truck right-turn behavior: As shown in FIG. 1, the present example provides an intelligent detection method for large truck right-turn behavior based on satellite positioning and geographic information system data. The method realizes fine classification of vehicle behavior by constructing a multi-dimensional feature vector and combining an integrated learning model.

[0073] Specifically, the following steps are included:

[0074] Step S1: data acquisition and trajectory point extraction, real-time acquisition of vehicle positioning trajectory data through a vehicle-mounted satellite positioning device. The system acquires basic information such as latitude and longitude coordinates, instantaneous speed, direction angle, and timestamp through real-time standard protocol data.

[0075] For a complete right-turn event, three key trajectory point sets distributed in time sequence are extracted , and the specific extraction strategy is as follows:

[0076] 1. Pre-turn point : defined as the trajectory point before the vehicle enters the right-turn area, preferably the data point at the start of the turning action, used to represent the initial state before turning;

[0077] 2. Turning midpoint : defined as the trajectory point of the vehicle at the triggering moment of the right-turn event, i.e., the real-time positioning data point recorded when the system detects that the vehicle enters the right-turn monitoring area and triggers an alarm. This point is automatically acquired by the vehicle-mounted satellite positioning device at the alarm triggering moment, and its timestamp is recorded as the accurate moment of alarm triggering. The latitude and longitude coordinates, instantaneous speed, and direction angle information of the turning midpoint are directly measured by the satellite positioning device, and are used to represent the motion state of the vehicle at the turning key moment;

[0078] 3. Post-turn point ​: defined as the trajectory point when the vehicle finishes the turning action and enters the straight driving state, preferably the data point at the moment when the turning action ends to represent the state after turning.

[0079] Each trajectory point contains a four-dimensional state vector , representing longitude, latitude, instantaneous speed, direction angle, and time stamp, respectively.

[0080] Technical explanation of three-point sampling sufficiency:

[0081] The invention adopts three key trajectory points instead of continuous trajectory sampling, based on the following technical considerations:

[0082] (1) Complete capture of time sequence changes: By calculating the time interval between adjacent trajectory points (time_diff_12, time_diff_23), the time consumption difference of the vehicle in the first half and the second half of the turning can be quantified, thereby indirectly reflecting the deceleration or parking behavior of the vehicle. If the vehicle parks near the midpoint of the turning, time_diff_12 or time_diff_23 will significantly increase.

[0083] (2) Representation of speed change trend: The instantaneous speeds of the three trajectory points (pre_speed, now_speed, next_speed) form a speed change sequence, which can represent different behavior patterns such as "deceleration-stop-acceleration", "uniform speed through", "slow deceleration", etc. The derived features such as speed difference and speed ratio generated by feature engineering further enhance the representation ability of the speed change trend.

[0084] (3) Quantitative analysis of direction change: By calculating the direction angle change (orientation_diff1, orientation_diff2) between adjacent trajectory points, the steering amplitude of the vehicle during the turning process can be quantified. In the normal right turning process, the direction angle change should show a regular progressive change; if the vehicle stops and waits, the direction angle change may be abnormal.

[0085] (4) Multi-dimensional representation of spatial position: In addition to the relative displacement (dis_12, dis_23) between the three trajectory points, the distances (dis_1p, dis_2p, dis_3p) from each trajectory point to the geofence are also calculated, thereby constructing a complete spatial position relationship atlas.

[0086] (5) Information enhancement by feature engineering: Through polynomial feature expansion, mathematical transformation, and feature interaction, 46-dimensional derived features are generated from the 13-dimensional original features, fully exploiting the implicit relationships and non-linear features between the three trajectory points, and making up for the deficiency of limited number of sampling points.

[0087] In summary, three-point sampling combined with multi-dimensional feature construction and feature engineering can effectively capture the dynamic behavior changes of vehicles during turning, ensuring computational efficiency while fully representing the turning behavior. The current actual test data shows that the recognition accuracy of this method reaches more than 92%, verifying the effectiveness of the three-point sampling scheme.

[0088] Step S2: Spatial relationship calculation:

[0089] Obtain intersection geofence information from GIS database , the fence boundary is usually defined as a polygon or linear equation . Calculate the spatial topological relationship of three key trajectory points with the geofence:

[0090] Vertical distance calculation: use the point-to-line distance formula to calculate the Euclidean distance from the trajectory point to the fence boundary : This distance is used to quantify the spatial proximity of the vehicle to the stop reference line.

[0091] Relative position determination: determine the inside-outside relationship of the trajectory point relative to the fence by algebraic sign method or ray method. Define the area orientation , distinguish whether the vehicle is located outside the fence (has not reached the stop line) or inside the fence (has crossed the stop line).

[0092] Step S3: Multi-dimensional feature construction: based on the extracted three key trajectory points and calculated spatial relationships, construct a high-dimensional feature vector containing 46 original and derived features , which includes the following five dimensions:

[0093] 1. Velocity Features: construct a velocity vector , corresponding to the instantaneous speed at three time points before, during and after turning, reflecting the dynamics of the vehicle throughout the turning process.

[0094] 2. Temporal Features: calculate the time interval between adjacent trajectory points, used to normalize behavior patterns under different sampling frequencies: and .

[0095] 3. Spatial Features:

[0096] 3.1. Moving distance: use geodesic distance algorithm (such as Geodesic algorithm) to calculate ​ to ) and ( to );

[0097] 3.2, Fence Distance: corresponding to the vertical distance from the three key points to the geofence.

[0098] 3, Heading Features: Calculate the change in the vehicle's heading angle and periodically normalize it to eliminate the 0° / 360° discontinuity effect.

[0099] 4, Spatial Region Features: This is one of the core features of the invention. Define the discrete feature variable extra_region (this variable is obtained by combining the results of comparing the Euclidean distance from the turning midpoint to the geofence with the preset distance threshold and the region orientation judgment result of the trajectory point relative to the fence. This feature discretizes continuous spatial position information into four levels to represent the vehicle's spatial state category relative to the stop reference line during turning). According to the distance from the turning midpoint to the fence and the region orientation , the four-level spatial state comprehensive judgment feature ( Figure 3 ) is:

[0100] 4.1, Level 0: and (outside region), indicating that the vehicle is far from the stop line;

[0101] 4.2, Level 1: and (inside region), indicating that the vehicle is approaching but not crossing the line;

[0102] 4.3, Level 2: , indicating that the vehicle is in the valid parking area;

[0103] 4.4, Level 3: Data missing or abnormal.

[0104] Step S4: Feature optimization processing: To improve model performance, as shown in Figure 4 , the following engineering processing is performed on the original feature vector:

[0105] 1, Feature Standardization: Use Z-score standardization method to map continuous numerical features to standard normal distribution.

[0106] 2. Feature expansion: Derivative features are generated through mathematical transformations and combinations to expand the feature space from the original dimension to 46 dimensions (e.g., speed change rate, average speed, nonlinear transformation, etc.).

[0107] 3. Feature selection: A variety of feature evaluation methods are used for feature screening, including feature filtering based on statistical distribution, correlation analysis based on the association between features and target variables, and feature importance evaluation based on machine learning models. Among them, feature filtering based on statistical distribution removes features with insufficient information by analyzing the distribution characteristics of each feature in the sample set; correlation analysis removes features with weak association with the classification target by calculating the statistical dependence between each feature and the target variable; feature importance evaluation based on machine learning models calculates the contribution of each feature to classification decisions by training a classification model. Finally, a comprehensive scoring mechanism is used to normalize the above evaluation indicators and weight them according to the preset weights to obtain a comprehensive quality score. According to this score, a stable feature subset with sufficient information, strong correlation with the target, and significant contribution to classification is selected, forming an optimal feature vector subset for subsequent classification prediction. The optimal feature subset includes: 、 、 , speed change rate, , orientation_diff1, , etc.

[0108] Step S5: Intelligent classification prediction: input the filtered feature vector into the pre-trained classification model. The model uses a multi-model integration strategy, which is implemented as follows:

[0109] Model integration: integrate multiple heterogeneous classifiers such as LightGBM, XGBoost, CatBoost, Random Forest, and LogisticRegression.

[0110] Training optimization: In the training phase, K-Fold (preferably K=5) cross-validation method is used to evaluate the performance of each base model, and grid search (Grid Search) is used to optimize hyperparameters. Finally, the model with the best performance is selected based on the F1 score or a weighted voting strategy is used.

[0111] State output: the model outputs one of the three parking states:

[0112] 1. State 0 (directly through): the vehicle maintains a high speed throughout the journey and no significant deceleration trend is detected;

[0113] 2. State 1 (slow driving): the vehicle shows significant deceleration, but the minimum speed does not meet the complete stop standard (e.g. );

[0114] 3. State 2 (Stop): There is at least one moment during the turning process that the vehicle speed is extremely low or zero (completely stopped).

[0115] Step S6: Result output and storage: The predicted classification results are structured and stored in the database, generating standardized labels in the format of "number-Chinese description" (such as "0-direct pass", "1-slow driving", "2-stop"). At the same time, the prediction timestamp, vehicle unique identifier, and corresponding original trajectory data are recorded.

[0116] Example 2: Large truck right turn behavior intelligent detection system:

[0117] As shown in FIG. 5, the present application also provides an intelligent recognition system based on the above method, and the logic architecture thereof includes the following core modules:

[0118] 1. Data acquisition module: responsible for real-time data acquisition and access. This module receives trajectory data uploaded by vehicle-mounted satellite positioning devices (supporting GPS, Beidou, etc.) through standard interfaces, and synchronizes intersection geofencing data from GIS servers, providing basic input for the system.

[0119] 2. Data preprocessing module: responsible for data cleaning and formatting. This module has a built-in JSON parsing engine for parsing raw trajectory data packets and extracting information at the three key time points (before, during, and after turning) defined in Example 1, while eliminating abnormal data such as GPS drift.

[0120] 3. Feature engineering module: this module is the core of the algorithm, including:

[0121] 3.1 Distance calculation unit: encapsulates the Geodesic algorithm and point-line distance formula to accurately calculate the vehicle movement distance and the perpendicular distance to the fence.

[0122] 3.2 Area discrimination unit: built-in spatial topology analysis logic, based on the preset distance threshold and orientation determination rules, outputs spatial area dimension features .

[0123] 3.3 Feature generation unit: performs polynomial expansion, mathematical transformation, and statistical aggregation to generate high-dimensional derivative features.

[0124] 3.4 Feature selection unit: maintains a feature importance list and real-time filters the highest contribution subset of features (such as 9-dimensional optimal features) to classification.

[0125] 4. Classification model module: responsible for loading and managing machine learning models. This module supports the serialized loading of model files (such as.pkl format), and supports the configuration of the optimal model. The module has a built-in inference engine, which supports fast classification prediction on input feature vectors.

[0126] 5. Real-time prediction module: uses a streaming processing architecture to interface with real-time data streams. This module calls the feature engineering module and the classification model module to complete the processing and prediction of newly arrived data within milliseconds, meeting the real-time needs of traffic regulation.

[0127] Result output module: responsible for the persistence and display of results. This module writes the prediction results to the database, generates "number-Chinese description" format classification labels, and can trigger real-time alerts or generate statistical reports for traffic management departments.

[0128] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the scope of protection of the present application patent should be subject to the appended claims.

Claims

1. A method for intelligent detection of right-turning behavior of large trucks, characterized in that, Specifically, the steps include the following: S1. Data Acquisition and Trajectory Point Extraction: Collect high-frequency continuous positioning data streams during the vehicle's right turn process, abstract the large truck into a moving mass point centered on the installation location of the on-board positioning terminal, and extract three key trajectory points distributed in time sequence before the turn, during the turn, and after the turn as a sample. Each trajectory point contains latitude and longitude coordinates, speed, timestamp, and direction angle information. S2. Spatial Relationship Calculation: Obtain geographic information system fence data and calculate the distance and relative positional relationship between three trajectory points and the geographic fence; S3. Multi-dimensional feature construction: Construct multi-dimensional feature vectors based on three key trajectory points, including velocity dimension features, time dimension features, distance dimension features, direction dimension features, and spatial region dimension features; S4. Feature optimization processing: After the data collected in step S1 is abstracted and extracted from trajectory points, the multi-dimensional feature vector corresponding to each group of three trajectory point data is calculated as a sample data. The feature vectors in all sample data are standardized in turn. The existing feature vectors are expanded with relevant features and the importance evaluation method is used to select features. Derived features are generated from the original features and the optimal feature vector subset is selected. S5. Intelligent classification prediction: Input the sample data corresponding to the optimal feature vector subset selected in step S4 into the classification model for prediction. The classification model is trained through a multi-model ensemble strategy. The training data labels are obtained by manual annotation of video surveillance data. The output is one of three parking states: pass directly, drive slowly, or stop. S6. Result Output and Storage: Store the prediction results in the database, generate classification labels with descriptions, and record the prediction timestamps.

2. The intelligent detection method for right-turning behavior of large trucks according to claim 1, characterized in that, In step S1, the three key trajectory points are obtained by identifying the continuous change characteristics of the vehicle's heading angle to lock in the right turn event, and extracting three key trajectory points that characterize the temporal state of the turning process: 1) Pre-turn trajectory point: The trajectory point where the vehicle is about to enter the turning area after completing straight-line driving; 2) Trajectory point during turning: The first trajectory point when entering the core area of ​​the intersection, that is, the location data point recorded when the vehicle first enters the preset monitoring range based on the virtual electronic fence; 3) Trajectory point after turning: The trajectory point where the vehicle resumes straight-line travel after completing the turning maneuver; The data for each trajectory point includes smoothed latitude and longitude coordinates, instantaneous velocity, timestamp, and heading angle information.

3. The intelligent detection method for right-turning behavior of large trucks according to claim 1, characterized in that, In step S3, the velocity dimension feature is the instantaneous velocity of the three trajectory points; the time dimension feature is the time interval between adjacent trajectory points; the distance dimension feature is the moving distance between adjacent trajectory points and the distance from each trajectory point to the fence; the direction dimension feature is the change in the direction angle between adjacent trajectory points; and the spatial region dimension feature is a comprehensive discrimination feature based on the distance threshold and the region location.

4. The intelligent detection method for right-turning behavior of large trucks according to claim 3, characterized in that, The spatial region dimension feature in step S3 is determined by comprehensively considering the distance from the current key trajectory point to the geofence and the region location. A four-level classification is performed based on distance and region affiliation: Level 0: The key trajectory point is more than a preset distance threshold from the geofence and is located in a preset region outside the geofence; Level 1: The key trajectory point is more than a preset distance threshold from the geofence and is located in a preset region inside the geofence; Level 2: The key trajectory point is less than or equal to a preset distance threshold from the geofence; Level 3: Data is abnormal or cannot be determined. The preset distance threshold is determined based on vehicle size and road conditions.

5. The intelligent detection method for right-turning behavior of large trucks according to claim 4, characterized in that, In step S4, feature expansion involves generating derived features through mathematical transformations and combinations, further expanding the feature space from the original dimensional features of step S3.

6. The intelligent detection method for right-turning behavior of large trucks according to claim 5, characterized in that, The importance assessment method in step S4 is to comprehensively use three indicators—variance threshold filtering, correlation analysis, and tree model-based feature selection—and adopt a weighted comprehensive scoring mechanism. After normalizing the above three indicators, they are weighted and summed according to preset weights to obtain a comprehensive quality score. Based on this score, a stable subset of features that are sufficiently informative, strongly correlated with the target, and significantly contribute to classification are selected.

7. An intelligent detection system for right-turning behavior of large trucks, characterized in that, include: The data acquisition module is used to collect vehicle positioning trajectory data and geographic information system fence data in real time; The data preprocessing module is used to parse trajectory data and extract key trajectory point information; The feature engineering module is used to calculate velocity, time, distance, direction, and spatial region features; The classification model module is used to load the trained model and make predictions. The real-time prediction module is used to classify new data in real time. The results output module is used to store the prediction results in the database and generate reports; Each module transmits information through the intelligent detection method for right-turning behavior of large trucks as described in any one of claims 1 to 6.

8. The intelligent detection system for right-turning behavior of large trucks according to claim 7, characterized in that, The feature engineering module includes: The distance calculation unit is used to calculate the movement distance between key trajectory points and the distance from key trajectory points to geofences; The region discrimination unit is used to determine the spatial region characteristics based on distance thresholds and region location. Feature generation unit, used to generate derived features; The feature selection unit is used to filter the optimal subset of features.

9. The intelligent detection system for right-turning behavior of large trucks according to claim 8, characterized in that, The classification label format generated by the result output module is "number-Chinese description", including "0 - pass directly", "1 - drive slowly" and "2 - stop".

10. An application of the intelligent detection method for right-turning behavior of large trucks, characterized in that, The system is used to monitor the parking behavior of large trucks when they make right turns in real time, identify violations, and generate alarms. Its application in vehicle safety assessment systems is used to evaluate drivers' driving behavior and safety awareness; Used to identify high-risk intersections with potential safety hazards; Used for statistical analysis of vehicle traffic conditions at intersections and optimization of traffic signal timing.

Images