CN-122024495-A - Driving style-based signal control intersection electric automobile track prediction method

CN122024495ACN 122024495 ACN122024495 ACN 122024495ACN-122024495-A

Abstract

The invention relates to the technical field of intelligent transportation and machine learning, and discloses a driving style-based signal control intersection electric automobile track prediction method, which comprises the steps of collecting high-definition videos of intersections through an unmanned aerial vehicle, extracting vehicle tracks through Datafromsky, and combining YOLOv to realize accurate identification of electric automobiles; the method comprises the steps of constructing an electric vehicle track database covering multi-dimensional running states and environment parameters, dividing driving styles by a K-Means algorithm, splicing time sequence track characteristics, driving style classification results and environment characteristics of the electric vehicle into an input characteristic set of a PINN-LSTM fusion model, converting kinematic constraint and dynamic constraint in the running process of the vehicle into physical loss terms to be fused into a model loss function, and finally outputting signals to control future time sequence track prediction results of the electric vehicle at an intersection.

Inventors

LI XIAOWEI
HE JIAQI
ZHAO XIANGMO
ZHAO YONGTONG
LI CHAO

Assignees

西安建筑科技大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (9)

1. The method for predicting the track of the electric automobile at the intersection based on the driving style signal control is characterized by comprising the following steps of: Collecting signal control intersection global video data through an unmanned plane; Extracting track data of all vehicles in the intersection from the video data based on Datafromsky software; identifying vehicles corresponding to the track data through YOLOv target detection models, and screening to obtain electric vehicle track data; Calculating running state parameters of the electric automobile based on the electric automobile track data, extracting environmental characteristics including signal lamp states, weather conditions and time periods through the video data, constructing time sequence track characteristics of the electric automobile based on continuous multi-frame track data and the running state parameters, and constructing an electric automobile running track database; Selecting speed, acceleration, headstock distance, headstock time distance and TTC from the running track database as core feature vectors, performing cluster analysis on the core feature vectors by adopting a K-Means algorithm, and classifying and dividing the driving style of the electric automobile according to average speed, speed standard deviation, average acceleration absolute value, maximum acceleration and average acceleration and deceleration change rate absolute value to obtain a driving style classification result; And splicing the time sequence track characteristics, the driving style classification result and the environmental characteristics of the electric automobile into an input characteristic set of a PINN-LSTM fusion model, extracting the time sequence dependency relationship of the vehicle track through an LSTM network, converting the kinematic constraint and the dynamic constraint in the running process of the vehicle into physical loss terms, merging the physical loss terms into a PINN-LSTM fusion model loss function, training a PINN-LSTM fusion model, and outputting a future time sequence track prediction result of the electric automobile through the trained model.
2. The method for predicting the track of the electric automobile at the intersection based on the signal control of the driving style, which is disclosed in claim 1, is characterized in that the flying height of the unmanned aerial vehicle is 60-120 m, the camera is perpendicular to the ground and is noded, the pitch angle error is not more than +/-1 degree, and after the acquisition, bilateral filtering denoising and SIFT feature point matching picture stabilization processing are carried out on the original video acquired by the unmanned aerial vehicle.
3. The method for predicting the track of the electric automobile at the intersection based on the signal control of the driving style according to claim 1 is characterized in that when track data are extracted based on Datafromsky software, a pixel scale is firstly configured, ground calibration is completed, a plane rectangular coordinate system with the center of the intersection as an origin, the east direction as the positive direction of the x axis and the north direction as the positive direction of the y axis is established, and then abnormal point elimination and moving average filtering smoothing processing are carried out on the extracted original track data.
4. The method for predicting the track of the electric automobile at the intersection based on the driving style signal control as claimed in claim 1, wherein when the automobile is identified by a YOLOv target detection model, only the identification result of the electric automobile with the confidence coefficient being more than or equal to 0.7 is reserved, and track data with the confidence coefficient being less than 0.7 is removed.
5. The method for predicting the track of the electric automobile at the intersection based on the signal control of the driving style according to claim 1, wherein before clustering by adopting a K-Means algorithm, the abnormal value processing by a box-line graph method and the Z-score standardization processing are carried out on the core feature vector, the clustering number is set to be 3, and the driving style is divided into a aggressive type, a steady type and a cautious type.
6. The driving style-based signal control intersection electric automobile track prediction method is characterized in that when an electric automobile running track database is constructed, abnormal track data with track length of <15 frames, speed of >60km/h or acceleration absolute value of >6m/s2 are removed, track data with state parameters or environment parameter deletion rate of >5% are removed, and Savitzky-Golay filtering smoothing is adopted for speed and acceleration sequences.
7. The method for predicting the track of the electric automobile at the intersection based on the signal control of the driving style according to claim 1, wherein the running track database contains multi-dimensional labeling information, the multi-dimensional labeling information comprises an intersection attribute, a vehicle running attribute and a traffic flow attribute, the intersection attribute comprises an intersection type, the number of entrance lanes and a signal timing scheme, the vehicle running attribute is an entrance lane type, and the traffic flow attribute is a traffic flow level.
8. The driving style-based signal control intersection electric automobile track prediction method of claim 1 is characterized in that the PINN-LSTM fusion model adopts a 2-layer LSTM network, the number of LSTM units in the first layer is 128, the number of LSTM units in the second layer is 64, the activation function is tanh, dropout rate=0.2, and the weight coefficient λ=0.3 of a physical loss term.
9. The method for predicting the track of the electric automobile at the intersection based on the signal control of the driving style according to claim 1, wherein an average absolute error MAE, a root mean square error RMSE and an average absolute percentage error MAPE calculated based on the deviation between the predicted coordinates and the real coordinates are adopted as prediction performance evaluation indexes, and when MAE is less than 0.3m, RMSE is less than 0.5m and MAPE is less than 3 percent, the model achieves effective prediction performance.

Description

Driving style-based signal control intersection electric automobile track prediction method Technical Field The invention relates to the technical field of intelligent transportation and machine learning, in particular to a driving style-based signal control intersection electric automobile track prediction method which is suitable for traffic flow optimization, safety early warning and intelligent scheduling scenes of urban signal control intersections. Background With the large-scale development of the electric automobile industry and the deep construction of an intelligent traffic system, a signal control intersection is used as a core node of an urban traffic network, and the traffic efficiency and the safety level of the signal control intersection directly influence the overall running quality of urban traffic. The electric automobile has substantial difference with the traditional fuel oil automobile due to the dynamic response characteristic and the energy management strategy, and has obvious specificity in the running behaviors of starting acceleration, following driving, steering avoidance, braking deceleration and the like at a signal control intersection. Meanwhile, the driving styles of different drivers can directly influence the running track of the electric automobile, and the complexity of intersection track prediction is further increased. The electric automobile track integrating the running state and the driving style of the automobile is accurately predicted, and the method has key significance in dynamically optimizing signal timing, improving the passing efficiency of intersections and reducing collision risks. The existing vehicle track prediction technology is mainly divided into a traditional physical model method and a data driving model method. The traditional physical model method (such as a dynamics model and a kinematic model) relies on accurate vehicle parameters and road environment modeling, the generalization capability is poor, the method is difficult to adapt to coupling scenes of complex traffic flows and diversified driving behaviors of an intersection, the data driving model method is superior to the traditional model in generalization capability through massive historical data training models such as LSTM, GNN, transformer, the method still has the core defects that firstly, the data acquisition means are single, the road side radars or vehicle-mounted sensors are relied on, complete coverage of global traffic flows of the intersection is difficult to achieve, the data integrity and objectivity are insufficient, secondly, the electric vehicles and the traditional fuel vehicles are not distinguished pertinently, the dynamic characteristic differences of the electric vehicles are ignored, track characteristic extraction deviation is caused, thirdly, the individualized influences of driving styles are not considered, vehicle tracks of different driving styles are mixed into one, prediction accuracy is limited, the existing model fusing physical constraints is not combined with time sequence dependency characteristics, the time continuity of tracks is difficult to be accurately captured, and an effective integration mechanism for driving styles is lacking, secondly, the electric vehicles with special and the driving styles are not used for supporting a data base. Disclosure of Invention The invention provides a driving style-based signal control intersection electric automobile track prediction method, which aims to solve the defects in the prior art and overcome the defects of single acquisition means, indistinguishable electric automobile specificity, neglecting driving style influence, lack of physical constraint and time sequence capturing capability of a model, lack of special database support and the like in the prior art. The technical scheme of the invention is that the method for predicting the track of the electric automobile at the intersection is controlled by signals based on driving style, and comprises the following steps: Collecting signal control intersection global video data through an unmanned plane; Extracting track data of all vehicles in the intersection from the video data based on Datafromsky software; identifying vehicles corresponding to the track data through YOLOv target detection models, and screening to obtain electric vehicle track data; Calculating running state parameters of the electric automobile based on the electric automobile track data, extracting environmental characteristics including signal lamp states, weather conditions and time periods through the video data, constructing time sequence track characteristics of the electric automobile based on continuous multi-frame track data and the running state parameters, and constructing an electric automobile running track database; Selecting speed, acceleration, headstock distance, headstock time distance and TTC from the running track database as core feature vectors, performing clust