Networked vehicle forced lane change duration prediction method for dynamic bus lane
By constructing the CNN-BiLSTM-Attention combined model, preprocessing and feature screening of forced lanes for bus lanes is solved, and the problem of inaccurate prediction of the duration of forced lanes for vehicles in the existing technology is solved, and the actual working conditions are achieved accurate prediction and resource utilization increase.
Patent Information
- Application Number
- CN202411483659.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2025-02-07
AI Technical Summary
The existing dynamic control method for bus lanes cannot achieve accurate prediction of the duration of forced lanes for vehicles, resulting in insufficient safety of wasted road resources and intelligent connected vehicles collaborative lanes for joint lanes.
Using the CNN-BiLSTM-Attention combined model, the original data set of forced lanes for bus lanes is preprocessed and feature screened, and a CNN-BiLSTM-Attention combined model is constructed and generated, for training and debugging, and finally the duration of forced lanes is realized. accurate prediction.
It realizes an accurate prediction of the length of forced lane change under actual working conditions, improves the resource utilization rate of dynamic bus lanes, and improves the safety of cooperative lane change in intelligent connected vehicles.
Smart Images
Figure CN119398247A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lane change time prediction, and in particular to a method for predicting the duration of forced lane change of a networked vehicle in a dynamic bus lane. Background Art
[0002] Dynamic bus lanes (also known as intermittent bus lanes) are a new type of dedicated lane that dynamically provides social vehicles with the right of way to drive on bus lanes while meeting the bus priority. However, the traditional dynamic control method of bus lanes often ensures bus priority by setting a long and fixed clearing distance in front of the bus, and a large part of the road resources are still wasted. With the rapid development of communication technology and vehicle networking technology, intelligent connected vehicles can obtain sensor data from other vehicles, providing technical support for the refined control of dynamic bus lanes. By establishing a dynamic clearing frame model that changes with the bus speed and the duration of forced lane change of the cleared vehicle in front of the bus, the utilization rate of bus lane resources can be further improved. As an important part of the dynamic clearing frame setting of the dynamic bus lane, the duration of forced lane change of the vehicle is an important part of the dynamic clearing frame setting of the dynamic bus lane. Its accurate prediction is not only crucial to improving the utilization rate of dynamic bus lanes, but also effectively improves the safety of cooperative lane change of intelligent connected vehicles.
[0003] At present, the prediction methods of lane change duration are mainly divided into trajectory modeling method and data-driven method. The trajectory modeling method can realize the prediction of lane change crossing time, but the existing research only focuses on the driving characteristics of the lane changing vehicle itself, and does not consider the influence of factors such as surrounding vehicles and the driver's reaction in a forced environment during the actual lane changing process. Secondly, most of these models are based on multiple assumptions, which makes it difficult to accurately predict the lane change duration during actual driving, and the robustness is also poor. The data-driven method can establish a lane change duration prediction model based only on the mapping relationship between the system input and output, without the need to establish a specific model for different driving conditions. However, most of the existing data-driven methods for lane change duration prediction rely on the state information of the vehicle at a certain moment during the lane changing process, without considering the dynamic characteristics of the state information during the lane changing process, and also ignore the spatial relationship between the various features. The prediction accuracy of the model still needs to be improved. Summary of the invention
[0004] The present invention solves the problem that traditional bus lane dynamic control methods cannot achieve accurate prediction, and proposes a method for predicting the duration of forced lane changes of networked vehicles in dynamic bus lanes, which can achieve accurate prediction of the duration of forced lane changes under actual working conditions, and provide decision-making reference for dynamic bus lane clearing distance setting and collaborative lane changes of networked intelligent vehicles.
[0005] In order to achieve the above object, the present invention adopts the following technical solution: a method for predicting the duration of forced lane change of a connected vehicle in a dynamic bus lane, comprising the following steps: S1, collecting the original data set of forced lane change in bus lanes, and preprocessing the original data set to obtain a feature data set; S2, screening the feature data contained in the feature data set, improving and constructing a CNN-BiLSTM-Attention combination model; S3, training and debugging the combined model to obtain optimal combined model parameters; predicting the forced lane change duration based on the trained combined model.
[0006] In this technical solution, firstly, an original data set of forced lane changes in bus lanes is established, and then the original data set is preprocessed. Based on the preprocessing, the features contained in the data set are screened; then, the structure of the traditional BiLSTM network is optimized, and the convolutional neural network and the attention mechanism are integrated; then, the model is trained using the forced lane change data set of bus lanes after extracting key features, and the trained model is embedded in the dynamic clearing system for bus lane reuse. By calling the data of the vehicle in front of the lane and its surrounding environment obtained by the networked bus, the lane change duration is finally predicted, providing a basis for the setting of the dynamic clearing box.
[0007] The present invention is further configured as follows: the process of obtaining the original data set is: S11, taking aerial photos of the bus lane to obtain complete video data of the driving process of the vehicle forced to change lanes to avoid the bus and the surrounding vehicles on the bus lane; S12, extracting the forced lane change trajectory point data from the complete video data, obtaining the forced lane change data of the bus lane to establish an original data set.
[0008] In the present technical solution, the process of collecting the original data set mainly includes the process of obtaining video data through aerial photography and the process of identifying and processing the video data; the object of the aerial photography is a section of a bus-only road; the video data identification and processing mainly rely on existing video recognition software; after completing the aerial photography and identification processing of the video data, the corresponding original data set can be obtained by sorting.
[0009] The present invention is further configured as follows: the preprocessing of the original data set includes: S101, smoothing the trajectory data; S102, performing differential processing on the lateral displacement coordinate sequence of the smoothed trajectory data to obtain a complete lane-changing trajectory and lane-changing duration of the target vehicle; S103, eliminating abnormal lane change data according to the lane change duration and trajectory; S104, extracting the time series of the target vehicle and its surrounding vehicles' features 1.2 seconds before the lane change according to the sliding window method, and fusing them to obtain a feature data set.
[0010] In this technical solution, the trajectory data is first smoothed using a corresponding algorithm, and then the lateral displacement coordinate sequence in the smoothed data is differentially processed. After extracting the corresponding time period data, the lane change trajectory and lane change duration completed by the target vehicle are obtained.
[0011] The present invention is further configured as follows: the improved and constructed CNN-BiLSTM-Attention combined model comprises: S201, adding a CNN layer to the hidden layer of the BiLSTM network, and finally extracting the key features and their spatial relationships in the input sequence through the convolution operation of the convolution kernel in the CNN layer; S202, input the output matrix of the BiLSTM layer into the attention mechanism, obtain the initial state matrix of the Attention layer, and finally obtain the output vector of the Attention layer.
[0012] In this technical solution, a convolutional neural network is used to extract high-dimensional features, reduce redundant information, and capture the spatial relationship between features. At the same time, a long short-term memory neural network is used to extract the key historical long-term and short-term dependencies of all values in the bidirectional time series from the feature matrix output by the CNN layer. The attention mechanism is used to focus on the features of important moments and extract the most critical state information that affects the duration of forced lane changes. A CNN-BiLSTM-Attention combined model is established to predict the duration of forced lane changes of vehicles.
[0013] The present invention is further configured as follows: the screening of the feature data contained in the feature data set includes: S21, based on the random forest feature selection algorithm, the Gini index is used to rank the relative importance of vehicle driving and surrounding environment information features; S22, features with a Gini index greater than 0.0025 are used as input, and lane densities with little impact on lane change duration and relatively low importance are eliminated.
[0014] In the present technical solution, the feature data set includes the driving characteristics of the lane-changing vehicle's longitudinal and lateral speeds, longitudinal and lateral accelerations, and lateral distance from the lane line, as well as the surrounding environmental information characteristics of the speed difference, distance, and lane traffic density with the front and rear vehicles in the current lane and the target lane. In order to avoid the influence of redundant or irrelevant features on the prediction accuracy, the features need to be screened, and the above-mentioned steps S21 and S22 are used to perform data screening and elimination.
[0015] The present invention is further configured as follows: the training and debugging of the combined model includes: The total samples are divided into training set and test set; a CNN-BiLSTM-Attention model is built based on the TensorFlow2.5 deep learning framework; the Adam optimizer is used as the training optimizer, the learning rate is set, the mean absolute error is selected as the training loss function, and the sequence step size is input; the model selects mean absolute error, root mean square error, and efficiency as evaluation indicators of the lane change duration prediction results, and finally obtains the optimal model parameters.
[0016] In this technical solution, model training parameters are set to complete model training and debugging and save the optimal model parameters.
[0017] The present invention is further configured as follows: the prediction of the forced lane change duration according to the trained combined model comprises: The trained combined model is embedded into the dynamic clearing system for bus lane reuse. By calling the data of vehicles in front of the lane and its surrounding environment obtained by the connected bus, the duration of forced lane change is predicted during driving.
[0018] In this technical solution, the above prediction can provide a basis for setting the dynamic clearing box.
[0019] The present invention is further configured such that the output vector Y of the Attention layer is expressed as: e t =tanh(V t A t +b t ) Where: V t is the weight coefficient matrix; b t is the offset; e t State A t The energy value determined.
[0020] In this technical solution, the corresponding weight is assigned according to the importance of each feature in the initial state matrix A, and finally different weights α are t Multiply and accumulate with the corresponding initial state vector to finally get the output vector Y of the Attention layer.
[0021] The present invention is further configured as follows: the hidden layer of the CNN-BiLSTM-Attention combination model includes not only a BiLSTM module, but also a CNN module and an Attention module. The CNN module extracts high-dimensional features, reduces redundant information, and captures the spatial relationship between each feature. The Attention module assigns weights to state feature vectors at different times, focusing on the feature vector at the last moment before the lane change begins, which has the greatest impact on the lane change duration.
[0022] In this technical solution, the CNN module extracts high-dimensional features, reduces redundant information, and captures the spatial relationship between each feature, and includes a one-dimensional convolutional layer; the Attention module assigns weights to the state feature vectors at different times, focusing on the feature vector at the last moment before the lane change begins, which has the greatest impact on the lane change duration, and includes a fully connected layer; the BiLSTM module extracts the key historical long-term and short-term dependencies in the bidirectional time series of the lane change duration and historical state information in the feature matrix output by the CNN layer, and the feature changes of future vehicles and the surrounding environment.
[0023] The present invention is further configured as follows: a sampling frequency for extracting forced lane change trajectory point data from the complete video data is 25 Hz.
[0024] In this technical solution, trajectory points are extracted from the original video at a fixed sampling frequency, and then the original data set is established.
[0025] The present invention can bring the following beneficial effects: The present invention relates to a method for predicting the duration of forced lane changes of connected vehicles in dynamic bus lanes. The method uses the CNN module's ability to extract high-dimensional features and their spatial relationships and the BiLSTM module's ability to memorize and represent temporal feature sequence information to obtain surrounding environment information and driver response features, as well as the temporal features of these information during the lane change process and the spatial relationships between the features. The method also uses the Attention mechanism to focus on important moments of temporal features, thereby achieving accurate prediction of the duration of forced lane changes under actual working conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a schematic diagram of vehicle driving characteristics collected by the present invention.
[0027] Figure 2 It is a schematic diagram of the surrounding environment characteristics of the present invention.
[0028] Figure 3 It is a schematic diagram of the relative importance of lane change duration according to the collected features of the present invention.
[0029] Figure 4 This is a one-dimensional convolutional network structure diagram of the present invention.
[0030] Figure 5 This is the CNN-BiLSTM-Attention network structure diagram of the present invention.
[0031] Figure 6 It is a schematic diagram of dynamic clearing of bus lane reuse according to the present invention. DETAILED DESCRIPTION
[0032] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0033] With the development of social economy and transportation demand, many bus lanes have been built on urban traffic roads. However, with the rapid growth of urban car ownership and transportation demand, traffic congestion has intensified, road resources have become increasingly scarce, and the low utilization rate of some bus lane resources has exacerbated this problem.
[0034] The dynamic bus lane proposed to address this problem is a new type of lane that dynamically provides social vehicles with the right of way to travel on the bus lane while meeting the bus priority requirement.
[0035] However, the traditional dynamic control method of bus lanes often ensures bus priority by setting a long and fixed clearing distance in front of the bus, and a large part of road resources are still wasted. With the rapid development of communication technology and vehicle networking technology, intelligent connected vehicles can obtain sensor data from other vehicles, providing technical support for the refined control of dynamic bus lanes. By establishing a dynamic clearing frame model that changes with the bus speed and the duration of the forced lane change of the cleared vehicle in front of the bus, the utilization rate of bus lane resources can be further improved.
[0036] The duration of vehicle forced lane change is an important part of the dynamic clearing frame setting of dynamic bus lanes. Its accurate prediction is not only crucial to improving the utilization rate of dynamic bus lanes, but also effectively improves the safety of collaborative lane changing of intelligent connected vehicles.
[0037] At present, the prediction methods of lane change duration are mainly divided into trajectory modeling method and data-driven method.
[0038] The trajectory modeling method can be used to predict the lane-changing crossing moment, but existing research only focuses on the driving characteristics of the lane-changing vehicle itself, and does not consider the influence of factors such as surrounding vehicles and the driver's reaction in a forced environment during the actual lane-changing process. Secondly, most of these models are based on multiple assumptions, and it is difficult to accurately predict the lane-changing duration during actual driving, and the robustness is also poor.
[0039] The data-driven method can establish a lane change duration prediction model based only on the mapping relationship between system input and output, without the need to establish specific models for different driving conditions. However, most of the existing data-driven methods for lane change duration prediction rely on the vehicle state information at a certain moment during the lane change process, without considering the dynamic characteristics of the state information during the lane change process, and also ignore the spatial relationship between the various features. The prediction accuracy of the model still needs to be improved.
[0040] Example 1 In view of the above technical problems, this embodiment proposes a method for predicting the duration of forced lane change of connected vehicles in dynamic bus lanes, referring to Figures 1 to 6 , which mainly includes the following steps.
[0041] Step S1, firstly, the original data set of forced lane change in the bus lane is collected, and after the collection is completed, the original data set is preprocessed to obtain a feature data set.
[0042] As for step S1, it mainly includes an acquisition step and a preprocessing step.
[0043] The acquisition step, i.e., the step of acquiring the original data set, specifically includes: Step S11, by taking aerial photos of the bus lane, complete video data of the vehicle that is forced to change lanes to avoid the bus and the driving process of the surrounding vehicles on the bus lane are obtained.
[0044] Step S12, based on the complete video data obtained above, extract the forced lane change trajectory point data therefrom to construct the forced lane change data of the bus lane, and finally construct the original data set.
[0045] In the above technical solution, the process of collecting the original data set mainly includes the process of obtaining video data through aerial photography and the process of identifying and processing the video data; the object of the aerial photography is a section of a bus-only road; the video data identification and processing mainly rely on existing video recognition software; after completing the aerial photography and identification processing of the video data, the corresponding original data set can be obtained by sorting.
[0046] In the above technical solution, a drone is specifically used to take aerial photos of the bus-only road section, and then complete video data of the vehicle that is forced to change lanes to avoid the bus and the driving process of the surrounding vehicles on the bus-only road are collected.
[0047] In the above technical solution, video recognition software is specifically used to extract the forced lane change trajectory point data. More comprehensively and specifically, in this embodiment, video recognition software is used to extract the forced lane change trajectory point data from the original video at a corresponding acquisition frequency to obtain the corresponding bus lane forced lane change data.
[0048] In the embodiment of the above technical solution, the sampling frequency of extracting the forced lane change trajectory point data from the complete video data is 25 Hz. The trajectory points are extracted from the original video at a fixed sampling frequency to establish the original data set.
[0049] The lane change duration prediction model requires a large amount of high-quality data for training to ensure good model prediction performance. The road conditions in the cities where the bus lanes are located are complex and the traffic flow fluctuates widely. At the same time, when the connected vehicle driver receives the prompt to enter the dynamic clear box, he is in a forced lane change state, which has obvious characteristic differences from the general free lane change. The existing dataset cannot reflect the above characteristics. Therefore, it is necessary to collect the forced lane change data of the bus lane to establish the original dataset.
[0050] The preprocessing step, that is, the process of preprocessing the original data set, includes the following contents.
[0051] Step S101, first, smoothing the trajectory data; using a corresponding filtering algorithm to perform smoothing.
[0052] Step S102: Subsequently, a differential process is performed on the lateral displacement coordinate sequence of the smoothed trajectory data to obtain a complete lane-changing trajectory and lane-changing duration of the target vehicle.
[0053] Step S103, then, according to the lane change duration and trajectory, the data of abnormal lane changes are eliminated; in this embodiment, the data of abnormal lane changes in step S103 mainly include data of continuous lane changes in a short period of time, lane change failures, and lane changes that violate traffic regulations.
[0054] Step S104: Finally, the time series of the target vehicle and its surrounding vehicles' features 1.2 seconds before the lane change are extracted according to the sliding window method, and the feature data set is obtained by fusion.
[0055] Specifically, the time series of the target vehicle and its surrounding vehicle features extracted by the sliding window method can be specifically referred to Figure 1 , Figure 2 and Table 1, Table 1 Characteristics of the target vehicle and its surrounding vehicles
[0056] In the above technical solution, the trajectory data is first smoothed using a corresponding algorithm, and then the lateral displacement coordinate sequence in the smoothed data is differentially processed. After extracting the corresponding time period data, the lane change trajectory and lane change duration completed by the target vehicle are obtained.
[0057] In the embodiment of the above technical solution, a symmetric exponential moving average filter algorithm (sEMA) is specifically used to smooth the trajectory data.
[0058] In the embodiment of the above technical solution, for step S102, in more detail, the lateral displacement coordinate sequence in the smoothed lane-changing vehicle trajectory data is differentially processed, and the lane-changing start and end times are extracted according to the characteristics of the continuous change of the lateral coordinates during the lane-changing process to obtain the complete lane-changing trajectory and lane-changing duration of the target vehicle.
[0059] Since the original data set of forced lane changes in bus lanes is automatically obtained based on video recognition software, there are generally system errors and detection errors, so it needs to be processed through a filtering algorithm. At the same time, some lane change data contain abnormal behaviors of illegal driving, which will definitely lead to abnormal recognition of the lane change process, so they are eliminated. After elimination, the driving characteristics and surrounding environment characteristic data are calculated and obtained, and the above preprocessing is performed on the original data set to obtain the final characteristic data set.
[0060] Step S2, screening the feature data contained in the feature data set, improving and constructing a CNN-BiLSTM-Attention combination model.
[0061] Step S2 mainly includes a data screening process and a model improvement process.
[0062] For the data screening process, refer to Figure 3 , which mainly includes the following steps.
[0063] Step S21, based on the feature selection algorithm of random forest, the relative importance of the vehicle driving and surrounding environment information features is sorted by Gini index. The sorting results can be referred to Figure 3 .
[0064] Step S22, after completing step S21, features with a Gini index greater than 0.0025 are used as input, and lane densities with little impact on the lane change duration and relatively low importance are eliminated.
[0065] In the present technical solution, the feature data set includes the driving characteristics of the lane-changing vehicle's longitudinal and lateral speeds, longitudinal and lateral accelerations, and lateral distance from the lane line, as well as the surrounding environmental information characteristics of the speed difference, distance, and lane traffic density with the front and rear vehicles in the current lane and the target lane. In order to avoid the influence of redundant or irrelevant features on the prediction accuracy, the features need to be screened, and the above-mentioned steps S21 and S22 are used to perform data screening and elimination.
[0066] The model improvement process mainly includes the following two steps.
[0067] Step S201, adding a CNN layer to the hidden layer of the BiLSTM network, through the convolution operation of the convolution kernel in the CNN layer, specifically, the feature sequence X={x t-S+1 , x t-(S-1)+1 , …, x t}, reconstruct the input sequence X to obtain a new feature sequence X out , extract key features and their spatial relationships in the input sequence, extract key features and their spatial relationships in the input sequence; for specific network results, please refer to Figure 4 .
[0068] Step S202: Input the output matrix of the BiLSTM layer into the attention mechanism to obtain the initial state matrix of the Attention layer, and finally obtain the output vector of the Attention layer. Specifically, the output matrix of the original BiLSTM layer H = [h1, h2, h3, ..., h t ] is input into the attention mechanism to obtain the initial state matrix A of the Attention layer = [a1, a2, a3, ...at ] Assign corresponding weights according to the importance of each feature in A, and finally assign different weights α t Multiply and accumulate with the corresponding initial state vector to finally get the output vector Y of the Attention layer.
[0069] Among them, the calculation process of the output vector Y of the Attention layer is expressed as: e t =tanh(V t A t +b t ) Where: V t is the weight coefficient matrix; b t is the offset; e t State A tThe energy value determined. According to the importance of each feature in the initial state matrix A, the corresponding weight is assigned, and finally the different weights α t Multiply and accumulate with the corresponding initial state vector to finally get the output vector Y of the Attention layer.
[0070] In this technical solution, the vehicle driving characteristics, surrounding environment information characteristics and lane change duration are highly complex and nonlinear. It is difficult to achieve fast and accurate prediction results using traditional prediction methods. Therefore, a convolutional neural network (CNN) is used to extract high-dimensional features, reduce redundant information, and capture the spatial relationship between each feature. At the same time, a long short-term memory neural network (BiLSTM) is used to extract the key historical long-term and short-term dependencies of all values in the bidirectional time series of the feature matrix output by the CNN layer, and an attention mechanism (Attention) is used to focus on the features of important moments and extract the most critical state information that affects the duration of forced lane changes. A CNN-BiLSTM-Attention combined model is established to predict the duration of forced lane changes of vehicles.
[0071] refer to Figure 5 , the hidden layer of CNN-BiLSTM-Attention consists of three main parts: CNN module, BiLSTM module and Attention module. Among them, CNN is used to extract high-dimensional features, reduce redundant information and capture the spatial relationship between features, including a one-dimensional convolution layer; BiLSTM extracts the key historical long-term and short-term dependencies in the bidirectional time series of the lane change duration and historical state information in the feature matrix output by the CNN layer, the feature changes of future vehicles and the surrounding environment; Attention assigns weights to the state feature vectors at different times, focusing on the feature vector at the last moment before the lane change begins, which has the greatest impact on the lane change duration, and includes a fully connected layer.
[0072] Step S3, training and debugging the combined model to obtain optimal combined model parameters; predicting the forced lane change duration based on the trained combined model.
[0073] As for step S3, it mainly includes a training process and a prediction process.
[0074] The training process mainly includes the following steps: The total samples are divided into training set and test set.
[0075] Build the CNN-BiLSTM-Attention model based on the TensorFlow2.5 deep learning framework.
[0076] The Adam optimizer is used as the training optimizer, the learning rate is set, the mean absolute error is selected as the training loss function, and the sequence step size is input; the model selects the mean absolute error, root mean square error, and efficiency as the evaluation indicators of the lane change duration prediction results, and finally the optimal model parameters are obtained.
[0077] In more detail, in this embodiment, the parameters used in the training process are set as follows: the total 12260 samples are randomly divided into training sets and test sets according to the ratio of 8:2, that is, the number of samples in the training set is 9808, and the number of samples in the test set is 2452. The CNN-BiLSTM-Attention combination model is built based on the TensorFlow2.5 deep learning framework. The Adam optimizer is used as the optimizer for training, and the learning rate is set to 0.001, the mean absolute error MAE is selected as the training loss function, and the input sequence step size is set to 30. The model selects the mean absolute error (MAE), the root mean square error (RMSE), and the efficiency (the proportion of samples in the test set with an absolute error of less than 1s between the predicted value and the true value to the total samples) as evaluation indicators for the lane change duration prediction results. The performance of the model is comprehensively evaluated in this way, and the optimal model parameters are saved.
[0078] The comparison between the model input features and the prediction results is shown in Table 2. Compared with the traditional driving feature moment state input, the introduction of surrounding environment features and time series state features significantly improves the model prediction effect. The response time of each model is shown in Table 3. The response time of a single sample is much less than 0.1 seconds, ensuring the prediction accuracy while meeting the real-time prediction requirements: Table 2 Comparison of model input features and prediction results Input Features MAE RMSE Efficiency Lane-changing vehicle driving + surrounding environment time series state characteristics 0.546 0.707 90.1％ Characteristics of driving sequence state of lane-changing vehicles 0.654 0.885 80.2％ Lane-changing vehicle driving + surrounding environment status characteristics at all times 0.695 0.881 75.0％ Table 3 Comparison of response time of each model Model Response time / s LSTM 0.460 BiLSTM 0.466 CNN-BiLSTM 0.464 CNN-BiLSTM-Attention 0.545 refer to Figure 6 , for the prediction process, it mainly includes the following steps: The trained combined model is embedded into the dynamic clearing system for bus lane reuse. By calling the data of vehicles in front of the lane and its surrounding environment obtained by the connected bus, the duration of forced lane change is predicted during driving. In this technical solution, the above prediction can provide a basis for the setting of the dynamic clearing box.
[0079] For a method for predicting the duration of forced lane changes of connected vehicles in a dynamic bus lane of the present embodiment, first, an original data set of forced lane changes in the bus lane is established, and then the original data set is preprocessed, and the features contained in the data set are screened based on the preprocessing; then, the structure of the traditional BiLSTM network is optimized, and the convolutional neural network and the attention mechanism are integrated; then, the model is trained using the forced lane change data set of the bus lane after extracting key features, and the trained model is embedded in the dynamic clearing system for bus lane reuse, and the data of the vehicle in front of the lane and its surrounding environment obtained by calling the connected bus is finally used to predict the duration of the lane change, providing a basis for setting the dynamic clearing box.
[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the duration of a forced lane change of a connected vehicle in a dynamic bus lane, characterized in that: The following steps are involved: S1, collecting the original data set of forced lane change in bus lanes, and preprocessing the original data set to obtain a feature data set; S2, screening the feature data contained in the feature data set, improving and constructing a CNN-BiLSTM-Attention combination model; S3, training and debugging the combined model to obtain optimal combined model parameters; predicting the forced lane change duration based on the trained combined model.
2. The method for predicting the duration of a forced lane change of a connected vehicle in a dynamic bus lane according to claim 1, characterized in that: The process of obtaining the original data set is as follows: S11, taking aerial photos of the bus lane to obtain complete video data of the driving process of the vehicle forced to change lanes to avoid the bus and the surrounding vehicles on the bus lane; S12, extracting the forced lane change trajectory point data from the complete video data, obtaining the forced lane change data of the bus lane to establish an original data set.
3. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 1 or 2, characterized in that: The preprocessing of the original data set includes: S101, smoothing the trajectory data; S102, performing differential processing on the lateral displacement coordinate sequence of the smoothed trajectory data to obtain a complete lane-changing trajectory and lane-changing duration of the target vehicle; S103, eliminating abnormal lane change data according to the lane change duration and trajectory; S104, extracting the time series of the target vehicle and its surrounding vehicles' features 1.2 seconds before the lane change according to the sliding window method, and fusing them to obtain a feature data set.
4. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 3, characterized in that: The improved and constructed CNN-BiLSTM-Attention combined model includes: S201, adding a CNN layer to the hidden layer of the BiLSTM network, and finally extracting the key features and their spatial relationships in the input sequence through the convolution operation of the convolution kernel in the CNN layer; S202, input the output matrix of the BiLSTM layer into the attention mechanism, obtain the initial state matrix of the Attention layer, and finally obtain the output vector of the Attention layer.
5. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 1, 2 or 4, characterized in that: The screening of the feature data contained in the feature data set includes: S21, based on the random forest feature selection algorithm, the Gini index is used to rank the relative importance of vehicle driving and surrounding environment information features; S22, features with a Gini index greater than 0.0025 are used as input, and lane densities with little impact on lane change duration and relatively low importance are eliminated.
6. The method for predicting the duration of a forced lane change of a connected vehicle in a dynamic bus lane according to claim 5, characterized in that: The training and debugging of the combined model includes: The total samples are divided into training set and test set; a CNN-BiLSTM-Attention model is built based on the TensorFlow2.5 deep learning framework; the Adam optimizer is used as the training optimizer, the learning rate is set, the mean absolute error is selected as the training loss function, and the sequence step size is input; the model selects mean absolute error, root mean square error, and efficiency as evaluation indicators of the lane change duration prediction results, and finally obtains the optimal model parameters.
7. The method for predicting the duration of a forced lane change of a connected vehicle in a dynamic bus lane according to claim 1 or 6, characterized in that: The predicting of the forced lane change duration according to the trained combined model includes: The trained combined model is embedded into the dynamic clearing system for bus lane reuse. By calling the data of vehicles in front of the lane and its surrounding environment obtained by the connected bus, the duration of forced lane change is predicted during driving.
8. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 4, characterized in that: The output vector Y of the Attention layer is expressed as: e t =tanh(V t A t +b t ) Where: V t is the weight coefficient matrix; b t is the offset; e t State A t The energy value determined.
9. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 1, 2 or 4, characterized in that: The hidden layer of the CNN-BiLSTM-Attention combination model includes not only a BiLSTM module, but also a CNN module and an Attention module. The CNN module extracts high-dimensional features, reduces redundant information, and captures the spatial relationship between each feature. The Attention module assigns weights to state feature vectors at different times, focusing on the feature vector at the last moment before the lane change begins, which has the greatest impact on the lane change duration.
10. The method for predicting the duration of a forced lane change of a networked vehicle in a dynamic bus lane according to claim 2, characterized in that: The sampling frequency of extracting the forced lane change trajectory point data from the complete video data is 25 Hz.