CN114612867A - BiLSTM-CRF model-based vehicle lane change intention prediction method - Google Patents

BiLSTM-CRF model-based vehicle lane change intention prediction method Download PDF

Info

Publication number
CN114612867A
CN114612867A CN202210058857.3A CN202210058857A CN114612867A CN 114612867 A CN114612867 A CN 114612867A CN 202210058857 A CN202210058857 A CN 202210058857A CN 114612867 A CN114612867 A CN 114612867A
Authority
CN
China
Prior art keywords
vehicle
intention
lane change
lane
change intention
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210058857.3A
Other languages
Chinese (zh)
Inventor
曾宪琳
王凯
方浩
陈仲瑶
窦丽华
杨庆凯
辛斌
陈杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202210058857.3A priority Critical patent/CN114612867A/en
Publication of CN114612867A publication Critical patent/CN114612867A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Traffic Control Systems (AREA)

Abstract

According to the method for predicting the lane-changing intention of the vehicle based on the BilSTM-CRF model, historical traffic data of the vehicle are traversed to obtain historical information of the lane-changing intention of the vehicle; preprocessing historical information of the vehicle lane-changing intention, and extracting vehicle lane-changing intention characteristics; forming a vehicle lane change intention characteristic sequence by the vehicle lane change intention characteristics according to the inducing rules; marking the vehicle lane change intention characteristic sequence to obtain a label of the vehicle lane change intention characteristic; after the label is fuzzified, dividing the lane change intention characteristics of the vehicle into a training set and a test set; training a BiLSTM-CRF model of a vehicle lane change intention prediction model by utilizing a training set of vehicle lane change intention characteristics and labels of the vehicle lane change intention characteristics; training a post-rule layer of a vehicle lane change intention prediction model by using an absolute rule; and testing the vehicle lane change intention prediction model by using the test set of the vehicle lane change intention characteristics. The method can solve the problem that prediction effect and credible interpretability are difficult to be considered simultaneously in prediction of the lane changing intention of the vehicles on the highway.

Description

BiLSTM-CRF model-based vehicle lane change intention prediction method
Technical Field
The invention belongs to the technical field of automatic driving intention prediction, and particularly relates to a method for predicting a lane-changing intention of a vehicle based on a BilSTM-CRF model.
Background
The automatic driving is in a high-speed development stage, and not only makes a major breakthrough in the academic research field continuously, but also more and more practical products are gradually put into use, such as tesla unmanned vehicles, hundred-degree automatic driving taxies and the like. With the increased speed of the commercialization model, the automatic driving is gradually entering the lives of people.
However, this technique is not yet mature. Safe and always is the foundation stone of the automatic driving automobile. Motion prediction for other vehicles may better assist an autonomous vehicle in making decision plans, and most advanced motion prediction methods are intent-based. OliverPentland used the pattern recognition method to realize the classification of driver intentions for the first time in 1996, which was the earliest study on driving behavior. Driver intent prediction has been extensively studied today, and most studies involve the use of heuristic methods or training classifiers. Heuristics may be used to determine what actions the target vehicle will perform in the short future. And the driving intentions are classified in more complex scenes, and discriminant learning algorithms are more popular. Mandalia HM et al propose an intention prediction method based on a support vector machine in the literature (Using support vector machines for lane change detection), which predicts the lane change intention of a target vehicle by Using position information of the vehicle from four different directions of a lane as a feature. Then, with the rise of the neural network, the means of intention prediction is further enriched. Abdelmoudjib Benterki et al in the literature (Prediction of Surrounding vehicle Lane Change Using Machine Learning) proposed a method for detecting motorway Lane Change maneuver Intention, Using the data of the vehicle preparation Lane Change stage to predict Lane Change, and tested two Lane Change Prediction methods of support vector Machine and artificial neural network on the public data set, verified the performance of the proposed method. Information cannot be transmitted between the neurons in the same layer of the neural network, which means that the interaction relationship between the neurons in the same layer cannot be fully considered, and the LSTM (long-short memory network) can better capture the information between the sequences. In the literature (Driver Lane Change Recognition of Intelligent Vehicle based on Long Short-Term Memory Network), the differences between the robustness and the prediction accuracy of the Lane Change Intention prediction by using three methods, namely a rule-based model, a support vector machine and an LSTM, are compared, and the LSTM is verified to have the capability of processing longer-time sequences. Another popular method for intent prediction is to use hidden markov equal probability map models, which can output the most probable label sequence corresponding to the observed sequence, with better interpretability.
The driving intention prediction is one of key technologies for developing an advanced driving assistance system, can effectively reduce traffic accidents caused by lane change, and guarantees driving safety. However, the intention prediction algorithm based on machine learning is poor in interpretability, so that drivers cannot trust unexplained prediction results, and people are not trusted; heuristic methods or probabilistic graphical models have strong interpretability but have poor intent prediction effect, while pure learning methods such as LSTM sometimes cause unexpected prediction errors due to their own structural features and without regular constraints, for example, predicting the lane change intent of the vehicle located at the leftmost side as lane change to the left, which further causes the uncertainty of the prediction result. In this way, it is difficult for the conventional intent prediction method to achieve both the prediction effect and the reliable interpretability.
Disclosure of Invention
The invention overcomes one of the defects of the prior art, provides a method for predicting the lane change intention of a vehicle based on a BilSTM-CRF (bidirectional long and short time memory network-conditional random field) model, and can solve the problem that the prediction effect and the credible interpretability are difficult to be considered simultaneously in the prediction of the lane change intention of the vehicle on the expressway.
According to one aspect of the disclosure, the invention provides a method for predicting lane change intention of a vehicle based on a BilSTM-CRF model, comprising the following steps:
traversing historical traffic data of the vehicle to obtain historical information of the lane changing intention of the vehicle;
preprocessing the historical information of the vehicle lane-changing intention, and extracting vehicle lane-changing intention characteristics reflecting the intention change of a driver;
forming the vehicle lane-changing intention characteristics into a vehicle lane-changing intention characteristic sequence according to an inducing rule;
labeling the vehicle lane change intention characteristic sequence to obtain a label of the vehicle lane change intention characteristic;
after the label of the vehicle lane change intention characteristic is subjected to fuzzification processing, dividing the vehicle lane change intention characteristic into a training set and a testing set;
training a BilSTM-CRF model of a vehicle lane change intention prediction model by utilizing the training set of the vehicle lane change intention characteristics and the labels of the vehicle lane change intention characteristics;
training a post-rule layer of the vehicle lane change intention prediction model by using an absolute rule;
and testing the vehicle lane change intention prediction model by using the test set of the vehicle lane change intention characteristics, and inputting the real-time vehicle lane change intention characteristics into the vehicle lane change intention prediction model to predict to obtain the vehicle lane change intention.
In one possible implementation, the history information H of the lane change intention of the vehicle includes: the distance X of the vehicle relative to the leftmost lane, the distance Y of the vehicle entering the starting acquisition point interface, and the vehicle's intent to maneuver.
In one possible implementation, the vehicle lane change intention feature reflecting the change of the driver intention includes: the sum of the squares of the transverse speeds of the vehicles in a preset time length, the sum of the turning angles and the left and right displacement times of the vehicles are calculated;
the sum of the squares of the lateral speeds and the sum of the turning angles is a basic characteristic of the lane change intention characteristic of the vehicle, and the number of times of the left and right displacement of the vehicle is a regular characteristic of the lane change intention characteristic of the vehicle.
In one possible implementation, the fuzzifying the tag of the lane change intention characteristic of the vehicle includes:
counting the label numerical values of the vehicle lane changing intention characteristics, and dividing N intervals according to the label numerical values of the vehicle lane changing intention characteristics;
judging the section where the label of the lane change intention characteristic of the vehicle is located according to the label numerical value of the lane change intention characteristic of the vehicle and the length of each section;
and taking the lower boundary of the interval as the value of the label fuzzified characteristic of the lane change intention of the vehicle.
In one possible implementation, the lane change intention characteristics of the vehicle are divided into two categories, namely basic characteristics and regular characteristics. The labels of the basic characteristics comprise four categories of left lane changing, right lane changing, straight line keeping or left lane changing, straight line keeping or right lane changing, and the labels of the rule characteristics are the compliance degree of the rules. .
In one possible implementation, the vehicle lane change intention prediction model comprises two BilSTM-CRF models, wherein the BilSTM-CRF models are divided into a BilSTM layer and a CRF layer;
wherein the BilSTM layer is used for outputting the score of each label of each lane change intention characteristic of the vehicle;
and the CRF layer is used for calculating the transition probability among the labels of the lane change intention characteristics of the vehicle.
In one possible implementation, one of the BilSTM-CRF models is used to predict a vehicle changing lanes to the left and keeping straight or changing lanes to the right, and the other BilSTM-CRF model is used to predict a vehicle changing lanes to the right and keeping straight or changing lanes to the left.
In one possible implementation, the absolute rules include traffic rules and physical constraints.
In one possible implementation, the sum of the squares of the lateral velocities
Figure BDA0003473608410000041
Sum of said yaw angles
Figure BDA0003473608410000042
In the formula, N is a positive integer and is the number of image frames of the target vehicle within the preset time;
VXis divided into VX_LAnd VX_RAnd represent the left lateral velocity and the right lateral velocity, respectively. Subtracting the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame from the distance X of the target vehicle relative to the leftmost lane in the (i) th frame, and if the obtained difference is positive, recording as VX_L(ii) a Otherwise, it is marked as VX_R；
VYSubtracting the distance Y of the target vehicle of the ith frame from the distance Y of the target vehicle of the (i + 1) th frame entering the interface of the initial acquisition point if VYA value of zero indicates that the target vehicle is in a stopped state or that the data set is erroneous.
In one possible implementation, the number of left and right vehicle displacements C is divided into the number of left displacements CLNumber of right-hand displacements CRWhen the distance X coordinate of the target vehicle relative to the leftmost lane in the ith frame is subtracted by the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame, if the obtained difference is positive, the left-direction displacement times CLPlus 1, otherwise the number of right-hand displacements CRAnd adding 1.
According to the method for predicting the lane changing intention of the vehicle based on the BilSTM-CRF model, historical information of the lane changing intention of the vehicle is obtained by traversing historical traffic data of the vehicle; preprocessing the historical information of the vehicle lane-changing intention, and extracting vehicle lane-changing intention characteristics reflecting the intention change of a driver; forming the vehicle lane-changing intention characteristics into a vehicle lane-changing intention characteristic sequence according to an inducing rule; labeling the vehicle lane change intention characteristic sequence to obtain a label of the vehicle lane change intention characteristic; after the label of the vehicle lane changing intention characteristic is subjected to fuzzification processing, dividing the vehicle lane changing intention characteristic into a training set and a testing set; training a BilSTM-CRF model of a vehicle lane change intention prediction model by utilizing the training set of the vehicle lane change intention characteristics and the labels of the vehicle lane change intention characteristics; training a post-rule layer of the vehicle lane change intention prediction model by using an absolute rule; and testing the vehicle lane changing intention prediction model by using the test set of the vehicle lane changing intention characteristics, and inputting the real-time vehicle lane changing intention characteristics into the vehicle lane changing intention prediction model to predict to obtain the vehicle lane changing intention. The method can solve the problem that prediction effect and credible interpretability are difficult to be considered simultaneously in prediction of the lane changing intention of the vehicles on the highway.
Drawings
The accompanying drawings are included to provide a further understanding of the technology or prior art of the present application and are incorporated in and constitute a part of this specification. The drawings expressing the embodiments of the present application are used for explaining the technical solutions of the present application, and should not be construed as limiting the technical solutions of the present application.
FIG. 1 illustrates a flow chart of a method for predicting lane change intention of a vehicle based on a BilSTM-CRF model according to an embodiment of the disclosure;
FIG. 2 illustrates a historical information collection schematic of a vehicle lane change intention, according to an embodiment of the present disclosure;
FIG. 3 illustrates a schematic diagram of a vehicle lane change intention prediction model according to an embodiment of the present disclosure;
FIG. 4 shows a schematic diagram of LSTM emission probability according to an embodiment of the present disclosure;
FIG. 5 shows a schematic diagram of a CRF transition probability matrix according to an embodiment of the present disclosure;
FIG. 6 is a diagram illustrating the predicted effect of a vehicle lane change intention prediction model according to an embodiment of the present disclosure.
Detailed Description
The following detailed description of the embodiments of the present invention will be provided with reference to the accompanying drawings and examples, so that how to apply the technical means to solve the technical problems and achieve the corresponding technical effects can be fully understood and implemented. The embodiments and the features of the embodiments can be combined without conflict, and the technical solutions formed are all within the scope of the present invention.
Additionally, the steps illustrated in the flow charts of the figures may be performed in a computer such as a set of computer-executable instructions. Also, while a logical order is shown in the flow diagrams, in some cases, the steps shown or described may be performed in an order different than here.
FIG. 1 shows a flow chart of a method for predicting lane change intention of a vehicle based on a BilSTM-CRF model according to an embodiment of the disclosure. As shown in fig. 1, the method may include:
step S1: and traversing historical traffic data of the vehicle to obtain historical information of the lane changing intention of the vehicle.
The BiLSTM-CRF model main body consists of a bidirectional long-short time memory network (Bi-LSTM) and a Conditional Random Field (CRF), the input of the model is a feature sequence, the output of the model is a prediction label corresponding to each feature, and the model is a bidirectional long-short time memory network-conditional random field model.
The history information H of the vehicle lane change intention may include: the distance (local X coordinate) of the vehicle relative to the leftmost lane, the distance (local Y coordinate) of the vehicle relative to the section of the entering starting acquisition point, the motor intention (manually marked according to the change of the lane where the vehicle is located), the vehicle identification number, the frame where the data is located, the vehicle instantaneous speed, the vehicle instantaneous acceleration, the current lane position (the 1 st lane is the leftmost lane), the vehicle driving direction, the global time, the vehicle identification number of the vehicle in front of the same lane, the vehicle identification number of the vehicle behind the same lane, the headway time of the vehicle in front, and the like.
In one example, the historical traffic data for the vehicle may be from the real traffic data set NGSIM (Next Generation Simulation), the selected road segment is an I-80 highway (U.S. san Francisco No. 80 interstate highway), and the time periods taken are 4:00-4:15 in the afternoon and 5:00-5:30 in the afternoon.
FIG. 2 illustrates a historical traffic data collection schematic of a vehicle lane change intent in accordance with an embodiment of the present disclosure.
And traversing the historical traffic data set to find a track containing the lane change intention of the vehicle. Wherein the vehicle lane change is defined as the target vehicle traveling from a current lane to an adjacent lane without collision. When the track of a certain vehicle is positioned to intersect with the lane line, historical information H of 45 frames (0.1 s of time per frame) before the intersection point is collected. Note that the intersection is G, and note that the frame where the intersection G is located is the 45 th frame, and the frame where the target vehicle is located 4.5s before the intersection with the lane is the first frame (see fig. 2).
Step S2: and preprocessing the historical information of the lane-changing intention of the vehicle, and extracting the features of the lane-changing intention of the vehicle reflecting the intention change of the driver.
Wherein, vehicle lane change intention characteristic includes: the sum of the squares of the lateral speeds of the vehicle within a preset time length, the sum of the turning angles, and the number of times of the left and right displacement of the vehicle.
Sum of squares of lateral velocities
Figure BDA0003473608410000071
Sum of yaw angles
Figure BDA0003473608410000072
In the formula, N is a positive integer and is the number of image frames of the target vehicle within the preset time;
VXis divided into VX_LAnd VX_RAnd represent the left lateral velocity and the right lateral velocity, respectively. Subtracting the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame from the distance X of the target vehicle relative to the leftmost lane in the (i) th frame, and if the obtained difference is positive, recording as VX_L(ii) a Otherwise, it is marked as VX_R；
VYSubtracting the distance Y of the target vehicle of the ith frame from the distance Y of the target vehicle of the (i + 1) th frame entering the interface of the initial acquisition point if VYA value of zero indicates that the target vehicle is in a stopped moving state or that the data set is erroneous.
The left and right displacement times C of the vehicle is divided into left displacement times CLNumber of right-hand displacements CRWhen the distance X coordinate of the target vehicle relative to the leftmost lane in the ith frame is subtracted by the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame, if the obtained difference is positive, the left-direction displacement times CLPlus 1, otherwise the number of right-hand displacements CRAnd adding 1.
For example, the preset time length is 1s (10 frames of images), and the sum of squares of the lateral velocities within the time length of 1s (i.e., 10 frames) is:
Figure BDA0003473608410000073
finally, the square sum V of the left lateral velocity in each 10 frames is calculatedSum_Sq_LSum of squares V of right lateral velocitySum_Sq_R。
Sum of yaw angles
Figure BDA0003473608410000074
Finally obtaining the sum A of the left steering angle in each 10 frames through calculationSum_LSum of right steering angle ASum_R。
The square sum of the transverse velocities in different time lengths can also be obtained by the formula
Figure BDA0003473608410000075
And sum of yaw angles
Figure BDA0003473608410000076
Where i is the number of frames of the start image of different preset time lengths.
Taking right turn as an example, the lane change process of the vehicle can be divided into three parts of lane keeping, lane changing and lane change completion. After calculating the above-mentioned features in the lane keeping and lane changing sections, it can be found that important changes that can reflect lane-changing intentions occur on average in the 15 th frame of the taken history sequence, and the data varies depending on the driving habits of the driver. In the actual prediction process, the value of the feature vector is updated with a time step of 0.1 s. In addition, the sequence length for prediction is 10 frames in the practical example, so that the trained model is sensitive to lane change of the vehicle and can quickly capture lane change behaviors.
Step S3: and forming the vehicle lane-changing intention characteristic sequence by using the vehicle lane-changing intention characteristics according to the induction rule. The rules are divided into absolute rules and inductive rules. The induced rule is different from the absolute rule in that the induced rule does not absolutely hold, but the vehicle lane change intention prediction model can be pushed in an induced mode to make a prediction conforming to the rule. For example, the inducing rule may be that, within a preset time interval, the target vehicle continues to move laterally and has a greater lateral speed, and the vehicle is considered to have a greater probability of changing lanes. The specific mode of action of the inducing rule is through transition probabilities between the corresponding labels of the rule features and the basic features.
Step S4: and marking the vehicle lane change intention characteristic sequence to obtain a label of the vehicle lane change intention characteristic. The lane change intention characteristics of the vehicle are divided into two types of basic characteristics and regular characteristics, labels of the basic characteristics comprise four types of lane change to the left, lane change to the right, lane change to keep straight or left, lane change to keep straight or right, and labels of the regular characteristics are the degree of compliance to the inductive rules under the values of the current basic characteristics and the regular characteristics, and the value range is [0,1 ].
Step S5: after the label of the vehicle lane change intention characteristic is subjected to fuzzification processing, the vehicle lane change intention characteristic is divided into a training set and a testing set.
In one example, the fuzzifying the tag of the lane change intention characteristic of the vehicle may include:
counting the label numerical values of the vehicle lane changing intention characteristics, and dividing N intervals according to the label numerical values of the vehicle lane changing intention characteristics;
judging the section where the label of the lane change intention characteristic of the vehicle is located according to the label numerical value of the lane change intention characteristic of the vehicle and the length of each section;
and taking the lower boundary of the interval as the value of the label fuzzified characteristic of the lane change intention of the vehicle.
Sum of squares V at left lateral velocitySum_Sq_LFor example, the sum of squares V of the lateral velocities in the left direction is countedSum_Sq_LAnd then carrying out interval division on the tag values lambda, wherein the more densely the tag values lambda are distributed, the smaller the length of the subinterval division in the region is, and the larger the length of the subinterval division is otherwise. Such as the sum of the squares V of the lateral velocities in the left directionSum_Sq_LAnd if the distribution of the tag values lambda is dense in the range of 0.01-1, the distribution interval of the tag values lambda is divided into 100 subintervals, and the length of the subintervals is 0.01. After fuzzification VSum_Sq_LIs taken to be the lower bound of the interval in which it lies, e.g. the sum of the squares V of the lateral velocities in the left directionSum_Sq_L0.1753 in subinterval [0.17, 0.18), the blurred value is the sum of the squares V of the lateral velocities to the leftSum_Sq_LThe value was 0.17. After the above treatment was completed, the ratio of 0.7: a ratio of 0.3 divides the training set and the test set.
Step S6: and training a BilSTM-CRF model of the vehicle lane-changing intention prediction model by utilizing the training set of the vehicle lane-changing intention characteristics and the labels of the vehicle lane-changing intention characteristics.
FIG. 3 shows a schematic diagram of a vehicle lane change intention prediction model according to an embodiment of the present disclosure.
As shown in FIG. 3, the vehicle lane change intention prediction model is divided into a BilSTM-CRF model and a post-rule layer. The BiLSTM-CRF model is divided into a BiLSTM layer and a CRF layer; wherein the BilSTM layer is used for outputting the score of each label of each lane change intention characteristic of the vehicle; the CRF layer is used for calculating transition probabilities among the labels of the lane change intention characteristics of the vehicle.
FIG. 4 shows a schematic diagram of LSTM emission probability according to an embodiment of the present disclosure; fig. 5 shows a CRF transition probability matrix diagram according to an embodiment of the present disclosure.
As shown in fig. 4, the LSTM network may output a score for each vehicle lane change intention feature on each tag, which is expressed in terms of a transmission probability. The BilSTM adds inverse operation on the basis of the LSTM network, and the final output is the simple superposition of the outputs of the forward LSTM and the inverse LSTM, so that the information between sequence contexts can be better considered. For example, the number of hidden layers of the BilSTM model is 20, and the most possible labels of each vehicle lane-changing intention characteristic are finally output and are independent of each other. After adding the CRF layer, as shown in fig. 5, the CRF layer can calculate the transition probability between labels, and the BiLSTM-CRF model outputs the optimal label sequence, not an independent label. In fig. 4 and 5, [0,0.5,1] indicates the degree of compliance with the induction rule, and [ L, S _ R ] indicates a vehicle intention label.
The BilSTM-CRF model can also score and identify the tag sequences to judge whether a certain tag sequence is the optimal tag sequence, and the scoring formula is as follows:
Score＝EmissionScore+TransitionScore
wherein EmissionScore is the score of the BilSTM model on the tag sequence, and TransitionScore is the sum of the corresponding state transition probabilities between the elements of the tag sequence.
The BilSTM-CRF model trained using the training set of vehicle lane-change intention features and the labels of the vehicle lane-change intention features enables the correct label sequence to be the highest of the scores in all possible sequences.
In one example, one BilSTM-CRF model is used to predict that a vehicle will lane left and stay straight or will lane right, and another BilSTM-CRF model is used to predict that a vehicle will lane right and stay straight or will lane left.
For example, two BilSTM-CRF models are respectively marked as M1And M2，M1For predicting lane changes to the left and not to the left (left-right and straight), M2For predicting right and non-right lane changes (straight and left lane changes are maintained). Feature sequence [ C ] for pairing target vehicles in predictionL,VSum_Sq_L，ASum_L]And [ CR,VSum_Sq_R，ASum_R]Separately input model M1And M2Then, the predicted results of the two models are compared to decideAnd (6) performing next action. Such as when M1Output lane left change and M2When the non-rightward lane change is output, the prediction result is the leftward lane change; when M is1Output lane left change and M2When the lane is changed to the right, the prediction result is uncertain, and the prediction result needs to be recalculated according to the characteristic sequence of the next time step. Thus model M1And M2The mutual restriction is realized, and the prediction accuracy is improved on the premise of sacrificing the prediction time.
Step S7: and training a post-rule layer of the vehicle lane change intention prediction model by using an absolute rule.
The absolute rule is a rule which must be observed under normal conditions, and can include traffic rules, physical constraints, turn signal and other rules, and a correction is made on the prediction result of the BilSTM-CRF model. For example, if the target vehicle is located in the leftmost lane or the rightmost lane, the vehicle cannot turn left or right any more. If the target vehicle is at rest, the vehicle is deemed not to have completed the lane change within 3s of the future.
Step S8: and testing a vehicle lane change intention prediction model by using a test set of vehicle lane change intention characteristics, and inputting the real-time vehicle lane change intention characteristics into the vehicle lane change intention prediction model to predict to obtain the vehicle lane change intention.
FIG. 6 is a diagram illustrating the predicted effect of a vehicle lane change intention prediction model according to an embodiment of the present disclosure. As shown in FIG. 6, parameters such as accuracy, precision, recall and F1 score can be used to measure the prediction effect of the training vehicle lane change intention prediction model.
The method for predicting the lane-changing intention of the vehicle based on the BiLSTM-CRF model is different from the traditional LSTM method, and the CRF layer enables the method to process the transfer relation among the characteristic sequence labels, so that the transfer probability matrix can be obtained. The interactive relationship among the features in the vehicle lane-changing intention feature sequence can be fully utilized, and the degree of adherence to the inductive rule under the current basic feature and rule feature values can be obtained as the labeling result of the rule features during prediction. The higher the rule adherence degree is, the greater the transition probability is in the intention prediction result corresponding to the rule, namely, the directional constraint is realized when the vehicle maneuvering intention is predicted through the transition probability matrix, and the prediction accuracy is improved. By such a method, the invention embeds the rules into the intention prediction model, and improves the interpretability of the prediction. And externally connecting a post-rule layer behind the BilSTM-CRF model. By combining with historical information, the post-rule layer can correct errors of the prediction result of the BilSTM-CRF model, eliminate errors of a machine by using traffic rules or human experience, and further improve the credibility of the prediction result.
Although the embodiments of the present invention have been described above, the above descriptions are only for the convenience of understanding the present invention, and are not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (10)

1. A method for predicting lane change intention of a vehicle based on a BilSTM-CRF model is characterized by comprising the following steps:
traversing historical traffic data of the vehicle to obtain historical information of the lane changing intention of the vehicle;
preprocessing the historical information of the vehicle lane-changing intention, and extracting vehicle lane-changing intention characteristics reflecting the intention change of a driver;
forming the vehicle lane-changing intention characteristics into a vehicle lane-changing intention characteristic sequence according to an inducing rule;
labeling the vehicle lane change intention characteristic sequence to obtain a label of the vehicle lane change intention characteristic;
after the label of the vehicle lane change intention characteristic is subjected to fuzzification processing, dividing the vehicle lane change intention characteristic into a training set and a testing set;
training a BilSTM-CRF model of a vehicle lane changing intention prediction model by utilizing the training set of the vehicle lane changing intention characteristics and the labels of the vehicle lane changing intention characteristics;
training a post-rule layer of the vehicle lane change intention prediction model by using an absolute rule;
and testing the vehicle lane change intention prediction model by using the test set of the vehicle lane change intention characteristics, and inputting the real-time vehicle lane change intention characteristics into the vehicle lane change intention prediction model to predict to obtain the vehicle lane change intention.
2. The method of vehicle lane change intention prediction according to claim 1, characterized in that the history information H of the vehicle lane change intention includes: the distance X of the vehicle from the leftmost lane, the distance Y of the vehicle entering the starting acquisition point interface, and the vehicle's intent to maneuver.
3. The method of vehicle lane change intention prediction according to claim 2, wherein the vehicle lane change intention feature reflecting the change in driver intention comprises: the sum of the squares of the transverse speeds of the vehicles in a preset time length, the sum of the turning angles and the left and right displacement times of the vehicles are calculated;
the sum of the squares of the lateral speeds and the sum of the turning angles is a basic characteristic of the lane change intention characteristic of the vehicle, and the number of times of the left and right displacement of the vehicle is a regular characteristic of the lane change intention characteristic of the vehicle.
4. The method of predicting the lane change intention of a vehicle of claim 1, wherein the fuzzifying the label of the lane change intention feature of the vehicle comprises:
counting the label numerical values of the vehicle lane changing intention characteristics, and dividing N intervals according to the label numerical values of the vehicle lane changing intention characteristics;
judging the section where the label of the lane change intention characteristic of the vehicle is located according to the label numerical value of the lane change intention characteristic of the vehicle and the length of each section;
and taking the lower boundary of the interval as the value of the label fuzzified characteristic of the lane change intention of the vehicle.
5. The method of predicting the lane change intention of a vehicle of claim 1, wherein the lane change intention characteristics of the vehicle are classified into two categories, basic characteristics and regular characteristics. The labels of the basic characteristics comprise four categories of left lane changing, right lane changing, straight line keeping or left lane changing, straight line keeping or right lane changing, and the labels of the rule characteristics are the compliance degree of the rules.
6. The method of claim 5, wherein the vehicle lane change intention prediction model comprises two BilSTM-CRF models, the BilSTM-CRF models being divided into a BilSTM layer and a CRF layer;
wherein the BilSTM layer is used for outputting the score of each label of each lane change intention characteristic of the vehicle;
and the CRF layer is used for calculating the transition probability among the labels of the lane change intention characteristics of the vehicle.
7. The method of claim 5, wherein one of the BilSTM-CRF models is used to predict that the vehicle will change lane to the left and remain straight or change lane to the right, and another of the BilSTM-CRF models is used to predict that the vehicle will change lane to the right and remain straight or change lane to the left.
8. The method of vehicle lane change intention prediction according to claim 1, wherein the absolute rules include traffic rules and physical constraints.
9. The method of vehicle lane change intention prediction according to claim 3, wherein the sum of squares of the lateral velocities
Figure FDA0003473608400000021
Sum of said yaw angles
Figure FDA0003473608400000022
In the formula, N is a positive integer, is the number of image frames of a target vehicle in a preset time, and i is the number of initial image frames in different preset time lengths;
VXis divided into VX_LAnd VX_RAnd represent the left lateral velocity and the right lateral velocity, respectively. Subtracting the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame from the distance X of the target vehicle relative to the leftmost lane in the (i) th frame, and if the obtained difference is positive, recording as VX_L(ii) a Otherwise, it is marked as VX_R；
VYSubtracting the distance Y of the target vehicle of the ith frame from the distance Y of the target vehicle of the (i + 1) th frame entering the interface of the initial acquisition point if VYA value of zero indicates that the target vehicle is in a stopped state or that the data set is erroneous.
10. The method according to claim 3, wherein the number of left-right vehicle displacements Cc is divided into a number of left-hand displacements CcLNumber of right-hand displacements CRWhen the distance X coordinate of the target vehicle relative to the leftmost lane in the ith frame is subtracted by the distance X of the target vehicle relative to the leftmost lane in the (i + 1) th frame, if the obtained difference is positive, the left-direction displacement times CLPlus 1, otherwise the number of right-hand displacements CRAnd adding 1.
CN202210058857.3A 2022-01-17 2022-01-17 BiLSTM-CRF model-based vehicle lane change intention prediction method Pending CN114612867A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210058857.3A CN114612867A (en) 2022-01-17 2022-01-17 BiLSTM-CRF model-based vehicle lane change intention prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210058857.3A CN114612867A (en) 2022-01-17 2022-01-17 BiLSTM-CRF model-based vehicle lane change intention prediction method

Publications (1)

Publication Number Publication Date
CN114612867A true CN114612867A (en) 2022-06-10

Family

ID=81858076

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210058857.3A Pending CN114612867A (en) 2022-01-17 2022-01-17 BiLSTM-CRF model-based vehicle lane change intention prediction method

Country Status (1)

Country Link
CN (1) CN114612867A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115512540A (en) * 2022-09-20 2022-12-23 中国第一汽车股份有限公司 Information processing method and device for vehicle, storage medium and processor

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115512540A (en) * 2022-09-20 2022-12-23 中国第一汽车股份有限公司 Information processing method and device for vehicle, storage medium and processor

Similar Documents

Publication Publication Date Title
Zyner et al. A recurrent neural network solution for predicting driver intention at unsignalized intersections
EP3699048B1 (en) Travelling track prediction method and device for vehicle
CN110297494B (en) Decision-making method and system for lane change of automatic driving vehicle based on rolling game
CN110796856B (en) Vehicle lane change intention prediction method and training method of lane change intention prediction network
CN113291308B (en) Vehicle self-learning lane-changing decision-making system and method considering driving behavior characteristics
US11555706B1 (en) Processing graph representations of tactical maps using neural networks
CN112793576B (en) Lane change decision method and system based on rule and machine learning fusion
CN109484404A (en) Controller of vehicle, control method for vehicle and storage medium
CN111930110A (en) Intent track prediction method for generating confrontation network by combining society
CN111845754B (en) Decision prediction method of automatic driving vehicle based on edge calculation and crowd-sourcing algorithm
CN112249008B (en) Unmanned automobile early warning method aiming at complex dynamic environment
Wang et al. Vehicle trajectory prediction by knowledge-driven LSTM network in urban environments
CN115523934A (en) Vehicle track prediction method and system based on deep learning
CN114212110A (en) Obstacle trajectory prediction method, obstacle trajectory prediction device, electronic device, and storage medium
Chen et al. Towards human-like speed control in autonomous vehicles: A mountainous freeway case
CN114612867A (en) BiLSTM-CRF model-based vehicle lane change intention prediction method
Peng et al. Driving maneuver detection via sequence learning from vehicle signals and video images
JP2006160032A (en) Driving state determination device and its method
Villagra et al. Motion prediction and risk assessment
CN115440041A (en) Method for predicting driving behavior of key vehicle under road side view angle
CN113424022B (en) Method for processing predicted motion trail, method and device for displaying constraint barrier
Zheng et al. An investigation on coordination of lane departure warning based on driver behaviour characteristics
Chamola et al. Overtaking Mechanisms Based on Augmented Intelligence for Autonomous Driving: Datasets, Methods, and Challenges
Wang et al. Deep-learning-based anomaly detection for lane-changing decisions
Chen et al. The Scanner of Heterogeneous Traffic Flow in Smart Cities by an Updating Model of Connected and Automated Vehicles

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination