A method for predicting vehicle lane change intention based on BiLSTM-CRF model
Through the BiLSTM-CRF model, the vehicle lane change intention characteristics and rules are solved, and the problem of difficult to balance the prediction effect and credibility in the prior art is achieved, and more accurate and reliable lane change intention prediction is achieved.
Patent Information
- Application Number
- CN202210058857.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-17
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2042-01-17
AI Technical Summary
The existing vehicle lane change intention prediction methods are difficult to take into account both the prediction effect and the credibility and interpretability, resulting in the unreliable prediction results.
Using the vehicle lane change intention prediction method based on BiLSTM-CRF model, the vehicle lane change intention characteristics are extracted by traversing historical traffic data, and the model is trained using inducible rules and absolute rules, and combined with the CRF layer to process the label transfer probability, improve the explanatory prediction Sex and accuracy.
It improves the accuracy and credibility of vehicle lane change intention prediction, ensures the interpretability of prediction results, and enhances drivers' trust in the autonomous driving system.
Smart Images
Figure CN114612867B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of automatic driving intention prediction, and in particular to a method for predicting vehicle lane change intention based on a BiLSTM-CRF model. Background Art
[0002] Autonomous driving is in a stage of rapid development. Not only has it made major breakthroughs in academic research, but more and more actual products are also being put into use, such as Tesla's driverless cars and Baidu's self-driving taxis. With the acceleration of commercialization, autonomous driving is gradually entering people's lives.
[0003] However, this technology is not yet mature. Safety has always been the cornerstone of self-driving cars. Motion prediction for other vehicles can better assist self-driving cars in making decision plans, and most advanced motion prediction methods are based on intention. Oliver Pentland first used pattern recognition methods to classify driver intentions in 1996, which was the earliest research on driving behavior. Today, driver intention prediction has been widely studied, and most studies involve the use of heuristic methods or training classifiers. Heuristic methods can be used to determine what operations the target vehicle will perform in the short future. In more complex scenarios, discriminative learning algorithms are more popular for classifying driving intentions. Mandalia HM et al. proposed an intention prediction method based on support vector machines in the literature (Using support vector machines for lane change detection). The position information of the vehicle in four different directions from the lane is used as a feature to predict the lane change intention of the target vehicle. Experiments have shown that this method has better prediction time and accuracy than the heuristic method. Then, with the rise of neural networks, the means of intention prediction have been further enriched. AbdelmoudjibBenterki et al. proposed a method to detect lane change intention on highways using artificial neural networks in the literature (Prediction of Surrounding Vehicles Lane Change Intention Using Machine Learning). The method used data from the vehicle's lane change preparation stage to predict lane changes, and tested two lane change prediction methods, support vector machine and artificial neural network, on public data sets to verify the performance of the proposed method. Neurons in the same layer of a neural network do not transmit information to each other, which means that it cannot fully consider the interaction between neurons in the same layer, while LSTM (Long Short-Term Memory Network) can better capture information between sequences. L Tang, H Wang et al. compared the robustness and prediction accuracy of three methods, rule-based models, support vector machines, and LSTM, in the literature (Driver Lane Change Intention Recognition of Intelligent Vehicle based on Long Short-Term Memory Network), and verified that LSTM has the ability to process longer time series. Another popular intention prediction method is to use hidden Markov and other probabilistic graph models, which can output the most likely label sequence corresponding to the observation sequence and have good interpretability.
[0004] Driving intention prediction is one of the key technologies for developing advanced driver assistance systems. It can effectively reduce traffic accidents caused by lane changes and ensure driving safety. However, the intention prediction algorithms based on machine learning today have poor interpretability, and drivers find it difficult to trust unexplainable prediction results, which leads to people not trusting machines. Heuristic methods or probabilistic graph models have strong interpretability, but the intention prediction effect is poor. Pure learning methods such as LSTM, due to their own structural characteristics and the lack of rule constraints, sometimes lead to unexpected prediction errors, such as predicting the lane change intention of the vehicle on the far left as a left lane change, which further leads to unreliable prediction results. In this way, it is difficult for existing intention prediction methods to balance prediction effect and trustworthy interpretability. Summary of the invention
[0005] The present invention overcomes one of the deficiencies of the prior art and provides a method for predicting vehicle lane-changing intention based on the BiLSTM-CRF (bidirectional long short-term memory network-conditional random field) model, which can solve the problem that it is difficult to strike a balance between prediction effect and credible interpretability in predicting lane-changing intention of vehicles on highways.
[0006] According to one aspect of the present disclosure, the present invention provides a method for predicting vehicle lane change intention based on a BiLSTM-CRF model, the method comprising:
[0007] Traversing the historical traffic data of the vehicle to obtain historical information of the vehicle's lane change intention;
[0008] Preprocessing the historical information of the vehicle's lane change intention to extract vehicle lane change intention features that reflect changes in the driver's intention;
[0009] According to an inductive rule, the vehicle lane change intention features are combined into the vehicle lane change intention feature sequence;
[0010] Labeling the vehicle lane change intention feature sequence to obtain a label of the vehicle lane change intention feature;
[0011] After fuzzifying the labels of the vehicle lane change intention features, the vehicle lane change intention features are divided into a training set and a test set;
[0012] A BiLSTM-CRF model of a vehicle lane change intention prediction model is trained using the training set of the vehicle lane change intention feature and the label of the vehicle lane change intention feature;
[0013] Using absolute rules to train the post-rule layer of the vehicle lane change intention prediction model;
[0014] The vehicle lane-changing intention prediction model is tested using the test set of the vehicle lane-changing intention features, and the real-time vehicle lane-changing intention features are input into the vehicle lane-changing intention prediction model to predict the vehicle lane-changing intention.
[0015] In a possible implementation, the historical information H of the vehicle's lane change intention includes: a distance X of the vehicle relative to the leftmost lane, a distance Y of the vehicle entering the starting collection point interface, and the vehicle's maneuvering intention.
[0016] In a possible implementation, the vehicle lane change intention feature reflecting the change in the driver's intention includes: the sum of the squares of the lateral speed of the vehicle within a preset time length, the sum of the turning angles, and the number of left and right displacements of the vehicle;
[0017] Among them, the sum of the squares of the lateral speed and the sum of the turning angles are basic features of the vehicle's lane change intention feature, and the number of left and right displacements of the vehicle is a regular feature of the vehicle's lane change intention feature.
[0018] In a possible implementation, the fuzzy processing of the label of the lane change intention feature of the vehicle includes:
[0019] Counting the label values of the vehicle lane-changing intention feature, and dividing the feature into N intervals according to the label values of the vehicle lane-changing intention feature;
[0020] Determining the interval where the label of the vehicle lane-changing intention feature is located according to the label value of the vehicle lane-changing intention feature and the length of each interval;
[0021] The lower boundary of the interval is used as the fuzzy value of the label of the vehicle lane change intention feature.
[0022] In a possible implementation, the vehicle lane change intention feature is divided into two categories: basic features and rule features. The labels of the basic features include four categories: changing lanes to the left, changing lanes to the right, keeping straight or changing lanes to the left, keeping straight or changing lanes to the right, and the labels of the rule features are the degree of compliance with the rules.
[0023] In a possible implementation, the vehicle lane change intention prediction model includes two BiLSTM-CRF models, and the BiLSTM-CRF model is divided into a BiLSTM layer and a CRF layer;
[0024] The BiLSTM layer is used to output the score of each label of each lane change intention feature of the vehicle;
[0025] The CRF layer is used to calculate the transition probability between labels of the vehicle lane change intention feature.
[0026] In a possible implementation, one BiLSTM-CRF model is used to predict whether the vehicle changes lanes to the left and keeps going straight or changes lanes to the right, and another BiLSTM-CRF model is used to predict whether the vehicle changes lanes to the right and keeps going straight or changes lanes to the left.
[0027] In a possible implementation, the absolute rules include traffic rules and physical constraints.
[0028] In a possible implementation, the sum of the squares of the lateral velocities The sum of the turning angles
[0029] Where N is a positive integer, which is the number of image frames of the target vehicle within the preset time;
[0030] V X Divided into V X_L With V X_R , respectively representing the left lateral speed and the right lateral speed. Subtract the distance X of the target vehicle from the leftmost lane in the i-th frame from the distance X of the target vehicle from the leftmost lane in the i+1-th frame. If the difference is positive, it is recorded as V X_L Otherwise, it is recorded as V X_R ;
[0031] V Y The distance Y of the target vehicle entering the starting acquisition point interface in the i+1th frame minus the distance Y of the target vehicle entering the starting acquisition point interface in the i-th frame. If V Y If it is zero, it means that the target vehicle is in a stopped state or the data set is wrong.
[0032] In a possible implementation, the number of left-right displacements C of the vehicle is divided into the number of left-right displacements C L The number of rightward displacements C R , when the X coordinate of the distance of the target vehicle in the i-th frame relative to the leftmost lane is subtracted from the distance X of the target vehicle in the i+1-th frame relative to the leftmost lane, if the difference is positive, the number of leftward displacements C L Add 1, otherwise the number of right shifts is C R Add 1.
[0033] The vehicle lane change intention prediction method based on the BiLSTM-CRF model disclosed in the present invention obtains the historical information of the vehicle's lane change intention by traversing the historical traffic data of the vehicle; pre-processes the historical information of the vehicle's lane change intention to extract the vehicle lane change intention features that reflect the driver's intention change; organizes the vehicle lane change intention features into the vehicle lane change intention feature sequence according to the inductive rule; annotates the vehicle lane change intention feature sequence to obtain the label of the vehicle lane change intention feature; after fuzzifying the label of the vehicle lane change intention feature, the vehicle lane change intention feature is divided into a training set and a test set; uses the training set of the vehicle lane change intention feature and the label of the vehicle lane change intention feature to train the BiLSTM-CRF model of the vehicle lane change intention prediction model; uses the absolute rule to train the post-rule layer of the vehicle lane change intention prediction model; uses the test set of the vehicle lane change intention feature to test the vehicle lane change intention prediction model, and inputs the real-time vehicle lane change intention feature into the vehicle lane change intention prediction model to predict the vehicle lane change intention. It can solve the problem that it is difficult to strike a balance between prediction effect and credible explainability in predicting lane change intentions of vehicles on highways. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] The accompanying drawings are used to provide a further understanding of the technical solution of the present application or the prior art, and constitute a part of the specification. Among them, the accompanying drawings expressing the embodiments of the present application are used together with the embodiments of the present application to explain the technical solution of the present application, but do not constitute a limitation on the technical solution of the present application.
[0035] Figure 1 A flow chart of a method for predicting vehicle lane change intention based on a BiLSTM-CRF model according to an embodiment of the present disclosure is shown;
[0036] Figure 2 A schematic diagram showing the collection of historical information of a vehicle's lane change intention according to an embodiment of the present disclosure is shown;
[0037] Figure 3 A schematic diagram of the structure of a vehicle lane change intention prediction model according to an embodiment of the present disclosure is shown;
[0038] Figure 4 A schematic diagram of LSTM emission probability according to an embodiment of the present disclosure is shown;
[0039] Figure 5 A schematic diagram of a CRF transition probability matrix according to an embodiment of the present disclosure is shown;
[0040] Figure 6 A schematic diagram showing the prediction effect of a vehicle lane change intention prediction model according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION
[0041] The following will describe the implementation methods of the present invention in detail with reference to the accompanying drawings and embodiments, so that the implementation process of how the present invention applies technical means to solve technical problems and achieve corresponding technical effects can be fully understood and implemented accordingly. The embodiments of the present application and the various features in the embodiments can be combined with each other without conflict, and the technical solutions formed are all within the protection scope of the present invention.
[0042] In addition, the steps shown in the flowchart of the accompanying drawings can be executed in a computer such as a set of computer executable instructions. Also, although a logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a sequence different from that here.
[0043] Figure 1 FIG. 1 shows a flow chart of a method for predicting a vehicle lane change intention based on a BiLSTM-CRF model according to an embodiment of the present disclosure. Figure 1 As shown, the method may include:
[0044] Step S1: Traverse the historical traffic data of the vehicle to obtain the historical information of the vehicle's lane change intention.
[0045] Among them, the main body of the BiLSTM-CRF model is composed of a bidirectional long short-term memory network (Bi-LSTM) and a conditional random field (CRF). The model input is a feature sequence, and the output is the predicted label corresponding to each feature. It is a bidirectional long short-term memory network-conditional random field model.
[0046] The historical information H of the vehicle's lane change intention may include: the distance of the vehicle relative to the leftmost lane (local X coordinate), the distance of the vehicle relative to the section of the entry starting collection point (local Y coordinate), the vehicle's maneuvering intention (manually marked according to the change of the lane where the vehicle is located), the vehicle identification number, the data frame, the vehicle's instantaneous speed, the vehicle's instantaneous acceleration, the vehicle's current lane position (with the first lane as the leftmost lane), the vehicle's driving direction, the global time, the vehicle identification number of the preceding vehicle in the same lane, the vehicle identification number of the following vehicle in the same lane, the headway time to the preceding vehicle, and other information.
[0047] In one example, the historical traffic data of the vehicle may come from a real traffic dataset NGSIM (Next Generation Simulation), the selected road section is the I-80 highway (Interstate 80 in San Francisco, USA), and the selected time periods are 4:00-4:15 pm and 5:00-5:30 pm.
[0048] Figure 2 A schematic diagram of historical traffic data collection of a vehicle's lane change intention according to an embodiment of the present disclosure is shown.
[0049] Traverse the above historical traffic data set to find trajectories containing vehicle lane change intentions. Among them, vehicle lane change is defined as the target vehicle driving from the current lane to the adjacent lane without collision. When the trajectory of a certain vehicle is located and intersects with the lane line, collect the historical information H of the 45 frames (each frame is 0.1s) before the intersection. The intersection is recorded as G, and the frame where the intersection G is located is recorded as the 45th frame, and the frame where the target vehicle is located 4.5s before the intersection with the lane is recorded as the first frame (such as Figure 2 ).
[0050] Step S2: pre-process the historical information of the vehicle's lane change intention and extract the vehicle's lane change intention features that reflect the driver's intention change.
[0051] The vehicle lane change intention characteristics include: the sum of the squares of the vehicle's lateral speed within a preset time length, the sum of the turning angles, and the number of left and right displacements of the vehicle.
[0052] The sum of the squares of the transverse velocities The sum of the turning angles
[0053] Where N is a positive integer, which is the number of image frames of the target vehicle within the preset time;
[0054] V X Divided into V X_L With V X_R , respectively representing the left lateral speed and the right lateral speed. Subtract the distance X of the target vehicle from the leftmost lane in the i-th frame from the distance X of the target vehicle from the leftmost lane in the i+1-th frame. If the difference is positive, it is recorded as V X_L Otherwise, it is recorded as V X_R ;
[0055] V Y The distance Y of the target vehicle entering the starting acquisition point interface in the i+1th frame minus the distance Y of the target vehicle entering the starting acquisition point interface in the i-th frame. If V Y If it is zero, it means that the target vehicle is in a stopped state or the data set is wrong.
[0056] The number of left and right displacements of the vehicle C is divided into the number of left displacements C L The number of rightward displacements C R , when the X coordinate of the distance of the target vehicle in the i-th frame relative to the leftmost lane is subtracted from the distance X of the target vehicle in the i+1-th frame relative to the leftmost lane, if the difference is positive, the number of leftward displacements C L Add 1, otherwise the number of right shifts is C R Add 1.
[0057] For example, if the preset time length is 1s (10 frames), the sum of the squares of the lateral velocity within the 1s time length (i.e. 10 frames) is: After calculation, the sum of the squares of the left lateral velocity in every 10 frames is finally obtained: Sum_Sq_L and the square of the right lateral velocity V Sum_Sq_R .
[0058] The sum of the turning angles After calculation, the sum of the left turning angles in every 10 frames is finally obtained. Sum_L and the right turning angle A Sum_R .
[0059] This formula can also be used to obtain the sum of the squares of the lateral velocities in different time lengths: and the turning angle Wherein, i is the frame number of the starting image of different preset time lengths.
[0060] Taking right turn as an example, the lane change process of a vehicle can be divided into three parts: lane keeping, lane change, and lane change completion. After calculating the above features in the lane keeping and lane change intervals, it can be found that the important changes that can reflect the intention of lane change occur on average in the 15th frame of the historical sequence collected, and this data varies depending on the driver's driving habits. In the actual prediction process, the value of the feature vector will be updated with the time step, and the time step is 0.1s. In addition, because the sequence length used for prediction in this example is 10 frames, the trained model is more sensitive to vehicle lane changes and can capture lane change behavior more quickly.
[0061] Step S3: According to the inductive rules, the vehicle lane change intention features are combined into a vehicle lane change intention feature sequence. Among them, the rules are divided into absolute rules and inductive rules. The difference between inductive rules and absolute rules is that they are not absolutely valid, but they can promote the vehicle lane change intention prediction model to make predictions that comply with the rules in an inductive way. For example, the inductive rule can be that within a preset time interval, if the target vehicle continues to move laterally and has a large lateral speed, it is considered that the vehicle has a high probability of changing lanes. The specific working mode of the inductive rule is through the transfer probability between the rule features and the labels corresponding to the basic features.
[0062] Step S4: Label the vehicle lane change intention feature sequence to obtain the label of the vehicle lane change intention feature. The vehicle lane change intention feature is divided into two categories: basic features and rule features. The labels of the basic features include four categories: change lanes to the left, change lanes to the right, keep going straight or change lanes to the left, keep going straight or change lanes to the right. The label of the rule feature is the degree of compliance with the inductive rule under the current basic feature and rule feature values, and the value range is [0,1].
[0063] Step S5: After fuzzifying the labels of the vehicle lane change intention features, the vehicle lane change intention features are divided into a training set and a test set.
[0064] In one example, fuzzifying the label of the vehicle lane change intention feature may include:
[0065] Counting the label values of the vehicle lane-changing intention feature, and dividing the feature into N intervals according to the label values of the vehicle lane-changing intention feature;
[0066] Determining the interval where the label of the vehicle lane-changing intention feature is located according to the label value of the vehicle lane-changing intention feature and the length of each interval;
[0067] The lower boundary of the interval is used as the fuzzy value of the label of the vehicle lane change intention feature.
[0068] The sum of the squares of the left lateral velocities V Sum_Sq_L For example, the sum of the squares of the left lateral velocities V Sum_Sq_L The label value λ distribution is then divided into intervals. The smaller the sub-interval length is in the area where the label value λ is more densely distributed, the smaller the sub-interval length will be. Sum_Sq_L In the range of 0.01-1, the label value λ is distributed densely, so the label value λ distribution interval is divided into 100 sub-intervals, and the sub-interval length is 0.01. Sum_Sq_L The value of is taken as the lower limit of its interval, such as the sum of the squares of the left lateral velocity V Sum_Sq_L =0.1753, and its subinterval is [0.17, 0.18), then the fuzzy value is the sum of the squares of the left lateral velocity V Sum_Sq_L The value is 0.17. After the above processing, the training set and the test set are divided into two parts with a ratio of 0.7:0.3.
[0069] Step S6: using the training set of the vehicle lane change intention features and the labels of the vehicle lane change intention features to train the BiLSTM-CRF model of the vehicle lane change intention prediction model.
[0070] Figure 3 A schematic diagram of the structure of a vehicle lane change intention prediction model according to an embodiment of the present disclosure is shown.
[0071] like Figure 3 As shown, 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 to output the score of each label of each of the vehicle lane change intention features; and the CRF layer is used to calculate the transition probability between the labels of the vehicle lane change intention features.
[0072] Figure 4 A schematic diagram of LSTM emission probability according to an embodiment of the present disclosure is shown; Figure 5A schematic diagram of a CRF transition probability matrix according to an embodiment of the present disclosure is shown.
[0073] like Figure 4 As shown in the figure, the LSTM network can output the score of each vehicle lane change intention feature on each label, which is expressed as the emission probability. BiLSTM adds a reverse operation on the basis of the LSTM network. The final output is a simple superposition of the outputs of the forward LSTM and the reverse LSTM, which can better consider the information between the sequence contexts. For example, the number of hidden layers of the BiLSTM model is 20, and the final output is the most likely label of each vehicle lane change intention feature, and the labels are independent of each other. After adding the CRF layer, as shown in Figure 5 As shown in , the CRF layer can calculate the transition probability between labels, so the BiLSTM-CRF model outputs the best label sequence, not independent labels. Figure 4 , 5 In , [0, 0.5, 1] is the degree of compliance with the inductive rule, and [L, S_R] is the vehicle intention label.
[0074] The BiLSTM-CRF model can also score and identify label sequences to determine whether a label sequence is the best label sequence. The scoring formula is as follows:
[0075] Score=EmissionScore+TransitionScore
[0076] The EmissionScore is the score of the label sequence given by the BilSTM model, and the TransitionScore is the sum of the corresponding state transition probabilities between the elements of the label sequence.
[0077] The BiLSTM-CRF model trained using the training set of vehicle lane change intention features and the labels of vehicle lane change intention features can make the correct label sequence have the highest score among all possible sequences.
[0078] In one example, one BiLSTM-CRF model is used to predict whether a vehicle changes lanes to the left and keeps going straight or changes lanes to the right, and another BiLSTM-CRF model is used to predict whether a vehicle changes lanes to the right and keeps going straight or changes lanes to the left.
[0079] For example, two BiLSTM-CRF models are recorded as M1 and M2, M1 is used to predict left lane change and non-left lane change (keep straight and change right lane), M2 is used to predict right lane change and non-right lane change (keep straight and change left lane). L ,V Sum_Sq_L , A Sum_L ] and [CR ,V Sum_Sq_R , A Sum_R ] Input models M1 and M2 respectively, and then compare the prediction results of the two models to decide the next behavior. For example, when M1 outputs a left lane change and M2 outputs a non-right lane change, the prediction result is a left lane change; when M1 outputs a left lane change and M2 outputs a right lane change, the prediction result is uncertain and needs to be recalculated based on the feature sequence of the next time step. In this way, models M1 and M2 constrain each other, which improves the prediction accuracy at the expense of the prediction time.
[0080] Step S7: using absolute rules to train the post-rule layer of the vehicle lane change intention prediction model.
[0081] Among them, the absolute rules are the rules that must be followed under normal circumstances, which may include traffic rules, physical constraints, turn signal signals, etc., to make a correction to the prediction results of the BiLSTM-CRF model. For example, if the target vehicle is in the leftmost lane or the rightmost lane, the vehicle can no longer change lanes to the left or right. If the target vehicle is stationary, it is considered that the vehicle will not complete the lane change in the next 3 seconds.
[0082] Step S8: Testing the vehicle lane-changing intention prediction model using a test set of vehicle lane-changing intention features, and inputting the real-time vehicle lane-changing intention features into the vehicle lane-changing intention prediction model to predict the vehicle lane-changing intention.
[0083] Figure 6 FIG. 2 shows a schematic diagram of the prediction effect of a vehicle lane change intention prediction model according to an embodiment of the present disclosure. Figure 6 As shown in the figure, the prediction effect of the training vehicle lane change intention prediction model can be measured by parameters such as accuracy, precision, recall and F1 score.
[0084] The vehicle lane change intention prediction method based on the BiLSTM-CRF model of the present invention is different from the traditional LSTM method. The CRF layer enables the method to process the transfer relationship between feature sequence labels, that is, to obtain the above-mentioned transfer probability matrix. Not only can the interactive relationship between the features in the vehicle lane change intention feature sequence be fully utilized, but also the degree of compliance with the inductive rules under the current basic features and rule feature values can be obtained as the labeling result of the rule features during prediction. The higher the degree of compliance with the rules, the greater the transfer probability of the intention prediction result corresponding to the rules, that is, the directional constraint is realized by the transfer probability matrix when predicting the vehicle maneuvering intention, thereby improving the accuracy of the prediction. Through such a method, the present invention embeds the rules into the intention prediction model to improve the interpretability of the prediction. A post-rule layer is externally connected to the BiLSTM-CRF model. Combined with historical information, the post-rule layer can correct the prediction results of the BiLSTM-CRF model, eliminate machine errors with traffic rules or human experience, and further enhance the credibility of the prediction results.
[0085] Although the embodiments disclosed in the present invention are as above, the above contents are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the patent protection scope of the present invention shall still be subject to the scope defined in the attached claims.
Claims
1. A vehicle lane change intention prediction method based on the BiLSTM-CRF model, characterized in that: The method comprises: Traversing the historical traffic data of the vehicle to obtain historical information of the vehicle's lane change intention; Preprocessing the historical information of the vehicle's lane change intention to extract vehicle lane change intention features that reflect changes in the driver's intention; According to an inductive rule, the vehicle lane change intention features are combined into the vehicle lane change intention feature sequence; Labeling the vehicle lane change intention feature sequence to obtain a label of the vehicle lane change intention feature; After fuzzifying the labels of the vehicle lane change intention features, the vehicle lane change intention features are divided into a training set and a test set; A BiLSTM-CRF model of a vehicle lane change intention prediction model is trained using the training set of the vehicle lane change intention feature and the label of the vehicle lane change intention feature; Using absolute rules to train the post-rule layer of the vehicle lane change intention prediction model; The vehicle lane-changing intention prediction model is tested using the test set of the vehicle lane-changing intention features, and the real-time vehicle lane-changing intention features are input into the vehicle lane-changing intention prediction model to predict the vehicle lane-changing intention.
2. The method for predicting vehicle lane change intention according to claim 1, characterized in that: The historical information H of the vehicle's lane change intention includes: the distance X of the vehicle relative to the leftmost lane, the distance Y of the vehicle entering the starting collection point interface, and the vehicle's maneuvering intention.
3. The method for predicting vehicle lane change intention according to claim 2, characterized in that: The vehicle lane change intention features reflecting the driver's intention change include: the sum of the squares of the lateral speed of the vehicle within a preset time length, the sum of the heading angles, and the number of left and right displacements of the vehicle; Among them, the sum of the squares of the lateral speed and the sum of the turning angles are basic features of the vehicle's lane change intention feature, and the number of left and right displacements of the vehicle is a regular feature of the vehicle's lane change intention feature.
4. The method for predicting vehicle lane change intention according to claim 1, characterized in that: The fuzzy processing of the label of the lane change intention feature of the vehicle includes: Counting the label values of the vehicle lane-changing intention feature, and dividing the feature into N intervals according to the label values of the vehicle lane-changing intention feature; Determining the interval where the label of the vehicle lane-changing intention feature is located according to the label value of the vehicle lane-changing intention feature and the length of each interval; The lower boundary of the interval is used as the fuzzy value of the label of the vehicle lane change intention feature.
5. The method for predicting vehicle lane change intention according to claim 1, characterized in that: The vehicle lane change intention features are divided into two categories: basic features and rule features; The labels of basic features include four categories: changing lanes to the left, changing lanes to the right, keeping straight or changing lanes to the left, and keeping straight or changing lanes to the right. The labels of rule features are the degree of compliance with the rules.
6. The method for predicting vehicle lane change intention according to claim 5, characterized in that: The vehicle lane change intention prediction model includes two BiLSTM-CRF models, and the BiLSTM-CRF model is divided into a BiLSTM layer and a CRF layer; The BiLSTM layer is used to output the score of each label of each lane change intention feature of the vehicle; The CRF layer is used to calculate the transition probability between labels of the vehicle lane change intention feature.
7. The method for predicting vehicle lane change intention according to claim 6, characterized in that: The one BiLSTM-CRF model is used to predict whether the vehicle changes lanes to the left and keeps going straight or changes lanes to the right, and the other BiLSTM-CRF model is used to predict whether the vehicle changes lanes to the right and keeps going straight or changes lanes to the left.
8. The method for predicting vehicle lane change intention according to claim 1, characterized in that: The absolute rules include traffic rules and physical constraints.
9. The method for predicting vehicle lane change intention according to claim 3, characterized in that: The sum of the squares of the transverse velocities The sum of the turning angles Wherein, N is a positive integer, i is the number of frames of the target vehicle image within the preset time, and n is the number of frames of the starting image of different preset time lengths; V X Divided into V X_L With V X_R , representing the left lateral speed and the right lateral speed respectively; subtract the distance X of the target vehicle in the i-th frame relative to the leftmost lane from the distance X of the target vehicle in the i+1-th frame relative to the leftmost lane. If the difference is positive, it is recorded as V X_L Otherwise, it is recorded as V X_R ; V Y The distance Y of the target vehicle entering the starting acquisition point interface in the i+1th frame minus the distance Y of the target vehicle entering the starting acquisition point interface in the i-th frame. If V Y If it is zero, it means that the target vehicle is in a stopped state or the data set is wrong.
10. The method for predicting vehicle lane change intention according to claim 9, characterized in that: The number of left and right displacements of the vehicle C is divided into the number of left displacements C L The number of rightward displacements C R , when the X coordinate of the distance of the target vehicle in the i-th frame relative to the leftmost lane is subtracted from the distance X of the target vehicle in the i+1-th frame relative to the leftmost lane, if the difference is positive, the number of leftward displacements C L Add 1, otherwise the number of right shifts is C R Add 1.