An autonomous driving vehicle behavior decision-making system and method

Abstract

The present invention relates to an autonomous vehicle behavior decision-making system, belonging to the field of intelligent vehicles. The system includes a data acquisition module that obtains surrounding environment information through sensors and constructs a road model; a spatio-temporal feature extraction module that uses a decision-making model based on LSTM-CNN to extract information from the road model to obtain spatio-temporal features; a driving behavior rule extraction module that analyzes the interactivity of the vehicle to obtain features generated by different behaviors, and constructs a driving behavior rule feature mining algorithm based on these features to obtain behavior rule feature encodings; a fusion decision module that uses a WIDE&DEEP fusion model to fuse behavior features and spatio-temporal features, and finally outputs a behavior decision result. The present invention also provides a method.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent vehicles and relates to an autonomous driving vehicle behavior decision-making system and method. Background Art

[0002] Behavior decision-making is a crucial link in autonomous driving vehicles. It makes decisions on the driving behavior of the vehicle based on environmental perception data to achieve the task goals of the vehicle.

[0003] Currently, the main autonomous driving vehicle behavior decision-making methods can be divided into two types: model-driven and data-driven behavior decision-making methods. However, both methods have certain deficiencies.

[0004] Chinese Patent Application: Vehicle Behavior Decision-Making Method and Device for Fusing Prediction Algorithm in Parking Lot Scenario (Application No.: CN202310045955.8) discloses a vehicle behavior decision-making method and device for fusing prediction algorithm in parking lot scenario, which is characterized by including: predicting the behavior of the target vehicle using the prediction algorithm; predicting the trajectories of the target vehicle respectively according to different prediction results. This method uses a behavior prediction model based on the long short-term memory network LSTM, and a single data-driven method has the problem of insufficient interpretability. The literature "A systematic solution of humandriving behavior modeling and simulation for automated vehicle studies" uses a decision tree model. By judging attributes such as vehicle spacing and whether at an intersection, the driving behavior of the autonomous driving vehicle is divided into different states such as parking, intersection handling, avoiding vehicles, overtaking and lane-changing, etc., and then the "IF-THRN" rule is used to model and describe these states until reaching the leaf node, finally realizing the behavior decision-making of the autonomous driving vehicle. However, it is difficult to handle complex traffic scenarios and abnormal situations, lacks the ability to handle uncertainties, and may lead to inaccurate and incomplete behavior decisions. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to propose a method that effectively combines the model-based and data-driven methods to improve the accuracy, generalization ability and prediction performance of the decision-making model, and improve the accuracy of autonomous driving vehicle behavior decision-making, aiming at the problems that the current model-driven method is difficult to model complex scenarios and has poor portability, and the data-driven behavior decision-making method lacks interpretability.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] On the one hand, the present invention provides an autonomous vehicle behavior decision-making system, including a data acquisition module, a spatio-temporal feature extraction module, a driving behavior rule extraction module, and a fusion decision-making module;

[0008] The data acquisition module obtains surrounding environment information through sensors and constructs a road model;

[0009] The spatio-temporal feature extraction module uses a decision-making model based on LSTM-CNN to extract information from the road model and obtain spatio-temporal features;

[0010] The driving behavior rule extraction module analyzes the interactivity of the vehicle to obtain the features generated by different behaviors, and constructs a driving behavior rule feature mining algorithm based on these features to obtain behavior rule feature encodings;

[0011] The fusion decision-making module uses a WIDE&DEEP fusion model to fuse behavior features and spatio-temporal features, and finally outputs a behavior decision result.

[0012] On the other hand, the present invention provides an autonomous vehicle behavior decision-making method, including the following steps:

[0013] S1: Data acquisition: Obtain the surrounding environment information of the vehicle through sensors and construct a road model, where the road model includes vehicle information V and lane information R;

[0014] S2: Spatio-temporal feature extraction: Process the vehicle information V and lane information R to extract spatio-temporal features SC;

[0015] S3: Driving behavior rule extraction: Use a driving behavior rule feature mining algorithm according to the vehicle information V and lane information R of the data acquisition module to construct behavior rule feature encodings and obtain driving behavior rule features D;

[0016] S4: Fusion decision-making: Use a WIDE&DEEP fusion model to fuse the driving behavior features D and spatio-temporal features SC to obtain the behavior decision Y of the current vehicle.

[0017] Further, in step S1, the road model is {V, R}, where V is the set of surrounding vehicles {V Ego , V Ego_front , V Ego_back , V Left_front , V Left_back , V Right_front , V Right_back}, which are the target vehicle, the vehicle in front, the vehicle behind, the vehicle in the front left, the vehicle in the rear left, the vehicle in the front right, and the vehicle in the rear right respectively; R is the road information {M, L id , F}, where M is the number of lanes, and L idis the current lane label, and F indicates whether lane changing is possible in the current lane. For vehicles in the leftmost lane, left lane changing is not possible, and for vehicles in the rightmost lane, right lane changing is not possible;

[0018] For each vehicle, there is vehicle state information {t, X, Y, Vel, Acc, θ, L, W, L id , X_dis, Y_dis}, which are the timestamp, lateral position, longitudinal position, speed, acceleration, steering angle, vehicle length, vehicle width, current lane, lateral offset from the target vehicle, and longitudinal offset from the target vehicle respectively; the greater the distance between vehicles, the smaller the interactive influence between vehicles. A larger value is used to indicate that there is no vehicle in a certain azimuth of the target vehicle; for the processing of default vehicles without vehicles around, 0 m / s is used to represent the speed of the default vehicle. For the lateral offset from the target vehicle and the longitudinal offset from the target vehicle, larger values of X_dis and Y_dis are used.

[0019] Furthermore, in step S2, a behavior decision-making model based on LSTM-CNN is built to extract spatio-temporal features. Several parallel LSTMs are used to extract the sequential features of the left-front vehicle Left_front and the left-rear vehicle Left_back, the right-front vehicle Right_front and the right-rear vehicle Right_back, the front vehicle Front, the rear vehicle Back, and the target vehicle information Ego respectively; finally, a fully connected layer is used to fuse the features to obtain the sequential feature S; a residual CNN network is used to extract the spatial features C of the surrounding environment, and finally the two are fused to obtain the spatio-temporal feature SC;

[0020] The sequential feature is expressed as follows:

[0021] S = F lstm (V, R, T, N)

[0022] where S is the sequential feature; F lstm is the LSTM network model; V is the set of surrounding vehicle states; R is the set of lane states; where T is the time step; N is the input information dimension.

[0023] Furthermore, for the extraction of the driving behavior rules described in step S3, the GBDT driving behavior rule mining algorithm is used for mining to obtain the driving behavior feature D; the mined rule attributes include the factors affecting the target vehicle's different driving behaviors, including the time, lateral coordinate, longitudinal coordinate, speed, acceleration, lateral offset, and longitudinal offset related features of the target vehicle and the surrounding environment vehicles; the feature encoding rule is: assume that the driving behavior rule mining algorithm iterates M times to obtain the final output Then M decision trees are constructed, and each sample is set to 1 when it falls to the leaf node of each decision tree, and the rest are set to 0. Furthermore, a set of 0-1 features is constructed using these features.

[0024] Furthermore, step S3 specifically includes the following steps:

[0025] S31: For the training set {(x 1 , y 1 ), (x 2 , y 2 ), (x 3 , y 3 ), …, (x N , y N )}, where x i = {V, R} is the corresponding decision in the current environment, N is the number of samples, and initialize a weak learner:

[0026]

[0027] where L is the squared loss function;

[0028] S32: Iterate from 1 to M times. Each iteration fits a tree. For each tree, first calculate the residual of each sample, that is:

[0029]

[0030] Use (x i , r im ) to iteratively update a decision tree. R jm is the leaf node region of the new decision tree, j = 1, 2, 3,.., J m , and J m is the total number of leaf nodes of the new decision tree;

[0031] For leaf nodes j = 1, 2, 3,.., J m calculate the best fit value:

[0032]

[0033] where I is 1 when the sample corresponds to the leaf node region R jm , otherwise 0;

[0034] S33: Output the final result

[0035]

[0036] When the error between the final output result of the sample x and the true result is smaller, the fitting effect is better.

[0037] Furthermore, in step S4, the spatio-temporal features obtained in the feature extraction stage are jointly trained with the driving behavior rule features, and finally the current behavior decision results, including lane keeping, left lane change, and right lane change, are output, specifically including the following steps:

[0038] S41: Implement the memory ability by introducing non - linear cross - features for the mined driving behavior rule feature information D through Wide, and obtain the driving behavior rule information WIDE d ;

[0039] WIDE d = W T *D + b

[0040] where W is the weight matrix and b is the offset;

[0041] S42: Extract the spatio - temporal feature SC by learning low - dimensional dense vectors, and explore the features that have not appeared or rarely appeared in historical data through the Deep component;

[0042] The spatio - temporal feature SC is implemented by LSTM - CNN, and DEEP s is the output obtained by processing the spatio - temporal feature SC by the DEEP network component;

[0043] DEEP s = f(SC), SC = f LSTM-CNN (x)

[0044] S43: Wide&Deep output:

[0045] After the forward calculations of the Wide component and the Deep component are completed, the outputs of the two parts are weighted and summed to obtain the vehicle decision result, expressed as:

[0046] Y = σ(WIDE d + DEEP s + b)

[0047] where σ and b are the activation function and the offset respectively, and Y represents the output probability of the n - dimensional behavior decision.

[0048] The beneficial effects of the present invention are as follows: Aiming at the problems that the current model - driven method has difficulties in modeling complex scenarios and poor portability, and the data - driven behavior decision method lacks interpretability, the present invention effectively combines the model - based and data - driven methods, and adopts a model - driven driving behavior rule mining algorithm to provide more prior knowledge for the LSTM - CNN behavior decision model and make up for the lack of interpretability. The present invention complements the advantages of the two methods, effectively improves the accuracy of the behavior decision of autonomous driving vehicles, and thus improves the driving safety of vehicles.

[0049] Other advantages, objects, and features of the present invention will be set forth in part in the following description, and in part will be obvious to those skilled in the art upon examination of the following, or may be learned by practice of the present invention. The objects and other advantages of the present invention may be realized and obtained by the following description of the specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0051] Figure 1 FIG. Figure 1 is the overall framework of a behavior decision-making system and method for an autonomous vehicle provided by the present invention;

[0052] Figure 2 is a schematic diagram of the road model described in the present invention;

[0053] Figure 3 is a schematic diagram of the data-driven model framework described in the present invention;

[0054] Figure 4 is a schematic diagram of the fusion decision-making stage described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0056] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation on the present invention; for better illustration of the embodiments of the present invention, some components in the drawings will be omitted, enlarged, or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0057] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and cannot be construed as a limitation of the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0058] Figure 1 The following shows the overall framework of a behavior decision-making system for an autonomous driving vehicle according to the present invention. The system includes four parts: a data acquisition module, a spatio-temporal feature extraction module, a driving behavior rule extraction module, and a fusion decision-making module.

[0059] The data acquisition module obtains the surrounding environment information through sensors and constructs a road model. This road model will be used as the input data for the entire decision-making system.

[0060] The spatio-temporal feature extraction module uses a decision-making model based on LSTM-CNN to extract spatio-temporal features from the road model information.

[0061] The driving behavior rule extraction module analyzes the interactivity of the vehicle to obtain the characteristics of different behaviors, and constructs a driving behavior rule feature mining algorithm based on these characteristics to obtain behavior rule feature codes.

[0062] The fusion decision-making module uses a WIDE&DEEP fusion model to fuse the behavior features and spatio-temporal features, and finally outputs the behavior decision result.

[0063] Figure 2 The following shows the schematic diagram of the road model of the invention. The road model uses the data collected by lidar as the input source. In order to extract the driving information of the surrounding vehicles for each vehicle's each piece of data, the following search strategy is adopted to index the surrounding vehicle information of the ego vehicle:

[0064] (1) Traverse the data to find the target vehicle and record the time stamp and the lane number where the vehicle is located.

[0065] (2) Search for other vehicles with the same time number as the target vehicle in the data and add them to the candidate surrounding vehicles, record their lane numbers, and then determine their relative position relationships with the target vehicle, such as left, right, front, rear, etc., according to the lane numbers and coordinate positions of these vehicles and the target vehicle.

[0066] (3) For each candidate vehicle, calculate its lateral deviation and longitudinal deviation from the target vehicle, and based on these distances, screen the six vehicles closest to the target vehicle, namely the vehicle in the front left, the vehicle directly in front of the target vehicle, the vehicle in the front right, the vehicle in the rear left, the vehicle directly behind the target vehicle, and the vehicle in the rear right, and record the status information of these vehicles.

[0067] (4) Incorporate the status information of all surrounding vehicles into the status information of the target vehicle to generate complete vehicle information.

[0068] The constructed road model is {V, R}, where V is the set of surrounding vehicles {V Ego , V Ego_front , V Ego_back , V Left_front , V Left_back , V Right_front , V Right_back}, which are the target vehicle, the vehicle in front, the vehicle behind, the vehicle in the front left, the vehicle in the rear left, the vehicle in the front right, and the vehicle in the rear right respectively. R is the road information {M, L id , F}, where M is the number of lanes, L id is the current lane label, and F indicates whether lane changing is possible on the current lane. For vehicles in the leftmost lane, left lane changing is not possible, and for vehicles in the rightmost lane, right lane changing is not possible.

[0069] For each vehicle, there is vehicle status information {t, X, Y, Vel, Acc, θ, L, W, L id , X_dis, Y_dis}, which are the timestamp, lateral position, longitudinal position, speed, acceleration, steering angle, vehicle length, vehicle width, current lane, lateral offset from the target vehicle, and longitudinal offset from the target vehicle respectively. There are not always several vehicles around each vehicle, and default vehicles will be generated. For the handling of default vehicles, a speed of 0 m / s is used to represent the speed of the default vehicle. For the lateral offset from the target vehicle and the longitudinal offset from the target vehicle, X_dis and Y_dis are used, and generally larger values are adopted.

[0070] Before using the vehicle trajectory data in the collected data for model training, first use the locally weighted scatter plot smoothing method to process the coordinate positions, speeds, and accelerations of the vehicles in the lidar collected data to reduce the influence of errors and improve the accuracy and reliability of the training model.

[0071] Such as Figure 3The following is a schematic diagram of the data-driven model framework of the present invention. After processing the data set, it is used for the training and testing of the autonomous driving behavior decision-making model built by the present invention. The spatio-temporal feature extraction module extracts the temporal feature S and the spatial feature C during the driving of the autonomous vehicle through the LSTM and CNN networks respectively, and uses a fully connected layer to fuse the temporal feature and the spatial feature to obtain the spatio-temporal feature SC.

[0072] The temporal feature is expressed as follows:

[0073] S = F lstm (V, R, T, N) (1)

[0074] where S is the temporal feature; F lstm is the LSTM network model; V is the set of surrounding vehicle states; R is the set of lane states; where T is the time step; N is the input information dimension. The GBDT driving behavior rule mining algorithm is used for mining to obtain the driving behavior feature D. The mined rule attributes include the factors that affect the target vehicle to take different driving behaviors, including but not limited to the time, lateral coordinate, longitudinal coordinate, speed, acceleration, lateral offset, longitudinal offset and other related features of the target vehicle and the surrounding environment vehicles. The feature encoding rule is: assume that the driving behavior rule mining algorithm iterates M times to obtain the final output Then M decision trees are constructed. Each sample is set to 1 when it falls into the leaf node of each decision tree, and the rest are set to 0. Furthermore, a set of 0-1 features can be constructed using these features. The specific steps are as follows:

[0075] (1) For the training set {(x 1 , y 1 ), (x 2 , y 2 ), (x 3 , y 3 ), …, (x N , y N )}, where x i = {V, R}, is the corresponding decision in the current environment, and N is the number of samples. Initialize a weak learner:

[0076]

[0077] where L is the square loss function.

[0078] (2) Iteration times (from 1 to M), each time an iteration fits a tree. For each tree, first calculate the residual of each sample, that is

[0079]

[0080] Use (x i , r im)Iteratively update a decision tree, R jm is the leaf node region of the new decision tree, j = 1, 2, 3,.., J m , J m is the total number of leaf nodes of the new decision tree.

[0081] For leaf nodes j = 1, 2, 3,.., J m Calculate the best fit value:

[0082]

[0083] where I is 1 when the node corresponding to the sample is the leaf node region R jm and 0 otherwise.

[0084] (3) Output the final result

[0085]

[0086] When the final output result of the sample x has a small error with the true result, the fitting effect is good.

[0087] Figure 4 The figure shows the schematic diagram of the fusion decision stage. The spatio-temporal features and driving behavior rule features obtained in the feature extraction stage are fused using the WIDE&DEEP framework and jointly trained, and finally the current behavior decision result Y is output. The specific steps are as follows:

[0088] (1) Introduce non-linear cross features through Wide for the mined driving behavior rule feature information D to achieve efficient memory ability, and obtain more prior driving behavior rule information WIDE d . Among them, the driving behavior feature D is mined through the GBDT driving behavior rule mining algorithm.

[0089] WIDE d = W T *D + b (6)

[0090] (2) For the extracted spatio-temporal features SC, by learning low-dimensional dense vectors, explore features that have not appeared or rarely appeared in historical data through the Deep component to enhance the generalization ability of the network model.

[0091] The spatio-temporal features SC are implemented by LSTM-CNN, and DEEP s is the output obtained by processing the spatio-temporal features SC by the DEEP network component.

[0092] DEEP s = f(SC), SC = f LSTM-CNN (x) (7)

[0093] (3) Wide&Deep Output:

[0094] After the forward calculations of the Wide component and the Deep component are completed, the outputs of the two parts are weighted and summed to obtain the vehicle decision result, expressed as:

[0095] Y = σ(WIDE d + DEEP s + b) (8)

[0096] In the formula, σ and b are the activation function and the offset respectively, and Y represents the output probability of the n-dimensional behavior decision. If the decision outputs are only three types: left turn lane, right lane change, and following the vehicle, then the decision vector is a 1x3 column vector.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for autonomous driving vehicle behavior decision-making, characterized in that: It includes the following steps: S1: Data collection: Obtain the environmental information around the vehicle through sensors and construct a road model, where the road model includes vehicle information V and lane information R; S2: Spatiotemporal feature extraction: By processing the vehicle information V and lane information R, the spatiotemporal feature SC is extracted; In step S2, a behavior decision-making model based on LSTM-CNN is built to extract spatiotemporal features. Several parallel LSTMs are used to extract the temporal features of the left-front vehicle Left_front and the left-rear vehicle Left_back, the right-front vehicle Right_front and the right-rear vehicle Right_back, the front vehicle Front, the rear vehicle Back, and the target vehicle information Ego respectively; finally, the fully connected layer is used to fuse the features to obtain the temporal feature S; the residual CNN network is used to extract the spatial feature C of the surrounding environment, and finally the two are fused to obtain the spatiotemporal feature SC; The temporal feature is expressed as follows: S = F lstm (V, R, T, N) Among them, S is the timing feature; F lstm is the LSTM network model; V is the set of surrounding vehicle states; R is the set of lane states; where T is the time step; N is the dimension of the input information; S3: Driving behavior rule extraction: According to the vehicle information V and lane information R of the data collection module, use the driving behavior rule feature mining algorithm to construct the behavior rule feature encoding and obtain the driving behavior rule feature D; S4: Fusion decision-making: Use the WIDE&DEEP fusion model to fuse the driving behavior feature D and the spatiotemporal feature SC to obtain the behavior decision Y of the current vehicle; in step S4, the spatiotemporal feature obtained in the feature extraction stage and the driving behavior rule feature are jointly trained, and finally the current behavior decision result is output, including lane keeping, left lane change, and right lane change, specifically including the following steps: S41: Introduce non-linear cross features through Wide for the mined driving behavior rule feature information D to achieve the memory ability, and obtain the driving behavior rule information WIDE d ; WIDE d = W T * D + b Where W is the weight matrix and b is the offset; S42: Pass the extracted spatiotemporal feature SC through learning the low-dimensional dense vector, and explore the features that do not appear or appear rarely in the historical data through the Deep component; The spatio-temporal feature SC is implemented by LSTM-CNN, DEEP s is the output obtained by processing the spatio-temporal feature SC by the DEEP network component; DEEP s = f(SC), SC = f LSTM-CNN (x) S43: Wide&Deep output: After the forward calculations of the Wide component and the Deep component are completed, the outputs of the two parts are weighted and summed to obtain the vehicle decision result, expressed as: Y = σ(WIDE d + DEEP s + b) Where σ and b are the activation function and the offset respectively, and Y represents the output probability of the n-dimensional behavior decision.

2. The method for autonomous driving vehicle behavior decision-making according to claim 1, characterized in that: The road model described in step S1 is {V, R}, where V is the set of surrounding vehicles {V Ego , V Ego_front , V Ego_back , V Left_front , V Left_back , V Right_front , V Right_back}, which are the target vehicle, the vehicle in front, the vehicle behind, the vehicle in the front left, the vehicle in the rear left, the vehicle in the front right, and the vehicle in the rear right respectively; R is the road information {M, L id , F}, where M is the number of lanes, L id is the current lane label, and F indicates whether lane change is possible in the current lane. For vehicles in the leftmost lane, left lane change is not possible, and for vehicles in the rightmost lane, right lane change is not possible; For each vehicle, the vehicle state information {t, X, Y, Vel, Acc, θ, L, W, L id , X_dis, Y_dis} are the timestamp, lateral position, longitudinal position, speed, acceleration, steering angle, vehicle length, vehicle width, current lane, lateral offset from the target vehicle, and longitudinal offset from the target vehicle respectively; the greater the distance between vehicles, the smaller the interactive influence between vehicles, and a value is used to indicate that there is no vehicle in a certain azimuth of the target vehicle; for the treatment of default vehicles without vehicles around, 0 m / s is used to represent the speed of the default vehicle, and for the lateral offset from the target vehicle and the longitudinal offset from the target vehicle, X_dis and Y_dis are used to represent them.

3. The method for autonomous driving vehicle behavior decision-making according to claim 1, characterized in that: For the driving behavior rule extraction in step S3, the GBDT driving behavior rule mining algorithm is used for mining to obtain the driving behavior feature D; the mined rule attributes include the factors that affect the target vehicle to take different driving behaviors, including the time, lateral coordinate, longitudinal coordinate, speed, acceleration, lateral offset, and longitudinal offset related features of the target vehicle and the surrounding environment vehicles; feature The encoding rule is as follows: assume that the driving behavior rule mining algorithm iterates M times to obtain the final output Then, M decision trees are constructed. For each sample, the value at the leaf node of each decision tree is set to 1, and the rest are set to 0. Then, a set of 0-1 features are constructed using these features.

4. The method for autonomous driving vehicle behavior decision-making according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31: For the training set {(x 1 , y 1 ), (x 2 , y 2 ), (x 3 , y 3 ), …, (x N , y N )}, where x i = {V, R} is the corresponding decision in the current environment, N is the number of samples, and initialize a weak learner: Where L is the squared loss function; S32: Iterate from 1 to M times, and fit a tree each time. For each tree, first calculate the residual of each sample, that is: Update a decision tree iteratively with (x i , r im ). R jm is the leaf node region of the new decision tree, j = 1, 2, 3,.., J m , J m is the total number of leaf nodes of the new decision tree; For leaf nodes j = 1, 2, 3,.., J m Calculate the best fit value: where I is 1 when the node corresponding to the sample is the leaf node region R jm and 0 otherwise; S33: Output the final result: When the final output result of the sample x The smaller the error from the true result, the better the fitting effect.

5. An autonomous vehicle behavior decision-making system, characterized in that: Based on the autonomous vehicle behavior decision-making method described in any one of claims 1-4, the system includes a data acquisition module, a spatio-temporal feature extraction module, a driving behavior rule extraction module, and a fusion decision-making module; The data acquisition module obtains surrounding environment information through sensors and constructs a road model; The spatio-temporal feature extraction module uses a decision-making model based on LSTM-CNN to extract information from the road model and obtain spatio-temporal features; The driving behavior rule extraction module analyzes the interactivity of the vehicle to obtain the features generated by different behaviors, and constructs a driving behavior rule feature mining algorithm based on these features to obtain behavior rule feature codes; The fusion decision-making module uses a WIDE&DEEP fusion model to fuse behavior features and spatio-temporal features, and finally outputs a behavior decision result.

Images