Method and system for predicting behavior of a motor vehicle on a regular road

Abstract

The present application relates to the technical field of vehicle automatic driving, in particular to a method and system for predicting the behavior of a motor vehicle on a regular road. By predicting the driving intention and driving trajectory of the target vehicle, sufficient response time can be reserved for the automatic driving control of the vehicle, ensuring safety and passenger comfort. The driving intention prediction method combines CNN, LSTM and MLP to efficiently extract feature information of the lane and vehicle, and then performs experiential training on the superimposed information model and combines Softmax multi-classification to accurately obtain the intention probability of each driving intention of the target vehicle. The driving trajectory prediction method performs segmented sampling on each driving trajectory and calculates its loss function, combines the intention probability of the driving intention, and obtains the trajectory probability of each driving trajectory. In combination with the Rule-Based method, the driving trajectory with abnormal characteristics is filtered out, and the most probable driving trajectory of the target vehicle can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle autonomous driving, and particularly relates to a method and system for predicting the behavior of a motor vehicle on a conventional road. Background Art

[0002] With the increasing maturity of current autonomous driving technology, autonomous driving vehicles will become more and more popular. In actual driving scenarios, the behaviors of pedestrians and other vehicles on the road are often unable to be accurately grasped. The vehicle in front may suddenly turn or brake, and the vehicle behind may suddenly overtake or have a rear-end collision due to too high speed. These scenarios are often encountered during daily driving. Therefore, autonomous driving vehicles especially need to accurately predict the movement trajectories of each object around the vehicle, and then make corresponding planning and control.

[0003] The existing vehicle autonomous driving control methods can only obtain real-time information such as the position and speed of the target objects around the vehicle based on sensors, and control the acceleration, deceleration, steering, etc. of the vehicle based on the real-time information of the target objects output by the perception fusion end; however, the existing technology cannot predict the future driving intentions and movement trajectories of the target objects, and the response efficiency to the sudden behaviors of the target objects is relatively low, making it difficult to avoid driving risks in advance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: aiming at the deficiencies of the existing technology, to provide a method and system for predicting the behavior of a motor vehicle on a conventional road, which can predict the driving trajectory of the target vehicle in advance, and then reserve sufficient response time for the vehicle autonomous driving control, ensuring safety and the comfort of passengers at the same time.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is:

[0006] I. A method for predicting the behavior of a motor vehicle on a conventional road

[0007] The present invention provides a method for predicting the behavior of a motor vehicle on a conventional road, which specifically includes predicting the driving intention of the target vehicle and predicting the driving trajectory of the target vehicle.

[0008] 1) The prediction of the driving intention of the target vehicle specifically includes the following steps:

[0009] S11, lane centerline feature extraction: adopting an image processing method based on a CNN convolutional neural network, extracting the feature information of each lane centerline from the road network image information obtained by the vision sensor;

[0010] S12, Motion state feature extraction of the target vehicle and its surrounding environment vehicles: The temporal feature extraction method based on LSTM recurrent neural network is adopted to extract the motion state feature information of the target vehicle and its surrounding environment vehicles from the motion state information of the target vehicle and its surrounding environment vehicles obtained by the environmental perception sensor.

[0011] S13, Overall lane network feature acquisition: The feature information of each lane centerline described in step S11 is processed by an MLP multi-layer recurrent neural network to obtain the MLP unit of each lane centerline feature information, and then processed by a GlobalPooling global pooling layer to output the overall lane network feature information.

[0012] S14, Overlay of lane and vehicle feature information: Overlay the motion state feature information of the target vehicle and its surrounding environment vehicles described in step S12 with the overall road network feature information of the lane described in step S13 to output a complete three-dimensional information model of the lane, and then use observation data of real vehicle behavior to perform empirical training on the model.

[0013] S15, Driving Intention Classification and Intention Probability Determination: The Softmax multi-classification method is used to classify driving intentions (lanes) in the complete three-dimensional information model of the lane described in step S14, and output the intention probability of each driving intention (each lane).

[0014] Optionally, the image processing method based on CNN convolutional neural network specifically involves: using road network image information as input matrix M, and using the feature information of a single lane centerline as convolution kernel filtering matrix F, performing CNN convolution operation with a preset stride S to obtain output matrix O, and extracting the feature information Feature_lane of each lane centerline from the output matrix O. The feature information of the lane centerline specifically includes: the lateral distance from the lane centerline to the lane boundary.

[0015]

[0016] In the formula, P is the number of zero-padding layers in the convolution operation.

[0017] Optionally, the temporal feature extraction method based on LSTM recurrent neural network specifically involves using the motion state information of the target vehicle and its surrounding vehicles as the input sequence X. t By using LSTM temporal feature extraction processing, an output sequence Y containing both time and spatial dimensions is obtained. t and from the output sequence Y tExtract the motion state feature information (Feature_target) of the target vehicle and its surrounding vehicles in each time period. The motion state feature information specifically includes: the lateral distance from the vehicle to the lane line boundary, the vehicle heading angle, the historical time interval, the lateral velocity and acceleration, and the longitudinal velocity and acceleration.

[0018] Y t =LSTM(X) t ×W+Y t-1 ×V+b)

[0019] In the formula, W and V are the weighting coefficients of the input sequence and the output sequence, respectively, and b is the correction coefficient.

[0020] Optionally, the GlobalPooling global pooling layer processing specifically includes:

[0021]

[0022] In the formula, a j T is the input set of MLP units for the centerline feature information of each lane. j The output set is Feature_global, which contains overall road network feature information for lanes. R is the number of roads, and d is the pooling coefficient.

[0023] Optionally, the complete three-dimensional information model of the lane includes historical information of the target vehicle, historical information of vehicles in the surrounding environment, and overall lane information;

[0024] Lane complete 3D information model E all =(Feature_target, Feature_global)

[0025] Optionally, the Softmax multi-classification method is as follows:

[0026]

[0027] In the formula, j specifically refers to 1, 2, and 3, prior1, prior2, and prior3 are the probabilities that the target vehicle intends to change lanes to the left, go straight, and change lanes to the right, respectively, E1, E2, and E3 are the three-dimensional information of the target vehicle's left lane, current lane, and right lane, respectively, and K is the number of categories in the multi-class classifier, K = 3.

[0028] 2) The prediction of the target vehicle's driving trajectory specifically includes the following steps:

[0029] S21, select N destination locations on each lane;

[0030] S22 uses Bézier curves to generate the driving trajectory at each endpoint;

[0031] S23, perform segmented sampling on each driving trajectory, and determine the acceleration acc of the target vehicle at each sampling point i. i Centripetal acceleration v i 2 r i -1 Collision distance d with the nearest surrounding vehicles i ;

[0032] S24, Calculate the loss function for each driving trajectory. The formula for calculating the loss function is as follows:

[0033] C=θ1C acc +θ2C centripetal_acc +θ3C collision

[0034] in:

[0035]

[0036]

[0037]

[0038] In the formula, n is the number of sampling points, θ1, θ2, and θ3 are the acceleration correction coefficient, centripetal acceleration correction coefficient, and collision distance correction coefficient, respectively, all of which are obtained through offline debugging or experimental verification, and Z1 and Z2 are regularization terms.

[0039] S25, based on the intent probability of each driving intent and the loss function of each driving trajectory, calculate the trajectory probability of the target vehicle selecting each driving trajectory, as shown in the following formula:

[0040]

[0041] In the formula, prior is the probability of intent in the lane where the end point of each driving trajectory is located, and z is the trajectory correction coefficient;

[0042] S26. Based on the Rule-Based method, driving trajectories with abnormal features are filtered out, and then the driving trajectory with the highest probability is selected as the predicted driving trajectory of the target vehicle.

[0043] II. A Motor Vehicle Behavior Prediction System

[0044] Based on the same inventive concept, this invention also provides a motor vehicle behavior prediction system for implementing the prediction method described above, mainly comprising:

[0045] 1) A visual sensor for acquiring road network image information, specifically including: lane line position information, road edge line position information, and lane center line position information;

[0046] 2) Environmental perception sensor, used to acquire motion state information of the target vehicle and vehicles in its surrounding environment, the motion state information specifically including: vehicle speed information, acceleration information, angle information between the vehicle and the lane line, and distance information between the vehicle and the lane line;

[0047] 3) The vehicle controller is used to determine the vehicle speed and receive information from the vision sensor and the environmental perception sensor to perform calculations and judgments in order to execute the prediction method described above.

[0048] Compared with the prior art, the present invention has the following main advantages:

[0049] 1. This invention proposes a method and system for predicting the behavior of motor vehicles on conventional roads. By predicting the driving intention and trajectory of the target vehicle, sufficient response time can be reserved for the autonomous driving control of the vehicle, ensuring both safety and passenger comfort.

[0050] 2. The driving intention prediction method adopted in this invention combines CNN convolutional neural network, LSTM recurrent neural network and MLP multilayer recurrent neural network to efficiently extract lane centerline feature information, motion state feature information of target vehicle and surrounding vehicles and overall road network feature information. Then, by empirically training the three-dimensional information model generated by superimposing all feature information and combining it with the Softmax multi-classification method, the intention probability of each driving intention of the target vehicle can be accurately obtained.

[0051] 3. The driving trajectory prediction method adopted in this invention samples each driving trajectory in segments and calculates the loss function of each driving trajectory. Combined with the intent probability of each driving intention, the trajectory probability of each driving trajectory can be obtained. Then, by combining the rule-based method to filter out driving trajectories with abnormal features, the driving trajectory with the highest probability of the target vehicle can be accurately predicted. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a target vehicle driving intention prediction scenario in an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of the driving intention prediction method in an embodiment of the present invention;

[0054] Figure 3 This is a model diagram of the driving intention prediction method in an embodiment of the present invention;

[0055] Figure 4This is a schematic diagram of a target vehicle trajectory prediction scenario in an embodiment of the present invention;

[0056] Figure 5 This is a flowchart of the driving trajectory prediction method in an embodiment of the present invention; Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0058] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0059] Example 1: This example provides a method for predicting the behavior of a motor vehicle on a conventional road. It is used to predict the driving intention and trajectory of a target vehicle on a conventional road (straight lane), thereby reserving sufficient response time for the vehicle's autonomous driving control, ensuring both safety and passenger comfort.

[0060] The prediction method specifically includes predicting the target vehicle's driving intention and predicting the target vehicle's driving trajectory.

[0061] like Figure 1 As shown, on a regular road (straight lane), the target vehicle predicted in this embodiment has three driving intentions: changing lanes to the left, going straight, and changing lanes to the right (lane 1, lane 2, lane 3). This application uses a driving intention prediction method to determine the probability of each driving intention of the target vehicle.

[0062] like Figures 2-3 As shown, the method for predicting the driving intention of the target vehicle specifically includes the following steps:

[0063] S11, Lane centerline feature extraction: An image processing method based on CNN convolutional neural network is used to extract the feature information of each lane centerline from the road network image information obtained by the visual sensor;

[0064] S12, Motion state feature extraction of the target vehicle and its surrounding environment vehicles: The temporal feature extraction method based on LSTM recurrent neural network is adopted to extract the motion state feature information of the target vehicle and its surrounding environment vehicles from the motion state information of the target vehicle and its surrounding environment vehicles obtained by the environmental perception sensor.

[0065] S13, Overall lane network feature acquisition: The feature information of each lane centerline described in step S11 is processed by an MLP multi-layer recurrent neural network to obtain the MLP unit of each lane centerline feature information, and then processed by a GlobalPooling global pooling layer to output the overall lane network feature information.

[0066] S14, Overlay of lane and vehicle feature information: Overlay the motion state feature information of the target vehicle and its surrounding environment vehicles described in step S12 with the overall road network feature information of the lane described in step S13 to output a complete three-dimensional information model of the lane, and then use observation data of real vehicle behavior to perform empirical training on the model.

[0067] S15, Driving Intention Classification and Intention Probability Determination: The Softmax multi-classification method is used to classify driving intentions (lanes) in the complete three-dimensional information model of the lane described in step S14, and output the intention probability of each driving intention (each lane).

[0068] The image processing method based on CNN convolutional neural network specifically involves: using road network image information as input matrix M, and using the feature information of a single lane centerline as convolution kernel filtering matrix F, performing CNN convolution operation with a preset stride S to obtain output matrix O, and extracting the feature information Feature_lane of each lane centerline from the output matrix O. The feature information of the lane centerline specifically includes: the lateral distance from the lane centerline to the lane boundary.

[0069]

[0070] In the formula, P is the number of zero-padding layers in the convolution operation.

[0071] Furthermore, the temporal feature extraction method based on LSTM recurrent neural network specifically involves using the motion state information of the target vehicle and its surrounding vehicles as the input sequence X. t By using LSTM temporal feature extraction processing, an output sequence Y containing both time and spatial dimensions is obtained. t and from the output sequence Y t Extract the motion state feature information (Feature_target) of the target vehicle and its surrounding vehicles in each time period. The motion state feature information specifically includes: the lateral distance from the vehicle to the lane line boundary, the vehicle heading angle, the historical time interval, the lateral velocity and acceleration, and the longitudinal velocity and acceleration.

[0072] Y t =LSTM(X) t ×W+Y t-1 ×V+b)

[0073] In the formula, W and V are the weighting coefficients of the input sequence and the output sequence, respectively, and b is the correction coefficient;

[0074] Using LSTM has the advantages of recurrent neural networks: it can pass the output of the current layer neuron to the next layer, and it can also pass it to itself for use; it also improves the problem that gradient vanishing and gradient explosion cause the weights of shallower networks in the neural network to not be updated.

[0075] Furthermore, during neural network training, numerous training data points are input into the model. Each data point consists of raw data and a corresponding label. For example, the input data might be an image, and the label might be a symbol containing a car or another symbol. The way a neural network learns from data is called backpropagation. First, the neural network receives the input and produces an output. Then, the computer compares the error between the output and the true value. This error is then propagated back to the entire network through backpropagation. The hidden layers adjust their intermediate values ​​(weights) based on the observed differences, thereby improving the accuracy of the neural network. A recurrent neural network with such a multi-layered structure is called an MLP unit.

[0076] The GlobalPooling global pooling layer processing specifically includes:

[0077]

[0078] In the formula, a j T is the input set of MLP units for the centerline feature information of each lane. j The output set is Feature_global, which contains overall road network feature information for lanes. R is the number of roads, and d is the pooling coefficient.

[0079] Furthermore, the complete three-dimensional information model of the lane includes historical information of the target vehicle, historical information of vehicles in the surrounding environment, and overall lane information;

[0080] Lane complete 3D information model E all =(Feature_target, Feature_global)

[0081] The vehicle history information (motion state feature information of the target vehicle and its surrounding environment vehicles) and the overall lane information (overall lane network feature information) are used as inputs;

[0082] This 3D information model is used to provide the probability that a target vehicle may take each lane sequence. To enable the model to learn new behaviors, it is empirically trained using actual observation data. During training, real vehicle behaviors are provided to the model, including not only the lane segments and the state of the object, but also which lane sequence the object ultimately chooses. As the records increase over time, the model can iteratively update itself, and the accuracy will continuously improve. Each record will consist of the lane segment sequence followed by the observed object and the object's related state. At each time point, the object occupies a segment and has a specific state. The entire record consists of a series of lane segments and the object's related states.

[0083] Furthermore, the Softmax multi-classification method is as follows:

[0084]

[0085] In the formula, j specifically refers to 1, 2, and 3, prior1, prior2, and prior3 are the probabilities that the target vehicle intends to change lanes to the left, go straight, or change lanes to the right, respectively, E1, E2, and E3 are the three-dimensional information of the target vehicle's left lane, current lane, and right lane, respectively, and K is the number of categories in the multi-class classifier. In this example, K = 3.

[0086] like Figure 4 As shown, each driving intention (lane 1, lane 2, lane 3) has multiple endpoint positions (endpoint positions A1, B1, C1, A2, B2, C2, A3, B3, C3), and each endpoint position corresponds to a different driving trajectory (dashed line, thin line, thick line). This application uses a driving trajectory prediction method to predict the most likely driving trajectory of the target vehicle.

[0087] like Figure 5 As shown, the method for predicting the trajectory of the target vehicle specifically includes the following steps:

[0088] S21, select N destination locations on each lane;

[0089] S22 uses Bézier curves to generate the driving trajectory at each endpoint;

[0090] S23, perform segmented sampling on each driving trajectory, and determine the acceleration acc of the target vehicle at each sampling point i. i Centripetal acceleration v i 2 r i -1 Collision distance d with the nearest surrounding vehicles i ;

[0091] The adoption of these three dimensions is based on the following considerations: appropriate acceleration and low centripetal acceleration can make the vehicle more comfortable; calculating the collision distance is for safety considerations; and combining safety and comfort is consistent with the actual performance of the vehicle.

[0092] S24, Calculate the loss function for each driving trajectory. The formula for calculating the loss function is as follows:

[0093] C=θ1C acc +θ2C centripetal_acc +θ3C collision

[0094] in:

[0095]

[0096]

[0097]

[0098] In the formula, n is the number of sampling points, θ1, θ2, and θ3 are the acceleration correction coefficient, centripetal acceleration correction coefficient, and collision distance correction coefficient, respectively, all of which are obtained through offline debugging or experimental verification, and Z1 and Z2 are regularization terms.

[0099] S25, based on the intent probability of each driving intent and the loss function of each driving trajectory, calculate the trajectory probability of the target vehicle selecting each driving trajectory, as shown in the following formula:

[0100]

[0101] In the formula, prior is the probability of intent in the lane where the end point of each driving trajectory is located, and z is the trajectory correction coefficient.

[0102] S26. Based on the Rule-Based method, driving trajectories with abnormal features are filtered out, and then the driving trajectory with the highest probability is selected as the predicted driving trajectory of the target vehicle.

[0103] For example, targets with longitudinal acceleration exceeding 3 m / s² (calibrable), lateral acceleration exceeding 2 m / s² (calibrable), or excessively small trajectory curvature radius can be filtered out.

[0104] The trajectory predicted by this method aligns with objective realities such as comfort and safety during driving, demonstrating good predictive performance and accuracy. Furthermore, based on the prediction results, the autonomous vehicle can steer, accelerate, or brake in advance, anticipating the actions of the target vehicle and improving the safety and intelligence of autonomous vehicles.

[0105] Example 2, based on the same inventive concept, also provides a motor vehicle behavior prediction system for implementing the prediction method described above, mainly including:

[0106] 1) A visual sensor for acquiring road network image information, specifically including: lane line position information, road edge line position information, and lane center line position information;

[0107] 2) Environmental perception sensor, used to acquire motion state information of the target vehicle and vehicles in its surrounding environment, the motion state information specifically including: vehicle speed information, acceleration information, angle information between the vehicle and the lane line, and distance information between the vehicle and the lane line;

[0108] 3) The vehicle controller is used to determine the vehicle speed and receive information from the vision sensor and the environmental perception sensor to perform calculations and judgments in order to execute the prediction method described above.

[0109] Example 3: Based on the same inventive concept, this example also provides a manual-automatic vehicle, which is equipped with the motor vehicle behavior prediction system described above.

[0110] Furthermore, all parts of this application that are not described in detail are the same as or implemented using existing technology.

[0111] In summary:

[0112] 1. This invention proposes a method and system for predicting the behavior of motor vehicles on conventional roads. By predicting the driving intention and trajectory of the target vehicle, sufficient response time can be reserved for the autonomous driving control of the vehicle, ensuring both safety and passenger comfort.

[0113] 2. The driving intention prediction method adopted in this invention combines CNN convolutional neural network, LSTM recurrent neural network and MLP multilayer recurrent neural network to efficiently extract lane centerline feature information, motion state feature information of target vehicle and surrounding vehicles and overall road network feature information. Then, by empirically training the three-dimensional information model generated by superimposing all feature information and combining it with the Softmax multi-classification method, the intention probability of each driving intention of the target vehicle can be accurately obtained.

[0114] 3. The driving trajectory prediction method adopted in this invention samples each driving trajectory in segments and calculates the loss function of each driving trajectory. Combined with the intent probability of each driving intention, the trajectory probability of each driving trajectory can be obtained. Then, by combining the rule-based method to filter out driving trajectories with abnormal features, the driving trajectory with the highest probability of the target vehicle can be accurately predicted.

[0115] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the behavior of motor vehicles on conventional roads, characterized in that: This includes methods for predicting the driving intention of the target vehicle and methods for predicting the driving trajectory of the target vehicle. The method for predicting the driving intention of the target vehicle includes the following strategies: An image processing method based on CNN convolutional neural networks is used to extract the feature information of the center line of each lane from the road network image information obtained by the visual sensor; A temporal feature extraction method based on LSTM recurrent neural network is adopted to extract the motion state feature information of the target vehicle and its surrounding environment vehicles from the motion state information of the target vehicle and its surrounding environment vehicles obtained by environmental perception sensors. The feature information of each lane centerline is processed by MLP multi-layer recurrent neural network and GlobalPooling global pooling layer to output the overall road network feature information of the lanes; The motion state feature information of the target vehicle and its surrounding vehicles is superimposed with the overall road network feature information of the lane to output a complete three-dimensional information model of the lane. The three-dimensional information model is then empirically trained using observation data of real vehicle behavior. The Softmax multi-classification method is used to classify driving intentions in the three-dimensional information model, and the probability of each driving intention is output. The method for predicting the trajectory of the target vehicle includes the following strategies: Multiple endpoint locations are selected on each lane, and a driving trajectory for each endpoint location is generated using Bézier curves; Each driving trajectory is sampled in segments to determine the acceleration acc of the target vehicle at each sampling point i. i Centripetal acceleration v i 2 r i -1 Collision distance d with the nearest surrounding vehicles i ; The loss function C for each driving trajectory is calculated using the following formula: C = θ1C acc + θ2C centripetal_acc +θ3C collision in: C acc = C centripetal_acc = C collision = In the formula, n is the number of sampling points, θ1, θ2, and θ3 are the acceleration correction coefficient, centripetal acceleration correction coefficient, and collision distance correction coefficient, respectively, all of which are obtained through offline debugging or experimental verification, and Z1 and Z2 are regularization terms; Based on the probability of each driving intention and the loss function of each driving trajectory, the trajectory probability of the target vehicle selecting each driving trajectory is calculated, as follows: TrackSelect= * prior * In the formula, Let z be the probability of the lane where the end point of each driving trajectory is located, and z be the trajectory correction coefficient. The rule-based method filters out driving trajectories with abnormal features, and then selects the driving trajectory with the highest probability as the predicted driving trajectory of the target vehicle.

2. The method for predicting the behavior of a motor vehicle on a conventional road according to claim 1, characterized in that, The image processing method based on CNN convolutional neural network is as follows: road network image information is used as input matrix M, and the feature information of the center line of a single lane is used as convolution kernel filtering matrix F. CNN convolution operation is performed in combination with a preset stride S to obtain output matrix O, and feature information of the center line of each lane is extracted from the output matrix O. The CNN convolution operation formula is as follows: In the formula, P is the number of zero-padding layers in the convolution operation.

3. The method for predicting the behavior of a motor vehicle on a conventional road according to claim 1, characterized in that, The temporal feature extraction method based on LSTM recurrent neural network specifically involves using the motion state information of the target vehicle and its surrounding vehicles as the input sequence X. t By using LSTM temporal feature extraction processing, an output sequence Y containing both time and spatial dimensions is obtained. t and from the output sequence Y t Extract the motion state characteristics of the target vehicle and its surrounding vehicles in each time period; The LSTM temporal feature extraction formula is as follows: In the formula, W and V are the weighting coefficients of the input sequence and the output sequence, respectively, and b is the correction coefficient.

4. The method for predicting the behavior of a motor vehicle on a conventional road according to claim 3, characterized in that, The motion state characteristic information specifically includes: the lateral distance from the vehicle to the lane line boundary, the vehicle heading angle, the historical time interval, the lateral velocity and acceleration, and the longitudinal velocity and acceleration.

5. The method for predicting the behavior of a motor vehicle on a conventional road according to claim 1, characterized in that, The GlobalPooling global pooling layer processing uses the following formula: In the formula, a j T is the input set of MLP units for the centerline feature information of each lane. j The output set contains overall road network feature information for lanes, where R is the number of roads and d is the pooling coefficient.

6. The method for predicting the behavior of a motor vehicle on a conventional road according to claim 1, characterized in that, The complete three-dimensional information model of the lane E all = (Feature_target, Feature_global), where Feature_target is the motion state feature information of the target vehicle and its surrounding vehicles, and Feature_global is the overall road network feature information of the lane; The specific formula for the Softmax multi-classification method is as follows: In the formula, j specifically refers to 1, 2, or 3. E1, E2, and E3 represent the probabilities that the target vehicle intends to change lanes to the left, go straight, or change lanes to the right, respectively. E1, E2, and E3 represent the three-dimensional information of the target vehicle's left lane, current lane, and right lane, respectively. K is the number of classes in the multi-class classifier, and K=3.

7. A motor vehicle behavior prediction system, used to implement the prediction method as described in any one of claims 1 to 6, characterized in that, include: A visual sensor is used to acquire road network image information, which specifically includes: lane line position information, road edge line position information, and lane center line position information. An environmental perception sensor is used to acquire motion state information of the target vehicle and other vehicles in its surrounding environment. The motion state information specifically includes: vehicle speed information, acceleration information, the angle between the vehicle and the lane line, and the distance between the vehicle and the lane line. A vehicle controller is used to determine the vehicle speed and to receive information from visual sensors and environmental perception sensors to perform calculations and judgments in order to execute the prediction method as described in any one of claims 1 to 6.

8. A non-transitory readable storage medium having a program stored thereon, characterized in that, When the program is executed by the motor vehicle behavior prediction system, it implements the prediction method as described in any one of claims 1 to 6.

9. A type of automatic transmission vehicle, characterized in that: Includes the motor vehicle behavior prediction system as described in claim 7.

Images