A method for jointly predicting driving intention and trajectory of surrounding vehicles by an autonomous vehicle

Abstract

A method for an autonomous driving vehicle to jointly predict the driving intention and trajectory of a surrounding vehicle, involving intelligent driving. 1) Processing of surrounding vehicle driving data: Collecting surrounding vehicle driving data, establishing rules to define the driving intention type of the surrounding vehicle, setting labels for each sequence data, and obtaining the data required to realize vehicle motion prediction; 2) Joint prediction of surrounding vehicle driving intention and trajectory: Constructing an intention recognition model based on BiLSTM to recognize the driving intention of the surrounding vehicle; Constructing a trajectory prediction model based on the encoder-decoder framework of LSTM, introducing an attention mechanism to improve the prediction effect of time series problems, and combining the position information of the surrounding vehicle and the driving intention to predict the longer time domain trajectory of the surrounding vehicle; 3) Providing decision information: Providing information basis for the subsequent operation of the main vehicle based on the results of the joint prediction. Save computing costs and improve the accuracy of long-term trajectory prediction; Reduce the impact of deep network gradient disappearance and network degradation problems. Provide decision information for vehicles to ensure safety and comfort.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving technology, and in particular to a method for an automatic driving vehicle to jointly predict the driving intention and trajectory of surrounding vehicles. Background Art

[0002] In recent years, autonomous vehicles have attracted more and more attention as an effective way to improve traffic safety and alleviate energy and environmental problems in intelligent transportation systems. Autonomous vehicles need to have the ability to predict the future movement of surrounding vehicles in order to plan safe driving paths in advance and effectively reduce traffic accidents. In the context of autonomous driving, early prediction of the movement of vehicles around the autonomous vehicle is a key factor in ensuring a high level of road safety. There are two main types of vehicle motion prediction: driving intention recognition and trajectory prediction.

[0003] With the rapid development of artificial intelligence, the latest progress in deep learning provides a powerful tool for solving the problem of vehicle motion prediction, and the vehicle motion prediction method based on deep learning has become mainstream. Reference [1] (Ding W, etc. Predicting Vehicle Behaviors Over An Extended Horizon Using Behavior Interaction Network, 2019 International Conference on Robotics and Automation (ICRA), Montreal, Canada. 2019) proposed a new vehicle behavior interaction network based on recurrent neural network (RNN) for vehicle interaction modeling to predict the cutting intention of surrounding vehicles. However, intention recognition cannot obtain clear trajectory information. Reference [2] (Yan J, etc. Trajectory Prediction for Intelligent Vehicles Using Spatial-Attention Mechanism [J]. IET Intelligent Transport Systems, 2020, 14 (13).) proposed two spatial attention mechanisms based on the LSTM encoder-decoder framework to achieve trajectory prediction. Although the prediction accuracy of the trajectory is improved, the effect of long-term trajectory prediction still has room for improvement because it only uses displacement and speed as prediction inputs. In addition, the above research only considers driving intention recognition or trajectory prediction for vehicle motion prediction, and the decision-making information provided to the subsequent main vehicle is relatively simple. Summary of the invention

[0004] The purpose of the present invention is to address the problems existing in the above-mentioned background technology, and to provide a method for an autonomous driving vehicle to jointly predict the driving intention and driving trajectory of a surrounding vehicle by using a deep learning network to collect the driving information of the surrounding vehicle equipped by the sensor of the autonomous driving vehicle, and to provide the vehicle with rich decision-making information to ensure the driving safety of the vehicle. The method considers the driving intention characteristics and introduces the attention mechanism to predict the trajectory, which can ensure the effect of long-term trajectory prediction; in addition, the joint prediction method can simultaneously obtain the driving intention of the surrounding vehicle and the subsequent trajectory of the surrounding vehicle in a longer time domain, providing the vehicle with rich decision-making information and ensuring the safety and comfort of the autonomous driving vehicle.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The present invention comprises the following steps:

[0007] 1) Processing of vehicle driving data: Process the vehicle driving data collected by the vehicle’s onboard sensors such as cameras and millimeter-wave radars, establish rules to define the driving intention type of the vehicle, and set labels for each sequence data to obtain the data required to achieve vehicle motion prediction;

[0008] 2) Joint prediction of the driving intention and trajectory of the surrounding vehicle: Construct an intention recognition model based on BiLSTM to identify the driving intention of the surrounding vehicle; construct a trajectory prediction model based on the LSTM encoder-decoder framework, and introduce an attention mechanism to improve the prediction effect of time series problems. At the same time, combine the location information and driving intention of the surrounding vehicle to predict the trajectory of the surrounding vehicle in a longer time domain;

[0009] 3) Providing decision-making information: Providing rich information basis for subsequent operations of the main vehicle based on the results of the joint prediction.

[0010] In step 1), the specific method of processing the peripheral vehicle travel data is:

[0011] (1) The driving data of the surrounding vehicle collected by the vehicle-mounted sensors such as the camera and millimeter-wave radar mainly include: the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration between the surrounding vehicle and the main vehicle, the azimuth angle of the surrounding vehicle relative to the main vehicle, and the azimuth angle change rate;

[0012] (2) Fill in the missing values ​​of the data. Use the nearest neighbor filling method to fill in single missing values, use the interpolation method to fill in multiple consecutive missing values, use the 3-σ rule to remove outliers in the data, and finally use the Savitzky-Golay filter to filter the data to obtain a relatively smooth data curve;

[0013] (3) Define a coordinate system to conveniently describe the position of the vehicle and calculate the position coordinates of the main vehicle and the surrounding vehicles at each moment;

[0014] (4) The driving intention of the vehicle is defined as changing lanes to the left, changing lanes to the right, and going straight, and an intention label is set for each segment of driving sequence data.

[0015] In step 2), the specific method for identifying the driving intention of the surrounding vehicle is:

[0016] (1) Driving intention recognition is a classification problem. The driving intention recognition model classifies the input multi-feature and multi-step long time series data.

[0017] (2) A driving intention recognition model for the surrounding vehicle is constructed based on the BiLSTM network. The BiLSTM network is composed of a forward LSTM and a backward LSTM. It can make full use of contextual information to improve the accuracy of time series prediction results.

[0018] (3) Inspired by ResNet, shortcut connections are added to the Bi-LSTM network to implement gradient stride calculation, which can effectively reduce the impact of gradient vanishing and network degradation problems in deep networks.

[0019] (4) Since the recognition of intention is different from trajectory prediction and requires the specific position information of the vehicle, the present invention uses eight features such as the surrounding vehicle driving data processed by step 1) (the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration between the surrounding vehicle and the main vehicle, the azimuth angle of the surrounding vehicle relative to the main vehicle, and the azimuth angle change rate) as the input of the intention recognition model, without considering the position information of the surrounding vehicle. This not only captures the main features of the driving intention, ensures the prediction accuracy, but also saves the computing cost.

[0020] (5) The eight feature sequences are input into the driving intention recognition model, and the driving intention probability vector of the surrounding vehicle is output. The driving intention with the highest probability is the final recognition result.

[0021] In step 2), the specific method for predicting the trajectory of the surrounding vehicle over a longer time domain is:

[0022] (1) In order to capture the temporal dependencies in the trajectory, a trajectory prediction model is established using an LSTM-based encoder-decoder framework, which is the most advanced architecture for handling trajectory prediction problems;

[0023] (2) The output of trajectory prediction is the specific location coordinates of the vehicle. In addition to the eight input features of the intention recognition model, the input of trajectory prediction will also add the location coordinate information of the surrounding vehicle trajectory;

[0024] (3) Add shortcut connections in the LSTM decoder;

[0025] (4) The LSTM encoder encodes the input historical feature information sequence into a fixed-length context vector. The LSTM decoder uses a soft attention mechanism during decoding. The dot product is used to calculate the attention value of the decoder hidden state at the previous moment and the hidden state of each historical moment in the encoder. The attention score of the hidden state at each historical moment is then obtained through the SoftMax function. Finally, the updated context vector is obtained through weighted summation.

[0026] (5) The LSTM decoder combines the driving intention probability vector obtained by the intention recognition model, the updated context vector and the trajectory position output predicted by the decoder at the previous time step as the input at the current moment to predict the trajectory position at the current time step.

[0027] In step 3), the specific method of providing decision information is as follows: through the joint prediction of driving intention and trajectory, the current driving intention of the surrounding vehicle and its trajectory position in the subsequent period of time are obtained; according to the identified driving intention, its behavior in a short period of time can be roughly judged, and reaction time is reserved for the main vehicle operation, thereby reducing the risk of vehicle collision and ensuring driving safety; the main vehicle makes path planning in advance according to the predicted trajectory position of the surrounding vehicle in the subsequent period of time, and obtains the optimal operation option to reach the destination, which not only ensures driving comfort but also reduces useless energy consumption.

[0028] The beneficial effects of the present invention are:

[0029] 1. The present invention uses a deep learning network to jointly predict the driving intention and trajectory of the surrounding vehicle, and can obtain the driving intention of the surrounding vehicle at the current moment and the subsequent trajectory information in a longer time domain: the recognition of driving intention provides the main vehicle with the current driving status of the surrounding vehicle, reserves operation reaction time for the driver, and ensures the safety of the main vehicle driving in the short term; the subsequent trajectory prediction in a longer time domain provides important information for the long-term decision-making planning of the main vehicle, which not only ensures the driving comfort of the main vehicle but also saves unnecessary energy consumption.

[0030] 2. Compared with traditional trajectory motion input using displacement and velocity, the present invention improves the prediction effect by utilizing additional information (acceleration, directional angle and directional angular velocity), and rationally selects different inputs for prediction according to the different characteristics of intention recognition and trajectory prediction, thereby saving computing costs and improving prediction efficiency.

[0031] 3. Since the driving intention of the vehicle is an important factor affecting the future driving trajectory of the vehicle, the present invention considers the driving intention characteristics for trajectory prediction. Based on the LSTM encoder-decoder framework, a soft attention mechanism is designed to focus on the impact of the historical information of the target vehicle at different times on its future trajectory, thereby improving the accuracy of long-term trajectory prediction. At the same time, the advantages of the ResNet network are learned, and shortcut connections are added to the prediction network, which can effectively reduce the impact of the deep network gradient disappearance and network degradation problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the scheme framework of the present invention.

[0033] Figure 2 It is a schematic diagram of the coordinate system of the present invention.

[0034] Figure 3 This is a schematic diagram of vehicle driving intention classification according to the present invention.

[0035] Figure 4 This is a schematic diagram of a model for joint prediction of the driving intention and trajectory of a surrounding vehicle according to the present invention. DETAILED DESCRIPTION

[0036] The following embodiments will further illustrate the present invention in conjunction with the accompanying drawings.

[0037] The scheme framework of the embodiment of the present invention is as follows Figure 1 As shown in the figure, it consists of three parts: data processing, joint prediction of surrounding vehicle intention recognition and trajectory prediction, and providing decision information, including the following steps:

[0038] Step 1: The data processing steps are as follows:

[0039] Step 1.1: Obtain the driving data of the surrounding vehicle collected by the vehicle-mounted sensors such as the camera and millimeter-wave radar equipped on the vehicle: the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration between the surrounding vehicle and the main vehicle, the azimuth angle of the surrounding vehicle relative to the main vehicle, and the azimuth angle change rate.

[0040] Step 1.2: For a single missing value in the data, the nearest neighbor filling method is used to fill it with the data of the previous time step or the next time step; for multiple consecutive missing values, the interpolation method is used to calculate the average of the previous and next values ​​of the missing value to fill it. For outliers in the data, the 3-σ criterion is used to remove the outliers, calculate the standard deviation σ and mean μ of each feature, and remove data with values ​​outside the interval (μ-3σ, μ+3σ). Finally, the Savitzky-Golay filter is used to filter the data. M sampling points are taken near the original data x(i), and a window of 2M+1 sampling points is constructed around x to fit the p-order polynomial y(i), which is expressed as follows:

[0041]

[0042]

[0043]

[0044] Wherein, y(i) is the processed data, i=-M, ..., 0, ..., M; E is the total sum of squared errors; p≤2M.

[0045] Step 1.3: Define the coordinate system, Figure 2 Schematic diagram of the coordinate system of the present invention; at the time t, the origin is fixed on the main vehicle, the x-axis points to the direction of motion of the highway, and the y-axis points to the direction perpendicular to it. This makes the data acquisition of the present invention better match the on-board sensors on the autonomous vehicle.

[0046] Step 1.4: According to the literature [3] (Toledo T, etc. Modeling duration of lane changes [J]. Transportation Research Record, 2007, 1999 (1).), the average duration of lane changes in urban areas is 5 seconds, and the average duration on highways is 5.8 seconds. Taking all these factors into consideration, the lane change time is set to 5.4 seconds. Figure 3 This is a schematic diagram of the vehicle driving intention classification of the present invention. As shown in the figure, the intersection of the vehicle trajectory and the left road dotted line is the left lane change point. The position of the vehicle at 2.7 seconds in the reverse time direction of the left lane change point is defined as the left lane change starting point, and the position of the vehicle at 2.7 seconds in the forward time direction of the left lane change point is defined as the left lane change end point; the left lane change segment in the trajectory sequence is defined as the left lane change intention; the right lane change intention is defined based on the same method; if the vehicle trajectory sequence does not contain a lane change segment, this trajectory segment is defined as the lane keeping intention. According to this rule, an intention label is set for each segment of driving data.

[0047] Step 2: The model diagram of joint prediction is as follows Figure 4 As shown in the figure, the driving intention and trajectory of the surrounding vehicle are jointly predicted. The implementation steps are as follows:

[0048] Step 2.1: Build an intention recognition model based on the BiLSTM network. The steps to identify the driving intention of the vehicle are as follows:

[0049] Step 2.1.1: Considering that a complete lane change takes about 5 seconds and the data acquisition frequency is 10Hz, the length of the input sequence is selected as 50. The vehicle driving intention is marked as changing lanes to the left, changing lanes to the right, and going straight. The vehicle intention recognition is a classification problem, and the input sequence needs to be labeled accordingly. The present invention uses 0, 1, and 2 to represent the above three intentions respectively.

[0050] Step 2.1.2: The input of the intention recognition model uses 8 features including the surrounding vehicle driving data processed in step 1) (the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration between the surrounding vehicle and the main vehicle, the azimuth angle of the surrounding vehicle relative to the main vehicle, and the azimuth angle change rate), as shown in the following formula:

[0051]

[0052] In the formula, X t is the input feature of the target vehicle at time t, t=1, 2, ..., 50. are the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, and lateral relative acceleration between the target vehicle and the main vehicle, θ t , They are the azimuth angle of the target vehicle relative to the main vehicle and the azimuth angle change rate.

[0053] Step 2.1.3: This model is based on the BiLSTM network. BiLSTM is composed of a forward LSTM and a backward LSTM. The forward and backward hidden layer vectors are concatenated. Its final output is the combined output of the forward and backward layers at each moment. In order to make full use of the input information, the input vector X t First, pass the fully connected layer FC to obtain the fixed-length embedding vector e t , e t As the input of the BiLSTM unit at time t. The hidden state h of the BiLSTM unit at time t t By the embedding vector e at the current moment t and the hidden state h at the previous moment t-1 Joint update. As follows:

[0054]

[0055] h t =BiLSTM(h t-1 , e t ; W l ) (6)

[0056] In the formula, is the FC layer, W emb is the weight of the FC layer, W l are the weights of the BiLSTM layer.

[0057] Step 2.1.4: This invention introduces shortcut connections and uses an FC layer to process the input vector X t , get a fixed-length vector r t , whose length is the same as the BiLSTM output vector h t By replacing the vector r t and the BiLSTM output vector h t Perform the add operation to build a shortcut connection, and then use ReLU to activate it to obtain the updated output vector As follows:

[0058]

[0059]

[0060] In the formula, is the FC layer, W r is the weight of the FC layer.

[0061] Step 2.1.5: Input the output vector of the last moment of the historical sequence, that is, the current moment t After being processed by an FC layer and a SoftMax function, an intention probability vector is obtained, as shown below:

[0062]

[0063] Where P = p i (i=0, 1, 2) represent the probabilities of the three intentions: changing lanes to the left, changing lanes to the right, and going straight. is the FC layer, W f is the weight of the FC layer.

[0064] Make a logical judgment on the output probability and set p 0、 The intention corresponding to the largest probability value among p1 and p2 is output as the result of intention recognition.

[0065] Step 2.2: The steps for predicting the trajectory of the surrounding vehicle are as follows:

[0066] Step 2.2.1: The length of the trajectory input sequence is also set to 50. In addition to the 8 input features of the driving intention recognition network, the trajectory position coordinates Y of the surrounding vehicle will also be added to the trajectory prediction input. t , as follows:

[0067] Y t =(x t ,y t ) (10)

[0068] In the formula, x t ,y t are the longitudinal and lateral coordinates of the target vehicle at time t respectively.

[0069] Step 2.2.2: The trajectory prediction model is based on the LSTM encoder-decoder framework; first, the input sequence is input into the LSTM encoder for encoding, and the encoder hidden state h at time t t The input vector X at the current moment t and the hidden state h at the previous moment t-1 Joint update. As follows:

[0070] h t =LSTM(h t-1 , X t ; W enc ) (11)

[0071] Where W enc are the weights of the LSTM encoder.

[0072] The hidden state obtained at the last moment of the input history sequence, i.e., the current moment t, is selected as the context vector C to complete the encoding process.

[0073] Step 2.2.3: Assume that the decoder is predicting the position coordinates at time t+p, H = (h1, h2, ..., h t ) is the hidden layer state of the encoder. The soft attention mechanism is introduced to consider the most relevant historical information at each prediction moment, and the hidden layer similarity is calculated by dot product to obtain the attention value s of the hidden state at each historical moment corresponding to the t+p moment. i，t+p ; Then the vector weight is obtained through the SoftMax function to obtain the hidden state attention score α at each historical moment i，t+p , and finally generate the most appropriate context vector C through weighted summation t+p As follows:

[0074]

[0075]

[0076]

[0077] Step 2.2.4: The target vehicle position coordinates predicted at time t+p-1 (When p = 1, the actual position coordinate Y is used t ), the generated context vector C t+p , the probability vector P obtained by the intention recognition module is concatenated to obtain the input U of the decoder's predicted position coordinate at time t+p t+p , the decoder jointly inputs U t+p and the hidden state h at time t+p-1 t+p-1 Get the hidden state h at the current moment t+p As follows:

[0078]

[0079]

[0080] h t+p =LSTM(h t+p-1 , U t+p ; W dec ) (17)

[0081] In the formula, are the predicted longitudinal and transverse coordinates of the target vehicle at time t+p-1, W dec is the weight of the LSTM decoder, and Concat is the concatenation operation.

[0082] Step 2.2.5: Add shortcut connections in the decoder and convert the input vector U t+p After passing through the FC layer, it is directly connected to the output of the LSTM encoder, the add operation is performed, and then ReLU is used for activation to obtain the updated hidden state As follows:

[0083]

[0084] In the formula, is the FC layer, W d is the weight of the FC layer.

[0085] Hide the state After a FC layer, linear mapping is performed to obtain the position coordinates at time t+p

[0086]

[0087] In the formula, is the FC layer, W pis the weight of the FC layer. The length of the output sequence is also set to 50, that is, p = 1, 2, ..., 50.

[0088] Step 3: Visualize the driving intention of the surrounding vehicle and its position coordinates in the subsequent period of time, so that the driver can intuitively perceive the movement of the surrounding vehicle, reserve reaction time for the main vehicle operation, and ensure driving safety; at the same time, the autonomous driving car can plan the path in advance based on the information obtained, while ensuring the comfort and energy saving of the vehicle driving.

[0089] In summary, the present invention provides a method for realizing the joint prediction of the driving intention and trajectory of the surrounding vehicle by an autonomous driving vehicle, which can simultaneously identify the driving intention of the surrounding vehicle and predict the trajectory of the surrounding vehicle in a longer time domain, providing rich decision-making information for the vehicle; considering the driving intention characteristics and the trajectory prediction of the attention mechanism, the effect of long-term trajectory prediction can be guaranteed; according to the different characteristics of intention recognition and trajectory prediction, different input features are reasonably selected for prediction, which not only saves computing costs but also improves prediction efficiency.

[0090] The above contents are further detailed descriptions of the present invention in combination with the preferred technical solutions, and the specific implementation of the invention cannot be limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as the protection scope of the present invention.

Claims

1. A method for jointly predicting the driving intention and trajectory of surrounding vehicles by an autonomous driving vehicle, characterized in that The following steps are involved: 1) Processing of vehicle driving data: Process the vehicle driving data collected by the camera and millimeter-wave radar on-board sensors equipped on the vehicle, establish rules to define the driving intention type of the vehicle, set labels for each sequence data, and thus obtain the data required to realize vehicle motion prediction; 2) Joint prediction of the driving intention and trajectory of the surrounding vehicle: Construct an intention recognition model based on BiLSTM to identify the driving intention of the surrounding vehicle; construct a trajectory prediction model based on the LSTM encoder-decoder framework, and introduce an attention mechanism to improve the prediction effect of time series problems. At the same time, combine the location information and driving intention of the surrounding vehicle to predict the trajectory of the surrounding vehicle in a longer time domain; The specific method for predicting the trajectory of the surrounding vehicle over a longer time domain is as follows: (1) The length of the trajectory input sequence is set to 50. In addition to the 8 input features of the driving intention recognition network, the trajectory position coordinates Y of the surrounding vehicle will also be added. t , as follows: AND t (x) t ,and t ) In the formula, x t ,y t are the longitudinal and transverse coordinates of the target vehicle at time t respectively; (2) The trajectory prediction model is based on the LSTM encoder-decoder framework. First, the input sequence is input into the LSTM encoder for encoding. The encoder hidden state h at time t is t The input vector X at the current moment t and the hidden state h at the previous moment t-1 The joint update is as follows: h t =LSTM(h t-1 ,X t ;W enc ) Where W enc is the weight of the LSTM encoder; Select the hidden state obtained at the last moment of the input history sequence, that is, the current moment t, as the context vector C to complete the encoding process; (3) Assume that the decoder is predicting the position coordinates at time t+p, H = (h1, h2, ..., h t ) is the hidden layer state of the encoder. The soft attention mechanism is introduced to consider the most relevant historical information at each prediction moment. The hidden layer similarity is calculated by dot product to obtain the attention value s of the hidden state at each historical moment corresponding to the t+p moment. i,t+p ; Then the vector weight is obtained through the SoftMax function to obtain the hidden state attention score α at each historical moment i,t+p , and finally generate the most appropriate context vector C through weighted summation t+p , as follows: (4) The target vehicle position coordinates predicted at time t+p-1 When p = 1, the actual position coordinate Y is used t , the generated context vector C t+p , the probability vector P obtained by the intention recognition module is concatenated to obtain the input U of the decoder's predicted position coordinate at time t+p t+p , the decoder jointly inputs U t+p and the hidden state h at time t+p-1 t+p-1 Get the hidden state h at the current moment t+p , as follows: h t+p =LSTM(h t+p-1 ,U t+p ;W dec ) In the formula, are the predicted longitudinal and transverse coordinates of the target vehicle at time t+p-1, W dec is the weight of the LSTM decoder, and Concat is the concatenation operation; (5) Add shortcut connections in the decoder to convert the input vector U t+p After passing through the FC layer, it is directly connected to the output of the LSTM encoder, the add operation is performed, and then ReLU is used for activation to obtain the updated hidden state As follows: In the formula, is the FC layer, W d is the weight of the FC layer; Hide the state After a FC layer, linear mapping is performed to obtain the position coordinates at time t+p In the formula, is the FC layer, W p is the weight of the FC layer, and the length of the output sequence is also set to 50, that is, p = 1, 2, ..., 50; 3) Providing decision-making information: Providing rich information basis for subsequent operations of the main vehicle based on the results of the joint prediction.

2. A method for jointly predicting the driving intention and trajectory of surrounding vehicles by an autonomous driving vehicle as claimed in claim 1, characterized in that In step 1), the specific method of processing the peripheral vehicle travel data is: (1) The driving data of the surrounding vehicle collected by the camera and millimeter-wave radar on-board sensor equipped on the vehicle mainly include: the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration between the surrounding vehicle and the main vehicle, the azimuth angle of the surrounding vehicle relative to the main vehicle, and the azimuth angle change rate; (2) Fill in the missing values ​​of the data. Use the nearest neighbor filling method to fill in single missing values, use the interpolation method to fill in multiple consecutive missing values, use the 3-σ rule to remove outliers in the data, and finally use the Savitzky-Golay filter to filter the data to obtain a relatively smooth data curve; (3) Define a coordinate system to conveniently describe the position of the vehicle and calculate the position coordinates of the main vehicle and the surrounding vehicles at each moment; (4) The driving intention of the vehicle is defined as changing lanes to the left, changing lanes to the right, and going straight, and an intention label is set for each segment of driving sequence data.

3. A method for jointly predicting the driving intention and trajectory of surrounding vehicles by an autonomous driving vehicle as claimed in claim 1, characterized in that In step 2), the specific method for identifying the driving intention of the surrounding vehicle is: (1) The driving intention recognition of Zhouche is a classification problem. The driving intention recognition model classifies the input multi-feature and multi-step long time series data; (2) A driving intention recognition model for the surrounding vehicles is constructed based on the BiLSTM network. The BiLSTM network is composed of a forward LSTM and a backward LSTM. It uses contextual information to improve the accuracy of time series prediction results. (3) Inspired by ResNet, shortcut connections are added to the Bi-LSTM network to implement gradient stride calculation, which can effectively reduce the impact of gradient vanishing and network degradation problems in deep networks; (4) Since the recognition of intention is different from trajectory prediction in that it requires the specific position information of the vehicle, the eight features of the surrounding vehicle driving data processed in step 1) are used as the input of the intention recognition model. The eight features are the longitudinal distance, lateral distance, longitudinal relative speed, lateral relative speed, longitudinal relative acceleration, lateral relative acceleration, azimuth angle and azimuth angle change rate of the surrounding vehicle relative to the main vehicle, without taking the position information of the surrounding vehicle into consideration. This not only captures the main features of the driving intention, ensures the prediction accuracy, but also saves the computing cost. (5) The eight feature sequences are input into the driving intention recognition model, and the driving intention probability vector of the surrounding vehicle is output. The driving intention with the highest probability is the final recognition result.

4. A method for jointly predicting the driving intention and trajectory of surrounding vehicles by an autonomous driving vehicle as claimed in claim 1, characterized in that In step 3), the specific method of providing decision information is as follows: through the joint prediction of driving intention and trajectory, the current driving intention of the surrounding vehicle and its trajectory position in the subsequent period of time are obtained; according to the identified driving intention, its behavior in a short period of time is judged, and reaction time is reserved for the main vehicle operation, thereby reducing the risk of vehicle collision and ensuring driving safety; the main vehicle makes path planning in advance according to the predicted trajectory position of the surrounding vehicle in the subsequent period of time, and obtains the optimal operation option to reach the destination, thereby ensuring driving comfort and reducing useless energy consumption.

Images