Automatic steering driving humanoid trajectory planning method and system

Abstract

The invention provides an automatic steering driving humanoid trajectory planning method and system. The method comprises the following steps: acquiring relative distance information between a vehicle and obstacles on two sides of a road and vehicle kinematics information by acquiring and processing human control data in a curve scene; non-linear features are extracted from the vehicle kinematics information based on a kernel principal component analysis (KPCA) method; the multi-state cycle prediction model is constructed based on an LSTM network, and the method comprises the following steps: carrying out weighted attention on key data of long and short term information by cell state information; the model inputs the nonlinear features, the relative distance information between the vehicle and obstacles on the two sides of the road and the kinematics information of the vehicle into a multi-state cycle prediction model, and predicts future driving track points. Therefore, the mapping precision of the network is improved, human manipulation behaviors are well learned, and the stability of the planned trajectory is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic driving trajectory planning, and in particular to a method and system for automatic steering driving humanoid trajectory planning. Background Art

[0002] In recent years, unmanned driving technology has been widely introduced in agriculture, environmental exploration, disaster response and other environments. However, there are no obvious traffic markings on the roads in the above scenarios. How to make vehicles turn safely and smoothly during automatic driving is an important issue that needs to be studied.

[0003] In the prior art, when applied, it is necessary to learn the human control characteristics based on learning algorithms, which can be roughly divided into indirect learning methods and direct learning methods. Under real driving conditions with complex uncertainty constraints, the optimization objectives and reward functions may not be comprehensive and accurate. In addition, the trajectories generated by indirect learning methods are usually selected from a set of candidate trajectories generated by a fixed model, or generated according to specific rules, and the selected trajectories may be different from the real trajectories of human drivers.

[0004] Therefore, how to enable the vehicle to plan a trajectory that conforms to human control behavior when turning in a scene without traffic markings is an urgent problem that needs to be solved. Summary of the invention

[0005] In view of this, the purpose of an embodiment of the present invention is to take vehicle speed, angular velocity, curvature of curve change, nonlinear characteristics between kinematic information, and the relative position relationship between obstacles on both sides of the road and the vehicle as inputs, and learn human control behavior from the collected data through a deep learning model, so as to plan a trajectory route that conforms to human control habits in the curve scene.

[0006] A first aspect of the present invention provides a method for human-like trajectory planning for automatic steering driving, the method comprising: S101, collecting and processing human control data in a curve scene, obtaining relative distance information between the vehicle and obstacles on both sides of the road and vehicle kinematic information; S102, extracting nonlinear features from the vehicle kinematic information based on a kernel principal component analysis method KPCA; S103, establishing and training a multi-state cycle prediction model; the multi-state cycle prediction model is constructed based on an LSTM network, including cell state information to pay weighted attention to key data of long-term and short-term information; the model includes a state output gate; S104, inputting the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict future driving trajectory points.

[0007] Furthermore, the curve scene includes a curve entry phase, a turning phase, and a curve exit phase; The vehicle kinematic information includes vehicle speed and vehicle angular velocity; The step S101 includes: calculating the distance between the vehicle and two obstacles on one side by using the relative position relationship between the current position of the vehicle and the obstacles on both sides.

[0008] Further, the S102 includes: mapping the input space data into the feature space, and then calculating the main components in the feature space; The S103 includes: after extracting the principal component through KPCA, the principal component and the relative distance information between the vehicle and the obstacles on both sides of the road and the vehicle kinematic information are used as the input of the deep network, and the multi-state cyclic prediction model is used to predict the future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

[0009] Further, in S103, the multi-state cycle prediction model includes: (10) in, is the input gate, is the sigmoid activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information; (11) in, is the sigmoid activation function, is the weight matrix, is the element-wise product, is the state cell, is the short-term memory output, is the bias vector, To input information, For the Gate of Oblivion; (12) in, For the Gate of Forgetfulness, is the element-wise product, is the state cell, is the output of the forget gate; (13) in, is a nonlinear transformation, is the state cell, is the weight matrix, is the activation function, is the element-wise product; (14) in, is a nonlinear transformation, is the state cell, is the element-wise product, is the activation function, is the weight matrix, is the state transfer gate; (15) in, is the activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information, is the output of the forget gate; (16) in, is the output of the forget gate, is the state transfer gate, is the input gate, is the element-wise product, Output for the upper layer; (17) in, is the sigmoid activation function, is the weight matrix, is the short-term memory output, To input information, is the output of the forget gate, is the bias vector, is the output gate; (18) in, is the activation function, For state cells.

[0010] Further, in S103, when training the multi-state cycle prediction model, the network is trained using mean square error (MSE) as a loss function, and the formula is: ; Where N is the number of samples in the training data set, is the actual trajectory coordinate value of the operator, is the predicted value of trajectory coordinates, and the Adamw optimizer is used to adaptively optimize the network parameters.

[0011] A second aspect of the present invention provides an automatic steering driving human-like trajectory planning system, the system comprising a collection and processing module, an analysis module, a model building module and a prediction module. The acquisition and processing module is used to acquire and process human control data in the curve scene, obtain the relative distance information between the vehicle and obstacles on both sides of the road, and the vehicle kinematics information; An analysis module, for extracting nonlinear features from the vehicle kinematic information based on a kernel principal component analysis method KPCA; A model building module, used to build and train a multi-state cycle prediction model; the multi-state cycle prediction model is built based on an LSTM network, including cell state information to pay weighted attention to key data of long-term and short-term information; the model includes a state output gate; The prediction module is used to input the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict the future driving trajectory points.

[0012] Furthermore, the curve scene includes a curve entry phase, a turning phase, and a curve exit phase; The vehicle kinematic information includes vehicle speed and vehicle angular velocity; The acquisition and processing module also includes a module for calculating the distance between the vehicle and two obstacles on one side by using the relative position relationship between the current position of the vehicle and the obstacles on both sides.

[0013] Furthermore, the analysis module also includes a method for mapping the input spatial data into a feature space and then calculating the main components in the feature space; The model building module also includes a method for extracting the principal components through KPCA, using the principal components together with the relative distance information between the vehicle and obstacles on both sides of the road and the vehicle kinematic information as the input of the deep network, and using the multi-state cyclic prediction model to predict future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

[0014] In addition, the third aspect of the present invention provides an electronic device, which includes: one or more processors, and a memory, wherein the memory is used to store one or more computer programs, and the computer programs are configured to be executed by the one or more processors, and the programs include steps for executing the automatic steering driving human-simulated trajectory planning method as described in the first aspect above.

[0015] In addition, the fourth aspect of the present invention provides a storage medium storing a computer program; the program is loaded and executed by a processor to implement the steps of the automatic steering driving human-simulated trajectory planning method as described in the first aspect above.

[0016] In the scheme of the present invention, by collecting and processing human control data in the curve scene, the relative distance information between the vehicle and the obstacles on both sides of the road and the vehicle kinematic information are obtained; nonlinear features are extracted from the vehicle kinematic information based on the kernel principal component analysis method KPCA; a multi-state cyclic prediction model is established and trained; the multi-state cyclic prediction model is constructed based on the LSTM network, including cell state information to pay weighted attention to the key data of long-term and short-term information; the model includes a state output gate; the nonlinear features, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information are input into the multi-state cyclic prediction model to predict future driving trajectory points. A human control data set is constructed for training and testing of deep networks. Then, the cell state information is introduced into the LSTM network to pay weighted attention to the key data of long-term and short-term information, and the gradient vanishing phenomenon in the training process is alleviated through the state output gate, thereby improving the mapping accuracy of the network, better learning human control behavior, and improving the stability of the planned trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 It is a flowchart of the steps of the human-simulated trajectory planning method for automatic steering driving disclosed in an embodiment of the present invention; Figure 2 It is a data acquisition device of a multimodal sensing system disclosed in an embodiment of the present invention; Figure 3 (a) is the road curvature of the curved road scene disclosed in the embodiment of the present invention; Figure 3 (b) is the vehicle speed information disclosed in the embodiment of the present invention; Figure 3 (c) is the vehicle angular velocity information disclosed in the embodiment of the present invention; Figure 3 (d) is the relative position information of the vehicle and the left obstacle disclosed in the embodiment of the present invention; Figure 3 (e) is the relative position information of the vehicle and the right obstacle disclosed in the embodiment of the present invention; Figure 3 (f) is the data collection site disclosed in the embodiment of the present invention; Figure 4 It is a schematic diagram of the network structure of the multi-state cyclic prediction model disclosed in an embodiment of the present invention; Figure 5 It is a comparison of parameter errors of various methods disclosed in the embodiments of the present invention; Figure 6 is the fitting situation of the predicted value and the true value of the proposed method disclosed in the embodiment of the present invention; Figure 7 This is a variable curvature curve experiment disclosed in an embodiment of the present invention; Figure 8 This is a straight-track-curve scenario experiment disclosed in an embodiment of the present invention; Fig. 9 This is a curve-straight road scenario experiment disclosed in an embodiment of the present invention; Fig.10 is the prediction error of each parameter after the migration weight disclosed in the embodiment of the present invention; Fig.11 is the prediction error of each parameter without migration weight disclosed in the embodiment of the present invention. DETAILED DESCRIPTION

[0019] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more comprehensive and complete and fully convey the concept of the example embodiments to those skilled in the art.

[0020] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0021] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0022] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0023] It should be noted that the “plurality” mentioned in this embodiment refers to two or more than two.

[0024] The implementation details of the technical solution of the embodiment of the present application are described in detail below: In this embodiment, a multimodal sensing system is used to collect human-controlled vehicle driving data, and nonlinear features are extracted from the vehicle kinematic information through the KPCA method. Then, the relative position relationship between the current position of the vehicle and the obstacles on both sides of the road is calculated, and the relative position distance is limited to the range of the two obstacles in front and behind. On this basis, after combining with the vehicle kinematic state information in a time series to screen the poor quality data, a human control data set is constructed for training and testing of deep networks. Then, cell state information is introduced into the LSTM network to weightedly focus on the key data of long-term and short-term information, and the gradient vanishing phenomenon during the training process is alleviated through the state output gate, thereby improving the mapping accuracy of the network. After comparing with a variety of methods, it can be seen from the Error, RMSE and MAE indicators of the planned trajectory and the true value that the method in this embodiment can better learn human control behavior, and the SD indicators of speed and angular velocity can show that the planned trajectory of the method in this embodiment has good stability.

[0025] like Figure 1 The figure shows a flow chart of the steps of the human-like trajectory planning method for automatic steering driving in this embodiment. The method comprises: S101, collecting and processing human control data in the curve scene, obtaining the relative distance information between the vehicle and obstacles on both sides of the road and the vehicle kinematic information.

[0026] Specifically, in this embodiment, a multimodal sensing system is used to collect the driving information of human-controlled vehicles and then process the human-controlled behavior data set, and then the future trajectory points are mapped through an improved deep learning network. The details are as follows: A. Human-controlled data collection and processing Figure 2 It is the data acquisition device of the multimodal sensing system of this embodiment, which collects human control data in the curve scene.

[0027] Furthermore, the curve scene includes a curve entry phase, a turning phase, and a curve exit phase; the vehicle kinematic information includes a vehicle speed and a vehicle angular velocity; and S101 includes: calculating the distance position between the vehicle and two obstacles on one side by using the relative position relationship between the vehicle's current position and obstacles on both sides.

[0028] Specific cornering scenarios such as Figure 3 shown. Figure 3 (a) is the road curvature of the curved scene; Figure 3(b) is the vehicle speed information; Figure 3 (c) is the vehicle angular velocity information; Figure 3 (d) is the relative position information between the vehicle and the obstacle on the left; Figure 3 (e) is the relative position information between the vehicle and the obstacle on the right; Figure 3 (f) is the data collection scene. This scene is roughly divided into the cornering stage - Turning phase - Corner exit , each stage presents a variable curvature shape, stage It is an approximate straight line section with a small change in curvature. The rear curvature increases significantly, and considering the stability and safety of the vehicle, the vehicle's driving radius increases first and then decreases, leaving Enter Afterwards, the operator controls the vehicle to slowly return to the virtual center line of the road (the test road in the figure has no center line or marking line) according to his own habits and environmental changes. The obstacles on both sides of the road are asymmetrically distributed. This embodiment uses the relative position relationship between the current position of the vehicle and the obstacles on both sides to calculate the distance between the vehicle and the two obstacles on one side. It can be seen from the figure that the distance between the obstacles on both sides changes regularly during the driving process of the vehicle. Therefore, this embodiment uses a deep network to learn human control habits from the processed state information.

[0029] S102, extracting nonlinear features from the vehicle kinematics information based on Kernel Principal Component Analysis (KPCA).

[0030] Furthermore, the S102 includes: mapping the input space data into the feature space, and then calculating the main components in the feature space.

[0031] Specifically, in this embodiment, since the collected human manipulation data presents a nonlinear relationship, in order to better learn human manipulation habits, this embodiment adopts the kernel principal component analysis (KPCA) method to map the input space data to the feature space while retaining the nonlinear characteristics of the original data as much as possible, and then calculate the main components in the feature space. , defined as, , (1) Here we assume and is a nonlinear mapping. In order to find the principal components, it is necessary to solve the eigenvalue problem in the feature space, as follows , (2) in , is an eigenvector and the coefficients , so we get the following formula , (3) Formula 2 is expanded into , (4) After combining formulas (1), (3) and (4), we obtain the following formula , (5) Afterwards, the inner product Depend on Replace and abbreviate multiplication , then formula (5) can be equivalent to: , (6) Formula (6) can be further rewritten as: (7) in , is an (n×n) matrix defined as . Then, the PCA algorithm is run in this embodiment to obtain the feature vector The eigenvalues ​​are . The first principal component of x Multiply , the formula is as follows, (8) The kernel function used in this embodiment is: (9) S103, establishing and training a multi-state cycle prediction model; the multi-state cycle prediction model is constructed based on an LSTM network, including cell state information to perform weighted attention on key data of long-term and short-term information; the model includes a state output gate.

[0032] Furthermore, in S103, when training the multi-state cycle prediction model, the network is trained using mean square error (MSE) as a loss function.

[0033] The S103 includes: after extracting the principal component through KPCA, the principal component and the relative distance information between the vehicle and the obstacles on both sides of the road and the vehicle kinematic information are used as the input of the deep network, and the multi-state cyclic prediction model is used to predict the future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

[0034] S104, inputting the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict future driving trajectory points.

[0035] Specifically, in this embodiment, the cell state information of the previous time step and the cell state information output by the forget gate in the current time step are referenced, and the long-term and short-term key information is weighted and paid attention to, and the gradient vanishing phenomenon in the training process is alleviated through the state transmission gate, thereby improving the mapping ability of the network. Specifically, the principal components are extracted through KPCA and the relative distance information between the vehicle and the obstacles on both sides of the road and the kinematic information of the vehicle processed in the previous section are used as the input of the deep network, and the improved multi-state cyclic prediction model is used to predict the future driving trajectory points, in which the state gate structure is added to transmit the memory state, thereby alleviating the gradient vanishing phenomenon in the training process and improving the mapping ability of the network.

[0036] like Figure 4 The diagram is a schematic diagram of the network structure of the multi-state cyclic prediction model of this embodiment. Among them, the state cells store long-term memory information, which plays an important role in sequence prediction, and the hidden state stores key short-term memory information. Therefore, this embodiment pays attention to the information in the state cells in the state update of LSTM, and introduces a state transmission gate to reduce the omission of the state information transmission process. Figure 4 The orange part in the middle is the main added structure of this embodiment, and the specific method is as follows: (10) in, is the input gate, is the sigmoid activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information; (11) in, is the sigmoid activation function, is the weight matrix, is the element-wise product, is the state cell, is the short-term memory output, is the bias vector, To input information, For the Gate of Oblivion; (12) in, For the Gate of Forgetfulness, is the element-wise product, is the state cell, is the output of the forget gate; (13) in, is a nonlinear transformation, is the state cell, is the weight matrix, is the activation function, is the element-wise product; (14) in, is a nonlinear transformation, is the state cell, is the element-wise product, is the activation function, is the weight matrix, is the state transfer gate; (15) in, is the activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information, is the output of the forget gate; (16) in, is the output of the forget gate, is the state transfer gate, is the input gate, is the element-wise product, Output for the upper layer; (17) in, is the sigmoid activation function, is the weight matrix, is the short-term memory output, To input information, is the output of the forget gate, is the bias vector, is the output gate; (18) in, is the activation function, For state cells.

[0037] After the model is established, this embodiment uses the processed human manipulation behavior data to train the improved network, and uses the Min-Max normalization method to scale the input data to the range of [0,1], and uses the mean square error (MSE) as the loss function to train the network. The formula is: (19) Where N is the number of samples in the training data set, is the actual trajectory coordinate value of the operator, The Adamw optimizer is used to adaptively optimize the network parameters for the predicted values ​​of trajectory coordinates. After the training is completed, the output data is rescaled back to normal values ​​when making predictions.

[0038] In order to verify the effectiveness of the method proposed in this embodiment, this embodiment conducted ablation experiment comparison, different curve scene experiment comparison, mainstream method and the latest time series prediction method experiment comparison, and used the Error, RMSE and MAE between the predicted value and the true value to evaluate the degree of anthropomorphism of the planned trajectory of each method, and evaluated the trajectory smoothness by the SD (standard deviation) of the predicted speed and angular velocity.

[0039] In the experiment, this embodiment analyzes the impact of each part of the method proposed in this embodiment through a series of experiments, and uses Error, RMSE and MAE for evaluation. For the convenience of description, this embodiment uses Proposed to represent the method in this embodiment, L represents LSTM network, LC represents LSTM+state information method, LCR represents LSTM+state information+relative position information, and LCRG represents LSTM+state information+relative position information+state gate structure.

[0040] like Figure 5 The figure shows the comparison of parameter errors of various methods in this embodiment; Figure 6 Table 1 shows the RMSE and MAE comparison of the predicted parameters of different ablation experimental methods in this embodiment.

[0041] Table 1 Comparison of RMSE and MAE of prediction parameters of different ablation experimental methods ; from Figure 5 , 6It can be seen from Table 1 that the method proposed in this embodiment is the best in terms of the prediction accuracy of trajectory value, speed and angular velocity. When the LSTM model (L) is used for prediction, the output error is large. After the forget gate output cell state cycle in the LSTM model is applied (LC), the overall output accuracy is improved by 4.2%, indicating that the improved LSTM cell state cycle model has a promoting effect on the prediction accuracy; after adding the relative distance constraint of obstacles on both sides of the road, the output accuracy is significantly improved by 41.5%, which shows that the data processing method in this embodiment has high effectiveness; after adding the state transmission gate (LCR) to the LSTM cell state cycle model, the output accuracy is improved by 47.7%, which shows that the cell state information can provide more historical information in the long-term and short-term prediction process, thereby enhancing the prediction ability of the model; after introducing the nonlinear feature constraint, the prediction accuracy is improved by 75.6%, which shows that the strong regularity of the processed nonlinear features can better assist the model to learn the human control trajectory, thereby predicting the trajectory information that conforms to the human control habits.

[0042] In order to verify the generalization of the method proposed in this embodiment, we conducted tests in three different curve scenarios: variable curvature curve, straight road-curve, and curve-straight road.

[0043] like Figure 7 The example is a variable curvature curve experiment of this embodiment; Figure 8 The straight road-curve scenario experiment of this embodiment is shown; Fig. 9 The following is a curve-straight road scenario experiment of this embodiment. Figure 7-9 As shown in the figure, we migrated the weights trained in the previous section and trained them. The amount of data was reduced to 1 / 3 of the original, greatly reducing the cost of data collection and processing. Fig.10 , 11 As can be seen from Table 2, the prediction accuracy after migrating the pre-training weights is 84.4% higher than that without migrating the pre-training weights, which verifies that the method in this embodiment has good generalization performance and indicates that the method in this embodiment can accurately predict trajectory information in the case of small batch data. Fig.10 is the prediction error of each parameter after the weights are migrated in this embodiment; Fig.11 is the prediction error of each parameter without migration weight in this embodiment.

[0044] In Table 2, Scene represents the RMSE and MAE of prediction when the pre-trained model is migrated to different scenes, and NScene represents the RMSE and MAE of prediction when the pre-trained model is not migrated to different scenes. Figure 7-9(a) shows the driving routes of vehicles controlled by different operators in a variable curvature curve scenario; (b) is a comparison chart of the curve curvature and the collected data curvature, and the yellow box is the location where the road curvature changes greatly; (c) is a schematic diagram of abnormal scenarios that are prone to danger; (d)-(k) are the normal driving process of the vehicle from the starting point to the end point.

[0045] Table 2 Comparison of RMSE and MAE of prediction parameters under different curve scenarios ; Since there are few studies on vehicle turning in an environment without traffic markings, the method in this paper belongs to the time series prediction problem. Therefore, we compare it with the mainstream method and the latest time prediction method, as follows: (a) GRU: This model solves the gradient vanishing or gradient exploding problem that occurs when the RNN network processes long sequence data by introducing update gates and reset gates, thereby capturing long-term dependencies and reducing reliance on unimportant information.

[0046] (b) Bilstm: This model consists of two LSTM layers, one for processing the forward sequence and the other for processing the reverse sequence, and predicts the future trajectory by considering the information of the previous and next time points.

[0047] (c) Netraj: This model adds a defined sliding temporal context window to the seq2seq network of the encoder-decoder structure to capture the long-term and short-term temporal dependencies between trajectories, and uses a spatial attention layer to find the spatial correlation of features to predict the trajectory.

[0048] (d) CAMD BiGRU: This model combines a novel restricted attention mechanism layer in the Bi-GRU module to extract the temporal attributes between data, and then uses Bi-GRU again to capture the periodic features in the data to predict future trajectory points.

[0049] It can be seen from Table 3 that the trajectory accuracy of the Chinese method in this paper is the best. The error between the trajectory information planned by the GRU method and the true value is large. This is because the network structure of this method is relatively simple and the ability to extract the dependency relationship between sequences is limited, resulting in poor quality of planned trajectories; the Bilstm method can better understand the time series attributes between trajectory points, and the prediction error is improved by 11.5%; the Netraj method significantly improves the model prediction accuracy by focusing on the important features of the time dimension and the space dimension, and the accuracy is improved by 32.8%; the CAMD BiGRU method captures the time attributes and periodicity between data, which improves the prediction accuracy by 27.6%; the method in this paper can extract correlation from the processed data through the improved model structure, and the output accuracy is improved by 72.3%, which shows that the method in this paper can well predict trajectory data that conforms to human control habits.

[0050] Table 3 RMSE and MAE comparison of prediction parameters of comparison methods ; Since the changes in the vehicle's speed and angular velocity can directly reflect the smoothness of the planned trajectory, this paper compares the standard deviations of the predicted vehicle speed and angular velocity. From Table 4, it can be seen that the planned trajectory of the method in this paper has obvious advantages in terms of stability. Figure 6 The fitting effect in the figure shows that the method in this paper can learn human control behavior habits very well, and the planned trajectory has good stability, which can ensure the safety and stability of the vehicle.

[0051] Table 4 Comparison of SD of predicted velocity and angular velocity ; In summary, in this embodiment, in order to make the vehicle plan a trajectory that conforms to human control behavior when turning in a scene without traffic markings, a human-like control trajectory planning method is proposed. After collecting data on human-controlled vehicle driving through a self-designed multimodal sensing system, the KPCA method is used to extract nonlinear features from the vehicle kinematic information, and then the relative distance between the current position of the vehicle and the obstacles on both sides of the road is calculated, and the relative distance is limited to the range of the two obstacles in front and behind. On this basis, the vehicle kinematic state information is combined in time series to filter the poor quality data, and a human control data set is constructed for deep network training and testing. Then, the cell state information is introduced into the LSTM network to weightedly focus on the key data of long-term and short-term information, and the gradient vanishing phenomenon in the training process is alleviated through the state output gate, thereby improving the mapping accuracy of the network. After comparing with a variety of methods, it can be seen from the Error, RMSE and MAE indicators of the planned trajectory and the true value that the method in this embodiment can better learn human control behavior, and the SD indicators of speed and angular velocity can show that the planned trajectory of the method in this embodiment has good stability. Therefore, the method in this embodiment can enable the vehicle to plan a safe and smooth trajectory route when turning in an environment without traffic markings.

[0052] A second aspect of this embodiment provides an automatic steering driving human-like trajectory planning system, the system comprising a collection and processing module, an analysis module, a model building module and a prediction module. The acquisition and processing module is used to acquire and process human control data in the curve scene, obtain the relative distance information between the vehicle and obstacles on both sides of the road, and the vehicle kinematics information; An analysis module, for extracting nonlinear features from the vehicle kinematic information based on a kernel principal component analysis method KPCA; A model building module, used to build and train a multi-state cycle prediction model; the multi-state cycle prediction model is built based on an LSTM network, including cell state information to pay weighted attention to key data of long-term and short-term information; the model includes a state output gate; The prediction module is used to input the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict the future driving trajectory points.

[0053] Furthermore, the curve scene includes a curve entry phase, a turning phase, and a curve exit phase; The vehicle kinematic information includes vehicle speed and vehicle angular velocity; The acquisition and processing module also includes a module for calculating the distance between the vehicle and two obstacles on one side by using the relative position relationship between the current position of the vehicle and the obstacles on both sides.

[0054] Furthermore, the analysis module also includes a method for mapping the input spatial data into a feature space and then calculating the main components in the feature space; The model building module also includes a method for extracting the principal components through KPCA, using the principal components together with the relative distance information between the vehicle and obstacles on both sides of the road and the vehicle kinematic information as the input of the deep network, and using the multi-state cyclic prediction model to predict future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

[0055] In addition, an embodiment of the present application also discloses an electronic device, which includes: one or more processors, and a memory, wherein the memory is used to store one or more computer programs, and the computer programs are configured to be executed by the one or more processors, and the programs include steps for executing the automatic steering driving human-simulated trajectory planning method as described above.

[0056] In addition, an embodiment of the present application further provides a storage medium storing a computer program; the program is loaded and executed by a processor to implement the steps of the automatic steering driving human-simulated trajectory planning method as described above.

[0057] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this embodiment can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0058] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0059] The units described as separate components may or may not be physically separated. As a unit, it can be appreciated by those skilled in the art that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this embodiment can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0060] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0061] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a grid device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0062] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A human-like trajectory planning method for automatic steering driving, characterized in that: The method comprises: S101, collecting and processing human control data in a curve scene, obtaining relative distance information between the vehicle and obstacles on both sides of the road and vehicle kinematic information; S102, extracting nonlinear features from the vehicle kinematic information based on a kernel principal component analysis method KPCA; S103, establishing and training a multi-state cycle prediction model; the multi-state cycle prediction model is constructed based on an LSTM network, including cell state information to pay weighted attention to key data of long-term and short-term information; the model includes a state output gate; S104, inputting the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict future driving trajectory points.

2. The method for human-like trajectory planning for automatic steering driving according to claim 1, characterized in that: The cornering scene includes a cornering phase, a turning phase, and a cornering phase; The vehicle kinematic information includes vehicle speed and vehicle angular velocity; The step S101 includes: calculating the distance between the vehicle and two obstacles on one side by using the relative position relationship between the current position of the vehicle and the obstacles on both sides.

3. The method for human-like trajectory planning for automatic steering driving according to claim 2, characterized in that: The S102 includes: mapping the input space data into the feature space, and then calculating the main components in the feature space; The S103 includes: after extracting the principal component through KPCA, the principal component and the relative distance information between the vehicle and the obstacles on both sides of the road and the vehicle kinematic information are used as the input of the deep network, and the multi-state cyclic prediction model is used to predict the future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

4. The method for human-like trajectory planning for automatic steering driving according to claim 3 is characterized in that: In S103, the multi-state cycle prediction model includes: （10） in, is the input gate, is the sigmoid activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information; （11） in, is the sigmoid activation function, is the weight matrix, is the element-wise product, is the state cell, is the short-term memory output, is the bias vector, To input information, For the Gate of Oblivion; （12） in, For the Gate of Forgetfulness, is the element-wise product, is the state cell, is the output of the forget gate; （13） in, is a nonlinear transformation, is the state cell, is the weight matrix, is the activation function, is the element-wise product; （14） in, is a nonlinear transformation, is the state cell, is the element-wise product, is the activation function, is the weight matrix, is the state transfer gate; （15） in, is the activation function, is the weight matrix, is the short-term memory output, is the bias vector, To input information, is the output of the forget gate; （16） in, is the output of the forget gate, is the state transfer gate, is the input gate, is the element-wise product, Output for the upper layer; （17） in, is the sigmoid activation function, is the weight matrix, is the short-term memory output, To input information, is the output of the forget gate, is the bias vector, is the output gate; （18） in, is the activation function, For state cells.

5. The method for human-like trajectory planning for automatic steering driving according to claim 4, characterized in that: In S103, when training the multi-state cycle prediction model, the network is trained using the mean square error (MSE) as the loss function, and the formula is: ； Where N is the number of samples in the training data set, is the actual trajectory coordinate value of the operator, is the predicted value of trajectory coordinates, and the Adamw optimizer is used to adaptively optimize the network parameters.

6. An automatic steering driving human-like trajectory planning system, characterized in that: The system includes a collection and processing module, an analysis module, a model building module and a prediction module, and is characterized by: The acquisition and processing module is used to acquire and process human control data in the curve scene, obtain the relative distance information between the vehicle and obstacles on both sides of the road, and the vehicle kinematics information; An analysis module, for extracting nonlinear features from the vehicle kinematic information based on a kernel principal component analysis method KPCA; A model building module, used to build and train a multi-state cycle prediction model; the multi-state cycle prediction model is built based on an LSTM network, including cell state information to pay weighted attention to key data of long-term and short-term information; the model includes a state output gate; The prediction module is used to input the nonlinear characteristics, the relative distance information between the vehicle and the obstacles on both sides of the road, and the vehicle kinematic information into the multi-state cyclic prediction model to predict the future driving trajectory points.

7. The automatic steering driving human-like trajectory planning system according to claim 6, characterized in that: The cornering scene includes a cornering phase, a turning phase, and a cornering phase; The vehicle kinematic information includes vehicle speed and vehicle angular velocity; The acquisition and processing module also includes a module for calculating the distance between the vehicle and two obstacles on one side by using the relative position relationship between the current position of the vehicle and the obstacles on both sides.

8. The automatic steering driving human-simulated trajectory planning system according to claim 7, characterized in that: The analysis module also includes a module for mapping the input spatial data into a feature space and then calculating the main components in the feature space; The model building module also includes a method for extracting the principal components through KPCA, using the principal components together with the relative distance information between the vehicle and obstacles on both sides of the road and the vehicle kinematic information as the input of the deep network, and using the multi-state cyclic prediction model to predict future driving trajectory points, wherein a state gate structure is added to transmit the memory state.

9. An electronic device, comprising: One or more processors, a memory, wherein the memory is used to store one or more computer programs; characterized in that the computer program is configured to be executed by the one or more processors, and the program includes steps for executing the automatic steering driving human-simulated trajectory planning method as described in any one of claims 1-5.

10. A storage medium storing a computer program; characterized in that: The program is loaded and executed by the processor to implement the steps of the automatic steering driving human-simulated trajectory planning method as described in any one of claims 1-5.

Images