Trajectory tracking error uncertainty quantification method based on deep neural network

Abstract

The invention discloses a trajectory tracking error uncertainty quantification method based on a deep neural network in the technical field of traffic and artificial intelligence crossing. The method comprises the following steps: acquiring an ideal speed curve on a running path of an object to be tracked; discretizing the operation path of the to-be-tracked object according to a fixed interval and representing the discretized operation path as a path center line reference point sequence; according to the ideal speed curve and the path midline reference point sequence, N trajectory position point sequences are sequentially generated based on a pre-trained reference trajectory generation network, and a reference trajectory is obtained; and inputting the reference trajectory into a pre-trained uncertainty quantization network, quantizing the uncertainty of the trajectory tracking error, and outputting trajectory error probability distribution of the object to be tracked relative to the reference trajectory. According to the method, accurate quantification of the trajectory tracking error uncertainty is realized, and the problems that in the prior art, operation is complex, the calculated amount is large, and trajectory distribution information contained in a large amount of recorded operation data cannot be fully utilized to quantify the trajectory tracking error uncertainty are solved.

Description

Technical Field

[0001] The present invention relates to a trajectory tracking error uncertainty quantification method based on a deep neural network, and belongs to the technical field of intersection of transportation and artificial intelligence. Background Art

[0002] In the current field of intelligent transportation systems, the prediction and quantification of trajectory tracking errors is one of the key technologies to ensure vehicle safety, optimize traffic flow management, and improve the performance of autonomous driving systems. Existing trajectory tracking prediction technologies mainly rely on high-precision simulation systems. Existing systems simulate the motion state of vehicles under different road conditions, weather conditions, and driving behaviors by building complex physical models. Existing models can accurately calculate the impact of factors such as vehicle dynamics, tire-ground interaction, and air resistance on vehicle trajectory, thereby achieving a more accurate prediction of the vehicle's future trajectory. In addition, some systems also combine advanced sensor technology and data processing algorithms to obtain vehicle status information in real time to further improve prediction accuracy.

[0003] Although high-precision simulation systems have made significant progress in trajectory tracking prediction, their inherent limitations cannot be ignored. First, such systems usually require complex condition settings and a large amount of computing resources to accurately simulate every detail of vehicle movement, which leads to high costs for system deployment and maintenance. Secondly, the construction of physical models is often based on a series of assumptions and simplifications, which makes it difficult to fully reflect the complexity and variability of the actual traffic environment, thereby limiting the universality and accuracy of the prediction results. More importantly, the existing systems fail to make full use of the large amount of accumulated operating data, which contains rich trajectory distribution information and potential driving modes, and is of great value for improving the prediction accuracy and uncertainty quantification ability of trajectory tracking errors. Therefore, in terms of uncertainty quantification of trajectory tracking errors, the existing technology still has significant deficiencies.

[0004] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgement or any form of suggestion that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the prior art is complex in operation, large in calculation, and fails to fully utilize the trajectory distribution information contained in a large amount of recorded operation data to quantify the uncertainty of trajectory tracking errors.

[0006] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0007] The present invention provides a trajectory tracking error uncertainty quantification method based on a deep neural network, comprising:

[0008] Obtaining the running path of the object to be tracked and the ideal speed curve on the running path of the object to be tracked;

[0009] The running path of the object to be tracked is discretized into a sequence of midline reference points of the path at fixed intervals;

[0010] According to the ideal speed curve and the path centerline reference point sequence, based on the pre-trained reference trajectory generation network, N trajectory position point sequences are generated in sequence to obtain the reference trajectory;

[0011] The reference trajectory is input into a pre-trained uncertainty quantification network, the uncertainty of the trajectory tracking error is quantified, and the trajectory error probability distribution of the object to be tracked relative to the reference trajectory is output.

[0012] Furthermore, the method generates N trajectory position point sequences in sequence based on the ideal speed curve and the path midline reference point sequence based on a pre-trained reference trajectory generation network to obtain a reference trajectory, including:

[0013] Constructing a first recurrent neural network and a second recurrent neural network based on a gated recurrent unit;

[0014] Encoding the ideal speed curve using a first recurrent neural network to obtain a first eigenvector;

[0015] Using a second recurrent neural network to encode the path midline reference point sequence to obtain a second feature vector;

[0016] Connecting the first eigenvector and the second eigenvector to obtain a connection vector;

[0017] After the connection vector is sent into the fully connected network, a fully connected vector is obtained;

[0018] The fully connected vector and the starting position of the ideal speed curve are input into the third recurrent neural network constructed based on the gated recurrent unit to generate N trajectory position point sequences in sequence to obtain the reference trajectory.

[0019] Furthermore, in order to pre-train the reference trajectory generation network and construct a discriminant network for adversarial training, after the fully connected vector and the starting position of the ideal speed curve are input into the third recurrent neural network constructed based on the gated recurrent unit, and N trajectory position point sequences are generated in sequence, it also includes:

[0020] Construct a discriminative network based on gated recurrent units and multi-layer perceptrons;

[0021] Input the trajectory position point sequence into the discriminant network, make a judgment, and output a judgment result of 0 or 1;

[0022] Among them, the judgment result of 1 indicates that the trajectory position point sequence is a reference motion trajectory generated by a high-precision simulation system, and the judgment result of 0 indicates that the trajectory position point sequence is a reference trajectory generated by a reference trajectory generation network.

[0023] Furthermore, the discriminant network includes:

[0024] A fourth recurrent neural network constructed based on a gated recurrent unit is used to obtain a sequence of trajectory position points, and update the feature vector in sequence according to the state of the gated recurrent unit to obtain an updated state vector;

[0025] A multilayer perceptron network with n fully connected layers is used to obtain the updated state vector, and after being processed by n layers of linear transformation and ReLU activation function, the discrimination result is output.

[0026] Furthermore, the training data labels of the discriminant network include:

[0027] From the trajectory data set, obtain the object's running path and the ideal speed curve on the object's running path;

[0028] The ideal speed curve is calculated by differentiation to obtain an acceleration curve, and converted into a braking and throttle opening curve;

[0029] According to the braking and throttle opening curves, based on a pre-built high-precision simulation system, simulation is performed to generate a reference motion trajectory;

[0030] The label of the reference motion trajectory is set to 0, and the label of the reference trajectory generated by the reference trajectory generation network is set to 1, so as to form a training data label containing positive and negative samples.

[0031] Furthermore, according to the discriminant network, the reference trajectory generation network is pre-trained using an adversarial training method; wherein the pre-training method of the reference trajectory generation network includes:

[0032] Repeat the following steps until the preset maximum number of adversarial training is reached:

[0033] Randomly select K samples from the trajectory data set, where the samples include the object motion path and the ideal speed curve on the object motion path;

[0034] Obtain a trajectory position point sequence according to the sample, input it into the discriminant network, and obtain K real sample discrimination results;

[0035] Obtain a path midline reference point sequence according to the sample, input the ideal speed curve on the object's running path and the path midline reference point sequence and the reference trajectory generation network to obtain K generated sample discrimination results;

[0036] According to the discrimination results of K real samples, calculations are performed to obtain the parameters of the discrimination network. The gradient of

[0037] According to the discriminant network parameters The gradient of the discriminant network parameters is adjusted using stochastic gradient ascent Perform m updates;

[0038] According to the K generated sample discrimination results, calculation is performed to obtain the network parameters for generating the reference trajectory. The gradient of

[0039] Generate network parameters based on the reference trajectory Gradient of the discriminant network parameters The network parameters are generated by stochastic gradient ascent on the reference trajectory. Update to obtain the pre-trained reference trajectory generation network.

[0040] Further, the discriminant network parameters The gradient of is expressed as:

[0041] (1);

[0042] In the formula, represents the discriminant network parameters, is the symbol of partial derivative, K represents the total number of real sample discrimination results, Indicates that the discriminant network contains the discriminant network parameters When processing the kth real sample The output, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, represents the reference trajectory generation network parameters, and log represents the logarithmic function.

[0043] Further, the reference trajectory generates network parameters The gradient of is expressed as:

[0044] (2);

[0045] In the formula, represents the reference trajectory generation network parameters, is the symbol of partial derivative, K represents the total number of real sample discrimination results, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, Represents the discriminant network parameters.

[0046] Furthermore, the reference trajectory is input into a pre-trained uncertainty quantification network to quantify the uncertainty of the trajectory tracking error, and the trajectory error probability distribution of the object to be tracked relative to the reference trajectory is output, including:

[0047] Input all the trajectory position points in the reference trajectory into the uncertainty quantification network for processing, and output the normal distribution parameters corresponding to each trajectory position point in turn;

[0048] Using the normal distribution parameters corresponding to each trajectory location point, a multivariate normal probability distribution is constructed;

[0049] Using the multivariate normal probability distribution as the trajectory error probability distribution of the object to be tracked relative to the reference trajectory;

[0050] The multivariate normal probability distribution is used to quantify the uncertainty of the motion of the object to be tracked relative to the reference trajectory.

[0051] Furthermore, before the reference trajectory is input into the pre-trained uncertainty quantification network, the uncertainty quantification network is trained and the parameters of the uncertainty quantification network are obtained through supervised learning using the loss function of the uncertainty quantification network, so that the uncertainty quantification network can accurately quantify the uncertainty, wherein the loss function of the uncertainty quantification network is expressed as:

[0052] (3);

[0053] Where, L Q represents the loss function of the uncertainty quantization network, N represents the total number of trajectory position point sequences, It represents the normal distribution parameter output by the trajectory position point at the current time t after being processed by the uncertainty quantification network, and log represents the logarithmic function.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The trajectory tracking error uncertainty quantification method based on deep neural network provided by the present invention can quickly generate a high-precision reference trajectory directly based on the ideal speed curve and the path centerline reference point sequence through a pre-trained reference trajectory generation network, significantly reducing the complex condition setting and calculation cost of the traditional high-precision simulation system. The method provided by the present invention not only simplifies the process of system deployment and maintenance, but also greatly improves the efficiency of trajectory prediction, making it possible to adjust and optimize the trajectory in real time and dynamically, effectively promoting the flexibility and response speed of the intelligent transportation system.

[0056] Furthermore, the present invention uses a pre-trained uncertainty quantification network to perform error analysis on the reference trajectory and outputs the probability distribution of the trajectory error of the object to be tracked relative to the reference trajectory. This innovative design makes full use of the large amount of accumulated operating data, deeply explores the trajectory distribution information and potential driving modes contained in the data, and realizes the accurate quantification of the uncertainty of trajectory tracking errors. This not only solves the problems of the existing technology that the operation is complex, the calculation is large, and the trajectory distribution information contained in the large amount of recorded operating data cannot be fully utilized to quantify the uncertainty of trajectory tracking errors, but also provides more reliable data support for subsequent intelligent traffic management, automatic driving decision-making, etc., and promotes the development of intelligent transportation systems in a more intelligent and autonomous direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a method for quantifying uncertainty of trajectory tracking errors based on a deep neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0059] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0060] Example 1

[0061] This embodiment introduces a trajectory tracking error uncertainty quantification method based on a deep neural network, including:

[0062] Obtain the running path of the object to be tracked and the ideal speed curve on the running path of the object to be tracked.

[0063] The ideal speed curve reflects the speed change that the object to be tracked, such as a vehicle or drone, should achieve on its running path under ideal conditions. It is usually determined based on factors such as traffic rules, road conditions, and vehicle performance. The ideal speed curve is used in the present invention to enable the generated reference trajectory to meet actual operating requirements, and also provides a benchmark for the quantification of trajectory tracking errors.

[0064] The running path of the object to be tracked is discretized at fixed intervals and represented as a sequence of reference points in the center of the path.

[0065] Since the actual path is often continuous, and computer processing requires discretized data, this embodiment discretizes the path into a series of reference points at fixed intervals. Discretization not only reduces the complexity of data processing, but also makes the path representation clearer and easier to operate.

[0066] According to the ideal speed curve and the path centerline reference point sequence, based on the pre-trained reference trajectory generation network, N trajectory position point sequences are generated in sequence to obtain the reference trajectory.

[0067] By using the pre-trained reference trajectory generation network, combined with the ideal speed curve and the path centerline reference point sequence, a series of possible trajectory position points can be generated to form a reference trajectory. The generated reference trajectory takes into account both speed changes and path constraints, providing a reliable reference for subsequent trajectory tracking and error quantification. At the same time, due to the use of a pre-trained network, it has higher efficiency and accuracy.

[0068] The reference trajectory is input into a pre-trained uncertainty quantification network, the uncertainty of the trajectory tracking error is quantified, and the trajectory error probability distribution of the object to be tracked relative to the reference trajectory is output.

[0069] Through the uncertainty quantification network, the deviation between the object to be tracked and the reference trajectory during actual operation can be evaluated, and the probability distribution of this deviation can be given. The output trajectory error probability distribution provides important information support for the intelligent transportation system. The trajectory error probability distribution can not only be used to evaluate the accuracy and reliability of trajectory tracking, but also provide a scientific basis for subsequent decision-making such as path adjustment and speed control. In addition, by quantifying uncertainty, the system's ability to respond to emergencies can be improved and the system's robustness can be enhanced.

[0070] Example 2

[0071] Based on the same inventive concept as Example 1, this example introduces a trajectory tracking error uncertainty quantification method based on a deep neural network, such as Figure 1 As shown, the following steps are included:

[0072] Step 1: Obtain the ideal velocity curve V on the running path P of the object to be tracked.

[0073] In this embodiment, the ideal speed curve , where v t represents the ideal speed at the current time t, Represents a positive integer.

[0074] Step 2: Discretize the path P of the object to be tracked into a sequence C of midline reference points of the path at fixed intervals.

[0075] In this embodiment, the path midline reference point sequence ,in, represents the coordinates of the midline reference point of the i-th running path, where represents the longitude of the midline reference point of the ith running path, Indicates the latitude of the midline reference point of the i-th running path.

[0076] Step 3: Based on the ideal speed curve and the path centerline reference point sequence, a pre-trained reference trajectory generation network is used to generate N trajectory position point sequences in sequence to obtain a reference trajectory, including:

[0077] Constructing a first recurrent neural network N1 and a second recurrent neural network N2 based on a gated recurrent unit;

[0078] The first recurrent neural network N1 is used to encode the ideal speed curve V to obtain the first feature vector h1;

[0079] Among them, the ideal speed of all the current time t in the ideal speed curve V is sequentially calculated by the gated recurrent unit. Processing is performed to obtain the state vector at the current time t ,

[0080] Where ⊙ represents the Hadamard product operator, represents the ideal speed of the linear layer at the current time t The feature vector after embedding, represents the state vector at the previous time t-1 in the first recurrent neural network N1, represents all parameters of the first recurrent neural network N1, represents the update gate of the first recurrent neural network N1 at the current time t, where , is the activation function, represents the ideal speed at the current time t in the first recurrent neural network N1 The feature vector after embedding The learnable parameters of the update gate that performs the linear transformation, Represents the state vector of the first recurrent neural network N1 at the previous time t-1 The learnable parameters of the update gate that performs the linear transformation, represents the learnable bias parameter of the update gate in the first recurrent neural network N1, is the internal function of the gated recurrent unit, , tanh represents the inverse sine function, represents the ideal speed at the current time t in the first recurrent neural network N1 The feature vector after embedding The learnable parameters of the internal function g(·) of the gated recurrent unit that performs the linear transformation, represents the internal function g(·) of the gated recurrent unit in the first recurrent neural network N1 The learnable parameters for the linear transformation, represents the learnable bias parameter of the internal function g(·) of the gated recurrent unit in the first recurrent neural network N1, represents the reset gate of the first recurrent neural network N1 at the current time t, where , is the activation function, represents the ideal speed at the current time t in the first recurrent neural network N1 The feature vector after embedding The learnable parameters of the reset gate that performs the linear transformation, Represents the state vector of the first recurrent neural network N1 at the previous time t-1 The learnable parameters of the reset gate that performs the linear transformation, Represents the learnable bias parameter of the reset gate in the first recurrent neural network N1.

[0081] The second recurrent neural network N2 is used to encode the path midline reference point sequence C to obtain the second feature vector h2;

[0082] Among them, the coordinates of all the i-th path centerline reference points in the path centerline reference point sequence C are processed in turn by the gated recurrent unit to obtain the state vector corresponding to the i-th running path point ,

[0083] Where ⊙ is the Hadamard product operator, Represented as the midline reference point of the linear layer for the i-th path The feature vector after embedding, represents the state vector corresponding to the midline reference point of the i-1th path in the second recurrent neural network N2, represents all parameters of the second recurrent neural network N2, represents the update gate corresponding to the second recurrent neural network N2 at the midline reference point of the i-th path, where, , is the activation function, Represents the midline reference point of the i-th path in the second recurrent neural network N2 The feature vector after embedding The learnable parameters of the update gate z that performs the linear transformation, Represents the state vector corresponding to the midline reference point of the i-1th path in the second recurrent neural network N2 The learnable parameters of the update gate z that performs the linear transformation, represents the learnable bias parameter of the update gate z in the second recurrent neural network N2, is the internal function of the gated recurrent unit, where , tanh represents the inverse sine function, Represents the midline reference point of the i-th path The feature vector after embedding The learnable parameters of the internal function g(·) of the gated recurrent unit that performs the linear transformation, Represents the pair in the second recurrent neural network N2 The learnable parameters of the internal function g(·) of the gated recurrent unit that performs the linear transformation, represents the learnable bias parameter of the internal function g(·) of the gated recurrent unit in the second recurrent neural network N2, represents the reset gate of the gated recurrent unit corresponding to the midline reference point of the i-th path in the second recurrent neural network N2, where, , is the activation function, Represents the midline reference point of the i-th path in the second recurrent neural network N2 The feature vector after embedding The learnable parameters of the reset gate that performs the linear transformation, Represents the state vector corresponding to the midline reference point of the i-1th path in the second recurrent neural network N2 The learnable parameters of the reset gate that performs the linear transformation, Represents the learnable bias parameter of the reset gate in the second recurrent neural network N2.

[0084] Connect the first eigenvector h1 and the second eigenvector h2 to obtain a connection vector h, where , Represents a concatenation operation.

[0085] After sending the connection vector h into the fully connected network, the fully connected vector z is obtained. , where F1(h) represents the first fully connected layer, ReLU represents the activation function, represents the learnable parameter for linearly transforming the connection vector h in the first fully connected layer, where the superscript T represents transposition and b1 represents the learnable bias parameter in the first fully connected layer.

[0086] Connect the full vector z to the starting position of the ideal velocity curve , input the third recurrent neural network N3 built based on the gated recurrent unit, and generate N trajectory position point sequences in sequence according to the autoregressive method , and obtain the reference trajectory , used to replace the reference trajectory R generated by the high-precision simulation system, where Represents the horizontal coordinate of the trajectory position point output at the current time t, Represents the ordinate of the trajectory position point output at the current time t.

[0087] At the present moment When t is t, the input of the third recurrent neural network N3 is the trajectory point position output at the previous time t-1. After the linear layer Embedded vector , get the state vector ,

[0088] Where ⊙ is the Hadamard product operator, represents the state vector at the last time t-1 in the third recurrent neural network N3, represents all parameters of the third recurrent neural network N3, represents the update gate of the third recurrent neural network N3 at the current time t in the third recurrent neural network N3, where, , is the activation function, Represents the vector after embedding the trajectory point position output at the previous time t-1 in the third recurrent neural network N3 The learnable parameters of the update gate that performs the linear transformation, Represents the state vector of the last time t-1 in the third recurrent neural network N3 The learnable parameters of the update gate that performs the linear transformation, represents the learnable bias parameter of the update gate in the third recurrent neural network N3, is the internal function of the gated recurrent unit, , tanh represents the inverse sine function, Represents the vector after embedding the trajectory point position output at the previous time t-1 The learnable parameters for the linear transformation, Express The learnable parameters for the linear transformation, Represents the internal function of the gated recurrent unit in the third recurrent neural network N3 The learnable bias parameters, represents the reset gate at the current time t in the third recurrent neural network N3, where , is the activation function, Represents the vector after embedding the trajectory point position output at the previous time t-1 in the third recurrent neural network N3 The learnable parameters of the reset gate that performs the linear transformation, Represents the state vector of the last time t-1 in the third recurrent neural network N3 The learnable parameters of the reset gate that performs the linear transformation, Represents the learnable bias parameter of the reset gate in the third recurrent neural network N3.

[0089] Combine the fully connected vector z with the state vector obtained by the third recurrent neural network N3 Input to a multilayer perceptron network consisting of n fully connected layers , the output layer generates a prediction result containing two components, which is processed by the softmax function and is used as the trajectory position point output at the current time t after normalization. . Follow the above steps to generate N trajectory position point sequences in sequence and obtain the reference trajectory .

[0090] In order to pre-train the reference trajectory generation network, a discriminant network for adversarial training is constructed, including:

[0091] Based on gated recurrent unit and multi-layer perceptron Build a discriminant network;

[0092] The trajectory position point sequence is input into the discriminant network, judged, and the judgment result 0 or 1 is output.

[0093] Among them, the judgment result of 1 indicates that the trajectory position point sequence is a reference motion trajectory generated by a high-precision simulation system, and the judgment result of 0 indicates that the trajectory position point sequence is a reference trajectory generated by a reference trajectory generation network.

[0094] The discriminant network comprises:

[0095] A fourth recurrent neural network N4 constructed based on a gated recurrent unit is used to obtain a trajectory position point sequence, and update the feature vector in sequence according to the state of the gated recurrent unit to obtain an updated state vector;

[0096] A multilayer perceptron network with n fully connected layers is used to obtain the updated state vector, and after being processed by n layers of linear transformation and ReLU activation function, the discrimination result is output.

[0097] Among them, the input of the fourth recurrent neural network N4 is a trajectory position point sequence.

[0098] At the present moment When , the state vector is updated in sequence according to the gated cycle unit state update formula , where ⊙ is the Hadamard product operator, Represents the coordinate vector of the trajectory position point at the current time t, represents the state vector at the previous time t-1 in the fourth recurrent neural network N4, represents all parameters of the fourth recurrent neural network N4, represents the update gate at the current time t in the fourth recurrent neural network N4, where , is the activation function, Represents the coordinate vector of the trajectory position point at the current time t in the fourth recurrent neural network N4 The learnable parameters of the update gate that performs the linear transformation, Represents the state vector of the fourth recurrent neural network N4 at the previous time t-1 The learnable parameters of the update gate that performs the linear transformation, represents the learnable bias parameter of the update gate in the fourth recurrent neural network N4, is the internal function of the gated recurrent unit, , tanh represents the inverse sine function, Represents the coordinate vector of the trajectory position point of the fourth recurrent neural network N4 at the current time t Gated recurrent unit internal function for linear transformation The learnable parameters of Represents the fourth recurrent neural network N4 Gated recurrent unit internal function for linear transformation The learnable parameters of Represents the internal function of the gated recurrent unit in the fourth recurrent neural network N4 Learnable bias parameters, represents the reset gate at the current time t in the fourth recurrent neural network N4, where , is the activation function, Represents the coordinate vector of the trajectory position point at the current time t in the fourth recurrent neural network N4 The learnable parameters of the reset gate that performs the linear transformation, Represents the state vector of the fourth recurrent neural network N4 at the previous time t-1 The learnable parameters of the reset gate that performs the linear transformation, Represents the learnable bias term parameters of the reset gate in the fourth recurrent neural network N4.

[0099] The state vector obtained after the last trajectory position point is input Input Multilayer Perceptron Network , after being processed by n layers of linear transformation and ReLU activation function, the discrimination result is output.

[0100] The training data labels of the discriminant network include:

[0101] From the trajectory data set, obtain the object's running path and the ideal speed curve on the object's running path;

[0102] The ideal speed curve is calculated by differentiation to obtain an acceleration curve, and converted into a braking and throttle opening curve;

[0103] According to the braking and throttle opening curves, based on a pre-built high-precision simulation system, simulation is performed to generate a reference motion trajectory;

[0104] The reference motion trajectory label of the high-precision simulation is set to 0, and the reference trajectory label generated by the reference trajectory generation network is set to 1, so as to form a training data label containing positive and negative samples.

[0105] According to the discriminant network, the reference trajectory generation network is pre-trained using an adversarial training method; wherein the pre-training method of the reference trajectory generation network includes:

[0106] Repeat the following steps until the preset maximum number of adversarial training is reached:

[0107] Randomly select K samples from the trajectory data set, where the samples include the object motion path and the ideal speed curve on the object motion path;

[0108] Obtain a trajectory position point sequence according to the sample, input it into the discriminant network, and obtain K real sample discrimination results;

[0109] Obtain a path midline reference point sequence according to the sample, input the ideal speed curve on the object's running path and the path midline reference point sequence and the reference trajectory generation network to obtain K generated sample discrimination results;

[0110] According to the discrimination results of K real samples, calculations are performed to obtain the parameters of the discrimination network. The gradient of

[0111] The discriminant network parameters The gradient of is expressed as:

[0112] (1);

[0113] In the formula, represents the discriminant network parameters, represents the sign of partial derivative, K represents the total number of real sample discrimination results, Indicates that the discriminant network contains the discriminant network parameters When processing the kth real sample The output, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, represents the reference trajectory generation network parameters, and log represents the logarithmic function.

[0114] According to the discriminant network parameters The gradient of the discriminant network parameters is adjusted using stochastic gradient ascent Perform m updates;

[0115] According to the K generated sample discrimination results, calculation is performed to obtain the network parameters for generating the reference trajectory. The gradient of

[0116] The reference trajectory generates network parameters The gradient of is expressed as:

[0117] (2);

[0118] In the formula, represents the reference trajectory generation network parameters, represents the sign of partial derivative, K represents the total number of real sample discrimination results, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, Represents the discriminant network parameters.

[0119] Generate network parameters based on the reference trajectory Gradient of the discriminant network parameters The network parameters are generated by stochastic gradient ascent on the reference trajectory. Update to obtain the pre-trained reference trajectory generation network.

[0120] Step 4: Input the reference trajectory into a pre-trained uncertainty quantification network, quantify the uncertainty of the trajectory tracking error, and output the trajectory error probability distribution of the object to be tracked relative to the reference trajectory, including:

[0121] Input all the trajectory position points in the reference trajectory into the uncertainty quantification network for processing, and output the normal distribution parameters corresponding to each trajectory position point in turn;

[0122] Using the normal distribution parameters corresponding to each trajectory location point, a multivariate normal probability distribution is constructed;

[0123] Using the multivariate normal probability distribution as the trajectory error probability distribution of the object to be tracked relative to the reference trajectory;

[0124] The multivariate normal probability distribution is used to quantify the uncertainty of the motion of the object to be tracked relative to the reference trajectory.

[0125] Before the reference trajectory is input into the pre-trained uncertainty quantification network, the uncertainty quantification network is trained and the parameters of the uncertainty quantification network are obtained through supervised learning using the loss function of the uncertainty quantification network, so that the uncertainty quantification network can accurately quantify the uncertainty, wherein the loss function of the uncertainty quantification network is expressed as:

[0126] (3);

[0127] Where, L Q represents the loss function of the uncertainty quantization network, N represents the total number of trajectory position point sequences, It represents the normal distribution parameter output by the trajectory position point at the current time t after being processed by the uncertainty quantification network, and log represents the logarithmic function.

[0128] In summary, the trajectory tracking error uncertainty quantification method based on deep neural network provided by the present invention can quickly generate a high-precision reference trajectory directly based on the ideal speed curve and the path centerline reference point sequence through a pre-trained reference trajectory generation network, significantly reducing the complex condition setting and calculation cost of the traditional high-precision simulation system. The method provided by the present invention not only simplifies the system deployment and maintenance process, but also greatly improves the efficiency of trajectory prediction, making it possible to adjust and optimize the trajectory in real time and dynamically, effectively promoting the flexibility and response speed of the intelligent transportation system.

[0129] Furthermore, the present invention uses a pre-trained uncertainty quantification network to perform error analysis on the reference trajectory and outputs the probability distribution of the trajectory error of the object to be tracked relative to the reference trajectory. This innovative design makes full use of the large amount of accumulated operating data, deeply explores the trajectory distribution information and potential driving modes contained in the data, and realizes the accurate quantification of the uncertainty of trajectory tracking errors. This not only solves the problems of the existing technology that the operation is complex, the calculation is large, and the trajectory distribution information contained in the large amount of recorded operating data cannot be fully utilized to quantify the uncertainty of trajectory tracking errors, but also provides more reliable data support for subsequent intelligent traffic management, automatic driving decision-making, etc., and promotes the development of intelligent transportation systems in a more intelligent and autonomous direction.

[0130] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0134] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A trajectory tracking error uncertainty quantification method based on deep neural network, characterized in that: include: Obtaining the running path of the object to be tracked and the ideal speed curve on the running path of the object to be tracked; The running path of the object to be tracked is discretized into a sequence of midline reference points of the path at fixed intervals; According to the ideal speed curve and the path centerline reference point sequence, based on the pre-trained reference trajectory generation network, N trajectory position point sequences are generated in sequence to obtain the reference trajectory; The reference trajectory is input into a pre-trained uncertainty quantification network, the uncertainty of the trajectory tracking error is quantified, and the trajectory error probability distribution of the object to be tracked relative to the reference trajectory is output.

2. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 1 is characterized in that: The method generates N trajectory position point sequences in sequence based on the ideal speed curve and the path midline reference point sequence and based on a pre-trained reference trajectory generation network to obtain a reference trajectory, including: Constructing a first recurrent neural network and a second recurrent neural network based on a gated recurrent unit; Encoding the ideal speed curve using a first recurrent neural network to obtain a first eigenvector; Using a second recurrent neural network to encode the path midline reference point sequence to obtain a second feature vector; Connect the first eigenvector and the second eigenvector to obtain a connection vector; After the connection vector is sent into the fully connected network, a fully connected vector is obtained; The fully connected vector and the starting position of the ideal speed curve are input into the third recurrent neural network constructed based on the gated recurrent unit to generate N trajectory position point sequences in sequence to obtain the reference trajectory.

3. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 2 is characterized in that: After the fully connected vector and the starting position of the ideal speed curve are input into the third recurrent neural network constructed based on the gated recurrent unit to generate N trajectory position point sequences in sequence, it also includes: Construct a discriminative network based on gated recurrent units and multi-layer perceptrons; Input the trajectory position point sequence into the discriminant network, make a judgment, and output a judgment result of 0 or 1; Among them, the judgment result of 1 indicates that the trajectory position point sequence is a reference motion trajectory generated by a high-precision simulation system, and the judgment result of 0 indicates that the trajectory position point sequence is a reference trajectory generated by a reference trajectory generation network.

4. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 3 is characterized in that: The discriminant network comprises: A fourth recurrent neural network constructed based on a gated recurrent unit is used to obtain a sequence of trajectory position points, and update the feature vector in sequence according to the state of the gated recurrent unit to obtain an updated state vector; A multilayer perceptron network with n fully connected layers is used to obtain the updated state vector, and after being processed by n layers of linear transformation and ReLU activation function, the discrimination result is output.

5. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 4 is characterized in that: The training data labels of the discriminant network include: From the trajectory data set, obtain the object's running path and the ideal speed curve on the object's running path; The ideal speed curve is calculated by differentiation to obtain an acceleration curve, and converted into a braking and throttle opening curve; According to the braking and throttle opening curves, based on a pre-built high-precision simulation system, simulation is performed to generate a reference motion trajectory; The label of the high-precision simulated motion trajectory is set to 1, and the label of the reference trajectory generated by the reference trajectory generation network is set to 0, so as to form a training data label containing positive and negative samples.

6. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 5 is characterized in that: According to the discriminant network, the reference trajectory generation network is pre-trained using an adversarial training method; wherein the pre-training method of the reference trajectory generation network includes: Repeat the following steps until the preset maximum number of adversarial training is reached: Randomly select K samples from the trajectory data set, where the samples include the object motion path and the ideal speed curve on the object motion path; Obtain a trajectory position point sequence according to the sample, input it into the discriminant network, and obtain K real sample discrimination results; Obtain a path midline reference point sequence according to the sample, input the ideal speed curve on the object's running path and the path midline reference point sequence and the reference trajectory generation network to obtain K generated sample discrimination results; According to the discrimination results of K real samples, calculations are performed to obtain the parameters of the discrimination network. The gradient of According to the discriminant network parameters The gradient of the discriminant network parameters is adjusted using stochastic gradient ascent Perform m updates; According to the K generated sample discrimination results, calculation is performed to obtain the network parameters for generating the reference trajectory. The gradient of Generate network parameters based on the reference trajectory Gradient of the discriminant network parameters The network parameters are generated by stochastic gradient ascent on the reference trajectory. Update to obtain the pre-trained reference trajectory generation network.

7. The method for quantifying uncertainty of trajectory tracking error based on deep neural network according to claim 6, characterized in that: The discriminant network parameters The gradient of is expressed as: （1）； In the formula, represents the discriminant network parameters, is the symbol of partial derivative, K represents the total number of real sample discrimination results, Indicates that the discriminant network contains the discriminant network parameters When processing the kth real sample The output, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, represents the reference trajectory generation network parameters, and log represents the logarithmic function.

8. The method for quantifying uncertainty of trajectory tracking error based on deep neural network according to claim 6, characterized in that: The network parameters are generated about the reference trajectory The gradient of is expressed as: （2）； In the formula, represents the reference trajectory generation network parameters, is the symbol of partial derivative, K represents the total number of real sample discrimination results, Represents the running path of the reference trajectory generation network based on the kth generated sample The ideal speed curve on the running path The generated reference trajectory, Represents the discriminant network parameters.

9. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 1, characterized in that: The reference trajectory is input into a pre-trained uncertainty quantification network, the uncertainty of the trajectory tracking error is quantified, and the probability distribution of the trajectory error of the object to be tracked relative to the reference trajectory is output, including: Input all the trajectory position points in the reference trajectory into the uncertainty quantification network for processing, and output the normal distribution parameters corresponding to each trajectory position point in turn; Using the normal distribution parameters corresponding to each trajectory location point, a multivariate normal probability distribution is constructed; Using the multivariate normal probability distribution as the trajectory error probability distribution of the object to be tracked relative to the reference trajectory; The multivariate normal probability distribution is used to quantify the uncertainty of the motion of the object to be tracked relative to the reference trajectory.

10. The trajectory tracking error uncertainty quantification method based on deep neural network according to claim 1, characterized in that: Before the reference trajectory is input into the pre-trained uncertainty quantification network, the uncertainty quantification network is trained and the parameters of the uncertainty quantification network are obtained through supervised learning using the loss function of the uncertainty quantification network, so that the uncertainty quantification network can accurately quantify the uncertainty, wherein the loss function of the uncertainty quantification network is expressed as: （3）； Where, L Q represents the loss function of the uncertainty quantization network, N represents the total number of trajectory position point sequences, It represents the normal distribution parameter output by the trajectory position point at the current time t after being processed by the uncertainty quantification network, and log represents the logarithmic function.

Images