CN115755134A - Vehicle positioning method and device based on Informer network and computer - Google Patents

Abstract

The invention relates to a vehicle positioning method, device and computer based on an Informer network. The method includes judging whether the satellite positioning function of the vehicle is normal; using a Gaussian mixture model to cluster the traffic environment image data of the vehicle to obtain multiple traffic environment clusters. class data set; sort the motion feature data of vehicles under each type of traffic conditions according to the order of vehicle travel time to obtain a motion feature time series data set; use the motion feature time series data sets corresponding to multiple types of traffic conditions to train separately Informer neural network model; when the satellite positioning function of the vehicle is abnormal, the bulldozer distance analysis method is used to determine the Informer neural network prediction model; the Informer neural network prediction model is used to predict the driving trajectory of the vehicle; the present invention uses the Informer network to realize the prediction of the vehicle driving trajectory , which effectively solves the problem of how to accurately predict the vehicle position when the GPS fails for a long time.

Description

A vehicle positioning method, device and computer based on Informer network

technical field

The invention relates to the technical field of vehicle navigation and positioning, in particular to a vehicle positioning method, device and computer based on an Informer network.

Background technique

Vehicle navigation and positioning technology is one of the key technologies of intelligent transportation systems, and many intelligent transportation system (ITS) applications and location-based services (LBS) need to use information about vehicle locations. Global Positioning System (GPS) motion relative positioning technology is widely used in vehicle position positioning as a high-precision positioning method. However, the traditional GPS pure motion relative positioning technology often fails to accurately locate in certain constrained urban road environments, such as road sections surrounded by tall buildings, road sections surrounded by dense forests, and tunnels, because satellite signals are blocked or unstable. It even loses the positioning ability, and cannot provide reliable and accurate positioning for the vehicle, resulting in the lack of vehicle trajectory.

Contents of the invention

In order to solve technical problems such as inability to provide reliable positioning when satellite signals are blocked or unstable in vehicle navigation and positioning technology, the present invention provides a vehicle positioning method, device and computer based on an Informer network.

The technical scheme that the present invention solves the problems of the technologies described above is as follows:

Determine whether the satellite positioning function of the vehicle is normal;

When the satellite positioning function of the vehicle is normal, the Gaussian mixture model is used to cluster the traffic environment image data of the vehicle to obtain multiple types of traffic environment clustering data sets; wherein, each type of traffic environment clustering data set corresponds to a class traffic conditions;

Sorting the plurality of motion feature data of the vehicles corresponding to each type of traffic condition according to the order of travel time of the vehicles to obtain multiple types of motion feature time series data sets; wherein, each type of motion feature time series data set Corresponding to a class of traffic conditions;

Respectively utilize the motion characteristic time series data in each class described motion characteristic time series data set to train the Informer neural network model, obtain a plurality of Informer neural network prediction models;

When the satellite positioning function of the vehicle is abnormal, use the bulldozer distance analysis method to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets to determine the current traffic environment image data of the vehicle. traffic status category, and select the corresponding Informer neural network prediction model according to the current traffic status category of the vehicle determined;

Using the selected Informer neural network prediction model to predict the vehicle's driving trajectory when the satellite positioning function is abnormal.

The beneficial effects of the present invention are: the present invention realizes the prediction of the vehicle trajectory by using the Informer network, and effectively solves the problem of how to accurately predict the vehicle position when the satellite positioning function fails for a long time. In the method of judging traffic conditions using the combination of Gaussian mixture model and KL divergence, the bulldozer distance analysis method is used instead of KL divergence, which solves the problem of KL divergence asymmetry, and when measuring non-overlapping distribution, the judgment effect is better than KL Divergence. By applying the Informer network model to predict the driving trajectory of the vehicle, the Informer network model's self-attention distillation technology, probabilistic sparse self-attention mechanism and generative decoder are used as the core technology of the basic network to improve the training speed and reasoning speed, and reduce the memory overhead of the network and improve prediction accuracy.

On the basis of the above technical solutions, the present invention can also be improved as follows.

Further, the Gaussian mixture model is used to cluster the traffic environment image data of the vehicle to obtain a multi-class traffic environment clustering data set, including the following steps:

Selecting feature vectors in the traffic environment image data;

calculating the probability distribution of the feature vector in the traffic environment image data of the vehicle by using a Gaussian mixture model;

Clustering all the traffic environment image data according to the corresponding probability distributions to obtain multiple types of traffic environment clustering data sets.

Further, the Gaussian mixture model is used to calculate the probability distribution of the feature vector in the traffic environment image data of the vehicle, including the following steps:

计算所述车辆的交通环境图像数据中的概率分布；其中，y表示所述特征向量，p(y)表示所述概率分布，N(y|μ_k,∑_k)表示所述高斯混合模型的第k个分量，π_k表示所述高斯混合模型每个分量的权重。use the formula

Calculate the probability distribution in the traffic environment image data of the vehicle; wherein, y represents the feature vector, p(y) represents the probability distribution, N(y|μ _k , ∑ _k ) represents the Gaussian mixture model The kth component, π _k represents the weight of each component of the Gaussian mixture model.

Further, using the bulldozer distance analysis method to classify all the traffic environment clustering data sets to obtain multiple types of traffic environment clustering data sets, including the following steps:

Using the bulldozer distance analysis method to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets, and determine the current traffic condition category of the vehicle, including the following steps:

Using the bulldozer distance analysis method, according to the distance value between the target probability distribution and the pre-stored probability distribution, determine the current traffic condition category of the vehicle; wherein, the target probability distribution is the feature vector in the current traffic environment of the vehicle The probability distribution in the image data, the pre-stored probability distribution is the probability distribution of the feature vector in the traffic environment image data of each type of the traffic environment clustering data set.

Further, using the bulldozer distance analysis method to determine the current traffic status category of the vehicle according to the distance value between the target probability distribution and the pre-stored probability distribution includes the following steps:

Using the bulldozer distance calculation formula to calculate the distance value between the target probability distribution and the pre-stored probability distribution; the bulldozer distance calculation formula is,

计算所述目标概率分布与预存概率分布之间的距离值；其中，s表示当前交通状况的目标概率分布，s(x)和

是定义在复平面Rⁿ上的概率分布，

是Rⁿ×Rⁿ上的联合分布，

为s和

组合起来的所有联合分布γ的集合，s和

是

的边缘分布，样本x和样本y是从联合分布γ中采样(x,y)～y得到的，

是关于x,y的核密度函数，对所有t＞0的情况，t＝1/α，α为1，p为2，

表示s与

所对应的预存概率分布之间的距离值；

Calculate the distance value between the target probability distribution and the prestored probability distribution; Wherein, s represents the target probability distribution of the current traffic situation, s(x) and

is a probability distribution defined on the complex plane R ⁿ ,

is the joint distribution on R ⁿ ×R ⁿ ,

for s and

The set of all joint distributions γ combined, s and

yes

The marginal distribution of the sample x and sample y is obtained by sampling (x,y)~y from the joint distribution γ,

It is the kernel density function about x, y, for all t>0 cases, t=1/α, α is 1, p is 2,

means s with

The distance value between the corresponding pre-stored probability distributions;

所对应的当前交通状况归类为

类；

表示预存的所述车辆的交通环境图像数据所对应的交通状况类别。will be minimum

The corresponding current traffic conditions are classified as

kind;

Indicates the traffic condition category corresponding to the prestored traffic environment image data of the vehicle.

The beneficial effect of adopting the above-mentioned further technical solution is that the improved bulldozer distance analysis method is used instead of KL divergence, which solves the problem of KL divergence asymmetry, and when measuring non-overlapping distribution, the judgment effect is better than KL divergence, compared with general The bulldozer distance distance can reduce the solution time while improving the solution accuracy.

Further, sorting the plurality of motion feature data of the vehicles corresponding to each type of traffic condition according to the order of the vehicle's travel time to obtain a multi-type motion feature time series data set, including the following steps:

Sorting the plurality of motion feature data of the vehicles corresponding to each type of traffic condition according to the order of the driving time of the vehicles to obtain multiple types of initial time series data sets;

Empty data and abnormal data in each type of initial time series data set are removed to obtain multiple types of motion feature time series data sets.

Further, utilize the motion characteristic time series data in each class described motion characteristic time series data set to train the Informer neural network model respectively, obtain a plurality of Informer neural network forecasting models, comprise the steps:

All the motion feature time series data in each type of motion feature time series data set are subjected to data normalization processing to obtain a plurality of normalized data sets; wherein, each of the normalized data sets corresponds to a Class described traffic conditions.

Utilize the data in a plurality of described normalized data sets to train described Informer neural network model respectively, obtain a plurality of Informer neural network prediction models;

The Informer neural network prediction model is verified and/or tested by using part of the normalized data set.

Further, the motion feature data of the vehicle includes the vehicle's driving time stamp, heading angle, pitch angle, latitude, longitude, altitude, and driving speed.

In order to solve the above-mentioned technical problems, the present invention also provides a vehicle positioning device based on the Informer network, and its specific technical scheme is as follows:

A vehicle positioning device based on the Informer network, comprising:

The function judging module is used to judge whether the satellite positioning function of the vehicle is normal;

The model training module is used for clustering the traffic environment image data of the vehicle using a Gaussian mixture model when the satellite positioning function of the vehicle is normal, so as to obtain a multi-class traffic environment clustering data set; A plurality of motion feature data of the vehicle corresponding to the condition is sorted according to the order of travel time of the vehicle to obtain a multi-type motion feature time series data set; using the motion feature time series in each type of the motion feature time series data set The data are used to train the Informer neural network model separately, and multiple Informer neural network prediction models are obtained; wherein, each type of traffic environment clustering data set corresponds to a type of traffic condition;

A trajectory judging module, used to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets by using the bulldozer distance analysis method when the satellite positioning function of the vehicle is abnormal , determine the current traffic status category of the vehicle, and select the corresponding Informer neural network prediction model according to the current traffic status category of the vehicle; Driving trajectory in case of malfunction.

In order to solve the above-mentioned technical problems, the present invention also provides a vehicle positioning device based on the Informer network, and its specific technical scheme is as follows:

A computer includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above-mentioned Informer network-based vehicle positioning method when executing the computer program.

Description of drawings

FIG. 1 is a flowchart of a method for vehicle positioning based on an Informer network in an embodiment of the present invention.

Detailed ways

The principles and features of the present invention are described below in conjunction with the accompanying drawings, and the examples given are only used to explain the present invention, and are not intended to limit the scope of the present invention.

As shown in Figure 1, the present embodiment provides a vehicle positioning method based on the Informer network, including the following steps:

S1. Judging whether the satellite positioning function of the vehicle is normal; wherein, the judgment of the satellite positioning function may be judging whether the GPS signal is available.

S2. When the satellite positioning function of the vehicle is normal, collect the traffic environment image data of the vehicle and the motion characteristic data of the vehicle in real time; use the Gaussian mixture model to cluster the traffic environment image data of the vehicle to obtain multiple types of traffic environments Clustering data set; wherein, each type of traffic environment clustering data set corresponds to a class of traffic conditions; the vehicle's motion feature data includes at least the vehicle's driving time stamp t, heading angle ψ, pitch angle θ, latitude B, longitude L, altitude Height H, and travel speed v.

Wherein, the Gaussian mixture model is used to cluster the traffic environment image data of the vehicle to obtain a multi-class traffic environment clustering data set, including the following steps:

S20, selecting a feature vector in the traffic environment image data of the vehicle;

S21. Using the Gaussian mixture model to calculate the probability distribution of the feature vector in the traffic environment image data of the vehicle; the specific steps are as follows:

图像数据中的概率分布；其中，y表示所述特征向量，p(y)表示所述概率分布，N(y|μ_k,∑_k)表示所述高斯混合模型的第k个分量，π_k表示所述高斯混合模型每个分量的权重。Using a Gaussian mixture model to calculate the feature vector in the traffic environment image data of the vehicle

Probability distribution in the image data; wherein, y represents the feature vector, p(y) represents the probability distribution, N(y|μ _k , ∑ _k ) represents the kth component of the Gaussian mixture model, π _k Indicates the weight of each component of the Gaussian mixture model.

S22. Clustering the traffic environment image data of all the vehicles according to the corresponding probability distributions to obtain multiple types of traffic environment clustering data sets.

S3. Sorting the plurality of motion characteristic data of the vehicles corresponding to each type of traffic condition according to the order of travel time of the vehicles to obtain a multi-type motion characteristic time series data set. Specific steps are as follows:

S30. Perform data normalization processing on all the motion feature time-series data in each type of motion feature time-series data set to obtain multiple normalized data sets; wherein, each of the normalized data sets Corresponding to a class of traffic conditions.

Among them, the specific calculation formula for normalizing the data is as follows:

In the formula, μ represents the mean value corresponding to each data in the above time series data set, σ represents the standard deviation corresponding to each data in the above time series data set, x represents each data in the above time series data set, and x ^* represents each data in the time series data set The normalized value of the data; include the normalized value in the normalized dataset.

S31. Using the data in multiple normalized data sets to train the Informer neural network models respectively to obtain multiple Informer neural network prediction models; specifically, select 80% of the normalized time series data sets respectively The data is used as training data to train the Informer neural network model to obtain multiple Informer neural network prediction models.

S32. Verify and/or test the Informer neural network prediction model by using part of the data in the normalized data set. Select 10% of the data in the time series data set after normalization processing as the verification data, and select the remaining 10% of the data in the time series data set after normalization processing as the test data, and the verification data is used to verify the trained Informer Neural network model, the test data is used to test the test accuracy of the trained Informer neural network model.

S4. Using the motion feature time series data in each type of motion feature time series data set to train the Informer neural network model respectively to obtain multiple Informer neural network prediction models;

After the data has been processed in the previous step, set the input sequence x ₁ ,x ₁ ,...,x _T . The sequence is multi-dimensional data, including the vehicle's driving time stamp t, heading angle ψ, pitch angle θ, latitude B, longitude L, Altitude H, and driving speed v, the initial parameters are set as follows: encoder input dimension is 7, encoder input dimension is 7, output dimension is 7, model size is 512, encoder layer number is 2, decoder layer number is 1, the activation function is the gelu function, and the learning rate is 0.0001. The batch size of the input training data is 32. The model training process is as follows:

局部时间戳PE和全局时间戳SE组成；转换公式为：The model input consists of the filtered smoothed feature scalar

Composed of local timestamp PE and global timestamp SE; the conversion formula is:

where: i ∈ {1,...,L _x }, α is a factor that balances the size between scalar mapping and local/global embedding.

具体操作为通过Conv1D将i维转换为512维向量。局部时间戳采用Transformer中的PositionalEmbedding，计算公式为：In the formula corresponding to the characteristic scalar

The specific operation is to convert the i dimension into a 512-dimensional vector through Conv1D. The local timestamp adopts PositionalEmbedding in Transformer, and the calculation formula is:

Where d _model is the feature dimension of the input, and the global timestamp uses a fully connected layer to map the input timestamp to a 512-dimensional Embedding.

Specific methods for generating encoders include:

输入到模型的编码器Encoder部分，首先在注意力模块对其进行稀疏性自注意力计算，每个Key只关注u个主要Query，Q为查询向量，K为键向量，V为值向量，计算公式如下所示：Unity Transformed Input

Input to the Encoder part of the model, first perform sparse self-attention calculation on it in the attention module, each Key only pays attention to u main Query, Q is the query vector, K is the key vector, V is the value vector, and the calculation The formula looks like this:

是大小与Q相同的稀疏矩阵且其只包含稀疏度量M(q_i,K)下的Top-u个Query。加入一个采样因子c，设置u＝clnLq。首先，为每个Query都随机采样c×lnL个Key，并计算每个Query的稀疏性得分M(q_i,K)。Q_i,K_i,V_i分别为Q,K,V的第i行，d为q_i的维度，而L_k＝L_q＝L，稀疏度量M(q_i,K)的近似计算公式为：in,

is a sparse matrix with the same size as Q and it only contains Top-u queries under the sparse metric M(q _i ,K). Add a sampling factor c, set u = clnLq. First, c×lnL keys are randomly sampled for each query, and the sparsity score M(q _i ,K) of each query is calculated. Q _i , K _i , V _i are the i-th row of Q, K, V respectively, d is the dimension of q _i , and L _k =L _q =L, the approximate calculation formula of the sparse measure M(q _i ,K) is :

Then, select the N queries with the highest sparsity score, and N defaults to c×lnL, and only calculate the dot product results of N queries and Key, and the remaining L-N queries are not calculated.

The output after sparse self-attention calculation has a redundant combination of V values, so the distillation operation is required to give higher weight to the dominant features with the main features, and generate a focused self-attention feature map in the next layer. Specifically, it is realized by four layers of Convld convolutional layers and a maximum pooling layer.

After multiple sparse self-attention layer calculations and a combination of distillation operations, the input of the decoder Decoder of the Informer neural network model is obtained. As for the decoder Decoder, the Decoder used by the Informer neural network model is similar to the traditional Decoder. In order for the algorithm to generate a long sequence of outputs, the Decoder needs the following inputs:

为输入Decoder的原始序列，

为需要预测序列(用0填充)，随后将序列通过一个基于掩码的稀疏性自注意力层，它可以防止每个位置都关注未来的位置，从而避免了自回归。将该层的输出以及Encoder的输出再传递给一个多头注意力层，经过一次计算输出结果。最终通过一个全连接层，得到最后的输出。将预测得到的输出和真实值进行损失函数Loss计算，损失函数采用MSE，计算公式如下：in,

For the original sequence input to Decoder,

For the sequence to be predicted (filled with 0s), the sequence is then passed through a mask-based sparsity self-attention layer, which prevents each position from focusing on future positions, thereby avoiding autoregression. The output of this layer and the output of the Encoder are passed to a multi-head attention layer, and the output result is calculated once. Finally, through a fully connected layer, the final output is obtained. Calculate the loss function Loss between the predicted output and the actual value. The loss function uses MSE, and the calculation formula is as follows:

为预测数据。不断迭代，直至训练条件终止，最终生成需要的模型。其中，训练条件终止具体为达到模型迭代次数或因MSE不下降而触发早停机制，最后得到损失函数最小的Informer预测模型。Among them, n is the number of samples, y _i is the real data,

for forecast data. Continue to iterate until the training conditions are terminated, and finally generate the required model. Among them, the termination of the training condition is specifically to reach the number of model iterations or trigger the early stop mechanism because the MSE does not decrease, and finally obtain the Informer prediction model with the smallest loss function.

S5. When the satellite positioning function of the vehicle is abnormal, use the bulldozer distance analysis method to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets, and determine the The current traffic condition category of the vehicle, and select the corresponding Informer neural network prediction model according to the determined current traffic condition category of the vehicle.

The specific steps are as follows: use the bulldozer distance analysis method to determine the current traffic status category of the vehicle according to the distance value between the target probability distribution and the pre-stored probability distribution; The probability distribution in the current traffic environment image data, the pre-stored probability distribution is the probability distribution of the feature vector in the traffic environment image data of each type of the traffic environment clustering data set.

Wherein, the bulldozer distance analysis method is used to determine the current traffic status category of the vehicle according to the distance value between the target probability distribution and the pre-stored probability distribution, including the following steps:

Utilize the bulldozer distance calculation formula

计算所述目标概率分布与预存概率分布之间的距离值；其中，s表示当前交通状况的目标概率分布，s(x)和

是定义在复平面Rⁿ上的概率分布，

是Rⁿ×Rⁿ上的联合分布，

为s和

组合起来的所有联合分布γ的集合，s和

是

的边缘分布，样本x和样本y是从联合分布γ中采样(x,y)～y得到的，

是关于x,y的核密度函数，对所有t＞0的情况，该函数是正定的，在实际求解过程中一般令t＝1/α，本发明中α的经验值为1，p的经验值为2，

表示s与

所对应的预存概率分布之间的距离值；

Calculate the distance value between the target probability distribution and the prestored probability distribution; Wherein, s represents the target probability distribution of the current traffic situation, s(x) and

is a probability distribution defined on the complex plane R ⁿ ,

is the joint distribution on R ⁿ ×R ⁿ ,

for s and

The set of all joint distributions γ combined, s and

yes

The marginal distribution of the sample x and sample y is obtained by sampling (x,y)~y from the joint distribution γ,

Be about x, the kernel density function of y, to all t＞0 situation, this function is positive definite, generally make t=1/α in the actual solution process, the empirical value of α among the present invention is 1, the empirical value of p value is 2,

means s with

The distance value between the corresponding pre-stored probability distributions;

所对应的当前交通状况归类为

类；

表示预存的所述车辆的交通环境图像数据所对应的交通状况类别。

分别为

以及

表示低复杂度交通状况类的预存概率分布，

表示中复杂度交通状况类的预存概率分布，

表示高复杂度交通状况类的预存概率分布。当车辆所处交通环境发生变化时，这段时间内收集的车辆运动特征数据将会被打包并标识，归类在与之对应的交通状况的车辆运动特征数据集中。will be minimum

The corresponding current traffic conditions are classified as

kind;

Indicates the traffic condition category corresponding to the prestored traffic environment image data of the vehicle.

respectively

as well as

Represents the pre-stored probability distribution of low-complexity traffic condition classes,

Represents the pre-stored probability distribution of medium-complexity traffic condition classes,

Represents a pre-stored probability distribution for a class of high-complexity traffic conditions. When the traffic environment where the vehicle is in changes, the vehicle movement characteristic data collected during this period will be packaged and marked, and classified into the vehicle movement characteristic data set corresponding to the traffic condition.

S6. Using the selected Informer neural network prediction model to predict the driving trajectory of the vehicle when the satellite positioning function is abnormal.

The embodiment of the present invention uses the Informer network to realize the prediction of the vehicle trajectory, effectively solving the problem of how to accurately predict the vehicle position when the satellite positioning function fails for a long time. In the method of judging traffic conditions using the combination of Gaussian mixture model and KL divergence, the improved bulldozer distance analysis method is used instead of KL divergence, which solves the problem of KL divergence asymmetry, and the judgment effect is excellent when measuring non-overlapping distributions. For the KL divergence, compared with the general bulldozer distance method, the improved bulldozer distance method adopted in the embodiment of the present invention can shorten the solution time while improving the solution accuracy. By applying the Informer network model to predict the driving trajectory of the vehicle, the Informer network model's self-attention distillation technology, probabilistic sparse self-attention mechanism and generative decoder are used as the core technology of the basic network to improve the training speed and reasoning speed, and reduce the memory overhead of the network and improve prediction accuracy.

Example 2

Based on Embodiment 1, this embodiment provides a vehicle positioning device based on the Informer network, including:

The function judging module is used to judge whether the satellite positioning function of the vehicle is normal;

The model training module is used for clustering the traffic environment image data of the vehicle using a Gaussian mixture model when the satellite positioning function of the vehicle is normal, so as to obtain a multi-class traffic environment clustering data set; A plurality of motion feature data of the vehicle corresponding to the condition is sorted according to the order of travel time of the vehicle to obtain a multi-type motion feature time series data set; using the motion feature time series in each type of the motion feature time series data set The data are used to train the Informer neural network model separately, and multiple Informer neural network prediction models are obtained; wherein, each type of traffic environment clustering data set corresponds to a type of traffic condition;

A trajectory judging module, used to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets by using the bulldozer distance analysis method when the satellite positioning function of the vehicle is abnormal , determine the current traffic status category of the vehicle, and select the corresponding Informer neural network prediction model according to the current traffic status category of the vehicle; Driving trajectory in case of malfunction. In this embodiment, the function judging module, the model training module and the trajectory judging module may be computer functional system modules or specific computer hardware with corresponding functions or specific computers with corresponding functions.

The embodiment of the present invention uses the Informer network to realize the prediction of the vehicle trajectory, effectively solving the problem of how to accurately predict the vehicle position when the GPS fails for a long time; at the same time, using the method of GMM and IMP-Wasserstein distance, it is possible to classify different vehicles. Traffic conditions, the purpose is to distinguish the trained model according to different traffic conditions. When GPS is unavailable, selecting the corresponding model according to the current traffic environment of the vehicle for prediction can greatly improve the accuracy rate. Among them, GMM represents the Gaussian mixture model, and the Wasserstein distance is the Wasserstein Distance, also known as the bulldozer distance, that is, Earth Mover’s distance, EMD. IMP-Wasserstein is an improved bulldozer distance, and the specific calculation formula of IMP-Wasserstein is referred to the bulldozer distance calculation formula in Embodiment 1. Using the self-attention distillation technology of the Informer network, the probability sparse self-attention mechanism and the generative decoder as the core technology of the basic network, the training speed and inference speed are improved, the memory overhead of the network is reduced, and the prediction accuracy is improved. The invention can be used to assist vehicle positioning and improve the positioning ability of the vehicle in complex environments.

Example 3

Based on Embodiment 1, this embodiment provides a computer, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, a vehicle positioning method based on the Informer network in Embodiment 1 is implemented A step of. The memory in this embodiment can be computer internal memory, computer external storage device, storage hard disk, mobile storage device, cloud storage, etc.; by implementing a vehicle positioning method based on the Informer network with a computer program, the computing efficiency is improved.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the concept and principles of the present invention shall be included in the protection of the present invention. within range.

Claims (10)

1. a vehicle location method based on Informer network, is characterized in that, comprises the steps: Determine whether the satellite positioning function of the vehicle is normal; When the satellite positioning function of the vehicle is normal, the Gaussian mixture model is used to cluster the traffic environment image data of the vehicle to obtain multiple types of traffic environment clustering data sets; wherein, each type of traffic environment clustering data set corresponds to a class traffic conditions; Sorting the plurality of motion characteristic data of the plurality of vehicles corresponding to each type of traffic condition according to the order of travel time of the vehicles to obtain a multi-type motion feature time series data set; wherein, each type of motion feature time series The data set corresponds to a class of said traffic conditions; Respectively utilize the motion characteristic time series data in each class described motion characteristic time series data set to train the Informer neural network model, obtain a plurality of Informer neural network prediction models; When the satellite positioning function of the vehicle is abnormal, use the bulldozer distance analysis method to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets to determine the current traffic environment image data of the vehicle. traffic status category, and select the corresponding Informer neural network prediction model according to the current traffic status category of the vehicle determined; Using the selected Informer neural network prediction model to predict the vehicle's driving trajectory when the satellite positioning function is abnormal. 2. the vehicle location method based on Informer network according to claim 1, is characterized in that, utilizes Gaussian mixture model to carry out clustering to the traffic environment image data of described vehicle, obtains multiclass traffic environment clustering dataset, comprises as follows step: Selecting feature vectors in the traffic environment image data; Using a Gaussian mixture model to calculate the probability distribution of the feature vector in the traffic environment image data; Clustering all the traffic environment image data according to the corresponding probability distributions to obtain multiple types of traffic environment clustering data sets. 3. the vehicle location method based on Informer network according to claim 2, is characterized in that, utilizes Gaussian mixture model to calculate the probability distribution of described feature vector in the traffic environment image data of described vehicle, comprises the steps: use the formula

Calculate the probability distribution in the traffic environment image data of the vehicle; wherein, y represents the feature vector, p(y) represents the probability distribution, N(y|μ _k , ∑ _k ) represents the Gaussian mixture model The kth component, π _k represents the weight of each component of the Gaussian mixture model. 4. a kind of vehicle positioning method based on Informer network according to claim 2, is characterized in that, utilizes bulldozer distance analysis method with the current traffic environment image data of described vehicle and the described traffic environment clustering data set of multiple classes The traffic environment image data is compared to determine the current traffic condition category of the vehicle, including the following steps: Using the bulldozer distance analysis method, according to the distance value between the target probability distribution and the pre-stored probability distribution, determine the current traffic condition category of the vehicle; wherein, the target probability distribution is the feature vector in the current traffic environment of the vehicle The probability distribution in the image data, the pre-stored probability distribution is the probability distribution of the feature vector in the traffic environment image data of each type of the traffic environment clustering data set. 5. the vehicle location method based on Informer network according to claim 4, is characterized in that, utilizes bulldozer distance analysis method, according to the distance value between target probability distribution and pre-stored probability distribution, determine the current traffic condition category of described vehicle , including the following steps: Using the bulldozer distance calculation formula to calculate the distance value between the target probability distribution and the pre-stored probability distribution; the bulldozer distance calculation formula is,

Among them, s represents the target probability distribution of the current traffic situation, s(x) and

is a probability distribution defined on the complex plane R ⁿ ,

is the joint distribution on R ⁿ ×R ⁿ ,

for s and

The set of all joint distributions γ combined, s and

yes

The marginal distribution of the sample x and sample y is obtained by sampling (x,y)~y from the joint distribution γ,

It is the kernel density function about x, y, for all t>0 cases, t=1/α, α is 1, p is 2,

means s with

The distance value between the corresponding pre-stored probability distributions; will be minimum

The corresponding current traffic conditions are classified as

kind;

Indicates the traffic condition category corresponding to the prestored traffic environment image data of the vehicle. 6. The vehicle positioning method based on the Informer network according to claim 1, wherein a plurality of motion characteristic data of the vehicles corresponding to the traffic conditions of each type are sorted according to the order of travel time of the vehicles , to obtain a multi-class motion feature time series data set, including the following steps: Sorting the plurality of motion feature data of the vehicles corresponding to each type of traffic condition according to the order of the driving time of the vehicles to obtain multiple types of initial time series data sets; Empty data and abnormal data in each type of initial time series data set are removed to obtain multiple types of motion feature time series data sets. 7. the vehicle positioning method based on Informer network according to claim 6, is characterized in that, utilizes the motion feature time-series data in each class described motion feature time-series data set to train Informer neural network model respectively, obtains a plurality of Informer neural networks The network prediction model includes the following steps: All the motion feature time series data in each type of motion feature time series data set are subjected to data normalization processing to obtain a plurality of normalized data sets; wherein, each of the normalized data sets corresponds to a traffic conditions as described in this category; Utilize the data in a plurality of described normalized data sets to train described Informer neural network model respectively, obtain a plurality of Informer neural network prediction models; The Informer neural network prediction model is verified and/or tested by using part of the normalized data set. 8. the vehicle location method based on Informer network according to claim 1, is characterized in that, the motion feature data of described vehicle comprises the running time stamp of vehicle, course angle, pitch angle, latitude, longitude, altitude and the travel time stamp. speed. 9. A vehicle positioning device based on the Informer network, characterized in that it comprises: The function judging module is used to judge whether the satellite positioning function of the vehicle is normal; The model training module is used for clustering the traffic environment image data of the vehicle using a Gaussian mixture model when the satellite positioning function of the vehicle is normal, so as to obtain a multi-class traffic environment clustering data set; A plurality of motion feature data of the vehicle corresponding to the condition is sorted according to the order of travel time of the vehicle to obtain a multi-type motion feature time series data set; using the motion feature time series in each type of the motion feature time series data set The data are used to train the Informer neural network model separately, and multiple Informer neural network prediction models are obtained; wherein, each type of traffic environment clustering data set corresponds to a type of traffic condition; A trajectory judging module, used to compare the current traffic environment image data of the vehicle with the traffic environment image data in multiple types of traffic environment clustering data sets by using the bulldozer distance analysis method when the satellite positioning function of the vehicle is abnormal , determine the current traffic status category of the vehicle, and select the corresponding Informer neural network prediction model according to the current traffic status category of the vehicle; Driving trajectory in case of malfunction. 10. A computer, characterized by comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1-8 when executing the computer program.

Images