CN111055849A - Intersection intelligent driving method and system based on support vector machine - Google Patents

Abstract

The invention provides an intelligent intersection driving method and system based on a support vector machine. In the support vector machine providing step, providing a support vector machine, wherein the support vector machine is subjected to a training process in advance, and in the training process, training data are provided for the support vector machine, and the training data are obtained by processing original data through a dimensionality reduction module and a time compensation module; in the data processing step, the p characteristics acquired by the environment sensing unit are provided to a support vector machine for classification after being processed by a dimensionality reduction module and a time compensation module; in the decision step, the driving behavior of the vehicle is determined according to the classification result of the support vector machine. Therefore, the decision accuracy of the support vector machine can be effectively improved.

Description

Intelligent driving method and system at intersection based on support vector machine

technical field

The present invention relates to an intelligent driving method at an intersection and a system thereof, and in particular to an intelligent driving method and system at an intersection based on a support vector machine.

Background technique

Generally speaking, there are multi-directional vehicle turns or straight intersections at intersections or road intersections. Therefore, when passing the intersection, it is necessary to accelerate, decelerate or drive at a constant speed according to the driving judgment. Once the driving judgment is wrong, traffic accidents will occur. occur. According to statistics from the US Bureau of Statistics, the proportion of traffic accidents at intersections or road junctions in 2008 was as high as 40%; according to the German Federal Statistics Office, the proportion of traffic accidents at intersections or road junctions in 2013 was as high as 47.5% , In some countries, the proportion of traffic accidents is even as high as 98%.

In order to assist in decision-making and judgment when driving through intersections, professionals/scholars have developed a highly automated vehicle (HAV), which includes artificial intelligence for machine learning to assist driving decision-making, and support vector machine is one of them. A learning approach that enables decision-making by constructing models to predict or estimate. For example, when to accelerate, decelerate, or stabilize when passing through an intersection.

In practical situations, driving will make appropriate driving behaviors based on the surrounding information in the past several units of time, that is to say, the real driving situation is time-dependent. However, in the training process of the conventional support vector machine, although the data used for training are recorded according to continuous time points, the data observed at each time is regarded as an independent data without considering The variables are time-dependent, so there is still room for improvement in the judgment accuracy.

In view of this, how to effectively improve the decision-making accuracy of support vector machines has become the goal of the relevant industry.

SUMMARY OF THE INVENTION

The present invention provides an intelligent driving method and system at an intersection based on a support vector machine, which can effectively improve the decision-making accuracy of the support vector machine through dimension reduction processing and time compensation processing.

An embodiment according to an aspect of the present invention provides a support vector machine-based intelligent driving method at an intersection, which is applied to a vehicle and includes a support vector machine providing step, a data processing step, and a decision-making step. In the step of providing the support vector machine, a support vector machine is provided, and the provided support vector machine has undergone a training process in advance. During the training process, a training data is provided to the support vector machine, and the training data is reduced by one dimension from an original data module and a time compensation module are processed and obtained, wherein the original data includes a plurality of training samples, each training sample includes a time total value passing through an intersection, and each sample in a plurality of sampling time points within the time total value The p features corresponding to the time point and the next decision; the dimension reduction module integrates the p features into k new features, the time compensation module provides a preset time, and the time compensation module samples any one of the training samples. The time point and the new features corresponding to other sampling time points before any of the aforementioned sampling time points are regarded as a sequence to be expanded. When the length of the sequence to be expanded is less than the number of sampling time points in the preset time, A new sequence to be expanded is formed after an estimated value is added to the sequence to be expanded, wherein the conditional distribution of the predicted value is obtained by the joint distribution of the sequence to be expanded and the data of all new eigenvalues obeying Gaussian distribution, and the time compensation module Reshape the new sequence to be expanded into an expanded sequence, and the length of the expanded sequence is equal to the number of sampling time points in a preset time, wherein p and k are positive integers, and p>k; in the data processing step, The p features obtained by an environment sensing unit are processed by the dimension reduction module and the time compensation module, and then provided to the support vector machine for classification; in the decision-making step, the driving behavior of the vehicle is determined by the classification result of the support vector machine.

In this way, after the training data and the features obtained while driving are processed by the dimension reduction module and the time compensation module, there will be a time dependency, which can improve the accuracy of the prediction results.

According to the foregoing embodiments of the support vector machine-based intelligent driving method at an intersection, the dimension reduction module may employ a principal component analysis method. Or the time compensation module can use a uniform scaling method. Or the preset time may be equal to the largest of the total time values.

According to various embodiments of the aforementioned SVM-based intelligent driving method at intersections, the p features may include a lateral speed of the vehicle relative to an oncoming vehicle, a lateral acceleration of the vehicle relative to an oncoming vehicle, and a longitudinal direction of the vehicle relative to an oncoming vehicle. Speed, a longitudinal acceleration of the vehicle relative to the oncoming vehicle, a distance between the vehicle and the oncoming vehicle, a distance between the vehicle and the intersection, and a speed of the oncoming vehicle. A plurality of features in the raw data can be obtained by an environment sensing unit, and the environment sensing unit includes at least one of a radar, a camera and a GPS positioning device.

Another embodiment according to an aspect of the present invention provides a support vector machine-based intelligent driving method at an intersection, which is applied to a vehicle and includes a support vector machine providing step, a data processing step, and a decision-making step. In the step of providing the support vector machine, a support vector machine is provided, and the provided support vector machine has undergone a training process in advance. During the training process, a training data is provided to the support vector machine, and the training data is reduced by one dimension from an original data module and a time compensation module are processed and obtained, wherein the original data includes a plurality of training samples, each training sample includes a time total value passing through an intersection, and each sample in a plurality of sampling time points within the time total value The p features corresponding to the time point and the next decision; the dimension reduction module integrates the p features into k new features, the time compensation module provides a preset time, and the time compensation module samples any one of the training samples. The time point and the new features corresponding to other sampling time points before any of the foregoing sampling time points are regarded as a sequence to be expanded, and when the length of the sequence to be expanded is less than the number of sampling time points in the preset time , a new sequence to be expanded is formed after the sequence to be expanded is filled with an estimated value, and the time compensation module reshapes the new sequence to be expanded into an expanded sequence, and the length of the expanded sequence is equal to the sampling time in the preset time. The number of points, wherein p and k are positive integers, and p>k; in the data processing step, the p features obtained by an environment sensing unit are processed by the dimension reduction module and the time compensation module, and then provided to The support vector machine is used for classification; in the decision-making step, the behavior of the vehicle is determined by the classification result of the support vector machine.

According to various embodiments of the aforementioned SVM-based intelligent driving method at an intersection, the time compensation module can use a principal component analysis method. Or the time complement module can use a uniform scaling method.

An embodiment according to another aspect of the present invention provides a support vector machine-based intersection intelligent driving system, which is applied to a vehicle. The support vector machine-based intersection intelligent driving system includes a processing unit and an environment sensing unit ; The processing unit is arranged in the vehicle and includes a dimension reduction module, a time compensation module and a support vector machine. The dimension reduction module integrates the p features corresponding to each sampling time point in the plurality of sampling time points into k new features, wherein p and k are positive integers, and p>k; the time compensation module provides a preset Time, the time complement value module regards any sampling time point and the new features corresponding to other sampling time points before any of the aforementioned sampling time points as a sequence to be expanded, and when the length of the sequence to be expanded is less than the preset time When there is the number of sampling time points in the to-be-expanded sequence, a new sequence to be expanded is formed after an estimated value is added to the sequence to be expanded, and the time complement value module reshapes the new sequence to be expanded into an expanded sequence, the length of the expanded sequence is equal to the number of sampling time points in the preset time; the support vector machine is trained by a training data, and the training data is obtained from a raw data after being processed by a dimension reduction module and a time compensation module, and the original data includes multiple training samples, Each training sample includes a time total for passing through an intersection, and p features and a current decision corresponding to each sampling time point within the time total. The environment sensing unit is disposed on the vehicle and is connected to the processing unit with a signal, and the environment sensing unit is used to obtain p features; wherein, the p features obtained by the environment sensing unit are processed by the dimension reduction module and the time compensation module of the processing unit. , provided to the support vector machine for classification, and the classification result of the support vector machine is used to determine the driving behavior of the vehicle.

According to various embodiments of the aforementioned SVM-based intersection intelligent driving system, the p features include: a lateral speed of the vehicle relative to an oncoming vehicle, a lateral acceleration of the vehicle relative to an oncoming vehicle, and a longitudinal direction of the vehicle relative to an oncoming vehicle. Speed, a longitudinal acceleration of the vehicle relative to the oncoming vehicle, a distance between the vehicle and the oncoming vehicle, a distance between the vehicle and the intersection and a speed of the oncoming vehicle. Or the environment sensing unit may include at least one of a radar, a camera and a GPS positioning device. Or the current decision may include at least one of acceleration, deceleration or constant speed.

Description of drawings

FIG. 1 is a flowchart illustrating a method for intelligent driving at intersections based on support vector machines according to the first embodiment of the present invention;

2 illustrates a first simulation training of the SVM-based intelligent driving method at intersections according to FIG. 1;

3 illustrates a second simulation training of the SVM-based intelligent driving method at intersections according to FIG. 1;

FIG. 4 shows a third simulation training of the intelligent driving method based on the support vector machine at intersection according to FIG. 1;

FIG. 5 illustrates the first accumulation rate of the first simulated training of FIG. 2;

FIG. 6 illustrates the first accumulation rate of the second simulation training of FIG. 3;

FIG. 7 illustrates the first accumulation rate of the third simulation training of FIG. 4; and

FIG. 8 is a block diagram illustrating an intelligent driving system at an intersection based on a support vector machine according to a third embodiment of the present invention.

Detailed ways

Embodiments of the present invention will be described below with reference to the accompanying drawings. For the sake of clarity, many practical details are set forth in the following description. The reader should understand, however, that these practical details should not be used to limit the invention. That is, in some embodiments of the present invention, these practical details are unnecessary. Furthermore, for the purpose of simplifying the drawings, some well-known and conventional structures and elements will be shown in the drawings in a simplified and schematic manner; and repeated elements will possibly be designated by the same or similar numerals.

Please refer to FIG. 1 , wherein FIG. 1 shows a flowchart of a method 100 for intelligent driving at intersections based on support vector machines according to the first embodiment of the present invention. The SVM-based intelligent driving method 100 at an intersection is applied to a vehicle and includes a SVM providing step 110 , a data processing step 120 and a decision-making step 130 .

In the support vector machine providing step 110, a support vector machine is provided. The support vector machine has undergone a training process in advance. During the training process, a training data is provided to the support vector machine. A time compensation module is processed and obtained, wherein the original data includes a plurality of training samples, each training sample includes a time total value passing through an intersection, and each sampling time point in a plurality of sampling time points within the time total value The corresponding p features and a current decision, the dimension reduction module integrates the p features into k new features, the time compensation module provides a preset time, and the time compensation module depends on any sampling time point in any training sample. And the new features corresponding to other sampling time points before any of the aforementioned sampling time points are a sequence to be expanded, and the time complement value module reshapes the sequence to be expanded into an expanded sequence, and the length of the expanded sequence is equal to the preset time The number of sampling time points in , where p and k are positive integers, and p>k.

In the data processing step 120, the p features obtained by an environment sensing unit are processed by the dimension reduction module and the time compensation module, and then provided to the support vector machine for classification.

In decision step 130, the driving behavior of the vehicle is determined based on the classification result of the support vector machine.

In this way, after the training data and the features obtained while driving are processed by the dimension reduction module and the time compensation module, there will be a time dependency, which can improve the accuracy of the prediction results. The details of the intelligent driving method 100 at an intersection based on the support vector machine will be described in detail later.

A support vector machine is a supervised machine learning classifier that can be used to assist in determining vehicle behavior. In the support vector machine providing step 110, the provided support vector machine is trained through the training process, and can determine the deceleration, acceleration or constant speed behavior of the vehicle when passing through the intersection.

In the training process, a simulation method can be used to simulate the situation of the vehicle passing through the intersection, so as to form a plurality of training samples. In the first embodiment, the simulated platform may be the PreScan advanced driver assistance system (Advanced Driver Assistance Systems; ADAS) developed by Tass International, which can build relevant intersection information to simulate vehicles passing through the intersection. In other embodiments, a plurality of training samples may be obtained on the actual road, or other simulation software may be used, which is not limited thereto.

When the vehicle passes through the intersection once, the data obtained can be regarded as a training sample. That is to say, when the vehicle passes through the intersection ten times, ten training samples can be obtained. Each training sample includes the total time of passing through the intersection, and the p features and a next decision corresponding to each sampling time point among a plurality of sampling time points within the time total value. For example, assuming that in the first training sample, the total time for passing the intersection is 2 seconds, and sampling is performed every 0.4 seconds, there will be 5 sampling time points, and each sampling time point will collect p features and a current decision, the current decision can be acceleration, deceleration or constant speed, and p features can include a lateral speed of the vehicle relative to an oncoming car, a lateral acceleration of the vehicle relative to the oncoming car, and a longitudinal direction of the vehicle relative to the oncoming car. Speed, a longitudinal acceleration of the vehicle relative to the oncoming vehicle, a distance between the vehicle and the oncoming vehicle, a distance between the vehicle and the intersection, and a speed of the oncoming vehicle.

The data of the first training sample can be shown in Table 1. Among them, in a single training sample, the p features obtained at q sampling time points can form an original feature matrix X, X=(x ₁ ,...,x _p ), where x _i =(x _i1 ,..., x _iq ) ^T , and the reader should know that when there are n training samples, there will be n original feature matrices X _l corresponding to different sampling time points q _l , where n and q are positive integers, and l is 1 A positive integer to n, where i is a positive integer from 1 to p. All the current decisions in the n training samples can form a current decision matrix ZZ, ZZ=(z ₁ ,...,z _n ). Hereinafter, T _lw represents the w-th sampling time point in the l-th training sample, w is a positive integer from 1 to q _l , x _lwi represents the i-th feature obtained at the sampling time point T _lw , and z _lw represents the The current decision obtained at the sampling time point T _lw . Therefore, T ₁₁ in Table 1 is the first sampling time point in the first training sample, which is 0.4 seconds in the first embodiment, and T ₁₂ is the second sampling time point in the first training sample , which is 0.8 seconds in the first embodiment, x ₁₁₁ represents the first feature obtained at the first sampling time point T ₁₁ in the first training sample, and x ₁₂₂ represents the second feature in the first training sample The second feature obtained at the sampling time point T ₁₂ , z ₁₃ represents the current decision obtained at the third sampling time point T ₁₃ in the first training sample, and so on, and will not be repeated.

Table 1. The first training sample

Also, suppose that in the second training sample, the total time for passing through the intersection is 2.4 seconds, and sampling is performed every 0.4 seconds, with a total of 6 sampling time points. Then the data of the second training sample can be shown in Table 2.

Table 2. The second training sample

Assuming that there are only 2 training samples, the original data includes the data in Table 1 and Table 2.

The above-mentioned raw data will be converted into training data after being processed by the dimension reduction module and the time compensation module. The dimension reduction module can use Principal Component Analysis (PCA), Partial Least Squares Regression (PLSR), Multidimensional Scaling (MDS), Projection Pursuit method ), Principal Component Regression (PCR), Quadratic Discriminant Analysis (QDA), Regularized Discriminant Analysis (RDA) and Linear Discriminant Analysis (LDA) Wait. Preferably, the time compensation module uses a principal component analysis method, and the relevant formulas of the principal component analysis method are as shown in formula (1), formula (2) and formula (3).

Y=a ^T X (1).

a为系数矩阵，a_ji表示第i个特征x_i所对应的系数。而第一实施例中，维度降低模块使各训练样本中的p个特征整合成1个新特征，也就是说，上述的k为1。因此，经维度降低模块处理后的第1个训练样本如表3所示，经维度降低模块处理后的第2个训练样本如表4所示。其中，y_ljw代表经重整后对应取样时点T_lw的第j个新特征，经维度降低模块处理后的数据为(Y_l,z_l)。The above formula is a formula based on one sampling time point in one training sample, so the variable l representing the number of training samples and the variable w representing the sampling time point are not included. Among them, Y represents the integrated new feature matrix, which contains k new features, namely Y=(y ₁ ,....,y _k ), where y _j represents the jth new feature, and j is a positive value from 1 to k Integer. And when considering n training samples and their corresponding sampling time points q _l , Y _l =(y ₁₁ ,....,y _lk ),

a is the coefficient matrix, and a _ji represents the coefficient corresponding to the i-th feature x _i . In the first embodiment, the dimension reduction module integrates p features in each training sample into a new feature, that is, the above k is 1. Therefore, the first training sample processed by the dimension reduction module is shown in Table 3, and the second training sample processed by the dimension reduction module is shown in Table 4. Wherein, y _ljw represents the j-th new feature corresponding to the sampling time point T _lw after reformation, and the data processed by the dimension reduction module is (Y _l , z _l ).

Table 3. The first training sample processed by the dimension reduction module

sampling time Y₁ z₁ T₁₁ y₁₁₁ z₁₁ T₁₂ y₁₁₂ z₁₂ T₁₃ y₁₁₃ z₁₃ T₁₄ y₁₁₄ z₁₄ T₁₅ y₁₁₅ z₁₅

Table 4. The second training sample processed by the dimension reduction module

sampling time Y₂ z₂ T₂₁ y₂₁₁ z₂₁ T₂₂ y₂₁₂ z₂₂ T₂₃ y₂₁₃ z₂₃ T₂₄ y₂₁₄ z₂₄ T₂₅ y₂₁₅ z₂₅ T₂₆ y₂₁₆ z₂₆

Then, the above data will be processed by the time compensation module for compensation. Since the classification data input by the supervised classifier must be of the same length and sequence data, the accumulated data of each unit time can be pulled into the same time length through the time compensation module.

The compensation method of the time compensation module may adopt a dynamic time warping (DTW) method or a uniform scaling method (Uniform scaling). Preferably, the time compensation module adopts a uniform scaling method.

When the uniform scaling method is used, the time compensation module can provide a preset time, wherein the preset time can be equal to the largest of the total time values. That is to say, in the first embodiment, the total time value of the first training sample is 2 seconds, the total time value of the second training sample is 2.4 seconds, and the maximum value is 2.4, so the preset time can be set to 2.4 seconds, the number of sampling time points in the preset time is 6.

Before performing the complement value, the time complement value module regards the new features corresponding to any sampling time point in any training sample and other sampling time points before the aforementioned any sampling time point as a sequence to be expanded. Table 5 is a list of numbers to be expanded, wherein L _ljw represents a sequence of numbers to be expanded, and L _ljw =(y _lj1 , . . . , y _ljw ). If the p features in each training sample are integrated into a new feature after the dimension reduction module, then j in Table 7 is all 1.

Table 5. List of numbers to be expanded

In more detail, in Table 5, the number to be expanded L _ljw includes all the j-th new features y _{lj1 ˜y ljw} corresponding to the first to w-th sampling time points T _l1 ˜T _lw in the l-th training sample . For example, for the extended sequence L ₁₁₁ , w=1, j=1, so the extended sequence L ₁₁₁ has a first new feature y ₁₁₁ at the first sampling time point T ₁₁ , and the to-be-extended sequence L ₁₁₁ has a length of 1, that is, it consists of a single value. For the to-be-expanded sequence L ₂₁₃ , w=3, j=1, so the expanded sequence L ₂₁₃ has three first to third sampling time points T ₂₁ to T ₂₃ in the second training sample. There are new features y ₂₁₁ -y ₂₁₃ , and the length of the sequence L ₂₁₃ to be extended is 3, that is, it consists of three numerical values, and so on. Therefore, the to-be-expanded sequence can be reformed into an expanded sequence through the complementary value, and the length of the expanded sequence is equal to the number of sampling time points in the preset time. Since in the first embodiment, the number of sampling time points in the preset time is 6, the extended sequence consists of 6 values. Therefore, after the time compensation module complements the value, all the extended arrays are composed of 6 values, as shown in Table 6, where L ^* _ljw represents the extended array. If the p features in each training sample are integrated into a new feature after the dimension reduction module, then j in Table 7 is all 1.

Table 6. List of expansion numbers

The extended sequence L ^* _ljw =(L ^* _ljw1 , . . . , L ^* _ljwr ) in Table 6 is the conversion of the to-be-extended sequence _L1jw through equations (4) and (5).

表示地板公式，且

Z为整数集，也就是说，将r×w/q_s的结果无条件舍去仅保留整数。 _{where r = 1} , . _s = max(q _l ).

represents the floor formula, and

Z is a set of integers, that is to say, the result of r×w/q _s is unconditionally discarded and only integers are retained.

For example, L ₁₁₃ = (y ₁₁₁ , y ₁₁₂ , y ₁₁₃ ) expands to L ^* ₁₁₃ = (L ^* ₁₁₃₁ , L ^* ₁₁₃₂ , L ^* ₁₁₃₃ , L ^* ₁₁₃₄ , L ^* ₁₁₃₅ , L ^* ₁₁₃₆ ), and L ^* ₁₁₃₁ = y ₁₁₁ , L ^* ₁₁₃₂ = y ₁₁₁ (because 2x3/6 = 1, so choose to fill y ₁₁₁ at the first position of L ₁₁₃ into the second position of L ^* ₁₁₃ ), L ^* ₁₁₃₃ = y ₁₁₁ (because 3x3/6=1.5, unconditional rounding only keeps the integer as 1, so choose to fill the y ₁₁₁ at the first position of L ₁₁₃ into the third position of L ^* ₁₁₃ ), L ^* ₁₁₃₄ = y ₁₁₂ (Because 4x3/6=2, so choose to fill y ₁₁₂ at the second position of L ₁₁₃ into the fourth position of L ^* ₁₁₃ ), L ^* ₁₁₃₅ = Y ₁₁₂ (because 5x3/6=2.5, unconditionally discarded Only the integer is kept as 2, so the selection will fill the fifth position of L ^* ₁₁₃ with the 2nd position y ₁₁₂ of L ₁₁₃ ), L ^* ₁₁₃₆ = y ₁₁₃ (since 6x3/6 = 3, the selection will be at L ₁₁₃ 3rd position y ₁₁₃ 's complement into L ^* ₁₁₃ 's sixth position). It should be noted here that, when the length of the above-mentioned sequence to be expanded is greater than the number of sampling time points in the preset time, that is, the length of the sequence to be expanded is greater than the length of the expanded sequence after expected reformation , the value can still be reduced according to formula (4) and formula (5), so as to achieve the purpose of the present invention.

训练数据如表7所示，其中最后一列因其不需补值而可以维持原数值，但为了方便表示其与支持向量机的关系，仍以L^* _ljwr表示。若经维度降低模块后使各训练样本中的p个特征整合成1个新特征，则表7中的j均为1。The original data processed by the dimension reduction module is (Y _l , z _l ), and (Y _l , z _l ) are processed by the time compensation module and then converted into training data (L ^* _l , z _l ), L ^* _l =(L ^* l1,...,L ^* _lk ),

The training data is shown in Table 7, in which the last column can maintain the original value because it does not need to be supplemented, but it is still represented by L ^* _ljwr for the convenience of expressing its relationship with the support vector machine. If the p features in each training sample are integrated into a new feature after the dimension reduction module, then j in Table 7 is all 1.

Table 7. Training data

The above training data can be provided to the support vector machine to find out the hyperplane. The relevant formulas of the support vector machine are shown in equations (6) to (9), and the starting target of the support vector machine is shown in equation (6) As shown, formula (6) is substituted into formula (7), and formulas (6) and (7) are rewritten as formula (8) according to calculus and the if-and-only principle.

Among them, C is the penalty coefficient (cost variable) and is greater than 0, W is the element coefficient (entries parameter), ξ _l is the slack variable (slack variable), b is the intercept term parameter (intercept term), α _d , α _e are the tension parameters The lagrange multiplier and Φ are radial bias function kernels, which extend the hyperplane to nonlinear cuts. L ^* _l , L ^* _d , and L ^* _e represent the above-mentioned extended value L ^* _ljwr , other variables are omitted for simplicity, and d and e are both variables.

After the support vector machine passes through this training data, it can find the hyperplane to assist in the decision-making.

In the data processing step 120, when the vehicle passes through the intersection, the environment sensing unit will collect p features in real time. The environment sensing unit may include multiple sensing devices such as radar, camera, GPS positioning device, etc. To detect distance, vehicle speed, etc., p features can be sensed, but the type and number of sensing devices are not limited to this. The p features collected by the vehicle at each sampling time point will enter the dimension reduction module and the time compensation module for processing, and the processing methods are as described above. Next, in the decision step 130, the processed data enters the support vector machine. Since the support vector machine has already gone through the training process and found the hyperplane, the immediately obtained p features are processed into the support vector machine, that is, A classification result can be generated, and the driving behavior of the vehicle, such as deceleration, acceleration or constant speed, can be determined based on the classification result.

Please refer to FIG. 2 , FIG. 3 and FIG. 4 , wherein FIG. 2 shows a first simulation training of the SVM-based intelligent driving method 100 at intersection according to FIG. 1 , and FIG. A second simulation training of the intelligent driving method 100 at the intersection, FIG. 4 shows a third simulation training of the intelligent driving method 100 based on the support vector machine at the intersection according to FIG. 1 . In the first simulation training, the second simulation training and the third simulation training, the vehicle V1 passes through a T-shaped intersection, and in the first simulation training, there is no oncoming vehicle on the lateral road R1; in the second simulation training, There is an oncoming vehicle V2 on the left side of the lateral road R1; during the third simulation training, there is an oncoming vehicle V2 on the right side of the lateral road R1.

The speed of the vehicle V1 is 40 kilometers per hour, and the speed of the oncoming vehicle V2 is between 15 kilometers per hour and 40 kilometers per hour. The collected 7 features are a lateral velocity of the vehicle V1 relative to the oncoming vehicle V2, a lateral acceleration of the vehicle V1 relative to the oncoming vehicle V2, a longitudinal velocity of the vehicle V1 relative to the oncoming vehicle V2, and a longitudinal acceleration of the vehicle V1 relative to the oncoming vehicle V2. , a distance between the vehicle V1 and the oncoming vehicle V2, a distance between the vehicle V1 and the intersection, and a speed of the oncoming vehicle V2. In addition, the first simulation training, the second simulation training and the third simulation training each contain 20 training samples. The current decision-making includes deceleration, constant speed, and acceleration. The preset time of the first simulation training is 16.9 seconds, and the second simulation training The preset time for the simulation training is 28.8 seconds, and the preset time for the third simulation training is 21.7 seconds.

Please refer to FIGS. 5 , 6 and 7 , wherein FIG. 5 shows the first accumulation rate of the first simulated training in FIG. 2 , FIG. 6 shows the first accumulation rate of the second simulated training in FIG. 3 , and FIG. 7 shows The first accumulation rate of the third simulation training of FIG. 4 . The average first accumulation rate of the first simulation training was 0.9619, the average first accumulation rate of the second simulation training was 0.7588, and the average first accumulation rate of the third simulation training was 0.8014. The above-mentioned first accumulation rates were all above 0.7. Indicates that the information after dimensionality reduction has explained the original data as much as possible and meets the requirements.

Table 8 shows the comparison between the decision result accuracy (AC) of the first comparative example and the intelligent driving method 100 based on the support vector machine at the intersection of the present case under the same situation as the first simulation training. It can be seen from Table 8 that using The decision-making accuracy of the intelligent driving method 100 at the intersection based on the support vector machine is relatively high. Among them, the first comparative example also uses the support vector machine for classification, but the difference is that the support vector machine of the first comparative example is only trained on the original data.

Table 8. Comparison of decision accuracy in the first simulated training situation

Table 9 shows the comparison between the decision result accuracy (AC) of the second comparative example and the intelligent driving method 100 based on the support vector machine at intersections of the present case under the same conditions as the second simulation training. It can be seen from Table 9 that using The decision-making accuracy of the intelligent driving method 100 at the intersection based on the support vector machine is relatively high. Among them, the second comparative example also uses the support vector machine for classification, but the difference is that the support vector machine of the second comparative example is only trained on the original data.

Table 9. Comparison of decision-making accuracy in the second simulated training situation

Table 10 shows the comparison of the decision result accuracy (AC) between the third comparative example and the intelligent driving method 100 based on support vector machine at intersections in this case under the same conditions as the third simulation training. The decision-making accuracy of the intelligent driving method 100 at the intersection based on the support vector machine is relatively high. Among them, the third comparative example also uses the support vector machine for classification, but the difference is that the support vector machine of the third comparative example is only trained on the original data.

Table 10. Comparison of decision-making accuracy in the third simulated training situation

It should be noted here that all the above-mentioned tests are the simulation results of PreScan, but can also be tested using actual roads.

In the second embodiment of the present invention, in the step of providing the support vector machine, the time compensation module provides a preset time, and the time compensation module depends on any sampling time point in any training sample and at any one of the aforementioned The new features corresponding to other sampling time points before the sampling time point are a sequence to be expanded, and when the length of the sequence to be expanded is less than the number of sampling time points in the preset time, the sequence to be expanded is filled with the aforementioned A new sequence to be expanded is formed after an estimated value of the next sampling time point at the sampling time point, and the time complement value module reshapes the new sequence to be expanded into an expanded sequence, and the length of the expanded sequence is equal to the length of the sequence in the preset time. the number of sampling time points.

In more detail, assuming that the original data includes the data in Table 1 and Table 2, the preset time is 2.4 seconds, and the number sequence to be expanded is shown in Table 5, because the accumulated sampling time points 0 to T ₁₁ in Table 5 correspond to the to-be-expanded The length of the sequence is 1, which is less than the number of sampling time points in the preset time (equal to 6), and the data of the next sampling time point _T12 has not been obtained from the accumulated sampling time points 0 to _T11 , so it can be An estimated value y' _1j2 is added to correspond to the sampling time point T ₁₂ . Similarly, in Table 5, the length of the sequence to be expanded corresponding to all the cumulative sampling time points is less than 6, so it is necessary to fill in the estimated value, then the new sequence to be expanded L' _ljw table is shown in Table 11, where y' _ljw represents the estimated value. It should be noted here that, in the new sequence to be expanded in Table 11, the length of the sequence to be expanded corresponding to the cumulative sampling time points 0 to T ₁₁ and 0 to T ₂₅ is 6, and there is no need to perform supplementation, but in order to The difference between the new to-be-expanded sequence and the to-be-expanded sequence is clearly shown, which is still listed, rather than limiting the present invention. In the second embodiment, the reformation of the new sequence to be expanded into the sequence to be expanded and the following support vector machine training methods are the same as those in the first embodiment, and will not be repeated.

Table 11. List of new to-be-expanded numbers

In the second embodiment, the conditional assignment of the estimated value can be obtained by using the joint assignment of the sequence to be expanded and the data of all new features y _ljw obeying the Gaussian assignment, and finally the estimated value can be obtained. Since the above-mentioned new features y _ljw after dimensionality reduction can be adapted to the data subject to Gaussian distribution (or Gaussian random process, marginal distribution), that is, y _ljw ~GP(μ _jw , Σ _jw ), where (μ _jw , Σ _jw ) can be estimated from equations (10) and (11). In addition, the joint distribution (jointdistribution) of other new features y _ljw (equivalent to the sequence to be expanded) before the estimated value can be obtained from equation (12), where w in equation (12) is greater than 2, and in equation (14) where c refers to the c-th training sample, and m refers to the m-th training sample.

y _lj [1:(w-1)]≡(y _lj1 ,...,y _lj(w-1) )～GP(μ _j[1:(w-1)] ,Σj _[1:(w-1 ) _)] ) (12).

(μ _j[1:(w-1)] , Σ _j[1:(w-1)] ) can be estimated by Equation (13), Equation (14) and Equation (15).

Finally, the conditional distribution of predicting the next unit time at the past time point can be obtained, that is, the conditional distribution of the predicted sampling time point _T12 at the cumulative sampling time point 0 to _T11 , the conditional distribution is as follows ( 16), as shown in formula (17).

可由

估算，有了条件分配，可以根据条件分配预测下一取样时点的预估值(其可根据Monte Carlo概念生成预估值)。and

by

Estimation, with conditional assignments, an estimate of the next sampling point in time can be predicted based on the conditional assignment (which can generate an estimate according to the Monte Carlo concept).

为y_ljw，y_lj[1:w-1]的变异数矩阵(covariance matrix)，GP(·，·)为高斯随机过程的缩写，μ为高斯随机过程平均数，Σ为高斯随机过程变异数矩阵(Covariance-variance矩阵)，

为变异数矩阵内的元素(Element)估计量。in,

is y _ljw , the covariance matrix of y _lj[1:w-1] , GP(·,·) is the abbreviation of Gaussian random process, μ is the mean of Gaussian random process, and Σ is the variance of Gaussian random process matrix (Covariance-variance matrix),

is an estimator of the elements in the variance matrix.

Table 12 shows the accuracy of decision-making results in the first simulated training situation after adding the estimated value, Table 13 shows the accuracy of the decision-making result in the second simulated training situation after adding the estimated value, and Table 14 shows the accuracy of the decision-making result in the second simulated training situation after adding the estimated value. The accuracy of the decision results in the third simulated training situation after evaluation. From the results in Table 12, Table 13 and Table 14, it can be seen that adding the estimated value can improve the decision accuracy.

Table 12. Decision Accuracy in the First Simulation Training Scenario

Table 13. Decision Accuracy in the Second Simulation Training Scenario

Table 14. Decision Accuracy in the Third Simulation Training Scenario

Please refer to FIG. 8 , wherein FIG. 8 is a block diagram of an intelligent driving system 200 based on a support vector machine at an intersection according to a third embodiment of the present invention. The SVM-based intersection intelligent driving system 200 is applied to a vehicle and includes a processing unit 210 and an environment sensing unit 220. The processing unit includes a dimension reduction module 211, a time compensation module 212 and a support vector machine 213, The dimension reduction module 211 integrates the p features corresponding to each sampling time point in the plurality of sampling time points into k new features, wherein p and k are positive integers, and p>k; the time compensation module 212 provides a In the preset time, the time compensation module 212 regards any sampling time point and the new features corresponding to other sampling time points before any of the foregoing sampling time points as a sequence to be expanded, and when the length of the sequence to be expanded is less than When there are the number of sampling time points in the preset time, a new sequence to be expanded is formed after an estimated value is added to the sequence to be expanded, and the time compensation module 212 reshapes the new sequence to be expanded into an expanded sequence, The length of the extended sequence is equal to the number of sampling time points in the preset time; the support vector machine 213 is trained by a training data, and the training data is obtained from a raw data after being processed by the dimension reduction module 211 and the time compensation module 212. The data includes a plurality of training samples, and each training sample includes a total time value of passing through an intersection, and p features and a current decision corresponding to each sampling time point in the total time value.

The environment sensing unit 220 is disposed in the vehicle and is connected to the processing unit 210 with a signal. The environment sensing unit 220 is used to obtain p features; wherein, the p features obtained by the environment sensing unit 220 are passed through the dimension reduction module 211 of the processing unit 210 and the After being processed by the time compensation module 212, it is provided to the support vector machine 213 for classification, and the classification result of the support vector machine 213 is used to determine the driving behavior of the vehicle.

In this way, the vehicle can be assisted in determining the acceleration, deceleration or constant speed behavior when passing through the intersection. The processing details of the dimension reduction module 211 and the time compensation module 212 and their relationship with the support vector machine 213 are as described above, and will not be repeated here. The p features may include a lateral velocity of the vehicle relative to an oncoming vehicle, a lateral acceleration of the vehicle relative to an oncoming vehicle, a longitudinal velocity of the vehicle relative to an oncoming vehicle, a longitudinal acceleration of the vehicle relative to an oncoming vehicle, and a distance between the vehicle and the oncoming vehicle , the distance between the vehicle and the intersection and the speed of the oncoming vehicle. The environment sensing unit 220 may include at least one of a radar, a camera, and a GPS positioning device. Or the current decision may include at least one of acceleration, deceleration or constant speed.

Although the present invention has been disclosed by the above embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the present invention The scope of protection shall be subject to the scope defined by the appended claims.

Claims (13)

1. a method for intelligent driving at intersection based on support vector machine, it is applied to a vehicle, it is characterized in that, this method for intelligent driving at intersection based on support vector machine comprises: A support vector machine providing step is to provide a support vector machine. The support vector machine has undergone a training process in advance. During the training process, a training data is provided to the support vector machine, and the training data is reduced from an original data by a dimension. module and a time compensation module to obtain after processing, wherein, the original data includes a plurality of training samples, each of the training samples includes a time total value passing through an intersection, and a plurality of sampling time points in the time total value Each of the p features corresponding to the sampling time point and a current decision, the dimension reduction module integrates the p features into k new features, the time compensation module provides a preset time, and the time compensation module will The new features corresponding to any sampling time point in any of the training samples and other sampling time points before the any sampling time point are regarded as a sequence to be expanded, when the length of the sequence to be expanded is less than When the number of the sampling time points in the preset time, a new sequence to be expanded is formed after the sequence to be expanded is supplemented with an estimated value, wherein the sequence to be expanded is based on the joint allocation of the sequence to be expanded and the new features The data of the value obeys the Gaussian distribution to obtain the conditional distribution of the estimated value, and the time compensation module reshapes the new to-be-expanded sequence into an expanded sequence, and the length of the expanded sequence is equal to the preset time. The number of these sampling time points, where p and k are positive integers, and p>k; a data processing step of processing the p features obtained by an environment sensing unit through the dimension reduction module and the time compensation module, and then providing the p features to the support vector machine for classification; and In a decision-making step, the driving behavior of the vehicle is determined based on the classification result of the support vector machine. 2 . The intelligent driving method at an intersection based on a support vector machine according to claim 1 , wherein the dimension reduction module adopts a principal component analysis method. 3 . 3 . The method for intelligent driving at intersections based on support vector machines according to claim 1 , wherein the time compensation module adopts a uniform scaling method. 4 . 4 . The intelligent driving method at an intersection based on a support vector machine according to claim 1 , wherein the preset time is equal to the maximum of the total time values. 5 . 5 . The intelligent driving method at an intersection based on support vector machine according to claim 1 , wherein the p features comprise a lateral speed of the vehicle relative to an oncoming car, a lateral acceleration of the vehicle relative to the oncoming car, 5 . A longitudinal speed of the vehicle relative to the oncoming vehicle, a longitudinal acceleration of the vehicle relative to the oncoming vehicle, a distance between the vehicle and the oncoming vehicle, a distance between the vehicle and the intersection and a speed of the oncoming vehicle. 6 . The method for intelligent driving at intersections based on support vector machines as claimed in claim 1 , wherein the features in the raw data are obtained by the environment sensing unit, and the environment sensing unit comprises a radar and a camera. 7 . and at least one of a GPS positioning device. 7. An intelligent driving method at an intersection based on a support vector machine, which is applied to a vehicle, wherein the intelligent driving method at an intersection based on a support vector machine comprises: A support vector machine providing step is to provide a support vector machine. The support vector machine has undergone a training process in advance. During the training process, a training data is provided to the support vector machine, and the training data is reduced from an original data by a dimension. module and a time compensation module to obtain after processing, wherein, the original data includes a plurality of training samples, each of the training samples includes a time total value passing through an intersection, and a plurality of sampling time points in the time total value Each of the p features corresponding to the sampling time point and a current decision, the dimension reduction module integrates the p features into k new features, the time compensation module provides a preset time, and the time compensation module will The new features corresponding to any sampling time point in any of the training samples and other sampling time points before the any sampling time point are regarded as a sequence to be expanded, and when the length of the sequence to be expanded When it is less than the number of the sampling time points in the preset time, a new sequence to be expanded is formed after the sequence to be expanded is supplemented with an estimated value, and the time complementing module rewrites the new sequence to be expanded. The whole is an extended sequence, and the length of the extended sequence is equal to the number of the sampling time points in the preset time, wherein p and k are positive integers, and p>k; a data processing step of processing the p features obtained by an environment sensing unit through the dimension reduction module and the time compensation module, and then providing the p features to the support vector machine for classification; and In a decision-making step, the driving behavior of the vehicle is determined based on the classification result of the support vector machine. 8 . The intelligent driving method at an intersection based on a support vector machine according to claim 7 , wherein the time compensation module uses a principal component analysis method. 9 . 9 . The intelligent driving method at an intersection based on a support vector machine according to claim 7 , wherein the time compensation module uses a uniform scaling method. 10 . 10. A support vector machine-based intersection intelligent driving system, which is applied to a vehicle, wherein the support vector machine-based intersection intelligent driving system comprises: a processing unit, disposed in the vehicle and comprising: The one-dimensional reduction module integrates the p features corresponding to each of the multiple sampling time points into k new features, where p and k are positive integers, and p>k; A time compensation module provides a preset time, and the time compensation module regards the new features corresponding to any sampling time point and other sampling time points before any sampling time point as a The sequence to be expanded, and when the length of the sequence to be expanded is less than the number of the sampling time points in the preset time, a new sequence to be expanded is formed after the sequence to be expanded is filled with an estimated value, and The time complement module reshapes the new sequence to be expanded into an expanded sequence, the length of the expanded sequence is equal to the number of the sampling time points in the preset time; and A support vector machine, trained on a training data, the training data is obtained from a raw data after being processed by the dimension reduction module and the time compensation module, the raw data includes a plurality of training samples, each of the training samples includes passing through an intersection A time total of , and the p features and an immediate decision corresponding to each of the sampling time points within the time total; and an environment sensing unit, disposed on the vehicle and connected to the processing unit with a signal; The p features obtained by the environment sensing unit are processed by the dimension reduction module and the time compensation module of the processing unit, and then provided to the support vector machine for classification, and the classification result of the support vector machine is used for Determines the driving behavior of the vehicle. 11. The intelligent driving system based on support vector machine of claim 10, wherein the p features comprise a lateral speed of the vehicle relative to an oncoming vehicle, a lateral acceleration of the vehicle relative to the oncoming vehicle, A longitudinal speed of the vehicle relative to the oncoming vehicle, a longitudinal acceleration of the vehicle relative to the oncoming vehicle, a distance between the vehicle and the oncoming vehicle, a distance between the vehicle and the intersection and a speed of the oncoming vehicle. 12 . The intelligent driving system at an intersection based on a support vector machine according to claim 10 , wherein the environment sensing unit comprises at least one of a radar, a camera and a GPS positioning device. 13 . 13 . The intelligent driving system at an intersection based on a support vector machine according to claim 10 , wherein the current decision includes at least one of acceleration, deceleration or constant speed. 14 .

Images