Automatic driving vehicle optimal path planning method based on fused traffic situation

Abstract

The invention discloses an automatic driving vehicle optimal path planning method based on a fused traffic situation, and belongs to the technical field of vehicle automatic driving track planning. Aiming at the problems of insufficient prediction precision of a traditional LSTM model, low search efficiency of an A * algorithm and the like, the method provides a multi-dimensional optimization strategy: firstly, constructing an adaptive probability YH-LSTM model, improving traffic situation prediction precision by dynamically adjusting neuron discarding probability, and optimizing hyper-parameters in combination with a particle swarm optimization (PSO); secondly, an A * algorithm is improved, a path steering angle sine factor is introduced to optimize a cost function, the search range is narrowed, and a road network layering strategy (the expressway / main road is a high layer and the secondary main road / branch is a low layer) is combined to improve the search efficiency; and meanwhile, the road network impedance is dynamically updated by fusing real-time and predicted traffic data, and the vehicle aggregation effect is quantified by adopting a BPR function, so that local congestion is avoided. According to the invention, an efficient and robust solution is provided for global optimal path planning in a complex dynamic traffic environment.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle autonomous driving trajectory planning, and particularly to an optimal path planning method for autonomous driving vehicles based on integrated traffic situation, which combines an improved deep learning model and a hierarchical optimization path search algorithm to achieve global optimal path planning in a dynamic traffic environment. Background Technique

[0002] With the rapid development of computer communication technology and mobile Internet technology, autonomous driving algorithms have gradually replaced traditional driving tasks. The path planning technology of autonomous driving vehicles is the core of the autonomous driving system, and the key lies in generating a global optimal path by combining real-time traffic information and prediction data.

[0003] In the prior art, long short-term memory neural network (LSTM) is usually used to predict traffic situation information, which mainly realizes traffic situation prediction by capturing the temporal dependence and non-linear characteristics of traffic data. However, in actual use, the traditional LSTM model usually adopts a dropout layer with a fixed probability and cannot adaptively adjust the neuron dropout strategy according to the dynamic characteristics of traffic data, resulting in limited prediction accuracy.

[0004] The path planning algorithm mainly provides an optimal path for an autonomous driving vehicle from the current position to the target position. Among them, the A* algorithm has become the mainstream method for path planning due to its high efficiency and optimality. However, its search efficiency is low in a large-scale road network, and it does not consider the impact of dynamic changes in traffic situation on path selection, resulting in the planning result may deviate from the global optimum. Summary of the Invention

[0005] In order to solve the deficiencies mentioned in the above background technique, the purpose of the present invention is to provide an optimal path planning method for autonomous driving vehicles based on integrated traffic situation, which improves the accuracy and efficiency of path planning for autonomous driving vehicles by adaptively adjusting neural network parameters, optimizing path search strategies, and dynamically updating road network impedance.

[0006] The purpose of the present invention can be achieved by the following technical solutions: An optimal path planning method for autonomous driving vehicles based on integrated traffic situation, comprising the following steps:

[0007] S1. Predict traffic situation based on an improved adaptive probability long short-term memory neural network (YH-LSTM), where the YH-LSTM adds a dropout layer to the traditional LSTM model and replaces the dropout neuron dropout probability value with the singular ratio of traffic data;

[0008] S2. Optimize the YH-LSTM model based on the particle swarm optimization algorithm (PSO) to obtain the PSO-YH-LSTM model;

[0009] S3. Weightedly fuse the traffic situation information predicted by PSO-YH-LSTM and the real-time traffic situation information to generate a fused passing cost and update the road network model;

[0010] S4. Improve the cost function of the A* algorithm, introduce the sine value of the steering angle between the path and the target point as the angle factor, and construct an improved A* algorithm model;

[0011] S5. Classify urban roads, and improve the A* algorithm layer by layer for path planning;

[0012] S6. According to the vehicle aggregation effect, convert the self-driving vehicles on the planned path into new traffic volume and dynamically update the road network impedance.

[0013] Furthermore, S1 specifically includes the following steps:

[0014] S101. Use an LSTM neural network to predict the future traffic situation. The data comes from the actual traffic speed data in the Q-Traffic dataset collected by Baidu Brain. Based on the characteristic analysis of urban traffic flow, the numerical values of adjacent traffic flows are determined as the input of the model.

[0015] S102. Adding a "dropout" layer to the traditional LSTM model can reduce the probability of the model overfitting. However, it does not determine how to calculate the probability value of neuron discarding according to the specific application scenario, but simply uses the previous empirical values. Therefore, replace the probability value of selectively discarding neurons by "dropout" with the traffic data singularity ratio, as follows:

[0016]

[0017] H j =H d +H w

[0018] It can be deduced from the above two formulas that:

[0019]

[0020] In the formula, H d represents the number of discarded nodes; H j represents the number of nodes in each layer; H q represents the number of singularities; H represents the number of nodes in a single-layer network; H w represents the number of nodes not discarded in a single-layer network; H qd represents the number of nodes discarded and belonging to noise in a single-layer network.

[0021] Furthermore, S2 specifically includes the following steps:

[0022] S201. The particle swarm optimization algorithm can effectively solve various optimization problems such as continuous, discrete, non-linear, and multi-objective problems. Since the influencing factors of urban road traffic flow are complex and the traffic flow state of the road network is unpredictable, showing obvious non-linear characteristics, the particle swarm optimization algorithm is used to optimize the YH-LSTM traffic situation prediction model.

[0023] The parameter values such as the number of neurons in the hidden layer of the YH-LSTM neural network, the number of iterations of the prediction model, and the learning rate are all optimized based on the particle swarm optimization algorithm. After optimization, the optimal hyperparameter settings of the PSO-YH-LSTM prediction model are obtained.

[0024] S202. Compare the prediction effect of the PSO-YH-LSTM model with that of the YH-LSTM model and the traditional LSTM model.

[0025] S203. The comparisons made in S202 are all horizontal comparisons based on the LSTM model. To further verify the prediction effect of the constructed PSO-YH-LSTM model, a GRU prediction model is used for vertical comparison.

[0026] Furthermore, S3 specifically includes the following steps:

[0027] S301. First, by abstracting the urban traffic network, a road network geometric model is established. The "graph" data structure is used to express the road network geometric structure, and the optimal path problem is studied based on graph algorithms. Therefore, a road network model is constructed as follows:

[0028]

[0029] In the formula: V represents the set of nodes; E represents the set of edges in the road network model, and (v i , v j ) and (v j , v i ) are two directions of the same road, with different weights; W represents the set of weights of each edge (v i , v j ), which can be replaced according to specific requirements.

[0030] Usually, the traffic flow conditions in opposite directions of the same road section in the urban traffic network are different, and even extreme situations may occur. Therefore, a topological graph that can express direction differences is used to describe the actual urban road network.

[0031] S302. Divide the urban road traffic into four levels, which are represented by the numbers 1 to 4 for smooth traffic, slow traffic, congestion, and severe congestion respectively. To ensure the original differences between data, standardize the traffic index data. During the process of calculating the road weight, make the passing weight 1 when the traffic is smooth through the standardization of traffic data, and define the passing weight values for slow driving, traffic congestion, and severe traffic congestion based on this, and their range is between [0, 1], specifically as follows:

[0032]

[0033] In the formula: f i represents the result after standardization; γ i represents the original value of the i-th traffic data; max(γ) represents the maximum value of the traffic data.

[0034] Reverse the four levels of traffic states so that the level is proportional to the value, and finally perform standardization processing on the value.

[0035] S303. Standardize the real-time situation information to obtain the urban road traffic conditions, and then obtain the future road traffic situation information according to the PSO-YH-LSTM prediction model, and then the corresponding traffic state can be obtained. Design a weighted fusion method for traffic situation information to fuse the two traffic state values, specifically as follows:

[0036] y = αy1 + (1 - α)y2

[0037] In the formula: y represents the fused traffic situation value; y1 represents the real-time traffic congestion value of the road section; y2 represents the predicted traffic condition value of the road section; α is an arbitrary constant (0 to 1).

[0038] In the research of path planning, the real-time situation information of the road has a greater impact than the predicted information. Therefore, the value of α should be greater than 0.5; after weighted fusion processing, the value range of the traffic situation value of each road section is [0, 1], and the fused traffic situation shows the reverse in describing the urban traffic congestion situation. The larger the value, the smaller the time cost of urban road passage, and vice versa.

[0039] Furthermore, S4 specifically includes the following steps:

[0040] S401. Modify the evaluation function of the A* algorithm by adding the sine value of the steering angle between the path and the target point to limit the search conditions of the algorithm, specifically as follows:

[0041]

[0042] Where: f(n) represents the estimated cost passing through node n; g(n) represents the actual cost from the nth node to the starting point; h(n) is the estimated cost from the nth node to the target point; sin(α) represents the vertex with the current node as the included angle vertex, and the sine value formed by this node, adjacent nodes, and the target point.

[0043] Introduce the sin(α) function into the estimated cost function. Since sin(α) is monotonically increasing in the interval and the function value is (0 - 1). Therefore, the larger the included angle between the current node, adjacent nodes, and the target point, the larger the sine value, which leads to an increase in the estimated time cost. So, by introducing the angle factor, it can prevent the autonomous driving vehicle from deviating from the target point direction, while reducing the search node range and improving the overall search efficiency of the algorithm.

[0044] Further, S5 specifically includes the following steps:

[0045] S501. Classify urban roads according to the differences in geometric dimensions, pavement driving quality, road driving speed, etc. of road sections.

[0046] S502. Use the plane search algorithm for road networks at each level, and finally obtain the global path through connection points.

[0047] Further, obtaining the global path by hierarchical processing of the road network specifically includes the following steps:

[0048] (1) Assume the starting point is S, the ending point is D, M1 and M2 are nodes on the middle layer, and M3 and M4 are nodes on the highest layer. Start searching for the path from the starting point S to the target point D.

[0049] (2) First, find the optimal path from the starting point S to the connection point M1 on the intersecting layer.

[0050] (3) Then, find the optimal path from M1 to the highest-level node M3 in the upper layer.

[0051] (4) And so on, then the optimal paths from the target point D to M2 and M4 can be found;

[0052] (5) Merge the optimal path segments of each level (S→M1→M3→M4→M2→D) to form the global optimal path.

[0053] The setting of the number of road levels is an important link in determining the success of path planning. Considering the balance between efficiency and accuracy, the roads are divided into two levels: upper layer (expressway, arterial road), lower layer (sub-arterial road, branch road).

[0054] After the S504 completes the hierarchical processing of urban roads, combined with the improved A* algorithm that introduces the angle factor in S401, a hierarchical improved A* algorithm is constructed.

[0055] Further, S6 specifically includes the following steps:

[0056] S601. To avoid the agglomeration effect brought by the path planning system, when constructing the path planning model, consider the real-time changing traffic conditions of the road network and the agglomeration effect problem caused by the path planning system. Convert the autonomous driving vehicles that have completed path planning into the actual traffic volume on the road and add it to the calculation process of the travel time cost of this path.

[0057] S602. When planning the path, two basic data (real-time traffic situation information and predicted traffic situation information) need to be considered to obtain the travel time cost. In view of this, improve the traditional road impedance model and incorporate the agglomeration effect of traffic flow in the road network into it. The model mainly consists of the estimated cost calculated by the evaluation function and the virtual impedance at the moment of path planning for autonomous driving vehicles. The estimated cost will be calculated according to the method constructed in S303; the virtual impedance will be calculated according to the BPR road impedance function proposed by the Federal Highway Administration of the United States, specifically as follows:

[0058]

[0059] In the formula: t represents the time actually required to pass through this section; t0 represents the free driving time of the section; α and β represent the undetermined parameters of the model, and the recommended values are 0.15 and 4 respectively; q represents the traffic volume of this section at that time, with the unit of vel / h; c represents the actual traffic capacity of the section, with the unit of vel / h.

[0060] Since the virtual impedance considers the influence of the newly added autonomous driving vehicles after the path planning is completed on the travel time cost on the basis of the estimated travel time cost of the integrated traffic situation. Therefore, only need to calculate the influence of the newly added traffic volume on the road traffic efficiency, so the road impedance function is modified, specifically as follows:

[0061]

[0062] In the formula: t z is the newly added time caused by the newly added traffic volume; q z is the newly added traffic volume of the section. Description of the Drawings

[0063] The following further describes the present invention in detail in conjunction with the drawings and specific embodiments. The drawings are only provided for reference and illustration, and are not used to limit the present invention.

[0064] Figure 1 It is the flowchart of the implementation method of the present invention;

[0065] Figure 2 It is the flowchart of LSTM neural network prediction in the present invention;

[0066] Figure 3 It is the flowchart of PSO-optimized YH-LSTM neural network in the present invention;

[0067] Figure 4 It is the prediction result of LSTM neural network in the present invention;

[0068] Figure 5 It is the prediction result of YH-LSTM neural network in the present invention;

[0069] Figure 6 It is the prediction result of PSO-YH-LSTM neural network in the present invention;

[0070] Figure 7 It is the change diagram of loss values of four types of prediction models in the present invention;

[0071] Figure 8 It is the prediction results of four types of prediction models in the present invention;

[0072] Figure 9 It is the schematic diagram of the road network model in the present invention;

[0073] Figure 10 It is the schematic diagram of path selection by traditional A* algorithm in the present invention;

[0074] Figure 11 It is the schematic diagram of hierarchical search of the road network in the present invention;

[0075] Figure 12 It is the flowchart of hierarchical improved A* algorithm in the present invention;

[0076] Figure 13 It is the path selection model of autonomous driving vehicles in the present invention;

[0077] Figure 14 It is the path planning results of each algorithm for nodes 1 to 50 in the example verification of the present invention;

[0078] Figure 15 It is the path planning results of each algorithm for nodes 15 to 39 in the example verification of the present invention;

[0079] Figure 16 It is the three path planning results under the agglomeration effect in the example verification of the present invention;

[0080] Figure 17 It is the comparison diagram of estimated travel time cost in the example verification of the present invention; Detailed implementation manners

[0081] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The accompanying drawings are only provided for reference and illustration, and are not intended to limit the present invention.

[0082] The present invention contemplates an optimal path planning method for autonomous vehicles based on integrated traffic situation, which combines an improved deep learning model and a hierarchical optimization path search algorithm to achieve global optimal path planning in a dynamic traffic environment. The specific process is as follows Figure 1 shown.

[0083] S1 specifically includes the following steps:

[0084] S101. Use an LSTM neural network to predict the future traffic situation. The data comes from the actual traffic speed data in the Q-Traffic dataset collected by Baidu Brain. Based on the feature analysis of urban traffic flow, the numerical values of adjacent traffic flows are determined as the input of the model. The process of using the LSTM neural network to predict the future traffic situation is as follows Figure 2 shown.

[0085] Adding a "dropout" layer to the traditional LSTM model can reduce the probability of the model overfitting, but it does not determine how to calculate the probability value of neuron dropout according to the specific application scenario, and only simply uses the previous empirical values. Therefore, the probability value of selectively discarding neurons by "dropout" is replaced by the traffic data singularity ratio, as follows:

[0086]

[0087] H j =H d +H w

[0088] It can be deduced from the above two formulas that:

[0089]

[0090] In the formula, H d represents the number of discarded nodes; H j represents the number of nodes in each layer; H q represents the number of singularities; H represents the number of nodes in a single-layer network; H w represents the number of nodes not discarded in a single-layer network; H qd represents the number of nodes discarded and belonging to noise in a single-layer network.

[0091] As Figures 3 - 8 shown,

[0092] S2 specifically includes the following steps:

[0093] S201. The particle swarm algorithm can effectively solve various optimization problems such as continuous, discrete, nonlinear, and multi-objective problems. Since the influencing factors of urban road traffic flow are complex and the traffic flow state of the road network is unpredictable, showing obvious nonlinear characteristics, the particle swarm algorithm is used to optimize the YH-LSTM traffic situation prediction model.

[0094] Each particle in the particle swarm continuously performs optimization, and the velocity and position of the individual particle are also constantly changing.

[0095] V in k+1 = ωV in k + c1r1(P in k - X in k ) + c2r2(P gn k - X in k )

[0096]

[0097] In the formula: c1 and c2 represent constantly changing constants; k represents the maximum number of iterations; n is the number of independent particles; X i = (x i1 , x i2 ,..., x iN ) T represents the position of the i-th particle; the velocity of the i-th particle is V i = (V i1 , V i2 ,..., V iN ) T ; its individual extreme value is P i = (P i1 , P i2 ,..., P iN ) T ; the global extreme value of the population is P g = (P g1 , P g2 , …P gN ) T ; ω represents the inertia weight; v in represents the velocity vector value of the particle at the current moment; r1 and r2 represent random numbers distributed between [0, 1].

[0098] To improve the search efficiency of the particle, the position and velocity of the particle will be restricted within the reasonable intervals [-X max , X max , [-V max , V max. The optimization of the PSO particle swarm algorithm is as follows:

[0099]

[0100] The formula can also be expressed as (i - particle, j - dimension), as follows:

[0101] v i,j v(t + 1) = ω·v(t)+c1·r(t)·[pBest(t)-x(t)]+c2·r(t)·[gBest(t)-x(t)] i,j (t)+c1·r i，j (t)·[pBest t,j (t)-x i,j (t)]+c2·r 2，j ·[gBest j (t)-x i,j (t)]

[0102] x i,j x(t + 1) = x(t)+v(t + 1) i,j (t)+v i，j (t+1)

[0103] In the formula, pBest(t) represents the component j of the local extreme value of the individual particle i; gBest(t) represents the component j of the global extreme value among all particles. i,j (t) represents the component j of the local extreme value of the individual particle i; gBest i,j (t) represents the component j of the global extreme value among all particles.

[0104] As Figure 3 shown in the figure is the process of optimizing the YH - LSTM prediction model by PSO, as follows:

[0105] (1) Preprocess the average vehicle speed data, add missing values, remove extreme values, divide the data set into an 80% training set and a 20% validation set, and perform standardization processing on this data;

[0106] (2) Determine the optimization range of the hyperparameters in the YH - LSTM neural network (number of neurons, number of iterations, batch size, learning rate);

[0107] (3) Initialize the parameters in the particle swarm algorithm, including the initial velocity, particle coordinates, etc.;

[0108] (4) Select the fitness function of the PSO algorithm and select the symmetric mean absolute percentage error value to search for the optimal parameter configuration of the particle swarm algorithm;

[0109] (5) Calculate and update the fitness value of the particle swarm algorithm;

[0110] (6) Determine whether the model operation has reached the number of iterations. If it has, retain the parameters. Then apply the optimal parameter configuration to the adaptive probability YH - LSTM model;

[0111] (7) The PSO-YH-LSTM predicts the traffic situation and outputs evaluation indicators. If the prediction result is good, the process ends; otherwise, return to step (5).

[0112] The parameter values such as the number of neurons in the hidden layer of the YH-LSTM neural network, the number of iterations of the prediction model, and the learning rate are optimized based on the particle swarm optimization algorithm. The selected interval is shown in Table 1.

[0113] Table 1 Optimization range of hyperparameters of the PSO-YH-LSTM prediction model

[0114]

[0115] After being optimized by the particle swarm optimization algorithm, the optimal hyperparameter settings of the PSO-YH-LSTM prediction model are shown in Table 2.

[0116] Table 2 Hyperparameter settings of the PSO-YH-LSTM prediction model

[0117]

[0118] S202. From Figures 4 - 6 It can be intuitively seen the prediction results of the three prediction models. We find that the coincidence degree between the true value and the predicted value in the prediction result of the PSO-YH-LSTM model is greater than that in the prediction results of the YH-LSTM model and the traditional LSTM model, indicating that the constructed PSO-YH-LSTM prediction model has higher accuracy.

[0119] The comparisons made in S202 are all horizontal comparisons based on the LSTM model. To further verify the prediction effect of the constructed PSO-YH-LSTM model, the GRU prediction model is used for vertical comparison, as Figure 7 and Figure 8 shown.

[0120] In the evaluation of the traffic flow prediction model effect, the following several commonly used error discrimination methods are usually adopted to reflect all aspects of the model prediction performance and objectively evaluate the improvement degree of the prediction accuracy of the optimized model, specifically as follows:

[0121] Symmetric mean absolute percentage error:

[0122]

[0123] Mean absolute percentage error:

[0124]

[0125] Mean absolute error:

[0126]

[0127] Root Mean Square Error:

[0128]

[0129] Mean Squared Error:

[0130]

[0131] Coefficient of determination:

[0132]

[0133] Where: n represents the sample size; y * represents the predicted value; y represents the actual value; represents the average predicted value.

[0134] In order to intuitively and clearly reflect the model prediction effect, calculate the six indicators of MSE, RMSE, MAE, MAPE, SMAPE and R 2 . The results are shown in Table 3.

[0135] Table 3 Comparison of evaluation indicators of four prediction models

[0136]

[0137] As can be seen from Table 2, all indicators of the PSO-YH-LSTM prediction model are better than the other three prediction models. Among them, the MAPE value is 2.8762%, that is, the prediction accuracy of the pre-model reaches 97.1238%, and the accuracy is improved by 2.4357% compared with the traditional LSTM; the goodness of fit reaches 0.9015, which is 0.1822 higher than the traditional LSTM prediction model, and the model has strong interpretability for data. It shows that the PSO-YH-LSTM prediction model can effectively complete the traffic situation prediction work.

[0138] S3 specifically includes the following steps:

[0139] S301. As shown by Figure 9 , first, abstract the urban traffic network to establish a road network geometric model. Use the "graph" data structure to express the road network geometric structure, and study the optimal path problem according to the graph algorithm. Therefore, construct a road network model as follows:

[0140]

[0141] Where: V represents the set of nodes; E represents the edge set in the road network model, and (v i , v j ) and (v j , v i) are two directions of the same road with different weights; W represents the set of weights of each edge (v i , v j ), which can be replaced according to specific requirements.

[0142] Usually, the traffic flow conditions in opposite directions of the same road section in the urban road network are different, and even extreme situations may occur. Therefore, a topological graph that can express direction differences is used to describe the actual urban road network.

[0143] S302. Divide urban road traffic into four levels, which are represented by the numbers 1-4 for smooth traffic, slow traffic, congestion, and severe congestion respectively. To ensure the original differences between data, standardize the traffic index data. During the process of calculating the road weight, the passing weight during smooth traffic is set to 1 by standardizing the traffic data, and based on this, the passing weights for slow driving, traffic congestion, and severe traffic congestion are defined, and their range is between [0, 1], specifically as follows:

[0144]

[0145] In the formula: f i represents the standardized result; γ i represents the original value of the i-th traffic data; max(γ) represents the maximum value of the traffic data.

[0146] Reverse the traffic states of the four levels so that the levels are proportional to the values, and finally perform standardization processing on the values. The standardized values are shown in Table 4.

[0147] Table 4 Reverse processing of traffic congestion index

[0148]

[0149] According to the provisions of the "Urban Road Traffic Congestion Evaluation Index System", correspond the congestion levels and speeds of various roads and correspond them to the congestion states in Table 3, as shown in Table 5.

[0150] Table 5 Road grade division and its speed correspondence

[0151]

[0152] S303. Standardize the real-time situation information to obtain the urban road traffic conditions, and then obtain the future road traffic situation information according to the PSO-YH-LSTM prediction model, and then the corresponding traffic state can be obtained. Design a weighted fusion method for traffic situation information to fuse the two traffic state values, specifically as follows:

[0153] y = αy1 + (1 - α)y2

[0154] Where: y represents the integrated traffic situation value; y1 represents the real-time traffic congestion value of the road section; y2 represents the predicted traffic condition value of the road section; α is an arbitrary constant (0 to 1).

[0155] In the research of path planning, the real-time situation information of the road has a greater impact than the predicted information. Therefore, the value of α should be greater than 0.5; after weighted integration processing, the value range of the traffic situation value of each road section is [0, 1]. The integrated traffic situation shows an inverse description of the urban traffic congestion situation. The larger the value, the smaller the time cost of urban road traffic, and vice versa.

[0156] S4 specifically includes the following steps:

[0157] S401. As Figure 10 is a schematic diagram of path selection by the traditional A* algorithm. In the figure, N and D are the starting point and the ending point, M3 is the next node selected by the traditional path planning algorithm, M1, M2, and M4 are three nodes adjacent to the starting point N, and α represents the included angle between the two road sections |M3N| and |DN|. If N and D points are used as two foci, and then an ellipse is drawn with |NM3| + |DM3| as the major axis. According to the properties of the ellipse, the positions of M1, M2, and M4 nodes are not inside the ellipse. And from Figure 10 the positions of each node, it can be seen that the included angle between |NM3| and |ND| is relatively large, and the included angles between the paths |NM4| and |NM2| and |ND| are relatively small; therefore, although selecting node M3 is better than M2 and M4 at the current moment, it is difficult to determine that this choice must be better than M2 and M4 for subsequent road section selection as time goes by. Therefore, the evaluation function of the A* algorithm is corrected by adding the sine value of the turning angle between the path and the target point to limit the search conditions of the algorithm, as follows:

[0158]

[0159] Where: f(n) represents the cost estimate passing through node n; g(n) represents the actual cost from the nth node to the starting point; h(n) is the estimated cost from the nth node to the target point; sin(α) represents the sine value formed by the current node as the vertex of the included angle, and this node, the adjacent node, and the target point.

[0160] Furthermore, the path planning process of the A* algorithm with the introduction of the angle factor is as follows:

[0161] (1) Initialize the road network information, determine S and D as the starting and ending points, and S as the initial node N;

[0162] (2) Traverse the other adjacent nodes of the current node, and then calculate the included angle α between the associated nodes M i , N, and Si (where \(i\) is the number of adjacent nodes, \(i = 1, 2,\cdots\)), thereby obtaining the sine value of the included angle;

[0163] (3) Calculate \(f(n)\) for each adjacent node according to the formula, and then compare and select the node with the smallest \(f(n)\) value as the node to be selected; if there is only a standby node among all the adjacent nodes of this node, then select the standby node;

[0164] (4) Perform iterative loop according to steps (2) and (3) until the end point \(D\) is found and then terminate, and plan the global shortest path.

[0165] Introduce the \(\sin(\alpha)\) function into the estimated cost function. Since \(\sin(\alpha)\) is monotonically increasing in the interval, and the function value is \((0\sim1)\). Therefore, the larger the included angle between the node at the current moment and the adjacent node and the target point, the larger the sine value, which will lead to an increase in the estimated time cost. So by introducing the angle factor, it can prevent the autonomous vehicle from deviating from the target point direction, and at the same time reduce the search node range and improve the overall search efficiency of the algorithm.

[0166] As Figure 11 and Figure 12 shown,

[0167] S5 specifically includes the following steps:

[0168] S501. Classify urban roads according to the differences in geometric dimensions, pavement driving quality, road driving speed, etc. of the road sections.

[0169] S502. As Figure 11 shown, adopt the plane search algorithm for road networks at each level, and finally obtain the global path through the connection points.

[0170] Further, obtaining the global path by hierarchical processing of the road network specifically includes the following steps:

[0171] (1) Assume that the starting point is \(S\), the end point is \(D\), \(M1\) and \(M2\) are nodes on the middle layer, and \(M3\) and \(M4\) are nodes on the highest layer. Start searching for the path from the starting point \(S\) to the target point \(D\).

[0172] (2) First, find the optimal path from the starting point \(S\) to the connection point \(M1\) on the intersecting layer.

[0173] (3) Then find the optimal path from \(M1\) to the highest-level node \(M3\) in the upper layer.

[0174] (4) And so on, then the optimal paths from the target point \(D\) to \(M2\) and \(M4\) can be found;

[0175] (5) Combine the optimal path segments at each level (S→M1→M3→M4→M2→D) to form the global optimal path.

[0176] S503 The setting of the number of road layers is an important link in determining the success of path planning. Considering the balance between efficiency and accuracy, the roads are divided into two levels: high-level (expressways, arterial roads), and low-level (sub-arterial roads, branch roads).

[0177] S504 After completing the hierarchical processing of urban roads, combined with the improved A* algorithm that introduces the angle factor in S401, construct a hierarchical improved A* algorithm.

[0178] By deeply analyzing the structural characteristics of the urban road network, it is found that the high-level road network is usually composed of more straight sections and relatively independent intersections. This structural characteristic makes the algorithm more efficient in finding the shortest path; while the low-level road network, due to the complexity of its structure, especially the existence of numerous adjacent nodes, increases the search space of the algorithm, thus having a negative impact on the calculation speed. In such cases, it is more appropriate to use the improved A* algorithm, with higher calculation efficiency.

[0179] Figure 12 It is the flowchart of the path planning for the hierarchical improved A* algorithm. The specific steps are as follows:

[0180] (1) Initialize the relevant data of the road network, and determine the starting point, target point S, D, the road network level, and coordinates;

[0181] (2) If both the starting and ending points are in the high-level, directly plan the shortest path through the traditional planar A*; otherwise, jump to step (3);

[0182] (3) If the starting point S is in the low-level road network and the target point D is in the high-level road network, at this time, it is necessary to obtain the intermediate point S' between the two points. The path planning from S to S' selects the hierarchical improved A*, and the path planning from S' to D selects the traditional planar A*. Connecting the two paths is the shortest path; otherwise, go to step (4);

[0183] (4) Calculate the value of |SD|. If |SD| < 3km, then plan the path according to the planar A* algorithm; otherwise, plan the shortest path through the hierarchical improved A*.

[0184] (5) Export the shortest path and end the process.

[0185] S6 specifically includes the following steps:

[0186] S601, such as Figure 13As shown in the figure, to avoid the agglomeration effect brought by the path planning system, when constructing the path planning model, consider the real-time traffic conditions of the road network and the agglomeration effect problem caused by the path planning system, convert the autonomous driving vehicles that have completed path planning into the actual traffic volume on the road, and add it to the calculation process of the time cost of this path.

[0187] S602. When path planning, two basic data (real-time traffic situation information and predicted traffic situation information) need to be considered to obtain the time cost of passing. In view of this, improve the traditional road impedance model and incorporate the agglomeration effect of traffic flow in the road network into it. The model is mainly composed of the estimated cost calculated by the evaluation function and the virtual impedance at the moment of path planning of autonomous driving vehicles. The estimated cost will be calculated according to the method constructed in S303; the virtual impedance will be calculated according to the BPR road impedance function proposed by the Federal Highway Administration of the United States, as follows:

[0188]

[0189] In the formula: t represents the time actually required to pass through this section; t0 represents the free driving time of the section; α and β represent the undetermined parameters of the model, and the recommended values are 0.15 and 4 respectively; q represents the traffic volume of this section at that time, with the unit of vel / h; c represents the actual passing capacity of the section, with the unit of vel / h.

[0190] Since the virtual impedance considers the impact of the newly added autonomous driving vehicles after path planning on the time cost of passing on the basis of the estimated passing time cost of the integrated traffic situation. Therefore, only need to calculate the impact of the newly added traffic volume on the road passing efficiency, so the road impedance function is modified, as follows:

[0191]

[0192] In the formula: t z is the newly added time caused by the newly added traffic volume; q z is the newly added traffic volume of the section.

[0193] Take a local area in Haidian District, Beijing as an example for verification:

[0194] Verification of the hierarchical improved A* algorithm path planning:

[0195] Select the roads around Beijing University of Posts and Telecommunications in Haidian District, Beijing as the experimental roads, and abstract 50 nodes and 79 road segments from the entire experimental road network. The vehicle speeds when the road traffic is in a smooth state are 50 km / h, 35 km / h, 25 km / h, and 25 km / h respectively. Then select two starting and ending points, and plan the paths through the Dijkstra algorithm (without considering the traffic situation), the traditional A* algorithm (without considering the dynamic update of the travel time cost), and the hierarchical improved A* algorithm (considering the dynamic update of the travel time cost). The evaluation indicators of the algorithms mainly include: the algorithm path search duration, the actual road mileage, the road travel time, and the proportion of high-grade road segments. Due to the contingency of the experiment, to prevent this situation, re-select the OD pairs with inconsistent starting and ending points for the experiment, and the traffic situation data corresponding to the morning rush hour is used in the experiment. The 2 search simulation paths are as shown in Figure 14 、 Figure 15 and the parameter results are shown in Tables 6 and 7.

[0196] Table 6 Comparison of path planning parameters of each algorithm for nodes 1 - 50

[0197]

[0198] Table 7 Comparison of path planning parameters of each algorithm for nodes 15 - 39

[0199]

[0200] As can be seen from Tables 6 and 7, the path planned by the hierarchical improved A* algorithm is slightly longer than the paths planned by the Dijkstra algorithm and the traditional A* algorithm in terms of road mileage. However, as can be seen from Figure 14 、 Figure 15 , the planned path of the hierarchical improved A* algorithm has fewer turns and is smoother. Figure 14It is shown that when the vehicle is at Node 3, the next node selected by the traditional A* algorithm is Node 4, while the hierarchical improved A* algorithm selects Node 6 which is closer to the target point direction, reducing the deviation angle from the target point; when at Node 1, the Dijkstra algorithm selects Node 10 to drive in a direction deviating from the target, while the hierarchical improved A* algorithm selects Node 2, and the sine value of the included angle between it and the target point direction is smaller, saving 341.79 s in the overall travel time. This shows that the accuracy of the path planning model can be effectively improved by introducing an angle factor to correct the evaluation function. As can be seen from Tables 5 and 6, the evaluation parameters of the hierarchical improved A* algorithm model are different from the planned paths of the Dijkstra algorithm and the traditional A* algorithm. Taking the path planning from the starting point 1 to the ending point 50 as an example, the actual road mileage of the path planned by the hierarchical improved A* algorithm is 8120 m, which is 590 m and 350 m longer than that of the Dijkstra algorithm and the traditional A* algorithm respectively. Although the path planned by the hierarchical improved A* algorithm is longer than the other two path planning structures, the paths planned by the improved algorithm mainly consist of high-grade roads, with fast vehicle driving speed, high comfort, and the ability to predict the future traffic state of the road in advance and avoid congested sections, greatly shortening the travel time, such as Figure 14 In the second cycle, the traffic state of the section from Node 3 to Node 4 on the North Second Ring Road changed from smooth to congested, and the travel time was no longer the shortest. However, the traditional A* algorithm could not predict the change of the traffic situation in this area and did not adjust the subsequent driving path; while the hierarchical improved A* algorithm could predict the change of the traffic state of this section in advance and re-plan the optimal path in advance, reducing the travel time cost. Among them, the travel time of the hierarchical improved A* algorithm is 911.83 s, saving 341.79 s and 152.75 s compared with the travel times of the Dijkstra algorithm and the traditional A* algorithm. As can be seen from Tables 6 and 7, the proportion parameters of the high-level road network in the two path plans of the improved A* algorithm are 100% and 80.1% respectively, which are about 20% higher than those of the other two algorithms, reflecting the superiority of road network layering.

[0201] Secondly, the hierarchical improved A* algorithm calculates the estimated cost value of the adjacent nodes of the current node more efficiently than the Dijkstra path planning algorithm. The hierarchical improved A* algorithm restricts the path search conditions by introducing an angle factor, thereby reducing some nodes and improving the efficiency of the algorithm in searching for nodes. Although calculating the sine value of the angle in the hierarchical improved path planning algorithm will increase the running time, compared with other search algorithms, the time required for calculating the sine value of the included angle is smaller. As can be seen from Table 6, the search duration of the improved A* algorithm is 1076 us, which is 923 us shorter than that of the Dijkstra algorithm, but the difference from the traditional A* algorithm is not significant. The reason is that the improved A* algorithm needs to consider the update of the road network travel cost when searching for the path, which will lead to an increase in the running time of the algorithm. The results in Table 7 can also confirm this conclusion.

[0202] Consider the verification of the agglomeration effect:

[0203] Assume that all the data of the path planning system of autonomous vehicles can be shared and used, that is, when the starting and ending points of an autonomous vehicle are determined, the path planned by the path planning system can be obtained by other vehicles. Considering the influence of the agglomeration effect brought by the path planning system, 200 autonomous vehicles will be added to the road network in this experiment. The starting and ending points of the vehicles are set to be the same, and they are placed in the road network according to the agglomeration effect coefficient. One path planning data is captured every 0.5 s, and the path finding results are presented every 10 s of simulation, with a total of 10 path finding results presented. Through the path results planned by the dynamic path planning model constructed by considering the agglomeration effect 10 times, a comprehensive analysis is carried out.

[0204] The path finding results are as Figure 16 shown. Set node 14 as the starting point and node 50 as the ending point. A total of 3 different paths are generated from the 10 path finding results. Among them, path A represents the path planning results of the 1st to 4th times, path B represents the path planning results of the 5th to 7th times, and path C represents the path planning results of the 8th to 10th times.

[0205] With the collection of the planned paths of other autonomous vehicles, all the vehicles that have completed path planning will be converted into the increment of the existing vehicles on the road according to the proportion of the agglomeration effect coefficient, resulting in an increase in the estimated travel time of the vehicles. When the estimated travel time exceeds the threshold, a shorter path will appear; at this time, if the planned path is not corrected, it is obviously unreasonable, and other path options need to be provided for autonomous vehicles with the same starting and ending points of travel. Therefore, this model provides 3 different path recommendations for 200 autonomous vehicles with the same starting and ending points.

[0206] Carry out a refined analysis of the experimental results, calculate the estimated travel time without considering the agglomeration effect and considering the agglomeration effect, and compare and analyze the two results. The comparison chart is as Figure 17 shown.

[0207] As Figure 17 can be seen, the two models choose the same path in the first 4 times, and the estimated travel time of the vehicles remains the same, and the curves coincide. However, as the number of autonomous vehicles increases, the path planning model considering the agglomeration effect provides other paths, and its estimated travel time is less than the estimated time of driving along the initial planned path: from the trend of the curve, it can be seen that as the number of autonomous vehicles increases, the advantage of the path planning model considering the agglomeration effect becomes more significant.

Claims

1. An optimal path planning method for autonomous vehicles based on integrated traffic situation, characterized in that, It includes the following steps: S1. Predict the traffic situation based on the improved adaptive probabilistic long short-term memory neural network (YH-LSTM). The YH-LSTM is obtained by adding a dropout layer to the traditional LSTM model and replacing the dropout neuron dropout probability value with the traffic data singularity ratio; S2. Optimize the YH-LSTM model based on the particle swarm optimization (PSO) algorithm to obtain the PSO-YH-LSTM model; S3. Weightedly fuse the traffic situation information predicted by the PSO-YH-LSTM and the real-time traffic situation information to generate a fused passing cost and update the road network model; S4. Improve the cost function of the A* algorithm, introduce the sine value of the turning angle between the path and the target point as the angle factor, and construct an improved A* algorithm model; S5. Classify urban roads, and improve the A* algorithm layer by layer for path planning; S6. According to the vehicle aggregation effect, convert the autonomous driving vehicles on the planned path into new traffic volume and dynamically update the road network impedance.

2. The optimal path planning method for an autonomous driving vehicle based on integrated traffic situation according to claim 1, wherein, S1 specifically includes the following steps: S101. Use the LSTM neural network to predict the future traffic situation. The data comes from the actual traffic speed data in the Q-Traffic dataset collected by Baidu Brain. Based on the characteristic analysis of urban traffic flow, the adjacent traffic flow value is determined as the input of the model. Based on the characteristic analysis of urban traffic flow, the adjacent traffic flow value is determined as the input of the model. S102. Adding a "dropout" layer to the traditional LSTM model can reduce the probability of the model overfitting. However, it is not determined how to calculate the neuron dropout probability value according to the specific application scenario, and only the previous empirical value is simply used. Therefore, the probability value of selectively discarding neurons in the "dropout" is replaced with the traffic data singularity ratio, as follows: H j = H d + H w It can be deduced from the above two formulas: where, H d represents the number of discarded nodes; H j represents the number of nodes in each layer; H q represents the number of singular points; H represents the number of nodes in a single-layer network; H w represents the number of non-discarded nodes in a single-layer network; H qd represents the number of discarded nodes that belong to noise in a single-layer network.

3. An optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 1, characterized in that, S2 specifically includes the following steps: S201. Use PSO to optimize the YH-LSTM model. The number of neurons in the hidden layer of the YH-LSTM neural network, the number of iterations of the prediction model, the learning rate and other parameter values are optimized based on the particle swarm algorithm.

4. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 1, characterized in that, S3 specifically includes the following steps: S301. Abstract the urban traffic network and establish a road network geometric model. Use the "graph" data structure to express the road network geometric structure, as follows: Where: V represents the set of nodes; E represents the set of edges in the road network model, and (v i , v j ) and (v j , v i ) are two directions of the same road, with different weights; W represents the set of weights of each edge (v i , v j ), which can be replaced according to specific requirements. S302. Divide the urban road traffic into four levels, which are represented by the numbers 1-4 for smooth traffic, slow traffic, congestion, and severe congestion respectively. In order to ensure the original differences between the data, standardize the traffic index data, as follows: where: f i represents the result after standardization; γ i represents the original value of the i-th traffic data; max(γ) represents the maximum value of the traffic data. Reverse the traffic states of the four levels, so that the level is proportional to the value, and finally perform standardization processing on the value. The standardized values are shown in Table 1. Table 1 Reverse processing of traffic congestion index According to the "Urban Road Traffic Congestion Evaluation Index System", correspond the congestion degree and speed of various levels of roads and correspond them to the congestion status in Table 1, as shown in Table 2. Table 2 Road grade division and its speed correspondence S303. Standardize the real-time situation information to obtain the urban road traffic conditions, and then obtain the future road traffic situation information according to the PSO-YH-LSTM prediction model, and the corresponding traffic state can be obtained. Design a weighted fusion method for traffic situation information to fuse the two traffic state values, as follows: y = αy1 + (1 - α)y2 Where: y represents the fused traffic situation value; y1 represents the real-time road section traffic congestion value; y2 represents the predicted road section traffic condition value; α is an arbitrary constant (0 to 1).

5. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 1, wherein S4 specifically includes the following steps: S401. Improve the A* algorithm by introducing an angle factor, as follows: Where: f(n) represents the cost estimate passing through node n; g(n) represents the actual cost from the nth node to the starting point 1; h(n) is the estimated cost from the nth node to the target point; sin(α) represents the sine value formed by the current node as the vertex of the included angle, and this node, the adjacent node, and the target point.

6. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 1, characterized in that, S5 includes the following specific steps: S501. Classify urban roads according to the differences in geometric dimensions, pavement driving quality, road driving speed, etc. of road sections. S502. Use a plane search algorithm for road networks at each level, and finally obtain the global path through connection points. The setting of the number of road layers is an important link in determining the success of path planning. Considering the balance between efficiency and accuracy, the roads are divided into two levels: high-level (expressways, arterial roads), low-level (sub-arterial roads, branch roads). After the urban roads are stratified, combine with the improved A* algorithm by introducing an angle factor in S401 to construct a hierarchical improved A* algorithm.

7. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 1, wherein S6 includes the following specific steps: S601. To avoid the agglomeration effect brought by the path planning system, when constructing the path planning model, consider the real-time traffic conditions of the road network and the agglomeration effect problem caused by the path planning system, convert the autonomous driving vehicles that have completed path planning into the actual traffic volume on the road, and add it to the calculation process of the passing time cost of this path. S602. When path planning, two basic data (real-time traffic situation information and predicted traffic situation information) need to be considered to obtain the passing time cost. In view of this, improve the traditional road impedance model and incorporate the agglomeration effect of traffic flow in the road network. The model is mainly composed of the estimated cost calculated by the evaluation function and the virtual impedance at the moment of autonomous driving vehicle path planning. The estimated cost will be calculated according to the method constructed in S303; the virtual impedance will be calculated according to the BPR road impedance function proposed by the US Federal Highway Administration, as follows: Where: t represents the actual time required to pass through this road section; t0 represents the free driving time of the road section; α and β represent the undetermined parameters of the model, and the recommended values are 0.15 and 4 respectively; q represents the traffic volume of this road section at that time, with the unit of vel / h; c represents the actual passing capacity of the road section, with the unit of vel / h. Since the virtual impedance is based on the estimated travel time cost of the known integrated traffic situation and takes into account the impact of newly added autonomous vehicles on the travel time cost after path planning is completed. Therefore, it is only necessary to calculate the impact of the newly added traffic volume on the road traffic efficiency, so the road resistance function is modified as follows: Where: t z is the additional time caused by the additional traffic volume; q z is the additional traffic volume of the road section.

8. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 3, characterized in that S201 specifically includes the following steps: (1) Preprocess the average vehicle speed data, add missing values, remove extreme values, divide the data set into an 80% training set and a 20% validation set, and perform standardization processing on this data; (2) Determine the optimization range of hyperparameters (number of neurons, number of iterations, batch size, learning rate) in the YH-LSTM neural network; (3) Initialize the parameters in the particle swarm optimization algorithm, and its parameters include the initial velocity, particle coordinates, etc.; (4) Select the fitness function of the PSO algorithm, and select the symmetric mean absolute percentage error value to search for the optimal parameter configuration of the particle swarm optimization algorithm; (5) Calculate and update the fitness value of the particle swarm optimization algorithm; (6) Determine whether the model operation has reached the number of iterations. If it has reached, retain the parameters. Then apply the optimal parameter configuration to the adaptive probability YH-LSTM model; (7) Use PSO-YH-LSTM to predict the traffic situation and output evaluation indicators. If the prediction result is good, end the process; otherwise, return to step (5).

9. A method for optimal path planning of an autonomous vehicle based on integrated traffic situation according to claim 5, characterized in that S401 specifically includes the following steps: (1) Initialize the road network information, determine S and D as the starting and ending points, and S as the initial node N; (2) Traverse other adjacent nodes of the currently located node, and then calculate the included angles α of associated nodes M i , N, and S i (i is the number of adjacent nodes, i = 1, 2,...), thereby obtaining the sine values of the included angles; (3) Calculate f(n) of each adjacent node according to the formula, and then compare and select the node with the smallest f(n) value as the node to be selected; if there is only a standby node among all adjacent nodes of this node, select the standby node; (4) Perform loop iteration according to steps (2) and (3) until the end point D is found and then terminate, and plan the global shortest path.

10. The optimal path planning method for an autonomous vehicle based on integrated traffic situation according to claim 6, wherein S502 specifically includes the following steps: (1) Assume that the starting point is S, the ending point is D, M1 and M2 are nodes on the middle layer, and M3 and M4 are nodes on the highest layer. Start searching for the path from the starting point S to the target point D. (2) First, find the optimal path from the starting point S to the connection point M1 of its intersecting layer. (3) Then find the optimal path from M1 to the highest-level node M3 in the high layer. (4) And so on, then the optimal paths from the target point D to M2 and M4 can be found; (5) Merge the optimal path segments of each level (S→M1→M3→M4→M2→D) to form the global optimal path.

11. A method for optimal path planning of an autonomous vehicle based on integrated traffic situation according to claim 6, characterized in that, S504 specifically includes the following steps: Furthermore, the flowchart of the hierarchical improved A* algorithm for path planning is as follows: (1) Initialize the relevant data of the road network, and determine the starting point, target point S and D, road network level and coordinates; (2) If both the starting and ending points are in the high layer, directly plan the shortest path through the traditional planar A*; otherwise, jump to step (3); (3) If the starting point S is located in the lower-level road network and the destination point D is located in the upper-level road network, it is necessary to obtain the intermediate point S' between the two points. For the path planning from S to S', the hierarchical improved A* algorithm is selected, and for the path planning from S' to D, the traditional planar A* algorithm is selected. Connecting the two paths gives the shortest path; otherwise, go to step (4); (4) Calculate the value of |SD|. If |SD| < 3 km, plan the path according to the planar A* algorithm; otherwise, plan the shortest path through the hierarchical improved A*. (5) Export the shortest path and end the process.

Images