Multi-target collision risk field establishment method based on graph model

Abstract

The invention discloses a multi-target collision risk field establishment method based on a graph model, and the method comprises the steps: building a vehicle body coordinate system, and obtaining a unit normal vector; scaling the relative distance vector according to the actual size of the vehicle to obtain an equivalent relative distance vector; regarding the geometric boundary of the vehicle as a rectangle, taking a rectangle circumscribed ellipse for parameter calculation, and calculating the risk value of the risk field around the vehicle based on the equivalent relative distance vector; establishing a self-vehicle feasible track cluster based on state and action space discretization sampling, wherein a surrounding vehicle track is input by an existing prediction module; directly representing vehicles in the scene through nodes in the graph model; determining directed edges through the positions of the own vehicle and the surrounding vehicles; constructing a weighted directed graph; carrying out importance sorting on the nodes, and carrying out peak value coordination processing on the independent anisotropic risk fields; and establishing a collision risk field of multi-target risk unified representation. According to the invention, a basis is provided for the risk index collision avoidance control of the vehicle based on the collision avoidance capability assessment, and the collision avoidance control performance of the autonomous vehicle is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and particularly relates to a method for establishing a multi-object collision risk field based on a graph model. Background Art

[0002] In recent years, with the advancement of infrastructure construction and urbanization, China's transportation industry has continued to develop rapidly. The increasing number of motor vehicles has improved the efficiency of economic and social operation and the quality of people's lives, but it has also brought great pressure to the efficient and safe operation of transportation facilities. As of the end of 2023, the number of deaths per 10,000 vehicles in road traffic accidents across the country throughout the year was 1.38, a year-on-year decrease of 5.5%. Through sorting, statistics, and analysis of the accident causes, among all road traffic accidents involving vehicles, accidents caused by collisions account for more than 95%. Therefore, improving the collision avoidance ability is the key to improving vehicle driving safety. In the ADAS system, collision avoidance algorithms such as forward collision warning, automatic emergency braking, and automatic emergency steering have been applied to mass-produced vehicles. According to the collision avoidance function, they can be divided into longitudinal collision avoidance functions and lateral collision avoidance functions.

[0003] The above-mentioned collision avoidance algorithms are mostly triggered urgently in a short time based on deterministic indicators. Whether it is a longitudinal collision avoidance function or a lateral collision avoidance function, due to the lack of surrounding vehicle observation information, these functions are difficult to identify potential risks and make reasonable collision avoidance decisions in advance. Therefore, the collision avoidance algorithm needs to design a suitable risk assessment method, which takes the surrounding vehicle trajectory prediction results and the self-vehicle motion state and other information as inputs and calculates the designed risk index results. In the multi-object risk assessment method, the method based on the potential field can effectively unify the representation of multi-object risks by using rich information. However, due to the lack of collision detection and consideration of the vehicle's collision avoidance ability, the rationality and stability of its risk assessment will deteriorate rapidly in case of emergency. Summary of the Invention

[0004] Aiming at the deficiencies existing in the prior art, the purpose of the present invention is to provide a method for establishing a multi-object collision risk field based on a graph model, which can solve the problems of lack of collision detection and consideration of the vehicle's collision avoidance ability in the multi-object risk assessment method. To achieve the above object and other advantages of the present invention, there is provided a method for establishing a multi-object collision risk field based on a graph model, including:

[0005] S1. Establish a vehicle body coordinate system with the longitudinal and lateral positions of the vehicle center point to obtain a unit normal vector;

[0006] S2. There is a relative distance vector of the points in the scene with respect to the vehicle center point in the vehicle body coordinate system. Scale the relative distance vector according to the actual size of the vehicle to obtain an equivalent relative distance vector;

[0007] S3. Consider the vehicle's geometric boundary as a rectangle, take the circumscribed ellipse of the rectangle for parameter calculation, and calculate the risk value of the risk field around the vehicle based on the equivalent relative distance vector;

[0008] S4. Establish a feasible trajectory cluster of the host vehicle based on discretized sampling of the state and action space, and the surrounding vehicle trajectories are obtained from the input of the existing prediction module;

[0009] S5. Represent the vehicles in the scenario directly through the nodes in the graph model, where the vehicles include the controlled vehicle and the vehicles around it;

[0010] S6. Take the nodes corresponding to the host vehicle and each vehicle in the scenario as the research objects, and determine the weights and directions of the edges between the nodes through quantitative description and calculation of the collision risk between the vehicles;

[0011] S7. Construct a corresponding weighted directed graph based on the information of the nodes, directed edges, and weights in the already determined graph model;

[0012] S8. Through sorting the importance of the nodes, perform peak coordination processing on each independent anisotropic risk field;

[0013] S9. Establish a collision risk field that uniformly represents multi-objective risks, providing a basis for the decision-making of the reference trajectory for vehicle collision avoidance.

[0014] Preferably, the risk value of the risk field around the vehicle in step S3 is specifically:

[0015]

[0016] where R k is the peak coefficient, used to directly adjust the peak size; μ0 is the road surface adhesion coefficient; is the equivalent moving mass, m k is the vehicle mass, v k is the vehicle speed, k m and k v are the mass influence coefficient and the speed influence coefficient respectively; and are the equivalent relative distance vectors in the O k X k direction and the O k Y k direction component sizes, θ ki,m and θ ki,m are respectively relative to the O k X k direction angle and relative to the O k Y k direction angle; a x,k and a y,kThe longitudinal acceleration and lateral acceleration of the vehicle in the X k O k Y k respectively, and λ k,x and λ k,y are the longitudinal acceleration influence coefficient and the lateral acceleration influence coefficient respectively. When the acceleration influence coefficient increases, the mean position of the risk field in the corresponding direction will have a greater offset as the absolute value of the acceleration in the corresponding direction increases.

[0017] Preferably, in step S6, the method of the graph model is used to represent the information of nodes, directed edges, and weights respectively to form a weighted directed graph. The specific steps are as follows:

[0018] S61. Define the weight between nodes as:

[0019] PTTC C,k,m,n = t collision - t k

[0020] represents the predicted time to collision of vehicle m and vehicle n at time t k , where t collision is the predicted collision time, which is determined by performing collision detection on the feasible trajectory of the host vehicle within the prediction time domain and the predicted trajectories of the surrounding vehicles input by the prediction module;

[0021] S62. Determine the start node and end node of the directed edge: Taking the host vehicle as the reference vehicle, when the surrounding vehicle is in the rear view area, there is only a directed edge with the surrounding vehicle as the start node and the host vehicle as the end node between the two nodes, indicating that the host vehicle can ignore the influence of the surrounding vehicle on itself; when the surrounding vehicle is in the front view area, left view area, or right view area, there are two directed edges between the two nodes, namely, a directed edge with the surrounding vehicle as the start node and the host vehicle as the end node and a directed edge with the host vehicle as the start node and the surrounding vehicle as the end node.

[0022] Preferably, step S8 ranks the importance of nodes based on three importance influence matrices of efficiency, the number of shortest paths based on the source node, and the number of shortest paths based on the target node. The specific steps are as follows:

[0023] S81. Construct an influence matrix based on efficiency;

[0024] S82. Construct an influence matrix centered on the source node and an influence matrix centered on the target node;

[0025] S83. Use the analytic hierarchy process based on the three-scale method to determine the weights of each matrix, perform weighted summation on each matrix, and obtain a multiple influence matrix;

[0026] S84. Calculate the node efficiency of each node;

[0027] S85. Weight the multiple influence matrices based on node efficiency to obtain a dependency matrix;

[0028] S86. Characterize the local influence of each node around its center and then perform normalization to obtain the standard importance of the node.

[0029] Preferably, in step S9, specifically, sort the standard importance values of each node in the graph model from large to small, and sequentially obtain a node list of importance;

[0030] Based on the node importance sorting result, the relative magnitude relationship of the peak values of the anisotropic risk fields corresponding to each vehicle for each node can be determined, thereby determining a complete collision risk field representing multi-object risks in the scenario.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: By defining a vehicle body coordinate system, calculating the relative distance vector of points in the scenario with respect to the vehicle center point, scaling the relative distance vector according to the actual vehicle size to obtain an equivalent relative distance vector, calculating the parameters by taking the circumscribed ellipse of a rectangle, and calculating the risk values of the risk fields around the vehicle based on the equivalent relative distance vector, the present invention can effectively unify the representation of multi-object risks by using rich information through a potential field-based method.

[0032] The present invention establishes a feasible trajectory cluster of the host vehicle based on the predicted trajectories and states of surrounding vehicles and discretized sampling of the state and action space, abstracts the collision avoidance scenario as a weighted directed graph according to the information such as nodes, directed edges, and weights in the calculated graph model, and coordinates the numerical magnitudes of the anisotropic risk fields corresponding to multiple vehicles based on the node importance sorting result to form a collision risk field for unified representation of multi-object risks. It provides a basis for collision avoidance control of risk indicators based on the collision avoidance ability evaluation of vehicles, enhances the collision avoidance control performance of autonomous vehicles, and has practical application value for the research of autonomous driving predictive control algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of a method for establishing a multi-object collision risk field based on a graph model according to the present invention;

[0034] Figure 2 It is an effect diagram of a method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the method for establishing a multi-object collision risk field based on a graph model according to the present invention;

[0035] Figure 3 It is an effect diagram of a method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the method for establishing a multi-object collision risk field based on a graph model according to the present invention;

[0036] Figure 4Effect diagram of the method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the present invention;

[0037] Figure 5 Effect diagram of the method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the present invention;

[0038] Figure 6 Effect diagram of the method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the present invention;

[0039] Figure 7 Effect diagram of the method for establishing a multi-object collision risk field based on a graph model according to Embodiment 1 of the present invention. Detailed implementation manners

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0041] Refer to Figure 1 , a method for establishing a multi-object collision risk field based on a graph model, including:

[0042] S1. Establish a vehicle body coordinate system with the longitudinal and lateral positions of the vehicle center point to obtain a unit normal vector; the unit normal vector is specifically:

[0043] i k = cos(θ) + x k , -sin(θ) + y k

[0044] j k = -sin(θ) + x k , cos(θ) + y k

[0045] where k is the given number of the predicted vehicle, and the longitudinal and lateral position coordinates corresponding to the center point O k are O k (x k , y k ), and X k O k Y k is the vehicle body coordinate system.

[0046] S2. There is a relative distance vector of the in-scene points with respect to the vehicle center point in the vehicle body coordinate system. Scale the relative distance vector according to the actual vehicle size to obtain an equivalent relative distance vector. The coordinates of a point M in the scene are M(x m , y m ). Then, the relative distance vector from this point to the vehicle center point can be obtained as d k,m = (x m - x k , y m - y k ). Since there are differences in the vehicle's geometric dimensions in the O k X k direction and the O k Y k direction in the vehicle body coordinate system, it is necessary to scale the relative distance vector d k,m to obtain an equivalent relative distance vector:

[0047]

[0048] where <d k,m , i k > and <d k,m , i k > represent the dot product operation results of d k,m and the unit normal vectors i k , j k . α k and β k are the semi-major axis and semi-minor axis of the circumscribed ellipse of the vehicle respectively. It should be noted that when the vehicle geometric boundary is regarded as a rectangle, there are theoretically countless circumscribed ellipses. For the unity of risk value calculation, the ellipse with the smallest area in the rectangle's circumscribed ellipses is taken for parameter calculation. Therefore, there are and where L k and W k are the length and width of the vehicle respectively.

[0049] S3. Regard the vehicle geometric boundary as a rectangle, take the circumscribed ellipse of the rectangle for parameter calculation, and calculate the risk value of the risk field around the vehicle based on the equivalent relative distance vector; based on the equivalent relative distance vector As Figure 3 shown, the risk value of the risk field around the vehicle can be obtained specifically as:

[0050]

[0051] where R k is the peak coefficient, used to directly adjust the peak size; μ0 is the road adhesion coefficient; is the equivalent moving mass, mk is the vehicle mass, v k is the vehicle speed, k m and k v are the mass influence coefficient and the speed influence coefficient respectively; and is the equivalent relative distance vector at O k X k direction and O k Y k direction component magnitudes, θ ki,m and θ ki,m then are respectively relative to O k X k direction angle and relative to O k Y k direction angle; a x,k and a y,k are respectively the longitudinal acceleration and the lateral acceleration of the vehicle under X k O k Y k , λ k,x and λ k,y are respectively the longitudinal acceleration influence coefficient and the lateral acceleration influence coefficient. When the acceleration influence coefficient increases, the mean position of the risk field in the corresponding direction will have a greater offset as the absolute value of the acceleration in the corresponding direction increases.

[0052] S4. Establish a feasible trajectory cluster of the ego vehicle based on state and action space discretization sampling, and the surrounding vehicle trajectories are obtained from the input of the existing prediction module; as Figure 4 shown, the feasible trajectories of the ego vehicle are obtained by using a sampling method that combines the longitudinal action space and the lateral state space. The expected states of the ego vehicle's expected trajectory cluster are specifically:

[0053]

[0054] where s(T), l(T), are respectively the longitudinal position, lateral position, longitudinal speed, lateral speed, longitudinal acceleration, and lateral acceleration at the termination moment of each trajectory, and l k are respectively the longitudinal acceleration sampling value and the lateral distance offset sampling value, and the total sampling duration is T = 5s.

[0055] Next, according to the current pose and motion state of the controlled vehicle, combined with the expected state after 5s, the parametric expressions of each feasible trajectory can be obtained using the curve interpolation method. From this, the transverse and longitudinal positions, speeds, accelerations, curvatures, etc. corresponding to each trajectory at each moment can be obtained, and these information are also the basis for the subsequent risk assessment index calculation and collision avoidance decision-making:

[0056]

[0057] S5. Represent the vehicles within the scene directly through the nodes in the graph model, where the vehicles include the controlled vehicle and the vehicles around it.

[0058] S6. Taking the nodes corresponding to the ego vehicle and each vehicle in the scene as the research objects, determine the weights and directions of the edges between the nodes through the quantitative description and calculation of the collision risks between the vehicles; using the method of the graph model, represent the information of the nodes, directed edges, and weights respectively to form a weighted directed graph, and define the weight between the nodes as:

[0059] PTTC C,k,m,n = t collision - t k

[0060] represents the predicted time to collision (PTTC) of vehicle m and vehicle n at time t k , where t collision is the predicted collision time, which is determined by performing collision detection on the feasible trajectories of the ego vehicle within the prediction horizon and the predicted trajectories of the surrounding vehicles input by the prediction module.

[0061] S7. Construct the corresponding weighted directed graph based on the information of the nodes, directed edges, and weights in the already determined graph model; determine the starting node and ending node of the directed edge: as Figure 5 shown, taking the ego vehicle as the reference vehicle, when the surrounding vehicle is in the rear view area, there is only a directed edge with the surrounding vehicle as the starting node and the ego vehicle as the ending node between the two nodes, indicating that the ego vehicle can ignore the effect of the surrounding vehicle on itself; when the surrounding vehicle is in the front view area, left view area, or right view area, there are two directed edges between the two nodes, namely the directed edge with the surrounding vehicle as the starting node and the ego vehicle as the ending node and the directed edge with the ego vehicle as the starting node and the surrounding vehicle as the ending node.

[0062] S8. Through sorting the importance of the nodes, perform peak coordination processing on each independent anisotropic risk field; sort the importance of the nodes based on three importance influence matrices of efficiency, the number of shortest paths based on the source node, and the number of shortest paths based on the target node, specifically as:

[0063] S81. The influence matrix based on efficiency is specifically:

[0064]

[0065] where n is the total number of nodes in the graph model, and e ij represents the reciprocal of the shortest distance d ij from node i to node j.

[0066] S821. The efficiency impact matrix E characterizes the importance between nodes from the perspective of the shortest distance. However, the interaction between any two nodes in the graph model is also related to the number of paths connecting them. For a weighted undirected graph, the element A k (i, j) of the k-th power of its adjacency matrix A is the number of paths of length k between node i and node j. This conclusion can also be extended to weighted directed graphs, that is, the total number of paths of length d ij between node i and node j in a weighted directed graph is Based on this inference, an influence matrix SIP centered on the source node and an influence matrix TIP centered on the target node are established:

[0067]

[0068] S83. Use the analytic hierarchy process based on the three-scale method to determine the weights of each matrix, and perform weighted summation on each matrix to obtain a multiple influence matrix:

[0069] M = (m ij )

[0070] M = 0.82E + 0.09SIP + 0.09TIP

[0071] Step 84. The node efficiency of each node is specifically:

[0072]

[0073] S85. When the node efficiency value is larger, it indicates that it is easier to transmit information from this node to other nodes in the graph model, and it also means that this node may be located in the central area of the graph model and plays a greater role in the information communication between nodes in the graph model. Therefore, the multiple influence matrix is weighted based on the node efficiency, and then the dependence matrix is obtained. The dependence matrix is specifically:

[0074]

[0075] where each element of the matrix is p ij = I i m ij , representing the quantitative result of the comprehensive influence of node i on node j, and can also characterize the degree of dependence of each node on the influence of other nodes in the graph model.

[0076] S86. Select the cross intensity to characterize the local influence of each node around its center:

[0077]

[0078] where λ∈(0,1) is a constant coefficient, is the in-strength of node i, is the out-strength of node i.

[0079] Then the importance of node i:

[0080]

[0081] After normalizing the importance, the final standard importance of the node is:

[0082]

[0083] S9. Establish a collision risk field that uniformly represents multi-object risks, providing a basis for the decision-making of the collision avoidance reference trajectory of the vehicle. Sort the standard importance values of each node in the graph model from large to small, and obtain a list of important nodes in turn. Based on the sorting result of node importance, the relative magnitude relationship of the peaks of the anisotropic risk fields of each vehicle corresponding to each node can be determined, so as to determine the complete collision risk field that represents the multi-object risks in the scenario.

[0084] To sum up, this technical solution proposes to use a graph model to represent the interaction relationship between multiple vehicles in a scenario through a weighted directed graph. First, based on the motion states of each vehicle in the scenario, its corresponding anisotropic risk field is established. Then, a weighted directed graph that represents the interaction relationship and strength between multiple vehicles in the scenario is established. After peak coordination processing of each independent anisotropic risk field based on the sorting result of node importance, a collision risk field that uniformly represents multi-object risks in the scenario is formed. It provides a basis for the collision avoidance control of the vehicle based on the risk index evaluated by the collision avoidance ability, enhances the collision avoidance control performance of the autonomous vehicle, and has practical application value for the research of the autonomous driving predictive control algorithm.

[0085] The number of devices and the processing scale described here are used to simplify the description of the present invention. The application, modification, and variation of the present invention are obvious to those skilled in the art. Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, other modifications can be easily achieved. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. A method for establishing a multi-objective collision risk field based on a graph model, characterized in that Including: S1. Establish a vehicle body coordinate system based on the longitudinal and lateral positions of the vehicle center point to obtain a unit normal vector; S2. There is a relative distance vector of the points in the scene with respect to the vehicle center point in the vehicle body coordinate system. Scale the relative distance vector according to the actual size of the vehicle to obtain an equivalent relative distance vector; S3. Consider the vehicle geometric boundary as a rectangle, take the circumscribed ellipse of the rectangle for parameter calculation, and calculate the risk value of the risk field around the vehicle based on the equivalent relative distance vector; S4. Establish a feasible trajectory cluster of the ego vehicle based on discretized sampling of the state and action space. The trajectories of surrounding vehicles are obtained from the input of the existing prediction module; S5. Represent the vehicles in the scene directly through the nodes in the graph model. The vehicles include the controlled vehicle and the surrounding vehicles; S6. Take the nodes corresponding to the ego vehicle and each vehicle in the scene as the research objects. Determine the weights and directions of the edges between the nodes through quantitative description and calculation of the collision risks between vehicles; S7. Construct a corresponding weighted directed graph based on the information of the nodes, directed edges, and weights in the determined graph model; S8. Perform peak coordination processing on each independent anisotropic risk field by sorting the importance of the nodes; S9. Establish a collision risk field that uniformly represents multi-objective risks, providing a basis for the decision-making of the collision avoidance reference trajectory of the vehicle.

2. The method for establishing a multi-objective collision risk field based on a graph model according to claim 1, wherein The risk value of the risk field around the vehicle in step S3 is specifically: where R k is the peak factor, used to directly adjust the peak value; μ0 is the road surface adhesion coefficient; is the equivalent moving mass, m k is the vehicle mass, v k is the vehicle speed, k m and k v are the mass influence coefficient and the speed influence coefficient respectively; and are the equivalent relative distance vectors in the O k X k direction and the O k Y k direction component magnitudes, θ ki,m and θ ki,m then are respectively the angles relative to the O k X k direction and the angles relative to the O k Y k direction; a x,k and a y,k are the longitudinal acceleration and the lateral acceleration of the vehicle in the X k O k Y k plane, λ k,x and λ k,y are the longitudinal acceleration influence coefficient and the lateral acceleration influence coefficient respectively. When the acceleration influence coefficient increases, the mean position of the risk field in the corresponding direction will have a greater offset as the absolute value of the acceleration in the corresponding direction increases.

3. The method for establishing a multi-objective collision risk field based on a graph model according to claim 1, characterized in that, In step S6, through the method of the graph model, represent the information of nodes, directed edges, and weights respectively to form a weighted directed graph. The specific steps are as follows: S61. Define the weight between nodes as: PTTC C,k,m,n = t collision -t k Indicates the expected time to collision between vehicle m and vehicle n at time t k where t collision is the expected collision time, determined by performing collision detection on the feasible trajectories of the host vehicle within the prediction horizon and the predicted trajectories of the surrounding vehicles input to the prediction module; S62. Determine the starting node and ending node of the directed edge: Taking the ego vehicle as the reference vehicle, when the surrounding vehicle is in the rear view area, there is only a directed edge with the surrounding vehicle as the starting node and the ego vehicle as the ending node between the two nodes, indicating that the ego vehicle can ignore the influence of the surrounding vehicle on itself; when the surrounding vehicle is in the front view area, left view area, or right view area, there are two directed edges with the surrounding vehicle as the starting node and the ego vehicle as the ending node and with the ego vehicle as the starting node and the surrounding vehicle as the ending node between the two nodes.

4. The method for establishing a multi-objective collision risk field based on a graph model according to claim 1, wherein In step S8, the importance of the nodes is sorted based on three importance influence matrices based on efficiency, based on the number of shortest paths from the source node, and based on the number of shortest paths to the target node. The specific steps are: S81. Construct an influence matrix based on efficiency; S82. Construct an influence matrix centered on the source node and an influence matrix centered on the target node; S83. Use the analytic hierarchy process based on the three-scale method to determine the weights of each matrix, perform weighted summation on each matrix to obtain a multiple influence matrix; S84. Calculate the node efficiency of each node; S85. Weight the multiple influence matrix based on the node efficiency to obtain a dependence matrix; S86. Characterize the local influence of each node around its center and then perform normalization processing to obtain the standard importance of the node.

5. The method for establishing a multi-objective collision risk field based on a graph model according to claim 4, wherein In step S9, specifically, sort the standard importance values of each node in the graph model from large to small, and obtain a node list of importance in sequence; Based on the sorting result of node importance, the relative size relationship of the peak values of the anisotropic risk fields of the vehicles corresponding to each node can be determined, so as to determine the complete collision risk field representing multi-objective risks in the scene.

Images