Highway commercial vehicle trajectory planning method fusing complex network theory and deep neural network and vehicle electronic device

Abstract

The application discloses a highway commercial vehicle trajectory planning method combining complex network theory and a deep neural network and vehicle electronic equipment, important vehicle nodes are evaluated and risk trees are extracted through the model, which is used to expand the action space of the trajectory planning of the ego vehicle, and various vehicle nodes are modeled in the form of fields, which is used to evaluate the collision risk of the planned trajectory. At the same time, the trajectory planning is directly performed in the XYT coordinate system in the highway scene, the predicted trajectories of other vehicles and the bounding boxes of other obstacles are mapped into the grid map generated according to the coordinate system, a deep neural network (DNN) is used for convolution operation, and a safe and feasible collision-free trajectory is obtained, so that the influence of commercial vehicles and passenger vehicles of different sizes on the trajectory planning is reduced, and the problems of difficult road modeling and excessive calculation amount are solved. In the whole intelligent driving system framework, a high-quality trajectory can be planned without coordinate system conversion, which can simultaneously consider safety, efficiency and flexibility.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving in artificial intelligence, specifically a method for trajectory planning of commercial vehicles on highways and an electronic vehicle that integrates complex network theory and deep neural networks. Background Technology

[0002] Commercial vehicles, with their superior load-bearing capacity and faster operating speeds, facilitate the rapid flow of logistics, people, and information, and now play an irreplaceable role in the comprehensive transportation system and engineering construction of society. Highways, with their relatively closed traffic environment, structured roads, and relatively simple road conditions, hold the promise of being the first to achieve large-scale application of highly automated driving commercial vehicles. However, due to their large size, weight, and significant driver blind spots, trajectory planning for commercial vehicles is more challenging than for passenger cars. Finding a high-quality trajectory that simultaneously balances safety, efficiency, and flexibility is a crucial issue in highway commercial vehicle trajectory planning.

[0003] The trajectory planning module generates the vehicle's trajectory by receiving and processing various sensory information and predicting the trajectories of other vehicles, making it a key technology in intelligent driving. To address the difficulties and excessive computational demands of directly modeling roads, current mainstream trajectory planning methods generally employ path-velocity decomposition, transforming the Cartesian coordinate system into a Frenet lane coordinate system. Then, the predicted trajectories of obstacles and other vehicles are projected onto the Frenet coordinate system for trajectory planning. However, current mainstream intelligent driving systems typically use multimodal trajectory prediction methods, simultaneously generating the probabilities of multiple predicted trajectories of other vehicles. Single-modal trajectory planning, on the other hand, requires the vehicle to simultaneously avoid multiple possible trajectories of other vehicles in the Frenet coordinate system, often leading to the vehicle choosing a very conservative planned trajectory. Furthermore, issues such as time misalignment may arise during the conversion between Cartesian and Frenet coordinate systems, and the quality of the trajectory generated in the Frenet coordinate system is significantly reduced when road curvature is high. Summary of the Invention

[0004] To address the aforementioned issues, this invention introduces cognitive theory based on complex networks to model the driving environment and guide the trajectory planning of autonomous vehicles. The model evaluates important vehicle nodes and extracts a risk tree to expand the action space for trajectory planning. Various vehicle nodes (commercial vehicles and passenger vehicles) are modeled as fields to assess the collision risk of the planned trajectory. Simultaneously, trajectory planning is performed directly in the XYT coordinate system (Cartesian coordinate system with a time dimension introduced) in the highway scenario. The predicted trajectories of other vehicles and the bounding boxes of other obstacles are mapped onto a grid map generated according to the coordinate system. Deep neural networks (DNNs) are used for convolution operations to derive a safe and feasible collision-free trajectory, thereby reducing the impact of different sizes of commercial vehicles and passenger vehicles on trajectory planning and solving the problems of difficult road modeling and excessive computation. This enables the planning of a high-quality trajectory that simultaneously considers safety, efficiency, and flexibility within the entire intelligent driving system framework without coordinate system transformation.

[0005] The purpose of this invention is to provide an intelligent driving trajectory planning method suitable for commercial vehicles on highways. It introduces cognitive theory based on complex networks to model the driving environment to guide the trajectory planning of the vehicle. At the same time, it uses deep neural networks (DNN) to perform convolution operations in the XYT coordinate system to generate safe and feasible collision-free trajectories, evaluates them, and finally selects a high-quality trajectory that can simultaneously take into account safety, efficiency and flexibility.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including the following steps:

[0007] Step 1: Apply cognitive theory based on complex networks to dynamically model the driving environment, including:

[0008] Step 1.1: First, based on complex network theory, we treat our own vehicle and other vehicles as network nodes and construct a dynamic complex network model:

[0009] G t = (P, E, W, Θ) t

[0010] Among them, G t For a dynamic complex network model; P = {p1, p2, ..., p...} N Let} be the set of nodes in the network, p i Let N be the number of nodes in the network; E = {e1, e2, ..., e...} K Let} be the set of edges of the nodes in the network, and e i Let W represent the edges of nodes in the network, and K be the number of edges of nodes in the network; W = {w1, w2, ..., w3}. . ., w k} represents the set of edge weights; Θ represents the movable region of a node.

[0011] Θ is modeled as a smooth bounded surface:

[0012]

[0013] Where Ω represents the boundary of a smooth, bounded surface.

[0014] Step 1.2: Based on the two-degree-of-freedom vehicle model and the motion point model, perform dynamic modeling of the nodes.

[0015] The two-degree-of-freedom model of the vehicle is as follows:

[0016]

[0017] The state variables of the model are the yaw rate r and the lateral velocity U. y a and b are the front and rear axle center of gravity, respectively; I z δ is the moment of inertia of yaw; m is the mass; δ is the front wheel steering angle; F yf and F yr These represent the forward and rearward lateral forces generated during driving; the longitudinal velocity U... x It is considered a time-varying parameter. When the tire cornering characteristics are within the linear range, the model can be expressed as:

[0018]

[0019] Among them, C f Cr and Cr are the front wheel lateral stiffness and rear wheel lateral stiffness, respectively.

[0020] Treating the relevant nodes as moving mass points with a center of gravity, we establish a moving point model relative to the desired path:

[0021]

[0022] Where Δψ and e are the heading angle deviation and lateral path deviation, respectively; s is the distance along the desired path; and k is the curvature of the desired path.

[0023] Step 1.3: Based on field theory, a variable Gaussian safety field model based on risk center transfer is proposed. The static safety field is described by a two-dimensional Gaussian function. Then the static safety field S... sta It can be represented as:

[0024]

[0025] Where x0 and y0 are the coordinates of the risk center (i.e., the origin of the coordinate system centered on the vehicle); C a a is the electric field strength coefficient; x and b yThese are the radii of the major and minor axes of the ellipse (the ellipse is an enlarged version of the inscribed ellipse of the vehicle's frame, which can be equivalently represented by the vehicle's aspect ratio).

[0026] As the vehicle moves, the risk center will shift accordingly, and the new risk center will be O′(x′0, y′0):

[0027]

[0028] in, k is the velocity vector. v It is a regulating factor, and 0 < k v <1 (forward) or -1 < k v <0 (backward); β is Angle with the x-axis; use a′ x and b′ y Let S represent the radii of the major and minor axes of the ellipse after the risk center shifts. Then the dynamic safety field S... dyn It can be represented as:

[0029]

[0030] Step 1.4: Based on the human driver's sensitivity to distance and reaction time, divide the area around the node into a first cognitive domain, a second cognitive domain, and an outer space.

[0031] The scope of the first cognitive domain is:

[0032] maxa′ x ≤s th1

[0033] s th1 =t c1 ·v e

[0034] Among them, s th1 The first threshold; t c1 The first cognitive reaction time of a human driver; v e This represents the maximum approach speed of other nodes in the environment.

[0035] The scope of the second cognitive domain is:

[0036] s th1 <maxa′ x ≤s th2

[0037] s th2 =t c2 ·v e

[0038] Among them, s th2 The second threshold; t c2This refers to the second cognitive reaction time of a human driver. The space outside the second cognitive domain is defined as the extra-domain space.

[0039] Within the framework of a variable Gaussian safety field, a risk perception function is established between nodes:

[0040]

[0041] in, It is node p i At node p j Field strength at the location, It is node p j scalar velocity, direction angle θ i，j It is node p j velocity vector With node p i field intensity vector The included angle (clockwise is positive), k c This is the risk perception adjustment coefficient.

[0042] Step 2: Based on the established model, generate a complex network and evaluate important nodes and extract risk trees.

[0043] Step 2.1: The evolution process of complex networks, including:

[0044] 1) Set the vehicle node as the master node, connect the master node and other nodes in the first cognitive domain, calculate the weight of the corresponding connection, and sort them according to the weight. For ease of calculation, if there are multiple nodes in the connection direction with the master node, take the nearest node as the node in the first cognitive domain, and regard other nodes that are far away as nodes in the second cognitive domain. Do the same for the nodes in the second cognitive domain.

[0045] 2) In the first cognitive domain, select environmental nodes in sequence according to the weight sequence determined in the previous step, calculate and sort their weights with other nodes, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0046] 3) Select the environmental nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pairs with the largest weights.

[0047] 4) In the second cognitive domain, select environmental nodes in sequence, calculate their weights relative to other nodes, sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0048] Step 2.2: Evaluate important nodes and generate a risk tree, including:

[0049] 1) Let node p iThe set of all neighboring nodes is N. i ; Node strength r i For node p i The sum of the weights of its neighboring nodes is then:

[0050]

[0051] 2) The average strength of all nodes in the network is defined as the network strength. <r>Then we have:

[0052]

[0053] 3) The node strength r i The ratio of the sum of the intensities of all nodes is defined as To assess the importance of a node, we have:

[0054]

[0055] Will The largest node is defined as an important node. Important nodes in the first cognitive domain are defined as first-class important nodes (nodes that are not main nodes). Important nodes in the second cognitive domain are defined as second-class important nodes.

[0056] 4) Define the tree generated by the master node, the first type of important nodes, and the second type of important nodes as a risk tree.

[0057] The above steps can be used to construct a cognitive model based on complex networks, while also evaluating important nodes in the network and extracting a risk tree.

[0058] Step 3: Map the predicted vehicle trajectory and the bounding boxes of other obstacles onto a raster map generated according to the XYT coordinate system.

[0059] Step 3.1: For other static obstacles, map them onto the grid map according to their original boundaries and assign a value of 1 to avoid collisions with the vehicle. For dynamic vehicles, expand the original rectangular boundary of the vehicle by a certain range according to the horizontal and vertical proportions of the variable Gaussian safety field constructed by each node in Step 1, forming a new circumscribed rectangle (the size of the expanded range can be adjusted to reflect different driving styles), defined as the Bounding Box. Simultaneously, convert it into a probability field, where the field strength represents the probability of collision between the vehicle and related vehicle node trajectories. The probability within the vehicle's original boundary is 1, and the probability within the region from the original boundary to the Bounding Box decreases as the field strength decreases (but is always greater than 0), ultimately converting the variable Gaussian safety field into a probability field.

[0060] Step 3.2: Let the current time in the coordinate system be t. i H represents the length of a planning period, where [t] i , t i+H The predicted trajectory of another vehicle in the interval is interpolated and converted into a bounding box form according to dt = 0.1s, and then projected onto t. i In the grid map at any given time, for vehicle nodes in the risk tree extracted in step 2, all possible predicted trajectories are simultaneously input into the grid map, while for other vehicle nodes, only the predicted trajectory with the highest probability is input into the grid map.

[0061] Step 4: Sampling trajectory points on the grid map, and treating the bounding box of the sampled points as a convolution kernel, performing convolution operations using a deep neural network (DNN) to obtain a safe and feasible collision-free trajectory, and evaluating and selecting the best trajectory.

[0062] Step 4.1: Use the Nonlinear Model Predictive Control (Nonlinear MPC) method to sample on the grid map to obtain a series of discrete trajectories with equal time intervals. The Nonlinear MPC model also adopts a two-degree-of-freedom vehicle model.

[0063] The sampled trajectory is in t i At time dt = 0.1s, at time [t] i , t i+H Interpolation is performed within the bounding box to obtain its bounding box, which is treated as a kernel with a value of 1. A DNN is used to convolve this bounding box with a grid map containing the predicted trajectories of other vehicles and the projections of obstacles, and the computation is accelerated by a GPU. If the convolution result is greater than 0, it is considered to have a collision risk and the trajectory is filtered out; otherwise, it is considered to have no collision risk. Thus, a series of safe and feasible collision-free discrete trajectories can be directly obtained in the XYT coordinate system.

[0064] Step 4.2: The cost of the planned trajectory is calculated using a comprehensive cost function F. total Sure:

[0065] F total =λ1F o +λ2F a +λ3F h +λ4F p +λ5F r

[0066] Where λ1, λ2, λ3, λ4, and λ5 are correlation coefficients; F o F is a distance function between a trajectory point and the predicted trajectories of other vehicles, used to penalize trajectories that are too close to the predicted trajectories of other vehicles; a F is a distance function between the trajectory point and the road boundary, used to penalize trajectories that are too close to the road boundary; h F is a distance function between the trajectory point and the vehicle in front, used to penalize trajectories that are too close to the vehicle in front; p F is a distance function between the trajectory points and the reference line, used to penalize trajectories that are too far from the reference line; r This is a function representing the degree of deviation between the trajectory point and the navigation direction, used to penalize trajectories with excessive lane changes. The weights of each component and the design of related functions are determined through inverse reinforcement learning and imitation learning.

[0067] The total cost of a trajectory is obtained by summing the costs of the trajectory points at each time step. The trajectory with the lowest cost is selected as the optimal trajectory. Nonlinear MPC is then used to optimize and smooth the discrete trajectories. Finally, an optimal trajectory that does not require coordinate transformation is output to the control module.

[0068] The beneficial effects of this invention are:

[0069] 1. This invention utilizes cognitive theory based on complex network theory to model various vehicle nodes, including commercial vehicles and passenger cars, in a highway driving environment. This model is used to identify important nodes, extract risk trees, and guide trajectory planning for intelligent driving vehicles. During the trajectory planning process, for vehicle nodes in the risk tree, the influence of all possible predicted trajectories is considered; for other vehicle nodes, only the influence of the predicted trajectory with the highest probability is considered. This expands the action space of the vehicle trajectory planning and improves the flexibility of the planned trajectory.

[0070] 2. This invention proposes a trajectory planning method for commercial vehicles on highways that integrates complex network theory and deep neural networks. Compared to current mainstream trajectory planning methods that decompose path velocity, the proposed method performs trajectory planning directly in the XYT coordinate system, achieving higher-quality planned trajectories while balancing safety, efficiency, and flexibility. This invention maps the predicted trajectory and obstacle bounding boxes onto a grid map and uses a DNN for convolution operations to obtain a collision-free safe trajectory. This reduces the impact of the size difference between commercial vehicles and passenger vehicles on trajectory planning, while also solving the problems of difficult road modeling and excessive computation in trajectory planning. Attached Figure Description

[0071] Figure 1 This is a flowchart of the method of the present invention.

[0072] Figure 2 Schematic diagram of a variable Gaussian safety field

[0073] Figure 3 A schematic diagram of the bounding box and probability field for a node vehicle.

[0074] Figure 4 A schematic diagram of the predicted trajectory mapping in a raster map. Detailed Implementation

[0075] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the content of the present invention is not limited thereto.

[0076] This invention provides an intelligent driving trajectory planning method suitable for commercial vehicles on highways, which can generate a high-quality trajectory that simultaneously considers safety, efficiency, and flexibility without the need for coordinate system transformation. Figure 1 As shown, the specific steps include:

[0077] First, the driving environment is modeled using cognitive theory based on complex networks. The real driving environment is abstracted into a dynamic network model, and important nodes are identified and risk trees are extracted.

[0078] (1) Based on complex network theory, a dynamic complex network model is constructed by treating the vehicle and other types of vehicles as network nodes:

[0079] G t = (P, E, W, Θ) t

[0080] Among them, G t For a dynamic complex network model; P = {p1, p2, ..., p...} N Let} be the set of nodes in the network, p i Let N be the number of nodes in the network; E = {e1, e2, ..., e...} K Let} be the set of edges of the nodes in the network, and e i Let W = {w1, w2, ..., w...} be the number of edges between nodes in the network; and K be the number of edges between nodes in the network. k } represents the weight of the edge; Θ represents the movable region of the node.

[0081] Θ is modeled as a smooth bounded surface:

[0082]

[0083] Where Ω represents the boundary of a smooth, bounded surface.

[0084] The state of a node can be modeled as follows:

[0085]

[0086] Among them, X i ∈R m For node p i State vector; U i ∈R q For the input vector; A i B is a dynamic matrix; i If the input matrix is ​​the control quantity, then node p i The output vector is:

[0087] Y i =f i (X i )

[0088] Then node p i and node p j The weighting function between them can be expressed as:

[0089] w ij =F(Y) i Y j )

[0090] Based on the vehicle's two-degree-of-freedom model and motion point model, dynamic modeling of nodes is performed. Simultaneously, related nodes are treated as moving mass points with a center of gravity, establishing motion point models relative to the desired path to predict their future trajectories in the absence of vehicle-to-vehicle communication and real-time state estimation methods. Through dynamic and kinematic modeling, state awareness of each node is achieved. Based on field theory, a variable Gaussian safety field model based on risk center transfer is proposed, revealing the dynamic field characteristics of the nodes.

[0091] The dynamic model used in this invention is a two-degree-of-freedom vehicle model, which can be expressed as:

[0092]

[0093] The state variables of the model are the yaw rate r and the lateral velocity U. y δ is the front wheel steering angle; a and b are the front and rear axle center of gravity, respectively; I z F is the moment of inertia of the yaw; m is the mass; F yf and F yr These represent the forward and rearward lateral forces generated during driving; the longitudinal velocity U... x It is considered a time-varying parameter. When the tire cornering characteristics are within the linear range, the model can be expressed as:

[0094]

[0095] Among them, C f and C r These are the front wheel lateral stiffness and the rear wheel lateral stiffness, respectively.

[0096] The kinematic model used in this invention can be represented as a motion point model. By treating relevant nodes as moving mass points with a center of gravity, a motion point model relative to the desired path is established:

[0097]

[0098] Where Δψ and e are the heading angle deviation and lateral path deviation, respectively; s is the distance along the desired path; and k is the curvature of the desired path.

[0099] (2) Based on the human driver’s sensitivity to distance and reaction time, the area around the node is divided into the first cognitive domain, the second cognitive domain and the outer space.

[0100] The scope of the first cognitive domain is:

[0101] maxa′ x ≤s th1

[0102] s th1 =t c1 ·v e

[0103] Among them, s th1 The first threshold; t c1 The first cognitive reaction time of a human driver; v e This represents the maximum approach speed of other nodes in the environment.

[0104] The scope of the second cognitive domain is:

[0105] s th1 <maxa′ x ≤s th2

[0106] s th2 =t c2 ·v e

[0107] Among them, s th2 The second threshold; t c2 This refers to the second cognitive reaction time of a human driver. The space outside the second cognitive domain is the extra-domain space.

[0108] Within the framework of a variable Gaussian safety field, a risk perception function is established between nodes:

[0109]

[0110] in, It is node p i At node p j Field strength at the location, It is node p j scalar velocity, direction angle θ i，j It is node p j velocity vector With node p i field intensity vector The included angle (clockwise is positive), k c This is the risk perception adjustment coefficient.

[0111] (3) Based on the established model, generate a complex network and evaluate important nodes and extract risk trees.

[0112] 1) Set the vehicle node as the master node, connect the master node and other nodes in the first cognitive domain, calculate the weight of the corresponding connection, and sort them according to the weight. For ease of calculation, if there are multiple nodes in the connection direction with the master node, take the nearest node as the node in the first cognitive domain, and regard other nodes that are far away as nodes in the second cognitive domain. Do the same for the nodes in the second cognitive domain.

[0113] 2) In the first cognitive domain, select environmental nodes in sequence according to the weight sequence determined in the previous step, calculate their weights with other nodes and sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0114] 3) Select the environmental nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pairs with the largest weights.

[0115] 4) In the second cognitive domain, select environmental nodes in sequence, calculate their weights relative to other nodes, sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0116] 5) Evaluate key nodes and generate a risk tree.

[0117] Let node p i The set of all neighboring nodes is N. i ; Node strength r i For node p i The sum of the weights of its neighboring nodes is then:

[0118]

[0119] The average strength of all nodes in the network is defined as the network strength. <r>Then we have:

[0120]

[0121] The node strength r i The ratio of the sum of the intensities of all nodes is defined as To assess the importance of a node, we have:

[0122]

[0123] Will The largest node is defined as an important node. Important nodes in the first cognitive domain are defined as first-class important nodes (nodes that are not the main node), and important nodes in the second cognitive domain are defined as second-class important nodes. The tree formed by the main node, first-class important nodes, and second-class important nodes is defined as a risk tree.

[0124] Furthermore, the predicted vehicle trajectories and obstacle bounding boxes are mapped onto a raster map generated based on the XYT coordinate system.

[0125] (1) For other static obstacles, their original boundaries are mapped onto the grid map and assigned a value of 1 to avoid collisions between the vehicle and them. For dynamic vehicles, the original boundaries of the vehicle are expanded by a certain range, and the bounding box is given a probability field form based on the field strength of the constructed variable Gaussian safety field to assess the collision risk between the vehicle and other vehicles.

[0126] (2) During the planning period, the predicted trajectory is interpolated at certain intervals and converted into the form of BoundingBox. It is then projected onto the grid map at the current time. At the same time, the predicted trajectory is pruned according to the different risk levels of the node vehicles to expand the vehicle's movement space and enable it to plan a more flexible trajectory.

[0127] Finally, by sampling trajectory points on the grid map and treating the bounding boxes of the sampled points as kernels, a safe and feasible collision-free trajectory is obtained by performing convolution operations using a DNN. The best trajectory is then evaluated and selected.

[0128] (1) The Nonlinear MPC method is used to sample on the grid map to obtain a series of discrete trajectories with equal time intervals. The Nonlinear MPC model also adopts a two-degree-of-freedom vehicle model.

[0129] The sampled trajectory is in t i At time dt = 0.1s, at time [t] i , t i+H Interpolation is performed within the bounding box to obtain its bounding box, which is treated as a kernel with a value of 1. A DNN is then used to convolve this bounding box with a grid map containing the predicted trajectories of other vehicles and the projections of obstacles, with computation accelerated by a GPU. If the convolution result is greater than 0, the trajectory is considered to have a collision risk and is filtered out; otherwise, it is considered to have no collision risk. This allows a series of safe and feasible collision-free discrete trajectories to be directly obtained in the XYT coordinate system.

[0130] (2) The cost of the planned trajectory is given by a comprehensive cost function F. total Sure:

[0131] F total =λ1F o +λ2F a +λ3F h +λ4F p +λ5F r

[0132] Where λ1, λ2, λ3, λ4, and λ5 are correlation coefficients; F o F is a distance function between a trajectory point and the predicted trajectories of other vehicles, used to penalize trajectories that are too close to the predicted trajectories of other vehicles; a F is a distance function between the trajectory point and the road boundary, used to penalize trajectories that are too close to the road boundary; h F is a distance function between the trajectory point and the vehicle in front, used to penalize trajectories that are too close to the vehicle in front; p F is a distance function between the trajectory points and the reference line, used to penalize trajectories that are too far from the reference line; r This is a function representing the degree of deviation between the trajectory points and the navigation direction, used to penalize trajectories with excessive lane changes. The weights of each component and the design of related functions are determined using IRL and IL.

[0133] The total cost of a trajectory is obtained by summing the costs of the trajectory points at each time step. The trajectory with the lowest cost is selected as the optimal trajectory. Nonlinear MPC is used to optimize and smooth the discrete trajectories, and finally, an optimal trajectory without coordinate transformation is output to the control module.

[0134] like Figure 2 As shown, this invention proposes a variable Gaussian safety field model based on risk center transfer. Under static conditions, the safety field model can be described by a two-dimensional Gaussian function. Then, the static safety field S... sta It can be represented as:

[0135]

[0136] Where x0 and y0 are the coordinates of the risk center (i.e., the origin of the coordinate system centered on the vehicle); C a a is the electric field strength coefficient; x and b y These are the radii of the major and minor axes of the ellipse (the ellipse is an enlarged version of the inscribed ellipse of the vehicle's frame, which can be equivalently represented by the vehicle's aspect ratio).

[0137] As the vehicle moves, the risk center will shift accordingly, and the new risk center will be O′(x′0, y′0):

[0138]

[0139] in, k is the velocity vector. v It is a regulating factor, and 0 < k v <1 (forward) or -1 < k v <0 (backward); β is Angle with the x-axis; use a′ x and b′ y Let S represent the radii of the major and minor axes of the ellipse after the risk center shifts. Then the dynamic safety field S... dyn It can be represented as:

[0140]

[0141] like Figure 3 As shown, for various dynamic vehicles, a new bounding rectangle is formed by expanding the original boundary of the vehicle according to the horizontal and vertical proportions of the variable Gaussian safety field constructed by each node in the cognitive model. This new bounding rectangle (different driving styles can be reflected by adjusting the coefficients of the expanded range) is defined as the Bounding Box. Simultaneously, it is transformed into a probability field, where the field strength represents the probability of the vehicle colliding with the trajectory of related nodes. The probability within the original boundary of the vehicle is 1, and the probability within the region from the original boundary to the Bounding Box decreases as the field strength decreases (but is always greater than 0). Ultimately, the variable Gaussian safety field is transformed into a probability field.

[0142] like Figure 4 As shown, the predicted trajectories of other vehicles within the planning period are interpolated at dt = 0.1s, transformed into the form of a bounding box, and projected onto the current time (t). i In a grid map with XY coordinates, for vehicle nodes (vehicle1, 2) in the risk tree, all possible predicted trajectories are input into the grid map simultaneously, while for other vehicle nodes (vehicle3, 4), only the predicted trajectory with the highest probability is input into the grid map.

[0143] In summary, this invention presents a trajectory planning method for commercial vehicles on highways that integrates complex network theory and deep neural networks. By constructing a cognitive model based on complex networks, it expands the action space for vehicle trajectory planning and assigns a field-like representation to each vehicle node. This allows the vehicle to generate a safe and collision-free trajectory directly in the XYT coordinate system using DNN convolution operations without needing coordinate system transformation. Compared to mainstream path velocity decomposition-based trajectory planning methods, the proposed method can achieve higher quality and flexibility in generating trajectories while ensuring safety.

[0144] Based on the above method, the present invention also proposes an automotive electronic device, which is a control device or a storage device, wherein the control device is equipped with the above method, and the storage device stores the program of the above method.

[0145] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.< / r> < / r>

Claims

1. A trajectory planning method for commercial vehicles on highways that integrates complex network theory and deep neural networks, characterized in that: Includes the following steps: Step 1: Dynamically model the driving environment based on the cognitive theory of complex networks; Step 2: Based on the model established in Step 1, generate a complex network and evaluate important nodes and extract a risk tree; Step 3: Map the predicted trajectory of other vehicles and the bounding boxes of other obstacles onto a raster map generated according to the XYT coordinate system; the XYT coordinate system is a coordinate system formed by adding a time dimension to the Cartesian coordinate system. Step 4: Sampling trajectory points on the grid map, and using the bounding box of the sampled points as the convolution kernel, performing convolution operations using a deep neural network (DNN) to obtain safe and feasible collision-free trajectories, and evaluating and selecting the best trajectory; The specific process of step 1 includes complex network modeling, specifically: Step 1.1: First, based on complex network theory, we treat our own vehicle and other vehicles as network nodes and construct a dynamic complex network model: in, It is a dynamic and complex network model; It is the set of nodes in the network. For nodes in the network, The number of nodes in the network; Let be the set of edges between nodes in the network. Let be the edges of the nodes in the network. This represents the number of edges between nodes in the network. The weight of the edge; This refers to the movable area of ​​the node; Modeled as a smooth bounded surface: in, The boundary of a smooth, bounded surface; The specific process of step 1 also includes dynamic modeling, specifically: Step 1.2: Based on the two-degree-of-freedom vehicle model and motion point model, perform dynamic modeling of the vehicle nodes; The two-degree-of-freedom model of the vehicle is as follows: The state variable of the model is the yaw rate. and lateral velocity ; This refers to the front wheel steering angle; and These are the forward and rearward lateral forces generated during driving; longitudinal velocity. It is considered a time-varying parameter. When the tire cornering characteristics are within the linear range, the model is expressed as: and These are the front wheel lateral stiffness and the rear wheel lateral stiffness, respectively. and These are the center of gravity of the front axle and the center of gravity of the rear axle, respectively. For quality; Treating vehicle nodes as moving mass points with a center of gravity, we establish a motion point model relative to the desired path: in, and These are heading angle deviation and lateral path deviation, respectively. This represents the distance along the desired path; Let be the curvature of the desired path; The specific process of step 1 also includes variable Gaussian safety field modeling, specifically: Step 1.3: Based on field theory, a variable Gaussian safety field model based on the transfer of risk centers is proposed. The static safety field is described by a two-dimensional Gaussian function: in, This is the electric field strength coefficient; and As the coordinates of the risk center, and These are the radii of the major and minor axes of the ellipse, respectively. The ellipse is an enlarged version of the inscribed ellipse of the vehicle's frame, and is equivalently represented by the vehicle's aspect ratio. When the vehicle moves, the risk center will shift accordingly, and the new risk center will be... : in, It is a velocity vector; As a regulating factor, and has or (; for and Given the included angle of the axes, the dynamic safety field can be expressed as: and This represents the radii of the major and minor axes of the ellipse after the risk center shifts. The specific process of step 1 also includes dividing the cognitive domain, specifically: Step 1.4: Based on the human driver's sensitivity to distance and reaction time, divide the area around the node into a first cognitive domain, a second cognitive domain, and an outer space; The scope of the first cognitive domain is: in, The first threshold; The first cognitive reaction time of a human driver; This represents the maximum approach speed of other nodes in the environment. The scope of the second cognitive domain is: in, The first threshold; The second cognitive reaction time of a human driver is defined as the space outside the second cognitive domain; Within the framework of a variable Gaussian safety field, a risk perception function is established between nodes: in, It is a node At the node Field strength at the location, It is a node scalar velocity, direction angle It is a node velocity vector With nodes field intensity vector The included angle, This is the risk perception adjustment coefficient.

2. The highway commercial vehicle trajectory planning method integrating complex network theory and deep neural networks according to claim 1, characterized in that, The specific process of step 2 includes: Step 2.1: The Evolution of Complex Networks 1) Set the vehicle node as the master node, connect the master node and other nodes in the first cognitive domain, calculate the weight of the corresponding connection, and sort them according to the weight. If there are multiple nodes in the connection direction with the master node, take the nearest node as the node in the first cognitive domain, and regard the other nodes that are farther away as nodes in the second cognitive domain. Do the same for the nodes in the second cognitive domain. 2) In the first cognitive domain, select environmental nodes in sequence according to the weight sequence determined in the previous step, calculate and sort their weights with other nodes, and connect the node pairs with the largest weights. If the weight is lower than the set threshold, do not connect them. If the edge already exists, do not connect it again. 3) Select the environmental nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pairs with the highest weights; 4) In the second cognitive domain, select environmental nodes in sequence, calculate their weights relative to other nodes, sort them, and connect the node pairs with the largest weights. If the weight is lower than the set threshold, do not connect them. If the edge already exists, do not connect it again. Step 2.2: Evaluate important nodes and generate a risk tree, including: 1) Let the node be The set of all neighboring nodes is Node strength For nodes The sum of the weights of its neighboring nodes is then: 2) The average strength of all nodes in the network is defined as the network strength. Then we have: 3) Node strength The ratio of the sum of the intensities of all nodes is defined as To assess the importance of a node, we have: Will The largest node is defined as an important node; important nodes in the first cognitive domain are defined as first-class important nodes; and important nodes in the second cognitive domain are defined as second-class important nodes. 4) Define the tree generated by the master node, the first type of important nodes and the second type of important nodes as a risk tree.

3. The highway commercial vehicle trajectory planning method integrating complex network theory and deep neural networks according to claim 1, characterized in that, The specific process of step 3 includes: Step 3.1: For other static obstacles, map them onto the grid map according to their original boundaries and assign a value of 1 to avoid collisions between the vehicle and them; for dynamic vehicles, expand the original rectangular boundary of the vehicle according to the horizontal and vertical proportions of the variable Gaussian safety field constructed by each node in Step 1 to form a new bounding rectangle, defined as the Bounding Box. At the same time, convert it into a probability field, where the field strength represents the probability of the vehicle colliding with the trajectory of related vehicle nodes. The probability within the original boundary of the vehicle is 1, and the probability within the region from the original boundary to the Bounding Box decreases as the field strength decreases. Finally, the variable Gaussian safety field is converted into a probability field. Step 3.2: Let the current time in the coordinate system be... , This indicates the length of a planning period. The trajectory of other vehicles predicted by the interval is based on Take the interpolation and convert it into the form of a Bounding Box, then project it onto... In the grid map at any given time, for vehicle nodes in the risk tree extracted in step 2, all possible predicted trajectories are simultaneously input into the grid map, while for other vehicle nodes, only the predicted trajectory with the highest probability is input into the grid map.

4. The highway commercial vehicle trajectory planning method integrating complex network theory and deep neural networks according to claim 1, characterized in that, The specific process of step 4 includes: Step 4.1: Use the Nonlinear Model Predictive Control (Nonlinear MPC) method to sample on the grid map to obtain a series of discrete trajectories with equal time intervals; The sampled trajectory is in Always in accordance with In time Interpolation is performed to obtain its BoundingBox, which is regarded as a kernel with a value of 1. A DNN is used to perform a convolution operation on it with a grid map containing the predicted trajectories of other vehicles and the projection of obstacles. If the convolution result is greater than 0, it is considered to have a collision risk and the trajectory is filtered out; otherwise, it is considered to have no collision risk. Thus, a series of safe and feasible collision-free discrete trajectories are directly obtained in the XYT coordinate system. Step 4.2: The cost of the planned trajectory is calculated using a comprehensive cost function. Sure: in, The correlation coefficient; This is a distance function between a trajectory point and the predicted trajectories of other vehicles, used to penalize trajectories that are too close to the predicted trajectories of other vehicles. This is a distance function between the trajectory point and the road boundary, used to penalize trajectories that are too close to the road boundary; This is a distance function between the trajectory point and the vehicle in front, used to penalize trajectories that are too close to the vehicle in front. This is a distance function between the trajectory points and the reference line, used to penalize trajectories that are too far from the reference line; A function representing the degree of deviation between trajectory points and navigation direction, used to penalize trajectories with excessive lane changes; the weights of each part and the design of related functions are determined through inverse reinforcement learning and imitation learning; The total cost of a trajectory is obtained by summing the costs of each trajectory point at each time step, and the trajectory with the lowest cost is selected as the optimal trajectory.

5. The highway commercial vehicle trajectory planning method integrating complex network theory and deep neural networks according to claim 4, characterized in that, The specific process of step 4 also includes: optimizing and smoothing the discrete trajectory using Nonlinear MPC to output an optimal trajectory that does not require coordinate transformation.

6. An automotive electronic device, characterized in that, The vehicle electronic device is a control device or a storage device, wherein the control device is equipped with the method described in any one of claims 1-5, and the storage device stores the program of the method described in any one of claims 1-5.

Images