Urban maneuvering target trajectory simulation method with coupled intention prediction and environmental constraints

Abstract

The present invention discloses a method for simulating the trajectory of an urban mobile target coupled with intention prediction and environmental constraints, and relates to the intersection of multi-objective planning and trajectory simulation generation. The present invention first extracts a set of interest areas of urban mobile targets in an application scenario, then extracts intersection nodes, geographical environmental elements of roads, and their own attributes, then constructs a neural network model, performs deep learning training, obtains the reward value of the intersection node, and then calculates the simulation set of mobile target trajectories based on the set of interest areas, the positions of each mobile target, and the reward value of each intersection node through a mixed integer linear programming method, and completes the simulation of the trajectory of an urban mobile target coupled with intention prediction and environmental constraints. The present invention overcomes the problems of low authenticity, weak scenario generalization ability, high requirements for training data, and low interpretability of current trajectory simulation methods, and effectively solves the problem of few real trajectories of multiple mobile target trajectories in special application scenarios.

Description

Technical Field

[0001] The present invention relates to the cross-field of multi-objective planning and trajectory simulation generation, and in particular to a method for simulating urban maneuvering target trajectories by coupling intention prediction and environmental constraints. Background Art

[0002] The trajectory data of mobile targets is the basis for studying human mobile behavior and decision-making logic, and is also the key to intelligent decision-making. It plays an important role in urban planning, traffic management, intelligent decision-making and other fields. Although with the development of acquisition equipment, the collection of trajectory data of mobile targets is not difficult, due to problems such as scenario restrictions, a large amount of real spatiotemporal trajectory data of mobile targets is still difficult to obtain and use, especially trajectory data in special scenarios. The behavior trajectory of mobile targets is urgently needed to support the control decision-making research in such special scenarios.

[0003] At present, the research on trajectory simulation can be divided into two types: rule-driven and data-learning-based. The rule-driven method focuses on statistically mining the macro-behavior patterns of mobile targets from a large amount of mobile data, and characterizes the motion characteristics of mobile targets by assuming the behavior patterns through complex mathematical statistical models. For example, Sibren Isaacman et al. (2012) split each trajectory into three decision-making processes, including determining the intention to move, determining the moving distance, and determining the number of moving steps. They only need to modify different moving intentions or moving distances to simulate the moving trajectory of the moving target under different intentions or different transportation modes; Xu Kang et al. (2020) split the trajectory generation task into three tasks: OD (origin-destination) selection, intermediate point discovery, and shortest path planning to simulate the trajectory of the moving target; Luca Pappalardo et al. (2018) generated paths by mining common paths in trajectory data. In addition, some scholars have proposed and continuously improved social force and potential field models that assume that the moving target is affected by various forces such as driving force, friction, repulsion, attraction, etc., and cellular automaton and agent models that assume that the moving target is affected by the neighborhood or other moving targets. Most of these models are based on the domain knowledge summarized and summarized by researchers, have good interpretability, and do not rely on or rarely rely on real data. However, due to the simplification of the behavior of the moving target, the simulated trajectories generally have authenticity problems.

[0004] Data-based learning mainly uses machine learning or deep learning to infer complex and potential behavior patterns from the mobile target's behavior trajectory, and simulates and generates approximately distributed mobile trajectories. Currently, there are two main implementation ideas for this method. One idea converts the mobile target behavior trajectory simulation task into an image generation task, and uses mature solutions in the field of image generation to solve the problem. For example, Chu Cao et al. (2021) and Kun Ouyang et al. (2018) converted the urban crowd movement trajectory into a picture-like matrix structure, and used the deep convolutional generative adversarial network (DCGAN) of the image generation task to simulate the generation of urban crowd trajectories. However, since this method only focuses on the spatial characteristics of the trajectory, the temporal characteristics of the simulated trajectory cannot be guaranteed. Another idea regards the mobile target's behavior trajectory as sequence data, and uses a temporal model to mine the mobile target's intrinsic behavior pattern. For example, Dou Huang et al. (2019) used convolutional neural networks and variational autoencoders to characterize the sequence characteristics of urban crowd trajectories, learn to estimate the distribution of original data, and generate corresponding simulated trajectories; Yu et al. (2017) regarded the training process of the adversarial generative network GAN as a reinforcement learning problem. SeqGAN regards the generator as an agent in a Markov decision process (MDP) and the discriminator as the discriminator of the MDP. The generator is optimized using a policy-based learning method to generate simulated trajectories of moving targets; Jie Feng et al. (2020) used a self-attention model to construct the generator in GAN and added prior expert domain knowledge to improve network performance; Yanbo Pang et al. (2020) used GAN and inverse reinforcement learning IRL to perform trajectory generation tasks, and the discriminator in GAN realized the process of constructing a reward function from the real trajectory. This type of model is more based on the spatiotemporal distribution characteristics of the real trajectory, and the simulated trajectory is more realistic, but the model has high requirements for training data and is not very interpretable.

[0005] In general, there is still a lack of trajectory simulation generation methods. The existing methods mainly have the following problems: ① The existing rule-driven trajectory simulation methods generate trajectories with low authenticity, rigid behavior and weak scenario generalization ability, and are extremely dependent on the assumption of the behavior rules of mobile targets; ② The trajectory generation methods based on data learning have extremely high requirements on the quality of training data and the simulated trajectories are not interpretable enough; ③ There are few real trajectories in special scenarios, and the sparse data infers the behavior patterns of mobile targets, which affects the quality of the generated trajectories. Summary of the invention

[0006] In view of this, the present invention proposes a method for simulating urban mobile target trajectories by coupling intention prediction and environmental constraints. This method overcomes the problems of low authenticity, weak scenario generalization ability, high requirements for training data and low interpretability of current trajectory simulation methods, and effectively solves the problem of few real trajectories of multiple mobile target trajectories in special application scenarios.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The urban maneuvering target trajectory simulation method coupled with intention prediction and environmental constraints includes the following steps:

[0009] Step 1: Based on the urban geographic environment dataset with unified spatiotemporal coordinates, combined with expert domain knowledge, considering the geographic environment elements of the urban built-up area, extract the set of interest areas that attract mobile targets;

[0010] Step 2: According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-regions, and the geographical environment elements of each intersection node and road are calculated;

[0011] Step 3, calculate the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency and the capacity of each adjacent road of each intersection node;

[0012] Step 4, representing each historical trajectory whose destination belongs to the set of interest areas as a set containing each intersection node and road in the historical trajectory;

[0013] Step 5, construct a neural network model. The input of the neural network model is the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency, capacity of each adjacent road, and geographical environment factors of each intersection node. The output of the neural network model is the reward value of each intersection node.

[0014] For each set in step 4, calculate its total reward value, generate all possible simulation paths according to the starting point and end point of the historical trajectory, and select the simulation path with the largest total reward value. Calculate the difference between the total reward value of the historical trajectory and the total reward value of the selected simulation path and record it as the loss value. According to the loss value, use the gradient descent method to backpropagate and update the parameters of the neural network model until the loss value converges or the maximum number of iterations is reached. End the training and save the reward values ​​of each intersection node output by the current neural network model to complete the deep learning training.

[0015] Step 6: Based on the set of interest areas, the positions of each maneuvering target, and the reward value of each intersection node, the maneuvering target trajectory simulation set is calculated through the mixed integer linear programming method to complete the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints.

[0016] Furthermore, the specific method of extracting the set of interest areas that attract the mobile target in step 1 is:

[0017] The set of interest areas to be visited is divided into ;

[0018] For a certain area of ​​interest to be visited , =1,2,3,......, , calculate the probability of the target accessing the area based on historical access records:

[0019] ;

[0020] ;

[0021] in, Indicates a maneuvering target Visit the area of ​​interest to be visited The probability of represents a maneuvering target, Indicates the statistical time window selected in the historical access records. Indicates that within the statistical time window, the maneuvering target A collection of areas of interest to be visited that have been signed in. For collection An element in Indicates the area of ​​interest to be visited geographical factors, Indicates the area of ​​interest to be visited geographical factors, It means to calculate the Wasserstein distance between two distributions. Indicates geographical environment elements: Indicates the characteristic distribution of local urban morphology, Represents the characteristic distribution of urban population and facility density, represents the characteristic distribution of urban functional types, Represents the characteristic distribution of urban terrain; Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The frequency, Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The number of times Indicates that within the statistical time window, the maneuvering target The total number of check-ins to visit the area of ​​interest;

[0022] Set the probability threshold for selecting an area , for mobile targets , select access probability greater than The area of ​​interest is used as a collection of areas of interest to attract them :

[0023] =1,2,3,......, ;

[0024] s is the number of regions of interest.

[0025] Furthermore, the specific method of step 2 is:

[0026] According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-areas:

[0027] Intersection Node Set ,in represents the e-th intersection node, and e represents the number of extracted intersection nodes;

[0028] Expressed as ,in, and Represents a junction node The two-dimensional position of Represents a junction node The adjacent roads Represents a junction node The number of adjacent roads; represents the adjacent spatial sub-region of the intersection node, Represents a junction node The number of adjacent spatial subregions of ;

[0029] road set ,in Indicates road, Indicates the number of extracted roads. The starting point and end point of each road are intersection nodes, and there are no intersection nodes inside the road.

[0030] Expressed as ,in Indicates the road Length, Indicates the road The number of lanes; Indicates the road The adjacent spatial subregion of Indicates the road The number of adjacent spatial subregions of ;

[0031] Spatial sub-region set ,in Indicates A spatial sub-region, is the number of sub-regions of the divided space;

[0032] Statistics are conducted on the characteristics of geographical environment elements in the spatial sub-areas. It is expressed as:

[0033] ;

[0034] in Represents a spatial sub-region The characteristic distribution of the local form of the inner city, Represents a spatial sub-region The characteristic distribution of inner city population and facility density, Represents a spatial sub-region The characteristic distribution of inner city functional types, Represents a spatial sub-region The characteristic distribution of inner city terrain;

[0035] The geographical environment elements of each road adjacent spatial sub-area are added to the road. For example:

[0036] ;

[0037] Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road Result;

[0038] Add the geographical environment element features of the spatial sub-regions adjacent to each intersection node to the intersection node. For example:

[0039] ;

[0040] Indicates that the intersection node The geographical environment features of the adjacent spatial sub-areas are added to the intersection node Result;

[0041] Add the geographical environment elements of the roads adjacent to each intersection node to the intersection node. For example:

[0042] ;

[0043] Represents a junction node Add the geographical environment features of adjacent roads to the intersection node As a result, Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road result.

[0044] Furthermore, the specific method of step 3 is:

[0045] Calculate the node degree of each intersection node:

[0046] ;

[0047] in, Represents the node degree of intersection node b, which belongs to the intersection node set , is the adjacency matrix of the road network. If the intersection node b and the road node c are directly connected by a road, then =1, otherwise = 0, intersection node c belongs to the intersection node set ;

[0048] Calculate the node betweenness of each intersection node:

[0049] ;

[0050] in, represents the node betweenness of intersection node b, Indicates the number of shortest paths between intersection node c and intersection node d, Indicates the number of paths that pass through intersection node b in the shortest path from intersection node c to intersection node d; intersection node b All belong to the intersection node set ;

[0051] Calculate the adjacent road capacity of each intersection node:

[0052] ;

[0053] in, represents the adjacent road capacity of intersection node b, U represents the set of intersection nodes directly adjacent to intersection node b through a road, is an element of the set U, Represents the intersection node b and the intersection node The length of the road between Represents the intersection node b and the intersection node The number of lanes of the road between;

[0054] Calculate the contribution of each intersection node to the global efficiency of the network:

[0055] - ;

[0056] ;

[0057] ;

[0058] in, represents the contribution of intersection node b to the global efficiency of the network, represents the global efficiency of the network, represents the global efficiency of the network after removing the intersection node b; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set ; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set , intersection node set is the intersection node set B after removing the intersection node b;

[0059] Calculate the contribution of each intersection node to the local efficiency of the network:

[0060] ;

[0061] in, Represents a junction node Contribution to local network efficiency, represents the number of elements in the set U, Represents the intersection nodes in set U With intersection node The shortest path length between .

[0062] Furthermore, the specific method of step 4 is:

[0063] End point The historical trajectory Take as an example, and represent it as a collection of intersection nodes and roads:

[0064] = { };

[0065] in, For historical trajectory The first intersection node, For historical trajectory The first path chosen, is the time at the first intersection node in the historical trajectory, is the speed at the first intersection node in the historical trajectory; Representing historical trajectory The end point is ; All belong to the intersection node set , All belong to the road set .

[0066] Furthermore, the neural network model in step 5 includes an input layer, a hidden layer and an output layer, wherein the input layer contains 9 neurons, receiving 9 dimensional features of each intersection node, including node degree, node betweenness, node contribution rate to network global efficiency, node contribution rate to network local efficiency, capacity of each adjacent road and geographical environment elements, wherein the geographical environment elements of the intersection nodes include the characteristic distribution of the local urban morphology corresponding to the intersection nodes, the characteristic distribution of urban population and facility density, the characteristic distribution of urban functional types and the characteristic distribution of urban terrain, the output layer includes 1 neuron, outputs the reward function value of each intersection node, the hidden layer includes 3 layers, including 128, 128 and 64 neurons respectively, and uses the relu function as the activation function.

[0067] Furthermore, the specific method of step 6 is:

[0068] First, calculate the reward function value matrix corresponding to the moving trajectory of the moving target :

[0069] ;

[0070] Among them, the reward function value matrix The Line Column Elements = , Indicates the area of ​​interest Move to area of ​​interest The maximum reward function value of and All belong to the area of ​​interest collection , for maneuvering targets in the region of interest Areas of interest Each path between calculates a corresponding total reward function value, and the maximum total reward function value is recorded as ;

[0071] Solve the following equation using mixed integer linear programming:

[0072] ;

[0073] The constraints are:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] in, and are all decision variables. If the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but = 1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0, k=1,2,3,......,K; represents the initial position of the kth maneuvering target, and Belongs to the area of ​​interest collection , if the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but = 1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0; and Respectively represent the kth maneuvering target visiting the area of ​​interest Areas of interest order;

[0080] According to the solution results of the mixed integer linear programming method, the movement planning scheme of each maneuvering target between each area of ​​interest is obtained, and for each path of the maneuvering target between two areas of interest, the path with the largest total reward function value is set as the simulation path of the maneuvering target between the two areas of interest, and finally the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints is obtained.

[0081] Due to the adoption of the above technical solution, the beneficial effects of the present invention compared with the prior art are:

[0082] 1. The present invention overcomes the problems of low authenticity, weak scenario generalization ability, high requirements for training data and low interpretability of current trajectory simulation methods, and effectively solves the problem of few real trajectories of multiple moving target trajectories in special application scenarios.

[0083] 2. The present invention can provide input for urban intelligent decision-making research, and is the key to understanding the mobile behavior of mobile targets and the recognition of decision-making logic. It has great application prospects and practical value in the research and application of control decision-making in special application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 Schematic diagram of the principle of the urban maneuvering target trajectory simulation method for coupling intention prediction and environmental constraints in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0086] Urban maneuvering target trajectory simulation method that couples intention prediction and environmental constraints, such as Figure 1 As shown, the following steps are included:

[0087] Step 1: Based on the urban geographic environment dataset with unified spatiotemporal coordinates, combined with expert domain knowledge, considering the geographic environment elements of the urban built-up area, extract the set of interest areas that attract mobile targets;

[0088] Specifically, in this embodiment, the geographic environment element data assimilation is processed by QGIS software, wherein the geographic environment data monitoring time is required to be consistent with the previous trajectory data, the resolution of the raster data (such as buildings, DEM, vegetation, population, etc.) is kept consistent, and the coordinate system 1 adopts the WGS-84 coordinate system.

[0089] Step 2: According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-regions, and the geographical environment elements of each intersection node and road are calculated;

[0090] Step 3, calculate the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency and the capacity of each adjacent road of each intersection node;

[0091] Step 4, representing each historical trajectory whose destination belongs to the set of interest areas as a set containing each intersection node and road in the historical trajectory;

[0092] Step 5, construct a neural network model. The input of the neural network model is the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency, capacity of each adjacent road, and geographical environment factors of each intersection node. The output of the neural network model is the reward value of each intersection node.

[0093] For each set in step 4, calculate its total reward value, generate all possible simulation paths according to the starting point and end point of the historical trajectory, and select the simulation path with the largest total reward value. Calculate the difference between the total reward value of the historical trajectory and the total reward value of the selected simulation path and record it as the loss value. According to the loss value, use the gradient descent method to backpropagate and update the parameters of the neural network model until the loss value converges or the maximum number of iterations is reached. End the training and save the reward values ​​of each intersection node output by the current neural network model to complete the deep learning training.

[0094] Step 6: Based on the set of interest areas, the positions of each maneuvering target, and the reward value of each intersection node, the maneuvering target trajectory simulation set is calculated through the mixed integer linear programming method to complete the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints.

[0095] Furthermore, the specific method of extracting the set of interest areas that attract the mobile target in step 1 is:

[0096] The set of interest areas to be visited is divided into ;

[0097] For a certain area of ​​interest to be visited , =1,2,3,......, , calculate the probability of the target accessing the area based on historical access records:

[0098] ;

[0099] ;

[0100] in, Indicates a maneuvering target Visit the area of ​​interest to be visited The probability of represents a maneuvering target, Indicates the statistical time window selected in the historical access records. Indicates that within the statistical time window, the maneuvering target A collection of areas of interest to be visited that have been signed in. For collection An element in Indicates the area of ​​interest to be visited geographical factors, Indicates the area of ​​interest to be visited geographical factors, It means to calculate the Wasserstein distance between two distributions. Indicates geographical environment elements: Indicates the characteristic distribution of local urban morphology, Represents the characteristic distribution of urban population and facility density, represents the characteristic distribution of urban functional types, Represents the characteristic distribution of urban terrain; Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The frequency, Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The number of times, Indicates that within the statistical time window, the maneuvering target The total number of check-ins to visit the area of ​​interest;

[0101] Specifically, in this embodiment, the characteristic distribution of the local urban morphology is divided into local urban building complexes based on six factors: sky view factor, aspect ratio, building surface score, impervious surface score, permeable surface score and roughness element height, wherein the sky view factor represents the ratio of the number of sky hemispheres visible from the ground to the number of unobstructed sky hemispheres; the aspect ratio represents the average height-to-width ratio of street canyons; the building surface score represents the ratio of the building plane area to the total plane area; the impervious surface score represents the proportion of the impervious plane area (paved, rock) to the total plane area; the permeable surface score represents the ratio of the permeable surface area (exposed soil, vegetation, water) to the total plane area; the roughness element height represents the geometric mean of the heights of buildings, trees / plants, as shown in Table 1:

[0102]

[0103] The distribution characteristics of urban population and facility density reflect the impact on the attractiveness of human activities, as shown in Table 2:

[0104]

[0105] According to the role of urban functions, they are divided into six categories: economy and commerce, administration and human services, energy, culture, communication and information, and transportation and distribution, and corresponding functional factors are set for each type of urban function. Among them, the economic and commercial category includes commercial and financial centers, including shops, restaurants, hotels, markets, banks, trading centers, business offices, entertainment facilities, surrounding industrial, mining and agricultural characteristic services and other facilities; the administrative and human services category includes urban administrative organizations and service functions related to public governance, health, safety and welfare in urban areas, including embassies, diplomatic agencies, judicial systems, hospitals and other medical service facilities, public housing and shelters, water supply systems, police, fire, rescue, prisons and welfare and social service systems and other facilities; the energy category provides basic services and resources for the urban population; the cultural category provides the urban population with its social identity and reflects its culture, including schools, museums and archaeological sites, historical sites, libraries and cinemas; the communication and information category controls the flow of information, including telecommunications, police fire and rescue communication systems, public broadcasting, postal systems, newspapers and magazines, the Internet and social media; the transportation and distribution category includes networked roads, railways, bridges, subways, tunnels, underground passages, overpasses, ferries and airports.

[0106] The characteristic distribution of urban terrain affects the movement behavior of mobile targets, including terrain factors such as slope, height difference, elevation, undulation frequency, and cutting depth;

[0107] ① Slope: Slope ratio, that is, the ratio of the vertical height h and the horizontal width l of the slope, that is of (tan∠a value, ∠a is the slope and the horizontal plane ), the slope mainly affects the action speed and traffic capacity of targets with different attributes, as shown in Table 3:

[0108]

[0109] ②Frequency of fluctuation: The number of times the terrain fluctuates within a unit distance in the direction of travel, which mainly affects the movement speed and travel capacity of targets with different attributes.

[0110] ③ Cutting depth: The vertical changes in the ground morphology (such as gullies, pits, steep cliffs, landslides and other micro-topography) caused by external forces or other influences on the ground surface. The difference between the average elevation of the neighborhood range of a point on the ground and the minimum elevation within the neighborhood range affects the movement speed and traffic capacity of targets with different attributes. The details are shown in Table 4:

[0111]

[0112] Set the probability threshold for selecting an area , for mobile targets , select access probability greater than The area of ​​interest is used as a collection of areas of interest to attract them :

[0113] =1,2,3,......, ;

[0114] s is the number of regions of interest.

[0115] Furthermore, the specific method of step 2 is:

[0116] According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-areas:

[0117] Intersection Node Set ,in represents the e-th intersection node, and e represents the number of extracted intersection nodes;

[0118] Expressed as ,in, and Represents a junction node The two-dimensional position of Represents a junction node The adjacent roads Represents a junction node The number of adjacent roads; represents the adjacent spatial sub-region of the intersection node, Represents a junction node The number of adjacent spatial subregions of ;

[0119] road set ,in Indicates road, Indicates the number of extracted roads. The starting point and end point of each road are intersection nodes, and there are no intersection nodes inside the road.

[0120] Expressed as ,in Indicates the road Length, Indicates the road The number of lanes; Indicates the road The adjacent spatial subregion of Indicates the road The number of adjacent spatial subregions of ;

[0121] Spatial sub-region set ,in Indicates A spatial sub-region, is the number of sub-regions of the divided space;

[0122] Statistics are conducted on the characteristics of geographical environment elements in the spatial sub-areas. It is expressed as:

[0123] ;

[0124] in Represents a spatial sub-region The characteristic distribution of the local form of the inner city, Represents a spatial sub-region The characteristic distribution of inner city population and facility density, Represents a spatial sub-region The characteristic distribution of inner city functional types, Represents a spatial sub-region The characteristic distribution of inner city terrain;

[0125] The geographical environment elements of each road adjacent spatial sub-area are added to the road. For example:

[0126] ;

[0127] Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road Result;

[0128] Add the geographical environment element features of the adjacent spatial sub-regions of each intersection node to the intersection node. For example:

[0129] ;

[0130] Indicates that the intersection node The geographical environment features of the adjacent spatial sub-areas are added to the intersection node Result;

[0131] Add the geographical environment elements of the roads adjacent to each intersection node to the intersection node. For example:

[0132] ;

[0133] Represents a junction node Add the geographical environment features of adjacent roads to the intersection node As a result, Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road result.

[0134] Furthermore, the specific method of step 3 is:

[0135] Calculate the node degree of each intersection node:

[0136] ;

[0137] in, Represents the node degree of intersection node b, which belongs to the intersection node set , is the adjacency matrix of the road network. If the intersection node b and the road node c are directly connected by a road, then =1, otherwise = 0, intersection node c belongs to the intersection node set ;

[0138] Calculate the node betweenness of each intersection node:

[0139] ;

[0140] in, represents the node betweenness of intersection node b, Indicates the number of shortest paths between intersection node c and intersection node d, Indicates the number of paths that pass through intersection node b in the shortest path from intersection node c to intersection node d; intersection node b All belong to the intersection node set ;

[0141] Calculate the adjacent road capacity of each intersection node:

[0142] ;

[0143] in, represents the adjacent road capacity of intersection node b, U represents the set of intersection nodes directly adjacent to intersection node b through a road, is an element of the set U, Represents the intersection node b and the intersection node The length of the road between Represents the intersection node b and the intersection node The number of lanes of the road between;

[0144] Calculate the contribution of each intersection node to the global efficiency of the network:

[0145] - ;

[0146] ;

[0147] ;

[0148] in, represents the contribution of intersection node b to the global efficiency of the network, represents the global efficiency of the network, represents the global efficiency of the network after removing the intersection node b; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set ; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set , intersection node set is the intersection node set B after removing the intersection node b;

[0149] Calculate the contribution of each intersection node to the local efficiency of the network:

[0150] ;

[0151] in, Represents a junction node Contribution to local network efficiency, represents the number of elements in the set U, Represents the intersection nodes in set U With intersection node The shortest path length between .

[0152] Furthermore, the specific method of step 4 is:

[0153] End point The historical trajectory Take as an example, and represent it as a collection of intersection nodes and roads:

[0154] = { };

[0155] in, For historical trajectory The first intersection node, For historical trajectory The first path chosen, is the time at the first intersection node in the historical trajectory, is the speed at the first intersection node in the historical trajectory; Representing historical trajectory The end point is ; All belong to the intersection node set , All belong to the road set .

[0156] Furthermore, the neural network model in step 5 includes an input layer, a hidden layer and an output layer, wherein the input layer contains 9 neurons, receiving 9 dimensional features of each intersection node, including node degree, node betweenness, node contribution rate to network global efficiency, node contribution rate to network local efficiency, capacity of each adjacent road and geographical environment elements, wherein the geographical environment elements of the intersection nodes include the characteristic distribution of the local urban morphology corresponding to the intersection nodes, the characteristic distribution of urban population and facility density, the characteristic distribution of urban functional types and the characteristic distribution of urban terrain, the output layer includes 1 neuron, outputs the reward function value of each intersection node, the hidden layer includes 3 layers, including 128, 128 and 64 neurons respectively, and uses the relu function as the activation function.

[0157] Furthermore, the specific method of step 6 is:

[0158] First, calculate the reward function value matrix corresponding to the moving trajectory of the moving target :

[0159] ;

[0160] Among them, the reward function value matrix The Line Column Elements = , Indicates the area of ​​interest Move to area of ​​interest The maximum reward function value of and All belong to the area of ​​interest collection , for maneuvering targets in the region of interest Areas of interest Each path between calculates a corresponding total reward function value, and the maximum total reward function value is recorded as ;

[0161] Solve the following equation using mixed integer linear programming:

[0162] ;

[0163] The constraints are:

[0164] ;

[0165] ;

[0166] ;

[0167] ;

[0168] ;

[0169] in, and are all decision variables. If the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but = 1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0, k=1,2,3,......,K; represents the initial position of the kth maneuvering target, and Belongs to the area of ​​interest collection , if the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but = 1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0; and Respectively represent the kth maneuvering target visiting the area of ​​interest Areas of interest order;

[0170] According to the solution results of the mixed integer linear programming method, the movement planning scheme of each maneuvering target between each area of ​​interest is obtained, and for each path of the maneuvering target between two areas of interest, the path with the largest total reward function value is set as the simulation path of the maneuvering target between the two areas of interest, and finally the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints is obtained.

[0171] In summary, the present invention overcomes the problems of low authenticity, weak scenario generalization ability, high requirements for training data and low interpretability of current trajectory simulation methods, and effectively solves the problem of few real trajectories of multiple moving targets in special application scenarios.

[0172] Those skilled in the art will appreciate that the embodiments described are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to the embodiments described. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for simulating urban maneuvering target trajectories by coupling intention prediction and environmental constraints, characterized in that: The following steps are involved: Step 1: Based on the urban geographic environment dataset with unified spatiotemporal coordinates, combined with expert domain knowledge, considering the geographic environment elements of the urban built-up area, extract the set of interest areas that attract mobile targets; Step 2: According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-regions, and the geographical environment elements of each intersection node and road are calculated; Step 3, calculate the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency and the capacity of each adjacent road of each intersection node; Step 4, representing each historical trajectory whose destination belongs to the set of interest areas as a set containing each intersection node and road in the historical trajectory; Step 5, construct a neural network model. The input of the neural network model is the node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency, capacity of each adjacent road, and geographical environment factors of each intersection node. The output of the neural network model is the reward value of each intersection node. For each set in step 4, calculate its total reward value, generate all possible simulation paths according to the starting point and end point of the historical trajectory, and select the simulation path with the largest total reward value. Calculate the difference between the total reward value of the historical trajectory and the total reward value of the selected simulation path and record it as the loss value. According to the loss value, use the gradient descent method to backpropagate and update the parameters of the neural network model until the loss value converges or the maximum number of iterations is reached. End the training and save the reward values ​​of each intersection node output by the current neural network model to complete the deep learning training. Step 6: Based on the set of interest areas, the positions of each maneuvering target, and the reward value of each intersection node, the maneuvering target trajectory simulation set is calculated through the mixed integer linear programming method to complete the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints.

2. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 1 is characterized in that: The specific method of extracting the set of interest areas that attract the mobile target in step 1 is: The set of interest areas to be visited is divided into ; For a certain area of ​​interest to be visited , =1,2,3,......, , calculate the probability of the target accessing the area based on historical access records: ； ； in, Indicates a maneuvering target Visit the area of ​​interest to be visited The probability of represents a maneuvering target, Indicates the statistical time window selected in the historical access records. Indicates that within the statistical time window, the maneuvering target A collection of areas of interest to be visited that have been signed in. For collection An element in Indicates the area of ​​interest to be visited geographical factors, Indicates the area of ​​interest to be visited geographical factors, It means to calculate the Wasserstein distance between two distributions. Indicates geographical environment elements: Indicates the characteristic distribution of local urban morphology, Represents the characteristic distribution of urban population and facility density, represents the characteristic distribution of urban functional types, Represents the characteristic distribution of urban terrain; Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The frequency, Indicates that within the statistical time window, the maneuvering target Check in to areas of interest to be visited The number of times Indicates that within the statistical time window, the maneuvering target The total number of check-ins to visit the area of ​​interest; Set the probability threshold for selecting an area , for mobile targets , select access probability greater than The area of ​​interest is used as a collection of areas of interest to attract them : =1,2,3,......, ； s is the number of regions of interest.

3. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 2 is characterized in that: The specific method of step 2 is: According to the urban road network, the city is divided into a set of intersection nodes, a set of roads, and a set of spatial sub-areas: Intersection Node Set ,in represents the e-th intersection node, and e represents the number of extracted intersection nodes; Expressed as ,in, and Represents a junction node The two-dimensional position of Represents a junction node The adjacent roads Represents a junction node The number of adjacent roads; represents the adjacent spatial sub-region of the intersection node, Represents a junction node The number of adjacent spatial subregions of ; road set ,in Indicates road, Indicates the number of extracted roads. The starting point and end point of each road are intersection nodes, and there are no intersection nodes inside the road. Expressed as ,in Indicates the road Length, Indicates the road The number of lanes; Indicates the road The adjacent spatial subregion of Indicates the road The number of adjacent spatial subregions of ; Spatial sub-region set ,in Indicates A spatial sub-region, is the number of sub-regions of the divided space; Statistics are conducted on the characteristics of geographical environment elements in the spatial sub-areas. It is expressed as: ； in Represents a spatial sub-region The characteristic distribution of the local form of the inner city, Represents a spatial sub-region The characteristic distribution of inner city population and facility density, Represents a spatial sub-region The characteristic distribution of inner city functional types, Represents a spatial sub-region The characteristic distribution of inner city terrain; The geographical environment elements of each road adjacent spatial sub-area are added to the road. For example: ； Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road Result; Add the geographical environment element features of the spatial sub-regions adjacent to each intersection node to the intersection node. For example: ； Indicates that the intersection node The geographical environment features of the adjacent spatial sub-areas are added to the intersection node Result; Add the geographical environment elements of the roads adjacent to each intersection node to the intersection node. For example: ； Represents a junction node Add the geographical environment features of adjacent roads to the intersection node As a result, Indicates that the road The geographical environment features of the adjacent spatial sub-areas are added to the road result.

4. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 3 is characterized in that: The specific method of step 3 is: Calculate the node degree of each intersection node: ； in, Represents the node degree of intersection node b, which belongs to the intersection node set , is the adjacency matrix of the road network. If the intersection node b and the road node c are directly connected by a road, then =1, otherwise = 0, intersection node c belongs to the intersection node set ; Calculate the node betweenness of each intersection node: ； in, represents the node betweenness of intersection node b, Indicates the number of shortest paths between intersection node c and intersection node d, Indicates the number of paths that pass through intersection node b in the shortest path from intersection node c to intersection node d; intersection node b All belong to the intersection node set ; Calculate the adjacent road capacity of each intersection node: ； in, represents the adjacent road capacity of intersection node b, U represents the set of intersection nodes directly adjacent to intersection node b through a road, is an element of the set U, Represents the intersection node b and the intersection node The length of the road between Represents the intersection node b and the intersection node The number of lanes of the road between; Calculate the contribution of each intersection node to the global efficiency of the network: - ； ； ； in, represents the contribution of intersection node b to the global efficiency of the network, represents the global efficiency of the network, represents the global efficiency of the network after removing the intersection node b; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set ; Represents a junction node With intersection node The shortest path length between intersection nodes With intersection node All belong to the intersection node set , intersection node set is the intersection node set B after removing the intersection node b; Calculate the contribution of each intersection node to the local efficiency of the network: ； in, Represents a junction node Contribution to the local efficiency of the network, represents the number of elements in the set U, Represents the intersection nodes in set U With intersection node The shortest path length between .

5. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 4 is characterized in that: The specific method of step 4 is: End point The historical trajectory Take as an example, and represent it as a collection of intersection nodes and roads: = { }； in, For historical trajectory The first intersection node, For historical trajectory The first path chosen, is the time at the first intersection node in the historical trajectory, is the speed at the first intersection node in the historical trajectory; Representing historical trajectory The end point is ; All belong to the intersection node set , All belong to the road set .

6. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 5 is characterized in that: The neural network model in step 5 includes an input layer, a hidden layer and an output layer, wherein the input layer contains 9 neurons, receiving 9 dimensional features of each intersection node, including node degree, node betweenness, node contribution rate to the global network efficiency, node contribution rate to the local network efficiency, capacity of each adjacent road and geographical environment elements, wherein the geographical environment elements of the intersection nodes include the characteristic distribution of the local urban morphology corresponding to the intersection nodes, the characteristic distribution of the urban population and facility density, the characteristic distribution of the urban functional type and the characteristic distribution of the urban terrain; the output layer includes 1 neuron, outputting the reward function value of each intersection node; the hidden layer includes 3 layers, including 128, 128 and 64 neurons respectively, and the relu function is used as the activation function.

7. The urban maneuvering target trajectory simulation method of coupling intention prediction and environmental constraints according to claim 6 is characterized in that: The specific method of step 6 is: First, calculate the reward function value matrix corresponding to the moving trajectory of the moving target : ； Among them, the reward function value matrix The Line Column Elements = , Indicates the area of ​​interest Move to area of ​​interest The maximum reward function value of and All belong to the area of ​​interest collection , for maneuvering targets in the region of interest Areas of interest Each path between calculates a corresponding total reward function value, and the maximum total reward function value is recorded as ; Solve the following equation using mixed integer linear programming: ； The constraints are: ； ； ； ； ； in, and are all decision variables. If the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but =1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0, k=1,2,3,......,K; represents the initial position of the kth maneuvering target, and Belongs to the area of ​​interest collection , if the kth maneuvering target moves from the area of ​​interest Move to area of ​​interest ,but =1, if the kth maneuvering target does not move from the area of ​​interest Move to area of ​​interest ,but =0; and Respectively represent the kth maneuvering target visiting the area of ​​interest Areas of interest order; According to the solution results of the mixed integer linear programming method, the movement planning scheme of each maneuvering target between each area of ​​interest is obtained, and for each path of the maneuvering target between two areas of interest, the path with the largest total reward function value is set as the simulation path of the maneuvering target between the two areas of interest, and finally the urban maneuvering target trajectory simulation with coupled intention prediction and environmental constraints is obtained.

Images