Vehicle continuous driving cycle generation method and system based on GIS road network and data-driven markov chain

Abstract

This invention relates to the field of automotive performance testing and autonomous driving virtual simulation technology, specifically a method and system for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains. The method includes four main steps: S1, road topology map construction and OD route planning based on GIS big data; S2, random traffic breakpoint simulation based on a probability distribution model; S3, kinematic envelope construction based on bidirectional physical propagation; and S4, data-driven Markov velocity generation with forward aiming. By constructing a smooth envelope that absolutely conforms to the vehicle's kinematic limits through a unique bidirectional physical propagation algorithm, the abnormal problems of velocity steps and acceleration / deceleration exceeding physical limits in the random generation process of traditional Markov chains are completely solved. The generated driving conditions can be directly used for simulation testing without manual correction.

Description

Technical Field

[0001] This invention relates to the field of automotive performance testing and autonomous driving virtual simulation technology, specifically to a method and system for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains. Background Technology

[0002] With the rapid development of the new energy vehicle industry and advanced autonomous driving technology, vehicle driving conditions, as core foundational data for energy consumption assessment, powertrain calibration, and autonomous driving algorithm verification, have become a hot topic in the industry. Currently, driving condition generation methods based on stochastic processes, especially Markov chain models, have received widespread attention and application due to their ability to effectively simulate the randomness of driving behavior.

[0003] A search revealed a prior art method, device, computer equipment, and storage medium for detecting vehicle traffic violations (such as Chinese Invention Patent Publication CN113256982A). This method collects vehicle driving data from actual roads, constructs a Markov transition probability matrix with vehicle speed and acceleration as state variables, and generates a continuous speed-time series through random sampling, thereby constructing driving conditions. This method overcomes, to some extent, the limitation of standard cyclic driving conditions (such as WLTC and NEDC) being too simplistic, and can generate speed curves with a certain degree of randomness.

[0004] However, existing Markov chain-based test case generation technologies, including CN113256982A, still have the following insurmountable technical defects in practical applications and simulation testing: (i) Inability to integrate real road network topology and geographical constraints The Markov chain sampling process in existing methods (such as CN113256982A) is essentially a purely mathematical random walk, and the generated velocity curves lack a correlation with the real Geographic Information System (GIS) road network. Specifically, this manifests as: the inability to perceive geographical topological features such as road geometry, intersection distribution, and changes in road segment length; the inability to distinguish the differentiated impacts of different driving conditions, such as highways, urban roads, and suburban roads, on driving behavior; and the inability of the generated conditions to effectively support the needs of long-distance, cross-scenario continuous simulation testing.

[0005] (ii) Lack of physically reasonable modeling for random traffic events Existing methods rely solely on statistical patterns from historical driving data for speed transfer, failing to incorporate explicit modeling of random events in real-world traffic environments (such as pedestrians crossing, roadside obstacles, and red lights). This results in a lack of reasonable traffic incentives for vehicle speed changes in the generated scenarios, and the generated deceleration and stopping behaviors do not align with the causal relationships of real-world traffic scenarios, reducing the physical plausibility and engineering credibility of the scenarios.

[0006] (iii) The generated operating conditions are prone to violating the physical limits of vehicle kinematics. Due to the random sampling characteristics of Markov chains, methods such as CN113256982A are prone to generating abnormal data such as velocity jumps, instantaneous accelerations, or decelerations exceeding the vehicle's physical limits during actual execution. Such data violates Newtonian mechanics and cannot be directly used for simulation testing or hardware-in-the-loop experiments. It must rely on extensive manual post-processing for correction, significantly reducing the efficiency and automation of test case generation.

[0007] (iv) Lack of simulation of forward aiming behavior of human drivers Existing Markov chain models only sample the current velocity and acceleration state for the next time step, representing a typical memoryless or short-memory process. However, real human drivers exhibit clear forward-looking behavior, anticipating changes in speed limits, traffic events, and intersection distribution to perform predictive acceleration or deceleration. Current technologies fail to simulate this behavioral characteristic; the generated speed curves often show abrupt decelerations before sudden speed limit drops or stopping points, significantly deviating from real driving behavior and reducing the credibility of algorithm verification in simulation tests.

[0008] In summary, although existing technologies, represented by CN113256982A, have attempted to solve the problem of random generation of driving conditions using Markov chains, they still have significant shortcomings in areas such as road network topology fusion, traffic event modeling, kinematic and physical compliance, and anthropomorphic driving behavior. Therefore, there is an urgent need to develop an automated method for generating continuous driving conditions that can simultaneously consider real-world road geographical constraints, the rationality of random traffic events, vehicle kinematic and physical limits, and forward-looking driving behavior, in order to solve the aforementioned technical problems. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains, comprising the following steps: S1. Road topology map construction and OD route planning based on GIS big data: Read GIS vector road network data, complete data cleaning and topology reconstruction, and construct a directed multigraph with geographical distance weights; generate legal start-end point pairs based on the directed multigraph, complete shortest path planning and lane-level trajectory offset processing, and output discretized path node sequences and corresponding road attribute information. S2. Simulation of random traffic breakpoints based on probability distribution model: Random traffic events are generated through a composite probability distribution model along the path node sequence generated in step S1, and speed constraint breakpoints along the path distance distribution are generated accordingly to obtain the initial speed limit sequence. S3. Construction of kinematic envelope based on bidirectional physical propagation: The initial speed limit sequence obtained in step S2 is discretized into a hard speed limit sequence of equidistant path nodes. The hard speed limit sequence is kinematically constrained and filtered by a bidirectional physical propagation algorithm of forward and backward propagation to generate a smooth kinematic limit envelope that conforms to the vehicle's physical limits. S4. Data-driven Markov velocity generation with forward preview: Within the kinematic limit envelope constraint range generated in step S3, based on the Markov chain model with forward preview mechanism, node-by-node velocity sampling is completed to generate a continuous velocity profile that conforms to real human driving behavior, and finally outputs the continuous driving conditions of the vehicle.

[0011] Preferably, step S1 specifically includes: S11. Data Cleaning and Topology Construction: Read GIS vector data, divide the road network into different working condition layers according to the road functional attributes, parse the vector line segments into topological nodes, and construct a directed multigraph with the real geographical distance as the edge weight. S12. Generation of valid OD pairs: In the maximum connected subgraph of the directed multigraph, the Monte Carlo random sampling method is used to generate the starting point O and the ending point D. The OD pairs are pruned by a preset straight-line distance threshold to remove invalid OD pairs that are too close or have no connected path. S13. Path Planning and Trajectory Discretization: Based on A The algorithm calculates the shortest path between OD point pairs, applies a lateral parallel offset to the path centerline based on the road type, and simulates the geometric trajectory of a real lane. The path is discretized into a sequence of nodes with equal step lengths, and the geographical coordinates, cumulative path distance, road type, and legal speed limit information of each node are output.

[0012] Preferably, the composite probability distribution model mentioned in step S2 includes Poisson distribution, Bernoulli distribution, and beta distribution; step S2 specifically includes: S21. Random Event Generation for Road Segments: For each independent road segment, the number of random events within the segment is generated using a Poisson distribution. The Poisson distribution formula is: N-Poisson(λ·L), where L is the length of the road segment, λ is the event occurrence rate of the corresponding road type, and N is the number of times the random event occurs within a given interval. The random events include pedestrian crossings, roadside obstacles, and slow-moving vehicles ahead. S22, Intersection traffic light event generation: For each signal-controlled intersection, Bernoulli distribution is used to determine whether a vehicle encounters a red light when it arrives at the intersection, and the corresponding stopping breakpoint is generated. S23. Event Impact Degree and Breakpoint Generation: For the various events generated in steps S21 and S22, the degree coefficient X of the driver's impact on the event is calculated using the beta distribution X-Beta(α,β), where X represents the domain of the beta distribution: X∈(0,1), α is the number of successes, and β is the number of failures. When X exceeds a preset threshold, a stopping breakpoint with a target speed limit of 0 is generated. When X does not exceed the preset threshold, a deceleration breakpoint with a target speed limit of the current road segment's legal speed limit multiplied by a reduction factor is generated. All breakpoints are sorted along the cumulative path distance to generate an initial speed limit sequence.

[0013] Preferably, the bidirectional physical propagation algorithm described in step S3 specifically includes: S31. Sequence Equal Distance Discretization: Discretize the initial speed limit sequence generated in step S2 into a hard speed limit sequence Limit[i] with equal distance nodes according to the preset path step size Δs, where i is the node number, i=0,1,2,…,n-1, and n is the total number of nodes in the path; S32. Forward Propagation Constraint Filtering: Traverse the hard speed limit sequence from the starting point to the ending point of the path, limiting the maximum acceleration of the vehicle, ensuring that the speed limit of the next node does not exceed the maximum speed that the vehicle can reach under the current node's maximum physical acceleration. The forward propagation formula is: , in The maximum permissible acceleration set for the vehicle; Limit[i+1] is the speed constraint for the (i+1)th segment; S33. Backpropagation Constraint Filtering: Traverse the speed limit sequence processed by forward propagation from the end point to the beginning point of the path, limiting the maximum deceleration of the vehicle to ensure that the speed limit of the current node can decelerate the vehicle to the speed limit requirement of the next node within the range of the vehicle's maximum braking deceleration. The backpropagation formula is: , in The maximum permissible braking deceleration set for the vehicle; S34, Kinematic envelope output: The speed limit sequence after double filtering through forward and backward propagation is the smooth kinematic limit envelope that conforms to the physical limits of vehicle kinematics.

[0014] Preferably, the Markov chain model with forward aiming mechanism described in step S4 specifically includes: S41. Construction of the prior transition probability matrix: Establish a matrix with current velocity v and current acceleration a as state variables, and the target velocity of the next node as the transition probability matrix. A three-dimensional Markov transition probability matrix for the target; Gaussian prior is injected into the transition probability matrix based on a real human driving dataset; S42. Dynamic forward aiming processing: At each current node i in the path, aim forward for a path interval of a preset distance, extract the minimum speed limit within the aiming interval as the near-end hard speed limit, and extract the speed limit at the end of the aiming interval as the far-end soft speed limit. S43. Dynamic adjustment of the transition probability matrix: including hard constraint masking and dynamic desire offset processing; S44. Velocity Sampling and State Update: Based on the dynamically adjusted transition probability matrix, the target velocity of the next node is obtained using the roulette wheel sampling method. ; Calculate the average speed and travel time Δt between the current node and the next node, update the actual acceleration a of the current vehicle, and complete one state update; S45. Iterative loop: Repeat steps S42 to S44 until speed sampling of all nodes on the path is completed, generating a complete speed-time series, i.e., the continuous driving conditions of the vehicle.

[0015] Preferably, the hard constraint masking process in step S43 is as follows: In the prior transition probability matrix, the transition probabilities corresponding to all candidate speeds of the next node that are greater than the current node's kinematic envelope speed limit or greater than the near-end hard speed limit are set to 0, thus completing probability truncation; the dynamic desire offset process is as follows: Based on the difference between the current speed and the far-end soft speed limit, a linear bias coefficient is applied to the probability vector corresponding to the acceleration / deceleration interval in the transition probability matrix. When the current speed is lower than the preset percentage threshold of the remote soft speed limit and the current speed is at least 5 km / h lower than the remote soft speed limit, the probability of transitioning to the acceleration zone is increased. When the current speed is within ±5 km / h of the far-end soft speed limit, increase the probability of transitioning to the constant speed range; When the current speed is at least 5 km / h higher than the far-end soft speed limit, the probability of transitioning to the deceleration zone is increased.

[0016] This invention also discloses a vehicle continuous driving condition generation system based on GIS road network and data-driven Markov chain, comprising: The road network processing and route planning module is used to read GIS vector road network data, complete topology construction, OD point pair generation, path planning and trajectory discretization, and output the discretized path node sequence; The traffic event simulation module communicates with the road network processing and route planning module. It is used to generate random traffic events and corresponding speed constraint breakpoints through a composite probability distribution model and output the initial speed limit sequence. The kinematic envelope generation module communicates with the traffic event simulation module and is used to filter the initial speed limit sequence through a bidirectional physics propagation algorithm to generate a kinematic limit envelope that conforms to the vehicle's physical limits. The forward-looking Markov velocity generation module communicates with the kinematic envelope generation module and is used to complete velocity sampling and generate continuous driving conditions through a Markov chain model with a forward-looking mechanism under the constraints of the kinematic limit envelope. The data output module communicates with the pre-aiming Markov speed generation module and is used to output and store the generated driving conditions according to a preset format.

[0017] Preferably, the traffic event simulation module incorporates a composite probability model of Poisson, Bernoulli, and Beta distributions to generate random events along the path and traffic light events at intersections, and generates stopping points or deceleration points based on the degree of impact of the events.

[0018] Preferably, the kinematic envelope generation module incorporates a bidirectional physical propagation algorithm, which includes: forward propagation constraint filtering by traversing the path in the forward direction to limit the maximum acceleration, and backward propagation constraint filtering by traversing the path in the reverse direction to limit the maximum braking deceleration.

[0019] Preferably, the pre-aiming Markov speed generation module has a built-in Markov chain model with a forward pre-aiming mechanism. The model includes: a priori transition probability matrix based on real driving data, a dynamic forward pre-aiming processing unit, a dynamic adjustment unit for transition probabilities for hard constraint masking and dynamic desire offset, and a roulette speed sampling unit.

[0020] The beneficial effects of this invention are as follows: Compared with the prior art, this invention has the following beneficial technical effects: First, a unique bidirectional physics propagation algorithm is used to construct a smooth envelope that perfectly conforms to the vehicle's kinematic limits, completely solving the abnormal problems of velocity steps and acceleration / deceleration exceeding physical limits in the random generation process of traditional Markov chains. The generated driving conditions can be directly used for simulation testing without manual correction. Second, it deeply integrates real GIS road network topology and OD routing planning, fully preserving geographical features such as road geometry, intersection distribution, and road type. It can generate long-distance continuous driving conditions across multiple scenarios including highways, suburbs, and cities, with significantly higher scene fidelity than pure mathematical modeling methods. Third, it introduces Poisson distribution and... The composite probability model of Nulli and Beta distributions simulates random traffic events, making the speed constraint breakpoints have realistic traffic causes and significantly improving the rationality of the operating conditions. In addition, a forward aiming mechanism is innovatively added to the Markov chain to simulate the driver's anticipatory acceleration and deceleration behavior in response to speed limits and traffic events ahead. The generated speed curves are smooth and natural, with a high degree of human-likeness. Finally, the parameterized and modular design of the entire process supports the batch generation of heterogeneous operating conditions in milliseconds, which greatly reduces the cost and cycle of operating condition construction and can fully meet the urgent needs of new energy vehicle energy consumption calibration and large-scale iterative verification of autonomous driving algorithms. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, specific technical features, structures, or methods in one or more embodiments described below can be arbitrarily combined or substituted without contradiction to form new implementation methods.

[0023] Example 1 This embodiment provides a method for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains. For example... Figure 1 As shown, this method comprises four main steps: S1, road topology map construction and OD route planning based on GIS big data; S2, random traffic breakpoint simulation based on probability distribution model; S3, kinematic envelope construction based on bidirectional physical propagation; and S4, data-driven Markov velocity generation with forward aiming. The following sections provide a detailed explanation of each step in conjunction with specific parameter settings and application scenarios.

[0024] Step S1: Construction of road topology map and OD routing planning based on GIS big data The core of this step is to construct road topology constraints based on real geographic information, providing a foundation of realistic road scenarios for subsequent work condition generation. Specifically, this involves reading GIS vector road network data, completing data cleaning and topology reconstruction, and constructing a directed multigraph with geographic distance weights. Based on the directed multigraph, valid start-end point pairs are generated, shortest path planning and lane-level trajectory offset processing are completed, and discretized path node sequences and corresponding road attribute information are output.

[0025] The specific steps are as follows: S11, Data Cleaning and Topology Construction This embodiment reads open-source GIS vector road network data of a city, in GeoJSON format. Based on the road function attribute labels defined in OpenStreetMap, the road network is divided into three types of layers: highway layer, urban arterial road layer, and suburban / residential road layer. Then, using Python's NetworkX library, the LineString vector segments in the road network are parsed into topological nodes. A directed multigraph (MultiDiGraph) is constructed using the actual geographic distance (in meters) of the line segments as the edge weight. For two-way roads, two directed edges in opposite directions are generated. For roads with multiple lanes but overlapping centerlines, they are distinguished through subsequent lateral offset processing, while the graph structure still maintains a single-edge representation to reduce topological complexity.

[0026] S12, Generation of valid OD point pairs Within the maximally connected subgraph of a directed multigraph, a Monte Carlo random sampling method is used to generate the starting point O and the ending point D. To ensure that the generated driving scenarios have sufficient effective mileage, a straight-line distance threshold of 5km-30km is set for OD point pairs. Point pairs with a straight-line distance less than 5km (resulting in excessively short scenarios) or greater than 30km (resulting in excessive computational cost) are discarded. Simultaneously, reachability checks in graph theory (such as calling the `has_path` function in NetworkX) are used to remove invalid OD point pairs without connected paths. Ultimately, this embodiment generates a valid OD point pair: the starting point O is located in the city center, and the ending point D is located at a suburban highway exit, with a straight-line distance of approximately 18.5km.

[0027] S13, Path Planning and Trajectory Discretization Using A The algorithm (using Euclidean distance as the heuristic function) calculates the shortest path between OD point pairs. Based on the road type, a lateral parallel offset of 1.75m is applied to the path centerline to simulate the actual driving trajectory of a single-lane road in one direction. The complete path is discretized into an equidistant node sequence with a step size of Δs=5m, and the geographic coordinates, cumulative path distance, road type, and legal speed limit of each node are output. In this embodiment, the total number of nodes in the path is n=4200, corresponding to a total mileage of 21km.

[0028] Step S2: Simulation of stochastic traffic breakpoints based on probability distribution model Along the path node sequence generated in step S1, random traffic events are generated using a composite probability distribution model, corresponding to speed constraint breakpoints along the path distance distribution, thus obtaining the initial speed limit sequence. The core of this step is to simulate random disturbances in real traffic flow and set reasonable speed constraints for the generation of operating conditions. The specific sub-steps are as follows: S21, Random Event Generation on Road Sections For each independent road segment (the section between two adjacent intersections or road attribute change points) in the path, the number of random events within the segment is generated using a Poisson distribution. The Poisson distribution formula is: N - Poisson(λ·L), where L is the road segment length (unit: kilometers), and λ is the event occurrence rate for the corresponding road type (unit: times / kilometer). In this embodiment, λ = 0.8 times / kilometer for urban roads, λ = 0.3 times / kilometer for suburban roads, and λ = 0.1 times / kilometer for highways. The generated random event types include: pedestrian crossing, roadside obstacles, and slow-moving vehicles, which are randomly assigned with equal probability. For example, if a certain urban road segment is 0.5 kilometers long, its event count N follows Poisson(0.4), and one event is actually sampled, with the type being "pedestrian crossing".

[0029] S22, Intersection Traffic Light Event Generation For each signalized intersection in the path (nodes with traffic light attributes extracted from GIS data), a Bernoulli distribution is used to determine whether a vehicle encounters a red light upon arrival at the intersection, generating a corresponding stopping breakpoint. In this embodiment, the probability of encountering a red light is set to 0.5. The path passes through 12 signalized intersections, and Bernoulli sampling identifies 5 intersections with red lights, generating 5 corresponding stopping breakpoints. Each stopping breakpoint is located 10 meters before the intersection node (considering the driver's reaction distance).

[0030] S23. Event Impact and Breakpoint Generation For the various events generated in steps S21 and S22 (including random events on road segments and red light events at intersections), the beta distribution X-Beta (α=2, β=5) is used to calculate the degree coefficient X of the driver's impact on the event. The expected value of this distribution is α / (α+β)=0.2857, indicating that most events have small impact coefficients, and only a few events have significant impacts. A threshold of 0.7 is set: when X>0.7, a stopping breakpoint with a target speed limit of 0 is generated; when X≤0.7, a deceleration breakpoint with a target speed limit of the current road segment's legal speed limit multiplied by a decrease factor of 0.6 is generated. For example, if an event with X=0.8 occurs on a road segment with a speed limit of 60km / h ahead, a stopping breakpoint with a target speed limit of 0km / h is generated; if X=0.5, a deceleration breakpoint with a target speed limit of 36km / h is generated. All breakpoints are sorted along the cumulative path distance, and an initial speed limit sequence distributed along the path is generated through linear interpolation.

[0031] Step S3: Construction of the kinematic envelope based on bidirectional physical propagation The initial speed limit sequence obtained in step S2 is discretized into a hard speed limit sequence of equidistant path nodes. A bidirectional physical propagation algorithm (forward and backward propagation) is used to filter the hard speed limit sequence based on kinematic constraints, generating a smooth kinematic limit envelope that conforms to the vehicle's physical limits. This step is one of the key innovations of this invention, constructing a kinematic envelope that absolutely conforms to the vehicle's physical limits using a bidirectional physical propagation algorithm. The specific sub-steps are as follows: S31, Sequence Equal Dispersion The initial speed limit sequence generated in step S2 (the distance along the path is not necessarily equal) is interpolated to the hard speed limit sequence Limit[i] corresponding to the path nodes with a step size of Δs=5m, where i=0,1,2,…,4199.

[0032] S32, Forward Propagation Constraint Filtering Traverse the hard speed limit sequence in the forward direction from the starting point (i=0) to the ending point (i=4199), applying the forward propagation formula:

[0033] in, The maximum allowable acceleration is set for the vehicle, and Limit[i+1] is the speed constraint for the (i+1)th segment; in this embodiment, it is taken as... =2m / s 2 This value corresponds to the typical acceleration capability of a regular passenger car. Forward propagation ensures that the acceleration process of the vehicle between any two adjacent nodes will not exceed the physical limits. This step effectively avoids the problem of "rapid acceleration violating the laws of physics" in subsequent speed sampling.

[0034] S33, Backpropagation Constraint Filtering Traverse the rate-limited sequence after forward propagation from the end point (i=4199) to the beginning point (i=0), applying the backpropagation formula:

[0035] in, The maximum permissible braking deceleration set for the vehicle is taken in this embodiment. =3m / s 2 Backpropagation ensures that the vehicle can safely decelerate from the current speed limit to the next speed limit within the maximum braking deceleration range. In other words, if the current speed limit is too high, making it impossible to achieve maximum deceleration within a step size Δs, then... If the speed is reduced to the speed limit of the next node, the speed limit of the current node will be reduced to a reasonable value. This step simulates the driver's anticipatory braking behavior and avoids "unrealistic sudden deceleration".

[0036] S34, Kinematic envelope output The speed-limited sequence, filtered by both forward and backward propagation, is the smooth kinematic limit envelope that conforms to the physical limits of vehicle kinematics. This envelope serves as the absolute constraint boundary for subsequent Markov velocity sampling; any velocity state exceeding this envelope will be forcibly prohibited.

[0037] Step S4: Data-driven Markov velocity generation with forward aiming Within the kinematic limit envelope constraint range generated in step S3, node-by-node velocity sampling is performed based on a Markov chain model with forward aiming mechanism to generate a continuous velocity profile that conforms to real human driving behavior, ultimately outputting the vehicle's continuous driving conditions. This step is another important innovation of this invention: generating a highly human-like velocity profile within the kinematic envelope using a Markov chain model with forward aiming. Specific sub-steps are as follows: S41. Construction of the Prior Transition Probability Matrix Establish a three-dimensional Markov transition probability matrix, with the state variables being: current velocity v (range 0-40 m / s, step size 0.5 m / s) and current acceleration a (range -3~2 m / s²). 2 Step length 0.2 m / s 2 ), the velocity of the next node (Values ​​range 0-40 m / s, step size 0.5 m / s) represents the three-dimensional Markov transition probability matrix of the target. Based on publicly available real human driving datasets, a Gaussian prior is injected into the transition probability matrix to construct a data-driven prior transition probability matrix. This prior reflects the general laws of human driving: the probability of acceleration is higher in the low-speed range, the probability of constant speed is higher in the high-speed range, and the probability of acceleration and deceleration approaching the physical limits is close to 0.

[0038] S42, Dynamic Forward Aiming Processing At each current node i (i=0,...,4199) on the path, a path interval with a preset distance is pre-aimed forward. In this embodiment, the pre-aiming distance is 150 meters (corresponding to 30 nodes, since Δs=5 meters). The minimum speed limit within the pre-aiming interval is extracted as the near-end hard speed limit. Simultaneously, the speed limit at the end of the pre-aiming interval (i.e., node i+30) is extracted as the far-end soft speed limit. The near-end hard speed limit represents an insurmountable speed obstacle ahead, while the far-end soft speed limit represents a desired speed target further ahead.

[0039] S43. Dynamic adjustment of the transition probability matrix This step includes two sub-processes: hard constraint masking and dynamic desire offset processing.

[0040] Hard constraint masking: In the prior transition probability matrix, all transition probabilities corresponding to the next node velocity candidate values ​​that are greater than the current node's kinematic envelope velocity limit (from step S34) or greater than the near-end hard velocity limit are set to 0, thus completing probability truncation. This ensures that the sampling rate is always within the physical limits.

[0041] Dynamic Desire Offset Processing: Based on the difference between the current speed and the far-end soft speed limit, a linear bias coefficient is applied to the probability vectors of the corresponding acceleration / deceleration intervals in the transition probability matrix: When the current speed is less than 80% of the far-end soft speed limit and the current speed is less than 5 km / h from the far-end soft speed limit, the probability vector of the acceleration interval is multiplied by an increasing bias coefficient of 1.2-2.0 to amplify the acceleration probability; when the current speed is within ±5 km / h of the far-end soft speed limit, the probability vector of the constant speed interval is multiplied by a bias coefficient of 1.5 to increase the transition probability of the constant speed interval; when the current speed is at least 5 km / h higher than the far-end soft speed limit, the probability vector of the deceleration interval is multiplied by an increasing bias coefficient of 1.5-2.5 to increase the deceleration probability of the deceleration interval. After biasing, the probability matrix is ​​normalized again.

[0042] S44. Velocity Sampling and State Update Based on the dynamically adjusted transition probability matrix, the target velocity of the next node is obtained using the roulette wheel sampling method. ; Calculate the average speed between the current node and the next node, and combine it with a step size of Δs=5m to calculate the travel time Δt, and then update the actual acceleration a of the current vehicle to complete one state update; S45, Iteration Repeat steps S42 to S44, processing nodes i=0,1,2,…,4198 sequentially until speed sampling of all 4200 nodes is completed. The resulting speed-time series represents a continuous driving scenario covering a total distance of 21km, encompassing urban, suburban, and highway environments. This scenario can be directly used for vehicle energy consumption simulation or virtual testing of autonomous driving algorithms.

[0043] Example 2 This embodiment provides a vehicle continuous driving condition generation system based on GIS road network and data-driven Markov chain, used to implement the method described in Embodiment 1. The system includes the following five modules: 1. Road Network Processing and Route Planning Module This module reads GIS vector road network data, performs topology construction, OD point pair generation, path planning, and trajectory discretization, and outputs a discretized path node sequence. It utilizes the NetworkX library to construct a weighted directed multigraph and provides a Monte Carlo OD sampling interface and A... The shortest path planning interface simulates lane trajectories by lateral offset based on road width and calls an equidistant discretization function to generate a node sequence. The module outputs a discretized list of path nodes, each containing attributes such as geographic coordinates, cumulative distance, road type, and legal speed limit.

[0044] 2. Traffic Incident Simulation Module This module communicates with the road network processing and route planning module to generate random traffic events and corresponding speed constraint breakpoints using a composite probability distribution model, and outputs an initial speed limit sequence. The traffic event simulation module incorporates a composite probability model of Poisson, Bernoulli, and Beta distributions to generate random events along the path and traffic light events at intersections, and generates stopping or deceleration breakpoints based on the degree of event impact. This module allows users to configure event occurrence rate λ, red light probability, Beta distribution parameters, and impact thresholds. The module outputs an initial speed limit sequence, which consists of several speed constraint breakpoints distributed along the path.

[0045] 3. Kinematic envelope generation module This module communicates with the traffic incident simulation module and is used to perform kinematic constraint filtering on the initial speed limit sequence using a bidirectional physics propagation algorithm to generate a kinematic limit envelope that conforms to the vehicle's physical limits. The module has a built-in bidirectional physics propagation algorithm, which includes forward propagation constraint filtering that traverses the path in the forward direction to limit the maximum acceleration, and backward propagation constraint filtering that traverses the path in the reverse direction to limit the maximum braking deceleration.

[0046] 4. Pre-aiming Markov velocity generation module This module communicates with the kinematic envelope generation module and is used to generate continuous driving conditions by performing velocity sampling through a Markov chain model with a forward aiming mechanism under kinematic limit envelope constraints. Internally, the module constructs a three-dimensional Markov transition probability matrix and supports loading real driving datasets to calculate prior probabilities. The model includes: a prior transition probability matrix based on real driving data, a dynamic forward aiming processing unit, a dynamic adjustment unit for transition probabilities used for hard constraint masking and dynamic desire offset, and a roulette-style velocity sampling unit.

[0047] 5. Data Output Module This module communicates with the pre-aiming Markov speed generation module to output and store the generated driving conditions according to a preset format. It supports CSV text format and MAT file format. This module can also call the matplotlib library to plot speed-time curves and speed-distance curves, and provide visualizations of driving condition statistics (average speed, maximum speed, acceleration / deceleration / constant speed time percentages, etc.).

[0048] The system described in this embodiment can be deployed on a cloud server or a local industrial control computer. All parameters can be adjusted through configuration files to achieve automated, batch generation of process conditions without manual intervention.

[0049] Those skilled in the art should understand that the parameters disclosed in the above embodiments (such as step size Δs=5m, aiming distance 150m, acceleration limit 2.0 m / s²) are not necessarily accurate. 2 Deceleration limit 3.0 m / s 2 (e.g., GIS + bidirectional physical propagation + pre-aiming Markov) are for illustrative purposes only and are not fixed. In practical applications, the parameters can be reasonably adjusted according to the specific test vehicle's power performance, simulation requirements, or road network characteristics. All simple parameter variations or equivalent substitutions based on the core idea of ​​this invention (GIS + bidirectional physical propagation + pre-aiming Markov) should fall within the protection scope of this invention.

Claims

1. A method for generating vehicle continuous driving cycle based on GIS road network and data-driven Markov chain, characterized in that, Includes the following steps: S1. Road topology map construction and OD route planning based on GIS big data: Read GIS vector road network data, complete data cleaning and topology reconstruction, and construct a directed multigraph with geographical distance weights; generate legal start-end point pairs based on the directed multigraph, complete shortest path planning and lane-level trajectory offset processing, and output discretized path node sequences and corresponding road attribute information. S2. Simulation of random traffic breakpoints based on probability distribution model: Random traffic events are generated through a composite probability distribution model along the path node sequence generated in step S1, and speed constraint breakpoints along the path distance distribution are generated accordingly to obtain the initial speed limit sequence. S3. Construction of kinematic envelope based on bidirectional physical propagation: The initial speed limit sequence obtained in step S2 is discretized into a hard speed limit sequence of equidistant path nodes. The hard speed limit sequence is kinematically constrained and filtered by a bidirectional physical propagation algorithm of forward and backward propagation to generate a smooth kinematic limit envelope that conforms to the vehicle's physical limits. S4. Data-driven Markov velocity generation with forward preview: Within the kinematic limit envelope constraint range generated in step S3, based on the Markov chain model with forward preview mechanism, node-by-node velocity sampling is completed to generate a continuous velocity profile that conforms to real human driving behavior, and finally outputs the continuous driving conditions of the vehicle.

2. The method of claim 1, wherein, Step S1 specifically includes: S11. Data Cleaning and Topology Construction: Read GIS vector data, divide the road network into different working condition layers according to the road functional attributes, parse the vector line segments into topological nodes, and construct a directed multigraph with the real geographical distance as the edge weight. S12. Generation of valid OD pairs: In the maximum connected subgraph of the directed multigraph, the Monte Carlo random sampling method is used to generate the starting point O and the ending point D. The OD pairs are pruned by a preset straight-line distance threshold to remove invalid OD pairs that are too close or have no connected path. S13, Path Planning and Trajectory Discretization: Based on The algorithm calculates the shortest path between OD point pairs, applies a lateral parallel offset to the path centerline according to the road type, and simulates the geometric trajectory of a real lane. The path is discretized into a sequence of nodes with equal step lengths, and the geographical coordinates, cumulative path distance, road type, and legal speed limit information of each node are output.

3. The method according to claim 1, characterized in that, The composite probability distribution model mentioned in step S2 includes the Poisson distribution, Bernoulli distribution, and beta distribution; step S2 specifically includes: S21. Random Event Generation for Road Segments: For each independent road segment, the number of random events within the segment is generated using a Poisson distribution. The Poisson distribution formula is: N-Poisson(λ·L), where L is the length of the road segment, λ is the event occurrence rate of the corresponding road type, and N is the number of times the random event occurs within a given interval. The random events include pedestrian crossings, roadside obstacles, and slow-moving vehicles ahead. S22, Intersection traffic light event generation: For each signal-controlled intersection, Bernoulli distribution is used to determine whether a vehicle encounters a red light when it arrives at the intersection, and the corresponding stopping breakpoint is generated. S23. Event Impact Degree and Breakpoint Generation: For the various events generated in steps S21 and S22, the degree coefficient X of the driver's impact on the event is calculated using the beta distribution X-Beta(α,β), where X represents the domain of the beta distribution: X∈(0,1), α is the number of successes, and β is the number of failures. When X exceeds a preset threshold, an event stopping breakpoint with a target speed limit of 0 is generated. When X does not exceed the preset threshold, a deceleration breakpoint with a target speed limit of the current road segment's legal speed limit multiplied by a reduction factor is generated. All breakpoints are sorted along the cumulative path distance to generate an initial speed limit sequence.

4. The method according to claim 1, characterized in that, The bidirectional physics propagation algorithm described in step S3 specifically includes: S31. Sequence Equal Distance Discretization: Discretize the initial speed limit sequence generated in step S2 into a hard speed limit sequence Limit[i] with equal distance nodes according to the preset path step size Δs, where i is the node number, i=0,1,2,…,n-1, and n is the total number of nodes in the path; S32. Forward Propagation Constraint Filtering: Traverse the hard speed limit sequence from the starting point to the ending point of the path, limiting the maximum acceleration of the vehicle, ensuring that the speed limit of the next node does not exceed the maximum speed that the vehicle can reach under the current node's maximum physical acceleration. The forward propagation formula is: ， in The maximum permissible acceleration set for the vehicle; Limit[i+1] is the speed constraint for the (i+1)th segment; S33. Backpropagation Constraint Filtering: Traverse the speed limit sequence processed by forward propagation from the end point to the beginning point of the path, limiting the maximum deceleration of the vehicle to ensure that the speed limit of the current node can decelerate the vehicle to the speed limit requirement of the next node within the range of the vehicle's maximum braking deceleration. The backpropagation formula is: ， in The maximum permissible braking deceleration set for the vehicle; S34, Kinematic envelope output: The speed limit sequence after double filtering through forward and backward propagation is the smooth kinematic limit envelope that conforms to the physical limits of vehicle kinematics.

5. The method according to claim 1, characterized in that, The Markov chain model with forward aiming mechanism mentioned in step S4 specifically includes: S41. Construction of the prior transition probability matrix: Establish a matrix with current velocity v and current acceleration a as state variables, and the target velocity of the next node as the transition probability matrix. A three-dimensional Markov transition probability matrix for the target; Gaussian prior is injected into the transition probability matrix based on a real human driving dataset; S42. Dynamic forward aiming processing: At each current node i in the path, aim forward for a path interval of a preset distance, extract the minimum speed limit within the aiming interval as the near-end hard speed limit, and extract the speed limit at the end of the aiming interval as the far-end soft speed limit. S43. Dynamic adjustment of the transition probability matrix: including hard constraint masking and dynamic desire offset processing; S44. Velocity Sampling and State Update: Based on the dynamically adjusted transition probability matrix, the target velocity of the next node is obtained using the roulette wheel sampling method. ; Calculate the average speed and travel time Δt between the current node and the next node, update the actual acceleration a of the current vehicle, and complete one state update; S45. Iterative loop: Repeat steps S42 to S44 until speed sampling of all nodes on the path is completed, generating a complete speed-time series, i.e., the continuous driving conditions of the vehicle.

6. The method according to claim 5, characterized in that, The hard constraint masking process in step S43 is as follows: In the prior transition probability matrix, the transition probabilities corresponding to all next node velocity candidate values ​​that are greater than the current node's kinematic envelope speed limit or greater than the near-end hard speed limit are set to 0, thus completing probability truncation; the dynamic desire offset process is as follows: Based on the difference between the current speed and the far-end soft speed limit, a linear bias coefficient is applied to the probability vector corresponding to the acceleration / deceleration interval in the transition probability matrix. When the current speed is lower than the preset percentage threshold of the remote soft speed limit and the current vehicle speed is at least 5 km / h lower than the remote soft speed limit, the probability of transitioning to the acceleration zone is increased. When the current speed is within ±5 km / h of the far-end soft speed limit, increase the probability of transitioning to the constant speed range; When the current speed is at least 5 km / h higher than the far-end soft speed limit, the probability of transitioning to the deceleration zone is increased.

7. A system for generating continuous vehicle driving conditions based on GIS road networks and data-driven Markov chains, characterized in that, include: The road network processing and route planning module is used to read GIS vector road network data, complete topology construction, OD point pair generation, path planning and trajectory discretization, and output the discretized path node sequence; The traffic event simulation module communicates with the road network processing and route planning module. It is used to generate random traffic events and corresponding speed constraint breakpoints through a composite probability distribution model and output the initial speed limit sequence. The kinematic envelope generation module communicates with the traffic event simulation module and is used to filter the initial speed limit sequence through a bidirectional physics propagation algorithm to generate a kinematic limit envelope that conforms to the vehicle's physical limits. The forward-looking Markov velocity generation module communicates with the kinematic envelope generation module and is used to complete velocity sampling and generate continuous driving conditions through a Markov chain model with a forward-looking mechanism under the constraints of the kinematic limit envelope. The data output module communicates with the pre-aiming Markov speed generation module and is used to output and store the generated driving conditions according to a preset format.

8. The system according to claim 7, characterized in that, The traffic event simulation module incorporates a composite probability model of Poisson, Bernoulli, and Beta distributions to generate random events along the path and traffic light events at intersections, and generates stopping or deceleration points based on the degree of impact of the events.

9. The system according to claim 7, characterized in that, The kinematic envelope generation module incorporates a bidirectional physical propagation algorithm, which includes forward propagation constraint filtering that traverses the path in the forward direction to limit the maximum acceleration, and backward propagation constraint filtering that traverses the path in the reverse direction to limit the maximum braking deceleration.

10. The system according to claim 7, characterized in that, The pre-aiming Markov velocity generation module has a built-in Markov chain model with a forward pre-aiming mechanism. The model includes: a priori transition probability matrix based on real driving data, a dynamic forward pre-aiming processing unit, a dynamic adjustment unit for transition probabilities for hard constraint masking and dynamic desire offset, and a roulette speed sampling unit.

Images