Lane-level navigation method based on lane traffic state recognition

Abstract

A lane-level navigation method based on lane traffic state recognition comprises the specific steps that a road structure of a whole road network is obtained through a high-precision map, and a layered map model is constructed; predicting a future traffic state according to the historical lane-level traffic flow; based on the current position of the vehicle, the target position and the predicted future traffic state, adopting an improved A * search algorithm to obtain a planned path of an initial road layer; mapping each road section of the road layer to a lane layer based on the constructed hierarchical map model; a lane layer path planning module is started, and the optimal path of the vehicle in the control point level network is searched; a vehicle sequentially passes through each control point in the path, the total passing time of the current road section is calculated, the updated total passing time of the current road section is fed back to the road layer, the road layer adopts an improved A * search algorithm to re-plan the path, and path guidance from top to bottom and passing feedback from bottom to top are realized. And the path planning precision, the passing efficiency and the traffic safety are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems technology, and relates to a lane-level navigation method for urban roads, specifically a lane-level navigation method based on lane traffic state recognition. Background Technology

[0002] The emergence of intelligent connected vehicle technology has provided new solutions and technical support for alleviating urban traffic pressure, improving road efficiency, and enhancing traffic safety. Through vehicle-road cooperation, information perception, and intelligent decision-making, intelligent connected transportation systems can achieve dynamic perception and intelligent control of traffic flow, promoting the transformation of traffic management from traditional models to intelligent and refined ones, and laying a solid foundation for building an efficient, green, and safe modern urban transportation system.

[0003] In existing urban road driving scenarios, traditional vehicle navigation systems and driver assistance technologies have significant limitations: Coarse route planning: Most navigation systems can only provide coarse route planning, unable to pinpoint specific lanes, making it difficult to meet the demands of modern traffic for refined driving control. Lack of real-time perception: These systems typically lack the ability to perceive and respond to road conditions in real time, failing to effectively consider dynamic traffic factors such as changes in traffic flow, road closures due to construction, and unexpected events. This results in a disconnect between the route planning results and actual road conditions, impacting traffic efficiency and driving experience.

[0004] Therefore, in order to overcome the above problems, there is an urgent need for a lane-level navigation method that can comprehensively perceive road information to improve the safety, comfort and efficiency of vehicles driving in complex road conditions and meet the growing demand for intelligent transportation. Summary of the Invention

[0005] In view of the shortcomings and deficiencies of the existing technology, the purpose of this invention is to provide a lane-level navigation method based on lane traffic state recognition. This method provides drivers with refined lane-level route planning services based on real-time perceived traffic conditions, effectively improving the accuracy of route planning, traffic efficiency and traffic safety, and is especially suitable for fine-grained navigation tasks in complex urban road network environments.

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

[0007] A lane-level navigation method based on lane traffic state recognition includes the following steps:

[0008] Step S1. Obtain the road structure of the entire road network through high-precision maps and collect historical lane-level traffic flow information;

[0009] Step S2. Construct a layered map model, which includes a road layer map model, a road-lane connection layer map model, and a lane layer map model;

[0010] Step S3. Run the lane-level navigation system to predict future traffic conditions based on historical lane-level traffic flow;

[0011] Step S4. Obtain location information;

[0012] Step S5. Based on the vehicle's current location, target location, and predicted future traffic conditions, an improved A* search algorithm is used to obtain the planned path for the initial road layer. The improved A* algorithm aims to minimize the cumulative travel time. Each time a node is expanded, the current cumulative travel time is recorded and used as the starting time t1 for entering the new road segment. The vehicle integrates on the predicted speed-time fitting curve of the new road segment starting from t1. The time corresponding to the integral value of the new road segment length L is the time t2 when the vehicle leaves the road segment. t2 plus the time to pass through the node is the current cumulative travel time of the new road segment.

[0013] Step S6. Based on the constructed hierarchical map model, map each road segment in the road layer to the lane layer, and determine the starting control point and ending control point of the lane-level path planning;

[0014] Step S7. Start the lane layer path planning module, which is based on a deep reinforcement learning framework to search for the optimal path of the vehicle in the control point level network;

[0015] Step S8. Based on lane-layer path planning, the vehicle starts from the starting control point of the path and sequentially passes through each control point in the path. Whenever the vehicle reaches the i-th control point in the path... At that time, the time consumption of all traffic segments is obtained, the remaining travel time of the road segment is estimated, the total travel time of the current road segment is calculated, and the updated total travel time of the current road segment is fed back to the road layer. The road layer uses an improved A* search algorithm to re-plan the path. If the optimal path is still the original path sequence, the lane layer path continues to be executed; if the path is updated, it returns to step S6 to relocate the starting control point and ending control point of the lane layer path planning, and performs lane layer path planning for the remaining control points.

[0016] Step S9. After completing the lane layer path planning and travel time feedback for the current road segment, the system automatically enters the planning process for the next road segment, repeating steps S6 to S8 until the end of the path is reached.

[0017] As a preferred embodiment of the present invention, the road layer map model is constructed in the form of a directed graph, with nodes being road intersections and edges being road segments with direction, length, speed, and flow rate; the road-lane connection layer map model is used to map road-level paths into a set of drivable lanes; the lane layer map model is used to describe each lane and its control point information, wherein the control points integrate future speed and flow rate information output by the traffic flow prediction model, and include lane change rules, left and right lane line types, and lane numbers.

[0018] As a preferred embodiment of the present invention, step S4 acquires the current position coordinates and target position coordinates of the vehicle in real time, and maps the coordinates to the starting node and target node of the road layer by means of map matching based on the current position coordinates and target position coordinates.

[0019] As a preferred embodiment of the present invention, step S5 includes the following steps:

[0020] Step S5-1. For any road segment, calculate the average speed of each lane based on the predicted speed of each lane control point, and determine the overall average speed of the road segment by using the predicted flow rate of the corresponding lane as the weight and the weighted average method.

[0021] Step S5-2. Fit the discrete speed prediction values ​​of each road segment to obtain the fitting curve of predicted speed-time for each road segment;

[0022] Step S5-3. Run the improved A* search algorithm on the constructed road layer map model with the goal of minimizing the cumulative travel time. Each time the improved A* algorithm expands a node, it records the current cumulative travel time and uses this time as the starting time t1 for entering the new road segment. It integrates the predicted speed-time fitting curve of the new road segment starting from t1. The time corresponding to the length L of the new road segment is the time t2 when the vehicle leaves the road segment. t2 plus the time to pass through the node is the current cumulative travel time of the new road segment.

[0023] As a preferred embodiment of the present invention, step S6 first queries the set of all feasible lanes corresponding to the road segment based on the constructed layered map model, clarifies the set of exit endpoints and the set of entry endpoints of the two intersections of the road segment, and then filters out lanes that meet the requirements of the target driving direction based on the topological connection relationship of the intersections, thereby determining the starting control point and the ending control point of the lane layer path planning.

[0024] As a preferred embodiment of the present invention, the deep reinforcement learning framework described in step S7 employs a deep Q-network, and state S t ={s cur ,s goal ,d dem ,v cur G pred ,b left,b right ,b forbid ,ρ risk}; where s cur Number the current control point, s goal Number the target control points for the current stage, d dem v is the remaining straight-line distance from the target control point. cur G is the current vehicle speed. pred b represents the predicted future traffic conditions for the current control point and its left and right adjacent control points. left b right : Boolean value, indicating whether a left / right lane change is currently allowed, b forbid ρ is a Boolean indicator to indicate whether the area is in a no-lane-changing zone. risk As a potential conflict risk index;

[0025] The set of actions to be performed is A = {a0, a1, a2} = {go straight, change lanes left, change lanes right};

[0026] Reward function R t =ω1(s t )R1+ω2(s t )R2+ω3(s t )R3+ω4(s t )R4+ω5(s t )R5+ω6(s t R6

[0027] Among them, R1 is the reward for approaching or reaching the target control point, R2 is the penalty for illegal lane change, R3 is the penalty for deviation from the target direction, R4 is the reward / penalty for target-oriented lane change, R5 is the reward for traffic conditions at the next control point, R6 is the penalty for potential conflict risk, and ω i (s t (i = 1, ..., 6) are dynamic weights.

[0028] As a preferred embodiment of the present invention, in step S8, when estimating the remaining travel time of the road segment, the straight-through time and lane-changing delay of the subsequent control point segment are calculated segment by segment based on the predicted speed and lane-changing behavior of each control point in the lane layer path.

[0029] As a further preferred embodiment of the present invention, the target is close to the reward weight. Weight of penalties for illegal lane changes Directional offset penalty weight Goal-oriented lane-changing weight Forward-looking transportation incentive weighting Risk penalty weight Where d0 is the characteristic attenuation distance, taken as 50m; γ1 is the enhancement coefficient, taken as 2.0; v maxThe road speed limit is set at 80 km / h; γ2 is the speed sensitivity coefficient, set at 1.5; (x) + =max(x,0); d ref The reference distance for lane-changing decisions is set to 200m; γ4 is the lane-changing guidance enhancement coefficient, set to 1.0. γ5 is the local predicted average speed; γ6 is the forward traffic sensitivity coefficient, taken as 1.2; γ7 is the speed risk coupling coefficient, taken as 1.0; δ is the risk weight gain, taken as 2.0; ρ risk ∈[0,1], which is the normalized risk index.

[0030] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the lane-level navigation method described above.

[0031] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the lane-level navigation method described above.

[0032] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0033] (1) This invention introduces a control point-based traffic flow prediction mechanism in the lane-level path planning process. By constructing a sequence of control points along the lane centerline and performing time-series predictions of speed and flow at each control point, fine-grained future traffic state information with spatiotemporal continuity is obtained. Compared to traditional path planning methods based on road segment averages or real-time states, control point-level prediction can more accurately capture micro-level traffic dynamics and local congestion trends, providing high-efficiency and high-resolution support for intelligent agents in lane selection and lane-changing decisions. This method effectively improves the accuracy, efficiency, and safety of path planning, and is particularly suitable for fine-grained navigation tasks in complex urban road network environments.

[0034] (2) This invention proposes a hierarchical path planning method that integrates road and lane layers. By constructing a map model that links road and lane levels, it achieves top-down path guidance and a bottom-up traffic feedback mechanism. During path execution, the lane layer completes path decisions based on detailed traffic conditions and feeds back travel time to the road layer, triggering dynamic updates to the upper-level path planning. This feedback iteration mechanism can dynamically adjust the overall path according to real-time traffic changes, improving the flexibility, stability, and global optimality of path planning, and overcoming the problems of static paths and delayed adjustments in existing technologies.

[0035] (3) The road layer of this invention is based on the improved A* search algorithm for path planning. This design can realize the linkage between road layer and lane level path planning, which is conducive to dynamically adjusting the overall path according to traffic changes.

[0036] (4) This invention constructs an intelligent agent decision-making framework (deep Q network) that integrates dynamic context perception and forward-looking risk assessment, which realizes the optimal path search of vehicles in the control point-level network in complex traffic environments, and has the advantages of safety, efficiency and lane-level path selection compliance.

[0037] (5) When constructing a deep Q-network, this invention designs a dynamic weighted, multi-objective coupled composite reward function that not only considers the immediate consequences of behavior, but also integrates traffic situation prediction and potential risk assessment, thereby achieving synergistic optimization of efficiency, safety and rule compliance.

[0038] (6) The lane-level navigation method provided by the present invention can accurately identify the specific lane position of the vehicle, and combined with high-precision maps, real-time traffic conditions and lane attribute information, provide more refined route guidance services, including recommending lane changing times, guiding early entry into turning lanes, and avoiding congested lanes. This plays a key role in improving driving safety, alleviating traffic congestion and optimizing travel experience.

[0039] (7) This invention constructs a hierarchical lane-level map model that integrates future traffic flow information. Through the collaborative expression of road layer, connection layer and lane layer, it achieves seamless connection between macro-path planning and micro-lane decision-making. Furthermore, it embeds dynamic traffic prediction and lane rules at the control point granularity, which significantly improves the foresight, safety and execution accuracy of path planning.

[0040] (8) This invention embeds traffic flow prediction depth into the map data structure itself, making control points the basic units that carry dynamic information. There is no need to call external prediction modules during planning, which significantly reduces system coupling complexity and improves real-time response capability.

[0041] (9) The reinforcement learning agent of the present invention uses the control point network as the basis of the action space, supports continuous path generation and local replanning. Compared with the action design based on discrete lane ID or grid, it has stronger generalization ability and environmental adaptability, and is especially suitable for urban scenarios with complex topology.

[0042] (10) When estimating the remaining travel time of a road segment, this invention abandons the traditional coarse-grained integral method based on road-level average speed or fitting function, and instead uses fine-grained predicted speed and lane-changing decision information of lane layer control point sequence to accurately accumulate the travel time of straight segments and lane-changing delay segment by segment, which significantly improves the accuracy and timeliness of the remaining travel time prediction and provides a high-fidelity feedback basis for dynamic replanning of the road layer.

[0043] (11) The present invention implements a dynamic rolling update mechanism with control points as the granularity during vehicle driving. Whenever the vehicle reaches any control point, the latest measured driving data and traffic prediction information are integrated in real time, the remaining path travel time is corrected in real time and road layer replanning is triggered, which significantly improves the system's response speed and path adaptability to sudden traffic events (such as accidents and sudden congestion), and overcomes the lag defect of traditional navigation systems that only update the path at fixed nodes or after long time intervals.

[0044] (12) This invention introduces a lane-change restricted area identification and path locking mechanism during path execution: when a vehicle approaches a solid lane line area or is less than the preset minimum safe distance for lane changing (e.g., 50 meters), the system determines that it has entered a lane-change restricted area, immediately freezes the current road segment's travel time feedback and replanning operation, and locks the existing lane-level path until it enters the next road segment. This mechanism effectively avoids repeated path switching caused by minor prediction fluctuations in the lane-change critical area, significantly improving driving stability and ride comfort. Attached Figure Description

[0045] Figure 1 This is a flowchart of the lane-level path planning method of the present invention;

[0046] Figure 2 The hierarchical map model structure diagram for this invention that integrates future traffic flow information;

[0047] Figure 3 This is a schematic diagram of lane-level hierarchical path planning according to the present invention;

[0048] Figure 4 This is a flowchart of the decision-making process based on Deep Q-Network (DQN) in the lane-level path planning module of this invention. Detailed Implementation

[0049] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0050] like Figures 1 to 4 As shown in the figure, this embodiment provides a lane-level navigation method based on lane traffic state recognition, which includes the following steps:

[0051] Step S1. Obtain the road structure of the entire road network through a high-precision map and collect historical lane-level traffic flow information; wherein, the road structure includes the road network, lane road topology, and lane network;

[0052] Step S2. Construct a layered map model (layered road network model), which includes a road layer map model, a road-lane connection layer map model, and a lane layer map model.

[0053] like Figure 2 As shown, in this embodiment, the road layer map model is constructed in the form of a directed graph, where nodes are road intersections and edges are road segments with attributes such as direction, length, speed, and flow rate, defined as W. a =(R a C a ); where R a =(P a E a Let P be the set of road segments. a E a Let C be the set of inbound and outbound endpoints. a It is a set of intersections, and each road segment has the starting and ending intersections and the speed and traffic forecast information for future times.

[0054] The road-lane connection layer map model is used to map road-level paths to a set of drivable lanes, establishing road-level entry and exit endpoints (P). a E a With lane level entry / exit endpoints (P) l E l The structural relationships between them.

[0055] The lane-layer map model describes each lane and its control points with a finer granularity, defined as W. l =(R l C l ); where R l For the set of lane-level road segments, C l This is a set of lane-level intersections, where each road segment contains several lanes, and each lane is controlled by a series of control points s. b The control points are used to characterize key locations on the lane centerline. They integrate future speed and traffic volume information from the traffic flow prediction model and include attributes such as lane change rules, left and right lane line types, and lane numbers.

[0056] Step S3. Obtain the time t when the vehicle V is running the lane-level navigation system through the real-time clock provided by the vehicle's local system or the cloud control platform. The cloud control platform calls the lane-level traffic flow parameter set G′ from the previous m time points. s =[g s (t-m+1),…,g s [t] is used as input. Let t be the velocity and flow rate observations at the control point. Represents observation speed, Representing the observed traffic flow, and combining the current location and the prediction time period, a lane-level traffic flow prediction model (which can use an existing model) is used to predict future traffic conditions. The output of the lane-level traffic flow prediction model is the set of traffic flow parameters G for each control point at the next k time points.s =[g s (t+1),…,g s (t+k)].

[0057] Step S4. Obtain the vehicle's current geographic coordinates (X) in real time through the vehicle positioning module. V =(x v ,y v The user's target location coordinates are X. D =(x d ,y d Based on the current location coordinates and the target location coordinates, the coordinates are mapped to the starting node c1 and the target node c5 of the road layer through map matching.

[0058] Step S5. Obtain the initial road layer path using the improved A* search algorithm, specifically including the following steps:

[0059] Step S5-1. For any road segment, calculate the average speed of each lane based on the predicted speed of each lane control point, and determine the overall average speed of the road segment using a weighted average method with the predicted traffic flow of the corresponding lane as the weight. The specific calculation formula is as follows:

[0060] Formula 1:

[0061] Formula 2:

[0062] Formula 3:

[0063] in, Let i be the average speed of the i-th lane. Let m be the predicted speed of the j-th control point in the i-th lane at time t. i ω represents the total number of control points for the i-th lane. i (t) represents the speed weight of the i-th lane during the conversion, calculated using the average predicted traffic flow of that lane. Let be the predicted traffic flow value at the j-th control point of the i-th lane at time t. Let t be the predicted road speed.

[0064] Step S5-2. Fit the discrete speed prediction values ​​of each road segment to obtain the fitting curve of the predicted speed of each road segment with respect to time (predicted speed-time fitting curve);

[0065] Step S5-3. Run the improved A* search algorithm on the constructed road layer map model with the goal of minimizing the cumulative travel time. The improved A* algorithm records the current cumulative travel time each time a node is expanded, and uses this time as the starting time t1 for entering the new road segment. The algorithm integrates the predicted speed-time fitting curve of the new road segment starting from t1. The time corresponding to the length L of the new road segment is the time t2 when the vehicle leaves the road segment. The time t2 plus the time to pass through the node (intersection) is the current cumulative travel time of the new road segment.

[0066] In this embodiment, the time to pass through the intersection is dynamically determined based on the traffic mode (straight, left turn, right turn) and the traffic light status obtained by the roadside equipment.

[0067] After searching with the improved A* algorithm, the output intersection sequence (c1, c2, c3, c4, c5) is the optimal intersection sequence at the road level, representing the current planned path from the start point to the end point. Here, road segments (c1, c2) represent a specific travel path between two adjacent intersections (e.g., ...). Figure 3 (As shown).

[0068] It should be noted that during the initial road layer path planning, the current cumulative travel time is mainly determined based on the road speed prediction value and the road segment length. After the path planning is carried out at the lane layer, the current cumulative travel time needs to be updated accordingly, and the road layer path planning needs to be redone based on the more accurate current cumulative travel time.

[0069] Step S6. To achieve more refined route navigation, each road segment in the road layer needs to be mapped to the lane layer (lane-level path), which specifically includes the following steps:

[0070] Step S6-1. Based on the pre-built layered map model, query the set of all feasible lanes {l1,l2,l3} corresponding to the road segment (c1,c2) ​​in the road-lane connection layer.

[0071] Step S6-2. Determine the set of lane exit endpoints E of intersection c1. c,1 =(E l,1,1 E l,1,2 E l,1,3 ), and the set of lane entry endpoints P of intersection c2. c,2 =(P l,2,1 ,P l,2,2 ,P l,2,3 ).

[0072] Step S6-3. Based on the topological connection relationship of intersection c2, select lane l1 that meets the left turn requirement of the target driving direction;

[0073] Step S6-4. Locate the starting control point for path planning. With the endpoint control point As the initial input for lane-level path planning.

[0074] Step S7. After mapping the road segments (c1, c2) in the road layer to the lane layer in step S6, the mapped start control point and end control point are used as input to start the lane layer path planning module. The lane layer path planning module is based on a deep reinforcement learning framework and preferably uses a deep Q network (DQN) to search for the optimal path of the vehicle in the control point level network. The goal is to minimize the cumulative travel time from the current control point to the target control point.

[0075] like Figure 4 As shown, in this embodiment, the network structure of the DQN includes:

[0076] The state input layer is used to receive current environmental state information, which includes at least the current vehicle position (current control point), target position (target control point), current vehicle speed, relative distance to surrounding obstacles, straight-line distance to the target, traffic state prediction values ​​of adjacent control points, traffic sign status, left and right lane alignment, and remaining distance.

[0077] The hidden layer, employing a fully connected neural network structure, is used for high-dimensional feature extraction and nonlinear mapping of input state information;

[0078] The action output layer is used to output the state-action value (Q value) corresponding to all optional actions;

[0079] The target network and experience replay mechanism employ a dual-network structure (main network and target network), and introduce an experience replay buffer to store the quadruple (S) during the agent's interaction with the environment. t ,a t ,R t ,S t+1 This is used for random sampling during the training phase to break data correlation and improve model stability and convergence efficiency.

[0080] Action selection strategy: Employ the ε-greedy strategy during training;

[0081] Loss function: The mean squared error (MSE) of the state-action value Q is used as the loss function to approximate the optimal policy Q*.

[0082] The lane-level path planning reinforcement learning method proposed in this example focuses on constructing an intelligent agent decision-making framework that integrates dynamic context awareness and forward-looking risk assessment to achieve safe, efficient, and compliant lane-level path selection in complex traffic environments. Its main elements are defined as follows:

[0083] (1) State-space design: The state S that the agent is in at each moment. t Acquired by a multi-source perception system deployed on the roadside or lane-level cloud control platform, defined as:

[0084] Formula 4: S t ={s cur ,s goal ,d dem ,v cur G pred ,b left ,b right ,b forbid ,ρ risk}

[0085] s cur : Current control point number, which indicates the vehicle's position within the lane grid on the lane layer map;

[0086] s goal : Current stage target control point number;

[0087] d dem The remaining straight-line distance from the target control point is used to determine whether to enter the lane change decision area.

[0088] v cur The current vehicle speed can be obtained from onboard equipment or by integrating roadside detectors;

[0089] G pred The predicted future traffic conditions for the current control point and its left and right adjacent control points, including speed and flow, are used to assess local traffic trends.

[0090] b left b right : Boolean value, indicating whether a lane change to the left or right is permitted, determined based on lane topology and solid line area;

[0091] b forbid : A Boolean indicator indicating whether the area is a no-lane-change zone, determined by the map's preset rules (such as ramps, tunnels, and solid line segments).

[0092] ρ risk The Potential Conflict Risk Index is based on the current position, speed, and acceleration information of surrounding traffic participants obtained by roadside units or vehicle-mounted sensing systems. It uses a short-term motion extrapolation model (such as the constant speed assumption) to estimate the minimum time to collision (TTC) or lateral safety margin within a preset future period, and quantifies the potential risk of the current lane-changing behavior accordingly.

[0093] (2) Action Space Design: Each time, the agent can choose from a set of actions A = {a0, a1, a2} = {go straight, change lanes to the left, change lanes to the right}. The feasibility of each action needs to be determined by considering the lane structure at the current location, traffic rule constraints of surrounding control points (such as whether it is in a no-lane-change zone), and road boundary conditions. During training and inference, infeasible actions will be disabled or given negative rewards to guide the agent to form a reasonable lane-changing strategy.

[0094] (3) Reward Function Design: The reward function quantifies the decision-making process, i.e., estimating the reward after performing an action in a given state. This embodiment designs a dynamically weighted, multi-objective coupled composite reward function, which not only considers the immediate consequences of actions but also integrates traffic situation prediction and potential risk assessment to achieve synergistic optimization of efficiency, safety, and rule compliance. This function is calculated after each action is performed, providing positive incentives or penalties for the agent's current behavior, and is formally expressed as follows:

[0095] Formula 5: R t =ω1(s t )R1+ω2(s t )R2+ω3(s t )R3+ω4(s t )R4+ω5(s t )R5+ω6(s t R6

[0096] R1: Reward for approaching or reaching the target control point. When the agent successfully approaches or reaches the target control point of the current stage, a large positive reward (+100) is given to reinforce the path completion behavior.

[0097] R2: Penalty for illegal lane change. If an agent attempts to change lanes in a prohibited lane-changing zone, a significant penalty (-20) is immediately imposed to ensure that it complies with traffic rules.

[0098] R3: Target direction deviation penalty. If the selected lane direction is inconsistent with the target control point direction, a moderate penalty (-2) is applied to encourage movement toward the target area.

[0099] R4: Goal-oriented lane change reward / penalty. Each time a lane change is executed, the system determines whether the target lane is closer to the current stage's target control point (s). goal Optimal lane sequence:

[0100] If the lane you are in after changing lanes is in the same direction as the lane where the target control point is located or is located on the recommended path lane, you will be given a positive reward (+1).

[0101] If the lane change deviates from the target direction or enters a non-recommended lane, a lane change penalty (-2) will be applied;

[0102] If redundant lane changes are performed without the need for lane changes (such as repeatedly hopping left and right in the target lane), a frequent lane change penalty (-1) is applied.

[0103] R5: Traffic condition reward at the next control point. The smoother the traffic conditions at the next control point along the agent's chosen path, the higher the reward (0 to +2), reflecting the agent's ability to predict future traffic conditions. This reward is calculated proportionally based on the pre-predicted lane speeds.

[0104] R6: Potential Conflict Risk Penalty. Based on the calculated probability of conflict with other traffic participants in the near future (e.g., 2 seconds), if the risk exceeds a threshold, a dynamic penalty (-5 to -10) is applied to achieve forward-looking safety constraints.

[0105] The dynamic weight ω i (s t (i = 1, ..., 6), based on the current state s t The traffic situation, mission phase, and safety risk level are calculated according to the following rules:

[0106] (1) Target proximity reward weight ω1(s) t )

[0107] Formula 6:

[0108] Where, d dem d0 is the remaining straight-line distance to the target control point; d0 is the characteristic attenuation distance, taken as 50m; γ1 is the enhancement coefficient, taken as 2.0.

[0109] (2) Weight of illegal lane change penalty ω2(s) t )

[0110] Formula 7:

[0111] Among them, v max The road speed limit is set at 80 km / h; γ2 is the speed sensitivity coefficient, set at 1.5.

[0112] (3) Directional offset penalty weight ω3(s) t )

[0113] Formula 8:

[0114] (4) Goal-oriented lane-changing weight ω4(s) t )

[0115] Formula 9:

[0116] Where, (x) + =max(x,0); dref The reference distance for lane-changing decisions is set to 200m; γ4 is the lane-changing guidance enhancement coefficient, set to 1.0.

[0117] (5) Weighting of forward-looking transportation incentives ω5(s) t )

[0118] Formula 10:

[0119] in, That is, the local predicted average speed; γ5 is the forward traffic sensitivity coefficient, which is taken as 1.2.

[0120] (6) Risk penalty weight ω6(s) t )

[0121] Formula 11:

[0122] Where γ6 is the velocity-risk coupling coefficient, taken as 1.0; δ is the risk weight gain, taken as 2.0; ρ risk ∈[0,1], which is the normalized risk index.

[0123] Finally, the reinforcement learning model outputs the optimal control point sequence for the lane layer, which is an optimal path sequence starting from the current control point, passing through several legal path control points, and finally reaching the target control point for the current stage.

[0124] Step S8. After completing the path planning for lane layer (c1, c2), vehicle V starts from the starting control point of this path. Starting from the designated point, the vehicle proceeds sequentially through each control point along the route. Each time the vehicle reaches the i-th control point on the route... At that time, perform the following operations:

[0125] 1. Lane-level travel time acquisition

[0126] Based on the vehicle's high-precision positioning and timestamp, the vehicle's V-axis is directly obtained. arrive Actual travel time

[0127] 2. Estimated travel time for the remaining section of the road

[0128] At the control point Based on the latest traffic flow forecast at the current moment (including control point speeds, lane-changing feasibility, etc.), the estimated flow rate from... To the final control point Lane-level predicted travel time The calculation method is as follows:

[0129] (1) For two adjacent control points in the straight section within the interval The passage time is:

[0130] Formula 12:

[0131] In the formula: v j Current control point The predicted velocity value, d j Control points To the next control point The length.

[0132] (2) If the journey from the current control point to the next control point requires crossing into an adjacent lane (i.e., a lane change), an additional lane change delay should be considered. This additional time delay includes two parts: the acceleration and deceleration time due to speed differences, and the time required to laterally cross the lane width. The specific calculation is as follows:

[0133] Changing lanes from a low-speed lane to a high-speed lane (V i+1 >V i Lane change delay time Δt lca for:

[0134] Formula 13:

[0135] Changing lanes from the high-speed lane to the low-speed lane (V i+1 ≤V i Lane change delay time Δt lca for:

[0136] Formula 14:

[0137] In the formula: V i+1 For the predicted speed information within the target control point range for lane changing, V i For the predicted velocity information within the current control point range, a r The acceleration (deceleration) values ​​of the vehicle during lane changing are obtained by the onboard detection system. w This represents the lane width and reflects the time required for vehicles to move laterally.

[0138] Therefore, from arrive The formula for calculating lane-level travel time is:

[0139] Formula 15:

[0140] Where: δ j =0,1, when δ j =1 indicates that a vehicle has changed lanes in this area, when δ j =0 indicates that no lane change occurred.

[0141] 3. Feedback the updated total travel time for the current road segment to the road layer:

[0142] Formula 8:

[0143] After receiving feedback, the road layer updates the travel time of the current road path in real time, and re-plans the path using an improved A* search algorithm starting from c2. If the optimal path is still the original path sequence (c1,c2,c3,…), the lane layer path continues to be executed; if the path is updated to (c1,c2,c6,…), it returns to step S6 to relocate the starting and ending control points of the lane layer path planning, and performs lane layer path planning for the remaining control points to ensure that the path planning result is consistent with the global objective.

[0144] In this embodiment, to avoid unnecessary path jitter caused by frequent feedback, when a vehicle approaches the solid line area of ​​the lane or the distance from the set minimum safe distance for lane changing is less than the threshold (50m), the system determines that it has entered the lane changing restricted area. At this time, the dynamic update of the current road segment's passage time is immediately stopped, and the current path is locked until the vehicle enters the next road segment (c2, c3).

[0145] Step S9. After vehicle V completes lane-level path planning and travel time feedback for the current road segment, the system will automatically enter the planning process for the next road segment until it reaches the end of the path. Specifically, steps S6 to S8 are repeated, that is, on each new road segment, lane start-end point mapping, lane-level path planning (DQN), travel time calculation and feedback are completed sequentially; if the path feedback does not cause a change in the road-level path structure, it continues to move forward; if the feedback causes the optimal road-level path to be updated, it immediately returns to step S6; the above process terminates when vehicle V reaches the final target intersection c5, and the system executes it in a loop to complete the entire path planning and traffic control task.

[0146] The high-precision map described in this embodiment refers to high-resolution map data relative to ordinary navigation maps. The high-precision map comprises two layers: a static high-precision map and a dynamic high-precision map. The static high-precision map is at the bottom layer, typically consisting of three types of vector information: lane models containing semantic information, road objects, and road attributes, as well as feature layers for multi-sensor positioning. The dynamic high-precision map is built upon the static high-precision map and mainly includes real-time dynamic information, including information on other traffic participants (such as road congestion, construction, traffic accidents, traffic control, and weather conditions) and information on traffic objects (such as traffic lights and pedestrian crossings).

[0147] The cloud control platform utilizes next-generation information and communication technologies to integrate dynamic basic data such as vehicle operation, infrastructure, traffic environment, and traffic management. It has basic service mechanisms such as high-performance information sharing, high real-time cloud computing, big data analysis, and information security. In this invention, it supports road traffic condition prediction and lane-level path planning.

[0148] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the lane-level navigation method described above.

[0149] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the lane-level navigation method described above.

[0150] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0151] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0152] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection described in the claims.

Claims

1. A lane-level navigation method based on lane traffic state recognition, characterized by, The method comprises the following steps: Step S1. Obtain the road structure of the entire road network through a high-precision map, and collect historical lane-level traffic flow information; Step S2. Construct a hierarchical map model, wherein the hierarchical map model comprises a road layer map model, a road-lane connection layer map model, and a lane layer map model; Step S3. Run a lane-level navigation system, and predict a future traffic state according to historical lane-level traffic flow; Step S4. Obtain positioning information; Step S5. Based on the current position of the vehicle, the target position, and the predicted future traffic state, an improved A* search algorithm is used to obtain a planned path of the initial road layer; the improved A* algorithm takes minimizing the cumulative travel time as the goal, and records the current cumulative travel time at each time of expanding a node, and takes this time as the starting time t1 of entering a new road section; the time corresponding to the integral value of the length L of the new road section on the predicted speed-time fitting curve starting from t1 is the time t2 at which the vehicle leaves the road section, and t2 plus the time of passing through the node is the current cumulative travel time of the new road section; Step S6. Based on the constructed hierarchical map model, each road section of the road layer is mapped to the lane layer, and the starting control point and the ending control point of the lane-level path planning are determined; Step S7. Start a lane layer path planning module, wherein the lane layer path planning module realizes optimal path search of the vehicle in the control point level network based on a deep reinforcement learning framework; Step S8. Based on the lane layer path planning, the vehicle starts from the starting control point of the path, and drives through each control point in the path in turn. When the vehicle reaches the i-th control point in the path , the time consumption of all traffic segments is obtained, the remaining traffic time of the road segment is estimated, the total traffic time of the current road segment is calculated, and the updated total traffic time of the current road segment is fed back to the road layer. The road layer re-plans the path using the improved A* search algorithm. If the optimal path is still the original path sequence, the lane layer path continues to be executed. If the path is updated, the starting control point and the ending control point of the lane layer path planning are repositioned, and the lane layer path planning of the remaining control points is performed. Step S9. After completing the lane layer path planning and travel time feedback of the current road section, automatically enter the planning process of the next road section, and repeatedly execute steps S6 to S8 until reaching the path ending point.

2. The lane-level navigation method based on lane traffic state recognition according to claim 1, characterized in that, The road layer map model is constructed in the form of a directed graph, wherein the nodes are road intersections, and the edges are road sections with direction, length, speed, and flow; the road-lane connection layer map model is used to map the road-level path to a drivable lane set; the lane layer map model is used to describe each lane and its control point information, wherein the control point fuses the future speed and flow information output by the traffic flow prediction model, and contains lane changing rules, left and right lane line types, and lane numbers.

3. The lane-level navigation method based on lane traffic state recognition according to claim 1, wherein, Step S4 obtains the current position coordinates and the target position coordinates of the vehicle in real time, and maps the above coordinates to the starting node and the target node of the road layer through map matching according to the current position coordinates and the target position coordinates.

4. The lane-level navigation method based on lane traffic state recognition according to claim 1, wherein, Step S5 comprises the following steps: Step S5-1. For any road section, calculate the average speed of each lane based on the predicted speed of each lane control point belonging to the road section, and determine the overall average travel speed of the road section by weighted average method with the predicted flow of the corresponding lane as the weight; Step S5-2. Fit the discrete speed prediction values of each road section to obtain the fitting curve of the predicted speed-time of each road section; Step S5-3. Running the improved A* search algorithm on the constructed road layer map model to minimize the cumulative travel time; the improved A* algorithm records the current cumulative travel time each time a node is expanded, and uses this time as the start time t1 of entering the new road segment, and then integrates on the predicted speed-time fitting curve of the new road segment from t1, and the integral value is the length L of the new road segment, and the time corresponding to the integral value is the time t2 of the vehicle leaving the road segment, and t2 plus the time through the node is the current cumulative travel time of the new road segment.

5. The lane-level navigation method based on lane traffic state recognition according to claim 1, wherein, Step S6 queries the entire set of drivable lanes corresponding to the road layer road segment based on the constructed hierarchical map model, and determines the out-point set and the in-point set of the two intersections of the road segment, and then screens the lanes that meet the target driving direction requirement according to the topological connection relationship of the intersections, so as to determine the start control point and the end control point of the lane layer path planning.

6. The lane-level navigation method based on lane traffic state recognition according to claim 1, wherein, The deep reinforcement learning framework described in step S7 adopts a deep Q network, state S t = {s cur ,s goal ,d dem ,v cur ,G pred ,b left ,b right ,b forbid ,ρ risk}; wherein s cur is the current control point number, s goal is the current stage target control point number, d dem is the remaining straight line distance from the target control point, v cur is the current vehicle speed, G pred is the future traffic state prediction value of the current control point and the left and right adjacent control points, b left , b right : Boolean value, indicating whether the current can left / right lane change, b forbid is a Boolean identifier indicating whether it is in a lane change prohibited area, and ρ risk is a potential conflict risk index; The set of actions A = {a0, a1, a2} = {straight, left lane change, right lane change} is executed. Reward function R t = ω1(s t )R1+ ω2(s t )R2+ ω3(s t )R3+ ω4(s t )R4+ ω5(s t )R5+ ω6(s t )R6 where R1is a reward for approaching or reaching a target control point, R2is a penalty for illegal lane change, R3is a penalty for target direction deviation, R4is a reward / penalty for target-oriented lane change, R5is a reward for traffic state of the next control point, R6is a penalty for potential conflict risk, ω i (s t ) are dynamic weights.

7. The lane-level navigation method based on lane traffic state recognition according to claim 1, wherein, The step S8 calculates the straight travel time and lane change delay of the subsequent control point segments based on the predicted speed and lane change behavior of each control point in the lane layer path.

8. The lane-level navigation method based on lane traffic state recognition according to claim 6, wherein, Target approach reward weight Illegal lane change penalty weight Direction offset penalty weight Target-oriented lane change weight Forward traffic reward weight Risk penalty weight where d0 is a feature attenuation distance, taking 50 m; γ1 is an enhancement coefficient, taking 2.0; v max is a road speed limit, taking 80 km / h; γ2 is a speed-sensitive coefficient, taking 1.5; (x) + = max(x, 0); d ref is a lane change decision reference distance, taking a value of 200 m; γ4 is a lane change guidance enhancement coefficient, taking 1.0; is a local prediction average speed; γ5 is a forward traffic-sensitive coefficient, taking 1.2; γ6 is a speed risk coupling coefficient, taking 1.0; δ is a risk weight gain, taking 2.0; ρ risk ∈ [0, 1], is a normalized risk index.

9. An electronic device comprising: One or more processors, memories; characterized in that the memory is used to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the lane level navigation method based on lane traffic state recognition according to any one of claims 1 to 8.

10. A computer readable medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the lane level navigation method based on lane traffic state recognition according to any one of claims 1 to 8.

Images