Obstacle avoidance and planning cooperative path generation method in urban complex environment

Abstract

The invention relates to the technical field of intelligent driving, in particular to a method for generating an obstacle avoidance and planning cooperative path in an urban complex environment, which comprises the following steps: acquiring environment real-time sensing data through a vehicle-mounted multi-sensor, and constructing a dynamic semantic traffic matrix in combination with high-precision map static semantic information; based on the matrix, a multi-objective optimization algorithm is adopted to calculate the security cost, the efficiency cost and the rule conformity cost of the path, and a global optimization path is generated; inputting the global optimization path and the dynamic obstacle motion vector into an intention prediction model, generating dynamic obstacle future trajectory probability distribution and interactive intention classification, and further generating an avoidance strategy and adjusting a local path in real time; and inputting the adjusted local path into a kinematics model to carry out kinematics feasibility verification, and outputting an executable track or path re-planning. According to the method, the integration of dynamic environment understanding, path planning and obstacle avoidance strategies is realized, and the method is suitable for the path generation task of an automatic driving system in an urban complex traffic scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, in particular to a method for obstacle avoidance and cooperative path planning in complex urban environments. BACKGROUND

[0002] Under the background of the rapid development of intelligent driving systems, urban traffic environments are considered as the core challenge scenario for automatic driving application landing due to their dense road structure, complex traffic participant behavior and variable traffic rules. To achieve safe and efficient automatic driving path planning, the system needs to handle multiple sensor inputs, dynamic obstacle perception, complex traffic semantic recognition and real-time path generation tasks simultaneously. Current mainstream path planning methods are mostly based on high-precision maps and static obstacle modeling, ignoring the deep modeling of dynamic traffic semantics such as traffic lights, lane rules and driving behavior intentions, resulting in poor adaptability, inconsistent behavior or lack of feasibility guarantee of path generation results in dynamic traffic environments, making it difficult to meet the comprehensive requirements of path safety, legality and executability in complex urban scenarios.

[0003] The existing technology mainly has the following problems: first, the path generation process lacks collaborative modeling of traffic semantics, dynamic obstacle behavior and rule constraints, making it difficult to comprehensively express the "traffic intention field" of urban road conditions; second, obstacle avoidance strategies are generally based on low-order features such as Euclidean distance or speed difference, lacking interactive behavior prediction mechanisms, resulting in delayed avoidance actions or misjudgments; third, path outputs are often not subjected to detailed kinematic feasibility verification, posing a risk of "plannable but not executable", especially in scenarios where vehicle dynamic performance is limited or path complexity is high, which can easily lead to system execution failure or abnormal behavior. SUMMARY

[0004] The present application provides a method for obstacle avoidance and cooperative path planning in complex urban environments, which integrates dynamic semantics, intention prediction and kinematic verification mechanisms to achieve closed-loop optimization from perception to executable path output.

[0005] A method for obstacle avoidance and cooperative path planning in complex urban environments, comprising the following steps: S1: Obtain real-time perception data of the environment through vehicle-mounted multi-sensor, combine with high-precision map static semantic information, and construct a dynamic semantic traffic matrix; S2: Taking the dynamic semantic traffic matrix as input, use a multi-objective optimization algorithm to simultaneously calculate the safety cost, efficiency cost and rule compliance cost of the path, and output a globally optimized path; S3: generating a dynamic obstacle avoidance strategy through an intention prediction model based on the global optimization path and the motion vector in the dynamic semantic passable matrix, adjusting the global optimization path in real time, and outputting an adjusted local path; S4: inputting the adjusted local path into a kinematics model for feasibility verification, triggering S2 to generate a path again if the verification fails, and outputting a final executable trajectory.

[0006] Optionally, the dynamic semantic passable matrix comprises spatial grid attributes, dynamic obstacle motion vectors, and scene rule labels.

[0007] Optionally, S1 comprises: S11: collecting environment real-time perception data through a vehicle-mounted multi-sensor, and loading pre-stored high-precision map static semantic information; S12: constructing a spatial grid map based on the environment real-time perception data, calculating the passability probability, terrain elevation, and road adhesion coefficient of each grid, and generating spatial grid attributes; S13: fusing the environment real-time perception data and the spatial grid attributes, identifying dynamic obstacles through a multi-target tracking algorithm, and extracting the position, speed, and heading angle parameters of the dynamic obstacles to output dynamic obstacle motion vectors; S14: matching the high-precision map static semantic information with the real-time perceived scene semantics, analyzing traffic signs, lane lines, and signal light states through a rule engine, and generating scene rule labels; S15: integrating the spatial grid attributes, dynamic obstacle motion vectors, and scene rule labels to construct a dynamic semantic passable matrix, wherein the spatial grid attributes serve as the matrix base layer, the dynamic obstacle motion vectors serve as the vector superposition layer, and the scene rule labels serve as the semantic constraint layer.

[0008] Optionally, in the constructed dynamic semantic passable matrix, the spatial grid attributes serve as the matrix base layer, the dynamic obstacle motion vectors serve as the vector superposition layer, and the scene rule labels serve as the semantic constraint layer.

[0009] Optionally, S2 comprises: S21: calculating the Euclidean distance between the candidate path point and the nearest dynamic obstacle based on the spatial grid attributes and the dynamic obstacle motion vectors in the dynamic semantic passable matrix, and generating a safety cost; S22: calculating the total length of the candidate path and the estimated passable time through each grid based on the spatial grid attributes and the scene rule labels of the dynamic semantic passable matrix, and generating an efficiency cost; S23: detecting the candidate path and the lane center line offset, the traffic sign violation state, and the signal light waiting time based on the scene rule labels in the dynamic semantic passable matrix, and generating a rule compliance cost.

[0010] Optionally, the S2 further comprises: S24: input the security cost, efficiency cost and rule compliance cost into a multi-objective optimization algorithm, and generate an initial path set based on the dynamic semantic traffic matrix; S25: perform Pareto front iterative optimization on the initial path set by the multi-objective optimization algorithm, and output a globally optimized path meeting the security cost threshold, efficiency cost threshold and rule compliance cost threshold.

[0011] Optionally, the S3 comprises: S31: input the globally optimized path and the dynamic obstacle motion vector in the dynamic semantic traffic matrix into an intention prediction model, and generate a future trajectory probability distribution of the dynamic obstacle and an interactive intention classification; S32: based on the future trajectory probability distribution and the interactive intention classification output by the intention prediction model, calculate the spatiotemporal conflict points of the globally optimized path and the dynamic obstacle, and generate a dynamic obstacle avoidance strategy; S33: adjust the local waypoint coordinates and speed profile of the globally optimized path according to the dynamic obstacle avoidance strategy, and output an adjusted local path, so as to ensure that the minimum safety distance between the adjusted local path and the dynamic obstacle motion vector is greater than a set threshold.

[0012] Optionally, the intention prediction model comprises a spatiotemporal feature encoder, an interactive intention classification head, a trajectory prediction head and a conflict detection module.

[0013] Optionally, the S4 comprises: S41: discretize the real-time adjusted local path into a continuous trajectory point sequence, extract the position coordinates, speed values and heading angles of each trajectory point, and constitute an input pose queue of the kinematic model; S42: based on the state transition equation of the kinematic model, calculate the curvature continuity, acceleration change rate and front wheel steering angle change rate of adjacent trajectory points in the pose queue point by point, and perform feasibility verification; S43: if the feasibility verification is passed, output the final executable trajectory; if the feasibility verification fails, trigger S2 to regenerate the path and update the globally optimized path.

[0014] The present application has the following advantages: The application can comprehensively express multi-dimensional semantic information such as road structure, obstacle behavior and traffic rules in the urban traffic environment by constructing a dynamic semantic passing matrix that fuses spatial grid attributes, dynamic obstacle motion vectors and scene rule labels, thereby providing a refined input basis for subsequent path generation. Compared with the traditional path planning method based on only static maps or obstacle positions, the present scheme can perceive key factors such as traffic signal light states, lane attributes, road adhesion conditions and the like in real time, thereby generating a path that conforms to the actual passing logic and improving the path rationality and traffic rule compliance degree of the system in semantic-intensive scenes such as red-green light intersections and lane-changing areas.

[0015] The application introduces an intention prediction model based on graph convolution and LSTM fusion, which can model the interaction relationship between dynamic obstacle motion vectors and globally optimized paths, output the future trajectory probability distribution of the obstacle and the interaction intention classification label. By further generating a dynamic obstacle avoidance strategy and adjusting the local path in real time, the method not only can predict potential conflict behaviors (such as accelerating to overtake or emergency braking), but also can implement avoidance actions in advance. Compared with the traditional obstacle avoidance method based on geometric distance or speed difference judgment, the present application can significantly enhance the predictability and active adaptability of the system to complex traffic interaction behaviors, and is especially suitable for dynamic decision-making in multi-vehicle intersection and high-density driving scenes.

[0016] The application uses a bicycle model as the kinematic model, and combines the state transition equation of the pose state vector and the control input vector to verify the locally adjusted path point by point, thereby ensuring that the path meets the vehicle chassis motion constraints and physical feasibility in terms of curvature continuity, acceleration change rate and front wheel angle change rate. When the verification fails, the system automatically triggers S2 to perform path re-planning, thereby realizing the path generation and execution closed loop. This design effectively prevents the model mismatch risk between path planning and vehicle execution, improves the stability, safety and continuity of path execution, and is especially suitable for application in the environment where the automatic driving vehicle has extremely high requirements for path executability, such as urban narrow lanes and sharp intersection. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the application, and those skilled in the art can also obtain other drawings according to these drawings without any creative effort.

[0018] Fig. 1 The method flowchart of the embodiment of the application is shown in the figure. Fig. 2 The S2 flowchart of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0019] The application will be described in greater detail in connection with the accompanying drawings and specific embodiments. It should be noted that the embodiments described below are best modes, preferred embodiments, and other alternative modes can be implemented by those skilled in the art for some well-known technologies; and the accompanying drawings are only used to more specifically describe the embodiments and are not intended to specifically limit the application.

[0020] It should be noted that the terms "one embodiment", "an embodiment", "exemplary embodiment", "some embodiments", etc. in the specification indicate that the described embodiment can include a specific feature, structure or characteristic, but not necessarily every embodiment includes the specific feature, structure or characteristic. In addition, when a specific feature, structure or characteristic is described in connection with an embodiment, it should be within the knowledge of those skilled in the art to implement such a feature, structure or characteristic in combination with other embodiments (whether or not explicitly described).

[0021] Generally, the terms can be understood at least in part from the context in which they are used. For example, depending on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular or can be used to describe combinations of features, structures, or characteristics in the plural. In addition, the term "based on" can be understood as not necessarily intended to convey exclusivity in a set of factors, but rather, at least in part, depending on the context, to allow for the existence of other factors not necessarily explicitly described.

[0022] As shown in Figs. 1-2 A method for obstacle avoidance and planning coordination path generation in urban complex environment, comprising the following steps: S1: obtaining real-time perception data of the environment through a vehicle-mounted multi-sensor, combining static semantic information of a high-precision map to construct a dynamic semantic traffic matrix, specifically: S11: collecting real-time perception data of the environment through a vehicle-mounted multi-sensor, while loading pre-stored static semantic information of a high-precision map; First, the vehicle-mounted system calls multi-type sensor components for environment perception. The vehicle-mounted multi-sensor includes a laser radar (LiDAR), a millimeter wave radar, an RGB camera, an infrared camera, and an inertial measurement unit (IMU). These sensors collect real-time environment perception data around the vehicle through a time synchronization mechanism, including obstacle point cloud, image frame, radar reflection echo, and vehicle motion posture.

[0023] At the same time, the current vehicle position is obtained by GNSS and IMU joint positioning, and the static semantic information of the corresponding area is loaded from the high-precision map database, including road boundaries, lane line types, traffic sign and signal light positions and attributes. To ensure information alignment, the system accurately registers perception data and high-precision map data in space through coordinate system conversion algorithm.

[0024] S12: Construct a spatial grid map based on real-time environmental perception data, calculate the passability probability, terrain elevation and road adhesion coefficient of each grid, and generate spatial grid attributes; After the registration of environmental perception data is completed, the system defines a perception area (e.g. 60 meters forward, 20 meters left and right) centered on the vehicle, and divides it into spatial grid units at a fixed resolution (e.g. 0.5m x 0.5m). Each grid calculates the following attributes based on sensor data: Passability probability: input the obstacle point cloud density in the current grid, historical pass records and radar reflection echoes into a Bayesian model for evaluation, reflecting the degree of safety in passing this area; Terrain elevation: reconstruct the ground model through laser radar point cloud, calculate the average elevation and slope of the grid, and identify terrain features such as steps and slopes; Road adhesion coefficient: combined with the road surface material type (such as dry asphalt, wet and slippery tiles) identified by the camera image and the millimeter wave radar signal reflection intensity, the corresponding adhesion coefficient is obtained by table lookup.

[0025] The integration of the above three types of data forms the spatial grid attributes, which serve as the basic information layer for subsequent semantic matrix construction.

[0026] S13: Fuse real-time environmental perception data and spatial grid attributes, identify dynamic obstacles through multi-target tracking algorithm, and extract their position, speed and heading angle parameters, output dynamic obstacle motion vector; Based on the constructed spatial grid map, the system further identifies the moving targets existing at the current time. By using the fused image sequence, radar echo and point cloud data, through multi-target tracking algorithm (such as joint probability data association filter JPDAF), continuous tracking of multiple types of dynamic obstacles is achieved.

[0027] For each tracking target, the system calculates its position coordinates in the vehicle coordinate system, estimates the velocity vector using time series changes, and further derives the heading angle parameter by combining the moving direction. The three together constitute the motion vector of the dynamic obstacle, which is used for dynamic prediction and passing avoidance.

[0028] S14: Match high-precision map static semantic information with real-time perceived scene semantics, analyze traffic signs, lane lines and signal light states through rule engine, and generate scene rule labels; To express the traffic environment constraints more comprehensively, the system will compare and match the high-precision map static semantic information loaded in S11 with the image perception results in S13.

[0029] Image processing uses deep convolutional neural networks for target detection and semantic segmentation, identifying traffic sign categories, lane line boundary types, and signal light current states. During the matching process, if the map and real-time perception results are inconsistent (such as temporary signs caused by construction), the system will prefer to use real-time perception results to correct map errors.

[0030] Combined with the reasoning structure of the rule engine, the following scenario rule labels are further extracted: The type of the current lane and the lane change constraints; The speed limit of each road segment; The phase and remaining switching time of the signal light; Prohibited access area identification; Priority of main and auxiliary road access.

[0031] This set of labels will be used as rule criteria input in subsequent path evaluation.

[0032] S15: Integrating spatial grid attributes, dynamic obstacle motion vectors, and scenario rule labels to construct a dynamic semantic access matrix; After obtaining complete environmental perception results, the system integrates the three types of information into a unified spatial grid framework, constructing a dynamic semantic access matrix with a multi-layer information structure.

[0033] Basic layer (spatial grid attributes): Expressing the static accessibility, terrain information, and ground adhesion performance of each location; Vector superposition layer (dynamic obstacle motion vector): In the path prediction area, assign dynamic risk factors to grids that potentially conflict with dynamic obstacle trajectories; Semantic constraint layer (scenario rule label): Set accessibility penalties or directly set to zero for areas that violate traffic rules (such as red light control areas and prohibited access zones).

[0034] Through the superposition of the above three layers, a dynamic semantic access matrix reflecting the current spatio-temporal environment, dynamic risk, and traffic rule constraints is finally formed, serving as the core input basis for path cost evaluation and generation algorithms.

[0035] S2: Using the dynamic semantic access matrix as input, a multi-objective optimization algorithm is used to simultaneously calculate the safety cost, efficiency cost, and rule compliance cost of the path, outputting a globally optimized path, specifically: S21: Based on the spatial grid attributes and dynamic obstacle motion vectors in the dynamic semantic access matrix, calculate the Euclidean distance between candidate path points and the nearest dynamic obstacle to generate a safety cost; Firstly, the spatial grid attributes covered by the candidate path are extracted from the dynamic semantic traffic matrix, and the dynamic obstacle motion vectors of each dynamic obstacle are obtained synchronously. For each candidate path point, the minimum Euclidean distance from the path point to each dynamic obstacle is calculated by taking the predicted position of the dynamic obstacle at the current time as the reference point, and the distance is taken as the risk indicator of the path point.

[0036] The calculation is as follows: ; wherein, is the Euclidean distance between the th path point of the candidate path and the predicted position of the th dynamic obstacle. If the distance is less than the set dynamic buffer radius (determined according to the obstacle speed and the minimum deceleration of the vehicle), the corresponding path point is marked as a high-risk area, and the value of the path point will be increased by a fixed penalty factor.

[0037] The safety cost of the entire candidate path is calculated by the weighted average of the risk weights of all path points: ; wherein represents the risk weight of the th path point, is the number of path points.

[0038] S22: Based on the spatial grid attributes of the dynamic semantic traffic matrix and the scene rule label, the total length of the candidate path and the estimated passing time through each grid are calculated to generate the efficiency cost; The total length of the path is calculated by accumulating the geometric distance between the path points.

[0039] At the same time, in each grid, the maximum safe driving speed is calculated according to the adhesion coefficient and slope information in the spatial grid attributes, and the speed limit of the corresponding grid in the scene rule label is combined to take the smaller value as the upper limit of the passing speed of the current grid.

[0040] For the candidate path , the passing time is calculated as follows: ; wherein, is the distance of the th segment in the path, is the speed limit value of the grid (from the scene rule label), is the maximum safe driving speed derived according to the adhesion coefficient.

[0041] The efficiency cost is calculated by the normalized total length of the path and the passing time: ; wherein is a system preset weight coefficient, are length and time normalization upper limits, respectively.

[0042] S23: Based on the scene rule label in the dynamic semantic pass matrix, the candidate path and the lane center line offset, the traffic sign violation state, and the signal lamp waiting time are detected, and a rule compliance cost is generated; The path compliance is comprehensively evaluated by the following three indicators: 1. Lane center line offset: The shortest distance between each point of the candidate path and the lane center line in the map is calculated. If the offset exceeds the set threshold (such as 0.5m), the cost is accumulated in proportion; 2. Traffic sign violation state: If the path passes through the rule label area of no passing, no left turn, etc., it is judged as a violation, and a fixed penalty value is given; 3. Signal lamp waiting time: If the path passes through the red light signal area, the additional waiting time is estimated according to the remaining red light time and the passing speed, and the penalty coefficient is converted.

[0043] The rule compliance cost of the path is finally calculated: ; wherein each represents a rule deviation factor of a single path point.

[0044] S24: Input the safety cost, efficiency cost and rule compliance cost into the multi-objective optimization algorithm, and generate an initial path set based on the dynamic semantic pass matrix; Combine the three types of costs of all candidate paths as a three-dimensional vector Input into the multi-objective optimization module. The optimization module adopts a hierarchical heuristic search strategy, taking the space grid allowed to pass in the dynamic semantic pass matrix as the node, expanding the path layer by layer and constructing the initial path set.

[0045] The specific process is as follows: From the current vehicle pose, the search tree root node is constructed; In each iteration, all passable adjacent grids are expanded from the current node to form path branches; Each path branch calculates the value of the three types of costs, and records its cost vector; If the path passes through an impassable area or the cost is too high, it is pruned; Repeat the above steps until the target point or the maximum iteration depth, generate multiple candidate path branches, and form an initial path set.

[0046] S25: Perform Pareto front iterative optimization on the initial path set through a multi-objective optimization algorithm, and output the globally optimized path that meets the safety cost threshold, efficiency cost threshold, and rule compliance cost threshold

[0047] In the initial path set, the system performs a Pareto front sorting algorithm to construct a multi-objective non-dominated solution set. The algorithm considers each path as a three-dimensional vector point and determines whether it is "dominated" by other paths in all dimensions.

[0048] Paths that meet any of the following conditions will be eliminated: Safety cost is higher than the threshold ; Efficiency cost is higher than the threshold ; Rule compliance cost is higher than the threshold ; Finally, among the remaining paths, calculate the comprehensive score: ; Where is the strategy parameter weight, which can be adjusted according to different urban road scenes (e.g., increase in rainy weather, increase on holidays).

[0049] The path with the lowest comprehensive score is selected as the globally optimized path, which is used for subsequent dynamic yield and dynamic semantics verification module calls.

[0050] S3: Based on the globally optimized path and the motion vector in the dynamic semantic traffic matrix, generate a dynamic obstacle avoidance strategy through an intention prediction model, adjust the globally optimized path in real time, and output the adjusted local path, specifically: S31: Input the globally optimized path and the dynamic obstacle motion vector in the dynamic semantic traffic matrix into the intention prediction model to generate the future trajectory probability distribution and interaction intention classification of the dynamic obstacle; First, align the trajectory point sequence of the globally optimized path (i.e., the spatial coordinates and speed corresponding to each timestamp in the vehicle's planned path) with the dynamic obstacle motion vector (including the position, speed, and heading angle of each obstacle) in the dynamic semantic traffic matrix, and input them as joint inputs into the intention prediction model for dynamic behavior analysis.

[0051] The intention prediction model consists of the following four structural modules: 1. Spatio-temporal feature encoder: used to extract the spatial topological relationship and temporal motion features between the ego vehicle and dynamic obstacles. Its internal structure includes: Graph Convolutional Layer: Encode the adjacency relationship between dynamic obstacles as a graph structure, and extract the relative position relationship, speed relationship and spatial adjacency between obstacles. The adjacency graph is defined by a distance threshold, and each obstacle is a node.

[0052] LSTM layer (Long Short-Term Memory network): Time series modeling of the historical state sequence of each obstacle and the ego vehicle, extracting the motion trend of the obstacle in consecutive time slices, and outputting a time series representation vector of the obstacle.

[0053] The outputs of the above two modules are concatenated to form a fused spatio-temporal feature vector, which is used as input for the downstream module.

[0054] 2. Interactive intent classification head: receives the output of the spatio-temporal feature encoder, extracts high-dimensional semantic features using a set of dense layers (DenseLayer), and then connects a softmax activation function for interactive intent classification.

[0055] The output interactive intent classification label is at least one of the following categories: accelerate overtaking, decelerate and yield, lane keeping, and emergency braking.

[0056] Each obstacle will output an interactive intent probability distribution and a main classification label corresponding to the maximum probability, representing its current behavior trend relative to the ego vehicle.

[0057] 3. Trajectory prediction head: based on the same set of spatio-temporal feature vectors, fuse the ego vehicle path and obstacle motion trend information, and through a spatio-temporal attention mechanism (Spatio-temporal Attention Module), strengthen the behavioral relevance between targets.

[0058] Finally, output the position probability distribution of each obstacle at future time points (such as 1s, 2s, 3s) in the form of a Gaussian mixture model, i.e. its future trajectory probability distribution, including mean, covariance matrix and mixing weight.

[0059] Conflict detection module: map the interactive intent classification label and the future trajectory probability distribution to the grid coordinate system of the dynamic semantic traffic matrix.

[0060] Take the globally optimized path of the ego vehicle as the reference benchmark, generate a spatio-temporal conflict heat map based on the spatial grid attributes. The value of this heat map represents the probability of conflict between the ego vehicle path and the predicted obstacle trajectory at the corresponding time and spatial location, providing a basis for subsequent path adjustment.

[0061] S32: Based on the future trajectory probability distribution and interactive intent classification output by the intent prediction model, calculate the spatio-temporal conflict points between the globally optimized path and dynamic obstacles, and generate a dynamic obstacle avoidance strategy; According to the conflict probability value of each grid in the conflict heat map, the high-risk path point area is extracted, and all time and space conflict points in the global optimization path that overlap or are adjacent to the predicted trajectory of the obstacle are identified.

[0062] For each conflict point, the system generates a corresponding avoidance priority rule by combining the interaction intention classification label of the corresponding obstacle (e.g., if it is "accelerate overtaking", the ego vehicle should actively decelerate to avoid; if it is "decelerate and yield", it can pass according to the original path).

[0063] According to the risk level, obstacle speed, expected path intersection time difference, and buffer distance, a set of dynamic obstacle avoidance strategies including avoidance direction, speed adjustment mode, and timing window are constructed.

[0064] The strategy can include: Active deceleration or acceleration to stagger through before the intersection point; Adjust the path by shifting to the lateral safety area; Wait for the obstacle to pass before rejoining the main path.

[0065] The output format of the strategy is "adjustment action instruction + corresponding path point index + execution time window".

[0066] S33: Adjust the local path point coordinates and speed profile of the global optimization path according to the dynamic obstacle avoidance strategy, and output the real-time adjusted local path to ensure that the minimum safety distance between the adjusted path and the dynamic obstacle motion vector is greater than the set threshold; Apply the avoidance strategy to the global optimization path, and adjust the coordinates and speed in units of local path segments. The constraints during the adjustment process include: The minimum safety distance between each adjusted path point and any predicted trajectory of a dynamic obstacle must be greater than a set threshold (e.g., 2.5m), otherwise the avoidance strategy is re-evaluated; All adjustment actions must satisfy the kinematic feasibility constraints of the vehicle (such as minimum turning radius, maximum acceleration and deceleration, etc.); The adjustment range is controlled within a limited local segment (such as the path within the next 20m), ensuring that the system maintains high-frequency updates.

[0067] The output result is a set of local path points, including a new sequence of path coordinates and corresponding speed profiles (speed value at each point), as the real-time adjusted local path at the current time.

[0068] S4: Input the adjusted local path into the kinematic model for feasibility verification, if the verification fails, trigger S2 to regenerate the path, until the final executable trajectory is output, which is: S41: discretize the real-time adjusted local path into a continuous sequence of trajectory points, extract the position coordinates, velocity values and heading angles of each trajectory point to form the input pose queue of the kinematic model; Firstly, the system performs equidistant interpolation resampling on the real-time adjusted local path output by S3 to obtain a discrete sequence of trajectory points with continuous dynamic characteristics. Assuming that the sampling step is 0.1 seconds, the path is refined in the time domain, and the following information is recorded: Two-dimensional position coordinates of each trajectory point ; Corresponding instantaneous velocity ; Heading angle derived from the curvature direction of the path , which can be approximated by the difference between adjacent position points.

[0069] Finally, a pose queue arranged in time sequence is formed: ∈1,2,..., ; This queue serves as the state input of the downstream kinematic model.

[0070] S42: based on the state transition equation of the kinematic model, calculate the curvature continuity, acceleration change rate and front wheel steering angle change rate of adjacent trajectory points in the pose queue point by point, and perform feasibility verification; The kinematic model is a bicycle model, and trajectory feasibility verification is performed based on the bicycle model. For each pair of consecutive pose points and , the following method is used for calculation and judgment: I. Trajectory curvature continuity analysis: first calculate the curvature of each trajectory point, which is estimated using the discrete formula: ; Compare the difference between adjacent curvatures pair by pair, if there is a sudden change (i.e. the curvature change is greater than the continuity threshold , for example 0.3 ), the trajectory is marked as discontinuous in curvature and temporarily determined as infeasible.

[0071] II. Acceleration change rate analysis uses the velocity difference to approximate the acceleration: ; And further obtain the acceleration change rate: ; If or exceeds the system physical control limit (for example , ), the acceleration is considered to not meet the vehicle dynamics characteristics.

[0072] III. Front wheel steering rate analysis: combined with the target curvature and speed , the required front wheel steering angle is calculated according to the kinematic model as follows: ; Further calculate the steering rate: ; And check the constraint boundary conditions: ; If there is a violation, it is determined that the trajectory cannot be safely executed by the current vehicle chassis.

[0073] All the above calculation processes are continuously evaluated for each pair of trajectory points, and if all the verification items in the entire trajectory pass, the trajectory is considered feasible.

[0074] S43: If the feasibility verification passes, output the final executable trajectory; if the feasibility verification fails, trigger S2 to regenerate the path and update the global optimization path; The above verification results are used as logical condition judgments: If the curvature continuity, acceleration rate and front wheel steering rate are within the constraint boundary of the kinematic model, the current trajectory is marked as the final executable trajectory and is handed over to the control system for execution; If any of them exceeds the limit, it is considered an infeasible path, and S2 will be automatically called to re-execute the multi-objective path cost optimization based on the latest dynamic semantic traffic matrix, generate a new global optimization path, and update the input of subsequent modules.

[0075] This verification-re-generation mechanism ensures the robustness and real-time response capability of the entire path planning system in complex urban environments.

[0076] The present application encompasses any substitutions, modifications, equivalent methods and schemes made on the essence and scope of the present application. In order for the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0077] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A method for generating collaborative paths for obstacle avoidance and planning in complex urban environments, characterized in that, Includes the following steps: S1: Real-time environmental perception data is acquired through vehicle-mounted multi-sensor data, and combined with static semantic information from high-precision maps to construct a dynamic semantic access matrix; S2: Using the dynamic semantic access matrix as input, a multi-objective optimization algorithm is used to simultaneously calculate the security cost, efficiency cost, and rule compliance cost of the path, and output the globally optimized path; S3: Based on the global optimized path and the motion vectors in the dynamic semantic passage matrix, a dynamic obstacle avoidance strategy is generated through the intent prediction model, the global optimized path is adjusted in real time, and the adjusted local path is output. S4: Input the adjusted local path into the kinematic model for feasibility verification. If the verification fails, trigger S2 to regenerate the path until the final executable trajectory is output.

2. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 1, characterized in that, The dynamic semantic access matrix includes spatial grid attributes, dynamic obstacle motion vectors, and scene rule labels.

3. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 2, characterized in that, S1 includes: S11: Collects real-time environmental perception data through vehicle-mounted multi-sensor systems, while simultaneously loading pre-stored high-precision map static semantic information; S12: Construct a spatial raster map based on real-time environmental perception data, calculate the accessibility probability, terrain elevation, and road surface adhesion coefficient of each raster, and generate spatial raster attributes; S13: Integrates real-time environmental perception data with spatial grid attributes, identifies dynamic obstacles through a multi-target tracking algorithm, extracts their position, velocity, and heading angle parameters, and outputs the motion vector of the dynamic obstacle; S14: Match the static semantic information of the high-precision map with the semantic information of the scene perceived in real time, and generate scene rule labels by parsing traffic signs, lane lines and traffic light status through the rule engine; S15: Integrate spatial grid attributes, dynamic obstacle motion vectors, and scene rule tags to construct a dynamic semantic access matrix.

4. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 3, characterized in that, In the constructed dynamic semantic access matrix, spatial grid attributes serve as the matrix base layer, dynamic obstacle motion vectors serve as the vector overlay layer, and scene rule labels serve as the semantic constraint layer.

5. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 4, characterized in that, S2 includes: S21: Based on the spatial grid attributes and dynamic obstacle motion vectors in the dynamic semantic passage matrix, calculate the Euclidean distance between candidate path points and the nearest dynamic obstacle, and generate the safety cost. S22: Based on the spatial grid attributes and scene rule labels of the dynamic semantic passage matrix, calculate the total length of the candidate path and the estimated passage time through each grid, and generate the efficiency cost; S23: Based on the scene rule labels in the dynamic semantic traffic matrix, detect the offset of candidate paths from lane centerlines, traffic sign violation status, and traffic light waiting time, and generate rule compliance costs.

6. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 5, characterized in that, S2 further includes: S24: Input the security cost, efficiency cost, and rule compliance cost into the multi-objective optimization algorithm to generate an initial path set based on the dynamic semantic passage matrix; S25: The initial path set is iteratively optimized using a multi-objective optimization algorithm through the Pareto front, and the output is a globally optimized path that satisfies the safety cost threshold, efficiency cost threshold, and rule compliance cost threshold.

7. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 6, characterized in that, S3 includes: S31: Input the global optimized path and the dynamic obstacle motion vectors in the dynamic semantic passage matrix into the intent prediction model to generate the future trajectory probability distribution of dynamic obstacles and the interaction intent classification; S32: Based on the probability distribution of future trajectories and the classification of interactive intentions output by the intent prediction model, calculate the spatiotemporal conflict points between the global optimized path and dynamic obstacles, and generate dynamic obstacle avoidance strategies. S33: Adjust the local waypoint coordinates and velocity profile of the global optimized path according to the dynamic obstacle avoidance strategy, and output the adjusted local path to ensure that the minimum safe distance between the adjusted local path and the dynamic obstacle motion vector is greater than the set threshold.

8. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 7, characterized in that, The intent prediction model includes a spatiotemporal feature encoder, an interaction intent classification head, a trajectory prediction head, and a conflict detection module.

9. The method for generating collaborative paths for obstacle avoidance and planning in complex urban environments according to claim 8, characterized in that, S4 includes: S41: Discretize the real-time adjusted local path into a continuous sequence of trajectory points, extract the position coordinates, velocity values ​​and heading angles of each trajectory point, and form the input pose queue of the kinematic model; S42: Based on the state transition equation of the kinematic model, calculate the curvature continuity, acceleration rate of change and front wheel angle rate of change of adjacent trajectory points in the pose queue point by point, and perform feasibility verification. S43: If the feasibility verification passes, output the final executable trajectory; if the feasibility verification fails, trigger S2 to regenerate the path and update the global optimized path.

Images