An unmanned driving planning method and system for complex road conditions adaptive obstacle avoidance

Abstract

The application discloses an unmanned driving planning method and system for complex road condition adaptive obstacle avoidance, relates to the technical field of vehicle active safety obstacle avoidance, and comprises the following steps: fusing multi-source sensing data, performing time-space alignment and feature fusion, and constructing a hierarchical high-precision real-time environment model containing road structure, obstacles, vehicle state and traffic flow information; extracting multi-dimensional road condition features from the model, matching a feature database by using a weighted Euclidean distance, and intelligently identifying a road condition complexity level; performing dynamic and static classification and situation prediction on obstacles, evaluating a risk level according to a collision probability and a predicted collision time; automatically matching a planning strategy according to the road condition complexity level, generating a global path and a speed sequence, and dynamically generating an obstacle avoidance trajectory such as an offset, a speed reduction, an emergency braking or a lane changing bypass according to the risk level, and performing deviation regression fusion with the global path to form a final trajectory; and simultaneously, dynamically adjusting the update frequency of the environment model according to the road condition complexity level and the highest risk level of the obstacles.

Description

Technical Field

[0001] This invention relates to the field of active safety obstacle avoidance technology for vehicles, and in particular to an autonomous driving planning method and system for adaptive obstacle avoidance in complex road conditions. Background Technology

[0002] As autonomous driving technology is gradually deployed in public transportation, park transportation, and urban services, path planning and active obstacle avoidance have become core technologies for ensuring the safe and efficient operation of driverless vehicles. However, existing obstacle avoidance planning methods for autonomous driving generally suffer from the following shortcomings: limited road condition adaptability, mostly only designed for structured standard roads, and rigid planning that is difficult to adjust in real time when faced with unstructured and dynamic complex road conditions such as missing lanes, road construction, non-motorized vehicle crossings, and pedestrian crossings; low obstacle classification accuracy, making it difficult to accurately distinguish between dynamic and static obstacles and predict their movement trends, often leading to over-avoidance or delayed response, making it difficult to balance smoothness and safety; decoupling of global path planning and local obstacle avoidance, making it impossible to quickly return to the optimal global path after local avoidance, while the planning algorithm has a large computational load and poor real-time performance, making it prone to planning failure when faced with sudden road conditions; and insufficient scene adaptability and versatility, mostly customized for single scenes, making it difficult to achieve adaptive switching between multiple road conditions and resulting in poor cross-scene operation stability. The aforementioned shortcomings mean that existing technologies cannot meet the safe driving requirements under complex, dynamic, and unstructured real-world road conditions. There is an urgent need for an autonomous driving technology solution that has the ability to adapt to various complex road conditions, accurately predict dynamic and static obstacles, and perform dynamic planning. Summary of the Invention

[0003] This invention provides a method and system for adaptive obstacle avoidance in autonomous driving under complex road conditions, enabling real-time planning for adaptive obstacle avoidance in autonomous driving under complex road conditions based on multi-source perception fusion and multi-level risk decision-making. The technical solution provided by this application is as follows: According to a first aspect of this application, an autonomous driving planning method for adaptive obstacle avoidance in complex road conditions is provided. This method includes: real-time acquisition of multi-source environmental data surrounding the vehicle using onboard sensors, and fusing the multi-source sensor data to construct a high-precision real-time environmental model. The multi-source environmental data includes road structure information, obstacle information, vehicle status information, and road condition and traffic flow information. The onboard sensors include LiDAR, high-definition cameras, millimeter-wave radar, vehicle body sensors, and an onboard positioning module. Based on the high-precision real-time environmental model and a preset road condition feature database, the current driving road condition is intelligently classified and identified to determine the road condition complexity level. The road condition feature database is used to represent... This system collects a set of multi-dimensional feature parameters representing the complexity of road conditions under different typical driving scenarios. It classifies and labels obstacles in a high-precision real-time environment model and predicts their situations. Based on the classification and prediction results, it classifies the risk levels of obstacles and determines their risk grades. It automatically matches corresponding obstacle avoidance strategies according to the road condition complexity level, generates real-time driving trajectories based on the planned obstacle avoidance strategies and obstacle risk levels, outputs vehicle control commands based on the real-time driving trajectory to drive the vehicle to perform obstacle avoidance driving, and dynamically adjusts the update frequency of the high-precision real-time environment model according to the road condition complexity level and obstacle risk level, achieving dynamic updating and adaptive adjustment of the real-time driving trajectory.

[0004] According to another aspect of this application, an autonomous driving planning system for adaptive obstacle avoidance in complex road conditions is provided. This system is applied to an autonomous driving planning method for adaptive obstacle avoidance in complex road conditions. The system includes: a multi-source fusion modeling module, used to collect multi-source environmental data around the vehicle in real time through onboard sensors, and to fuse and process the multi-source sensor data to construct a high-precision real-time environmental model. The multi-source environmental data includes road structure information, obstacle information, vehicle's own state information, and road condition and traffic flow information; the onboard sensors include LiDAR, high-definition cameras, millimeter-wave radar, vehicle body sensors, and an onboard positioning module; and a road condition complexity classification module, used to intelligently classify and identify the current driving road conditions based on the high-precision real-time environmental model and a preset road condition feature database, determining the complexity of the road conditions, etc. The system comprises: a road condition feature database, a multi-dimensional set of feature parameters representing the complexity of road conditions under different typical driving scenarios; an obstacle risk determination module, used to classify, label, and predict the situation of obstacles in the high-precision real-time environment model, and to classify the risk level of obstacles based on the classification and prediction results; and an adaptive obstacle avoidance control module, used to automatically match the corresponding planned obstacle avoidance strategy according to the road condition complexity level, generate a real-time driving trajectory based on the planned obstacle avoidance strategy and obstacle risk level, output vehicle control commands based on the real-time driving trajectory to drive the vehicle to perform obstacle avoidance driving, and dynamically adjust the update frequency of the high-precision real-time environment model according to the road condition complexity level and obstacle risk level to achieve dynamic updating and adaptive adjustment of the real-time driving trajectory.

[0005] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By spatiotemporal fusion of multi-source data from various vehicle sensors, a high-precision hierarchical real-time environment model is constructed. Combined with weighted Euclidean distance, a four-level intelligent classification of road conditions is achieved. Based on the dynamic and static classification of obstacles, collision probability, and predicted collision time, the risk is accurately classified, thereby enabling refined and comprehensive perception and judgment of complex driving environments and obstacle risks. This allows the autonomous driving system to accurately adapt to various structured and unstructured complex road conditions and implement proactive risk prevention.

[0006] 2. By adaptively matching differentiated obstacle avoidance strategies based on the complexity level of road conditions and the risk level of obstacles, combined with traffic priority determination, horizontal and vertical trajectory coupling optimization and trajectory deviation regression fusion mechanism, and dynamically adjusting the update frequency of the environmental model, the system can balance vehicle driving safety, trajectory smoothness and traffic efficiency, thereby achieving real-time dynamic adaptive optimization and stable and reliable execution of unmanned driving trajectory in complex dynamic traffic flow scenarios.

[0007] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0008] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is a flowchart of an autonomous driving planning method for adaptive obstacle avoidance in complex road conditions provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an autonomous driving planning system for adaptive obstacle avoidance in complex road conditions, provided by an embodiment of the present invention. Detailed Implementation

[0009] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0010] This invention provides an autonomous driving planning method for adaptive obstacle avoidance in complex road conditions. For example... Figure 1The flowchart shown illustrates an autonomous driving planning method for adaptive obstacle avoidance in complex road conditions. This method's processing flow can include the following steps: Real-time acquisition of multi-source environmental data surrounding the vehicle via onboard sensors; fusion processing of the multi-source sensor data to construct a high-precision real-time environmental model; multi-source environmental data including road structure information, obstacle information, vehicle's own state information, and road and traffic flow information; onboard sensors including LiDAR, high-definition cameras, millimeter-wave radar, vehicle body sensors, and an onboard positioning module; intelligent classification and identification of the current road conditions based on the high-precision real-time environmental model and a pre-set road condition feature database to determine the complexity level of the road conditions; the road condition feature database is used for... This system employs a multi-dimensional set of feature parameters to characterize the complexity of road conditions under different typical driving scenarios. It classifies and labels obstacles in a high-precision real-time environment model and predicts their situations. Based on the classification and prediction results, it assigns risk levels to obstacles and determines their risk grades. The system automatically matches corresponding obstacle avoidance strategies according to the road condition complexity level, generates real-time driving trajectories based on the planned obstacle avoidance strategies and obstacle risk levels, outputs vehicle control commands based on the real-time driving trajectory to drive the vehicle to perform obstacle avoidance, and dynamically adjusts the update frequency of the high-precision real-time environment model according to the road condition complexity level and obstacle risk level, achieving dynamic updating and adaptive adjustment of the real-time driving trajectory.

[0011] Furthermore, the road condition feature database collects a large amount of multi-dimensional road condition feature data from typical scenarios. After manually labeling the road condition complexity levels, the database calculates the average value or cluster centers for the feature data at each level to obtain the standard feature vector for each level. The database can be built and stored offline on an in-vehicle computing platform, or it can be updated via the cloud.

[0012] In this embodiment, multi-source environmental data around the vehicle is collected by various types of vehicle-mounted sensors, and after fusion processing, a high-precision real-time environmental model is constructed. At the same time, a pre-set road condition feature database is used to complete intelligent road condition classification, and the risk level of obstacles is determined. Finally, based on the road condition level and obstacle risk level, an obstacle avoidance planning strategy is adaptively matched, a driving trajectory is generated, and the model update frequency is dynamically adjusted. This achieves full-process adaptive linkage of perception modeling, road condition recognition, risk assessment, trajectory planning, and perception update, which can be adapted to various complex driving scenarios. It significantly improves the obstacle avoidance safety, driving stability, and real-time environmental perception of autonomous vehicles in complex road conditions, and avoids problems such as obstacle avoidance failure and driving bumps caused by fixed planning strategies being unable to adapt to dynamic road conditions.

[0013] Preferably, the multi-source sensor data is fused to construct a high-precision real-time environment model. Specifically, this includes: spatiotemporal alignment of the raw data collected by LiDAR, high-definition camera, millimeter-wave radar, vehicle body sensors, and vehicle positioning module; timestamp synchronization of each sensor's data, using the GPS second pulse of the vehicle positioning module as the global time reference, and using linear interpolation to align the sampling data of LiDAR, millimeter-wave radar, and camera to the same timestamp, with the synchronization error controlled within 10ms, and unified to the same time reference; the detection results of each sensor are transformed from their respective sensor coordinate systems to a unified vehicle body coordinate system through preset sensor extrinsic parameters, and combined with vehicle positioning data to be transformed to the global world coordinate system. The preset sensor extrinsic parameters are the pre-calibrated installation pose parameters of each sensor relative to the vehicle body coordinate system, including the three-dimensional translation vector of the sensor origin relative to the vehicle's center of mass, and the rotation matrix of the sensor coordinate axes relative to the vehicle body coordinate axes, used to characterize the sensor's installation position and installation angle on the vehicle, and to achieve spatial coordinate unification of multi-source sensing data.

[0014] Obstacle feature fusion is performed based on spatiotemporally aligned multi-source data: the position and 3D size information of obstacles are obtained through LiDAR point cloud detection, the speed and direction of movement of obstacles are obtained through millimeter-wave radar detection, and the outline boundary of obstacles is corrected through high-definition camera visual detection; the multi-source detection results of the same obstacle are correlated and matched, the bounding boxes of obstacles detected by each sensor are projected onto the vehicle body coordinate system, the IOU value of each pair of bounding boxes is calculated, and the detection results with an IOU greater than 0.5 are determined to be the same obstacle, and unified obstacle attributes are output to obtain obstacle information, including obstacle position, size, speed and direction of movement.

[0015] Road structure fusion is performed based on high-definition camera image data after spatiotemporal alignment and high-precision map data matched by the vehicle positioning module: the high-definition camera identifies the lane lines and road boundaries of the current lane, and performs correction and fusion by combining the road reference data in the high-precision map to output road structure information, including lane lines and road boundaries; the wheel speed detection data of the vehicle body sensors and the attitude determination data of the vehicle positioning module are fused to calculate the vehicle's own state information, including vehicle speed and heading angle.

[0016] Based on the fused dynamic obstacle information, the dynamic obstacle density, average traffic flow speed, and speed variance within the current preset perception range (a rectangular area of ​​80m in front of and behind the vehicle, and 20m to the left and right) are statistically analyzed to obtain road condition and traffic flow information. Centered on the vehicle's current position, the road structure information, obstacle information, vehicle status information, and road condition and traffic flow information are mapped layer by layer onto a grid map with a preset resolution such as 0.2m. This constructs a high-precision real-time environment model that includes a road structure layer, an obstacle layer, a vehicle status layer, and a road condition and traffic flow statistics layer.

[0017] In this embodiment, by aligning multi-sensor data in a unified coordinate system, multi-dimensional feature fusion of obstacles, road structure, vehicle status, and traffic flow is achieved. Finally, a grid environment model containing multi-layered information is constructed. Through complementary correction of multi-source data, blind spots and detection errors of single sensors are eliminated, achieving comprehensive and accurate modeling of roads, obstacles, vehicle status, and traffic flow. This ensures the high precision, completeness, and real-time performance of the environment model, providing accurate and comprehensive data support for subsequent road condition classification, obstacle risk assessment, and trajectory planning, thereby improving the accuracy of obstacle avoidance planning for autonomous driving from the source of perception.

[0018] Preferably, based on a high-precision real-time environment model and a preset road condition feature database, intelligent classification and identification of the current driving road conditions are performed to determine the road condition complexity level. Specifically, this includes: extracting a multi-dimensional road condition feature vector of the current scene from the high-precision real-time environment model. The multi-dimensional road condition feature vector includes lane line integrity, static obstacle coverage, dynamic obstacle density, average traffic flow speed, and speed variance. Each feature is defined as follows: Lane line integrity: the ratio of the effective lane line length to the total road length within the current sensing range, ranging from 0 to 1; Static obstacle coverage: the ratio of the number of grid cells occupied by static obstacles within the sensing range to the total number of grid cells in the road area; Dynamic obstacle density: the number of dynamic obstacles per unit area within the sensing range, measured in vehicles / Square kilometers; a corresponding weight coefficient is preset for each feature component, set sequentially as 0.3, 0.25, 0.25, 0.1, and 0.1, corresponding to cumulative path length, path curvature change, inverse distance to obstacles, and lane centerline distance, respectively. For each standard feature vector of each complex scenario in the preset road condition feature database, the difference between each component of the current feature vector and the corresponding component of the standard feature vector is calculated. Each complex scenario includes Level 1, Level 2, Level 3, and Level 4 complex road conditions. Level 1 complex road condition: standard structured road, light traffic flow, normal driving scenario; Level 2 complex road condition: partial lane loss, a small number of lane-occupying obstacles, low-speed mixed pedestrian and vehicle scenario; Level 3 complex road condition: road construction, large-scale lane occupation, dense non-motorized vehicle crossing, congested and slow-moving scenario; Level 4 complex road condition: unstructured road without lane lines, sudden obstacles, high-risk scenario of disorderly crossing by multiple pedestrians and vehicles. The typical values ​​of the corresponding standard feature vectors are as follows: Level 1 Complex Road Conditions: [0.95, 0.02, 20, 60km / h, 5 (km / h)] Level 2 Complex Road Conditions: [0.7, 0.1, 80, 30 km / h, 20 (km / h)] Level 3 Complex Road Conditions: [0.4, 0.25, 200, 15km / h, 40 (km / h)] Level 4 complex road condition: [0.1, 0.4, 350, 8km / h, 60 (km / h); Squaring each difference and multiplying it by the corresponding weight coefficient; Summing all weighted squares and taking the square root to obtain the weighted Euclidean distance between the current scene and the standard scene; If the minimum weighted Euclidean distance is less than or equal to the preset maximum similarity threshold (e.g., 0.15), the complexity level of the standard feature vector corresponding to the minimum weighted Euclidean distance is determined as the complexity level of the current road condition; If the minimum weighted Euclidean distance is greater than the preset maximum similarity threshold, it is directly determined as a Level 4 complex road condition.

[0019] In addition, before calculating the difference for each component, a normalization step is included: dividing each component of the current feature vector by the preset normalization benchmark value corresponding to that component to obtain the normalized current component; dividing each component of the standard feature vector by the preset normalization benchmark value corresponding to that component to obtain the normalized standard component; calculating the difference between the normalized current component and the normalized standard component as the normalized difference; squaring the normalized difference to obtain the squared difference; multiplying the squared difference with the weight coefficient corresponding to that component to obtain the weighted squared value; summing all the weighted squared values ​​and taking the square root to obtain the weighted Euclidean distance between the current scene and the standard scene.

[0020] In this embodiment, by extracting multi-dimensional road condition feature vectors and using weighted Euclidean distance matching to preset four-level standard road condition scenarios, the complexity level of the current road condition is accurately determined. At the same time, the classification criteria for the four types of gradient road condition scenarios are clarified. Relying on multi-dimensional quantitative features and weighted matching algorithms, the complexity of road conditions is classified in a refined, intelligent, and standardized manner. The classification results are consistent with real road scenarios, providing accurate scenario basis for subsequent differentiated obstacle avoidance strategy matching, and ensuring the pertinence and adaptability of obstacle avoidance planning strategies under different complex road conditions.

[0021] Preferably, obstacles in the real-time environment model are classified, labeled, and their situations predicted. Specifically, this includes: obtaining the average speed of each obstacle over multiple consecutive time frames; if the average speed is lower than a preset static speed threshold (e.g., 0.5 m / s), the obstacle is classified as a static obstacle; otherwise, it is classified as a dynamic obstacle. For static obstacles, the collision probability of each static obstacle is calculated based on the shortest relative distance between the vehicle's current planned driving path and the obstacle's occupied area, the vehicle's current speed, and the obstacle's size. Specifically, this includes: obtaining the shortest Euclidean distance between the path point sequence of the vehicle's current planned driving path and the boundary of the static obstacle's occupied area, as the shortest relative distance; and calculating the emergency braking safety distance by combining the vehicle's current speed with a preset road adhesion coefficient of 0.7 (typical value for dry asphalt roads; for wet and slippery roads, this can be corrected to 0.3~0.5): where is the system reaction time, taken as 0.3 s, and g... The gravitational acceleration is 9.8 m / s². The emergency braking safety distance is added to half the dimension of the obstacle along the path to obtain the critical safety distance. The ratio of the shortest relative distance to the critical safety distance is calculated. This ratio is input into a preset collision probability function, which uses a piecewise S-shaped decreasing function: when the ratio is 0.8, the collision probability is 1; when the ratio is 1.5, the collision probability is 0; when 0.8 < ratio < 1.5, the collision probability is calculated by the difference between the ratio and the lower threshold of 0.8 to obtain the first difference. The first difference is divided by the transition interval length of 0.7 to obtain the normalization parameter. The normalization parameter is multiplied by to obtain the input angle of the cosine function. The cosine value of the input angle is taken and then added to 1 to obtain the transition value. The transition value is divided by 2 to obtain the collision probability, thus generating the collision probability of each static obstacle.

[0022] For dynamic obstacles, based on the current position, velocity and movement direction obtained by tracking in the high-precision real-time environment model, a constant velocity model is used to predict the movement trajectory forward for a duration of 3s, with a prediction time step of 0.1s, so as to obtain the spatio-temporal position sequence of the dynamic obstacle; the spatio-temporal position sequence of the planned driving path of the vehicle itself for a future period of time is obtained; at each prediction time step, the relative distance between the vehicle position and the predicted position of the obstacle is calculated, and the minimum relative distance and the corresponding relative velocity within the entire prediction period are recorded; the minimum predicted time-to-collision between the vehicle and the dynamic obstacle is calculated based on the minimum relative distance and the relative velocity. If the relative velocity is close to zero or there is no intersection between the two trajectories, the predicted time-to-collision is set to infinity; the minimum predicted time-to-collusion is input into a preset collision probability mapping function, and the function adopts an increasing Sigmoid function: when the predicted time-to-collision TTC ≤ 1s, the collision probability is 1; when TTC ≥ 5s, the collision probability is 0; when 1s<TTC<5s, the collision probability is [missing expression in original text], the function outputs a collision probability that increases monotonically as the predicted time-to-collision decreases, generates the collision probability of the dynamic obstacle, and outputs the shortest relative distance and the minimum predicted time-to-collision as the calculation basis for the predicted time-to-collision.

[0023] In this embodiment, static and dynamic obstacles are distinguished by a velocity threshold, and targeted collision probability calculation logics are designed respectively. For static obstacles, collision probability is calculated in combination with vehicle distance, braking distance and obstacle size; for dynamic obstacles, collision probability is mapped through trajectory prediction and minimum time-to-collision; differentiated and refined situation assessment and collision risk quantification for two types of obstacles are realized, which can accurately capture potential collision risks of different types of obstacles, avoid missed judgment and misjudgment of risks caused by unified prediction, and provide accurate quantitative data support for obstacle risk classification.

[0024] Preferably, the obstacle risk level is classified based on the classification label and situation prediction results to determine the obstacle risk level. Specifically, the obstacle risk level includes low, medium, high, and extremely high risk levels; the collision probability of the obstacle is obtained; wherein the predicted collision time of static obstacles is calculated based on the shortest relative distance, the current speed of the vehicle, and the preset emergency braking deceleration, according to a uniform deceleration motion model, and the predicted collision time of dynamic obstacles is the minimum predicted collision time; when the collision probability is greater than or equal to a preset high collision probability threshold, such as 0.8, and the predicted collision time is less than a preset emergency collision time threshold, such as 2s, the obstacle risk level is determined to be extremely high. Risk: When the collision probability is greater than or equal to the preset medium collision probability threshold (e.g., 0.5) and the predicted collision time is less than the preset safe collision time threshold (e.g., 4s), or when the collision probability is greater than or equal to the preset high collision probability threshold and the predicted collision time is greater than or equal to the preset emergency collision time threshold, the obstacle risk level is determined to be high risk; when the collision probability is greater than or equal to the preset low collision probability threshold (e.g., 0.2) and less than the preset medium collision probability threshold, and the predicted collision time is greater than or equal to the preset safe collision time threshold, the obstacle risk level is determined to be medium risk; when the collision probability is less than the preset low collision probability threshold, the obstacle risk level is determined to be low risk.

[0025] In this embodiment, a four-level obstacle risk classification rule is established by combining the dual quantitative indicators of obstacle collision probability and predicted collision time, accurately classifying four risk levels: low, medium, high, and extremely high. Through dual-parameter coupling judgment, the level of obstacle risk is refined and standardized, accurately distinguishing the degree of danger and urgency of different obstacles. This provides a direct risk basis for subsequent differentiated obstacle avoidance trajectory generation and dynamic trajectory adjustment, ensuring that obstacle avoidance operations accurately match the obstacle danger level.

[0026] Preferably, the corresponding obstacle avoidance strategy is automatically matched according to the level of road condition complexity. Based on the obstacle avoidance strategy and the obstacle risk level, a real-time driving trajectory is generated. Specifically, this includes: taking the vehicle's current position as the starting point and the destination as the ending point, marking the grids occupied by static obstacles as impassable areas and the remaining grids as passable areas on a high-precision real-time environment model grid map; within the passable areas, expanding the search layer by layer from the starting point, using the vehicle's current heading as the reference, and generating multiple candidate path points for the next layer within the turning angle range allowed by the vehicle's minimum turning radius, according to a preset angle resolution such as 5.

[0027] For each candidate path point, a comprehensive cost value is calculated. The comprehensive cost value is obtained by weighted summation of the first cost term, the second cost term, the third cost term, and the fourth cost term, with corresponding weight coefficients of 0.3, 0.2, 0.3, and 0.2, respectively. The first cost term is the cumulative path length between the candidate path point and the starting point; the second cost term is the change in path curvature formed by the candidate path point and the path point in the previous layer; the third cost term is the reciprocal of the distance from the candidate path point to the nearest static obstacle grid; and the fourth cost term is the perpendicular distance from the candidate path point to the center line of the lane in the same direction. In each layer of expansion, only a preset number of candidate path points with the minimum comprehensive cost value, such as 20, are retained as surviving nodes. The expansion continues to the next layer until a candidate path point enters the preset reachable range of the destination. The surviving nodes of each layer are backtracked to obtain a global driving path composed of ordered path points. The radius of the arc formed by any three adjacent points on this path is not less than the minimum turning radius of the vehicle.

[0028] Along the global driving path, the road speed limit value at each waypoint is obtained sequentially, and the radius of curvature of the path at each waypoint is calculated. The radius of curvature is then compared with the preset comfort lateral acceleration limit of 2m / s². 2 Substituting into the centripetal acceleration formula, the safe speed for passing through the curve is calculated, where is the radius of curvature of the path and is the upper limit of the comfortable lateral acceleration. The smaller value between the safe speed for passing through the curve and the road speed limit is taken as the target speed limit for that point. Starting from the starting point, with the preset upper limit of the comfortable acceleration of 2m / s and the upper limit of the comfortable deceleration of 3m / s as constraints, the speed change between adjacent path points is adjusted in turn to generate a reference speed sequence with a smooth speed curve and continuous acceleration.

[0029] The system acquires traffic rule information for the current lane at intersections, ramps, and merging zones. Combined with the distance and speed of other vehicles relative to the vehicle, it determines the vehicle's priority by comparing the arrival time of the vehicle at the conflict zone with that of other vehicles. Traffic priorities are categorized as priority passage, yielding, and waiting. Specifically, this includes: defining the spatial boundaries of conflict zones where traffic flows from different directions converge based on road structure information and high-precision map data; calculating the estimated arrival time of the vehicle to the initial boundary of the conflict zone based on its current position, current speed, and a reference speed sequence; and calculating the estimated arrival time of each dynamic obstacle to the initial boundary of the conflict zone based on its current position and speed.

[0030] The estimated arrival time of this vehicle is compared with the estimated arrival times of all dynamic obstacles one by one. If the vehicle's arrival time is earlier than the arrival times of all dynamic obstacles, and the time difference is greater than the preset safe meeting time threshold of 3 seconds, and the traffic rules do not require this vehicle to yield, then the vehicle's priority is determined to be priority to pass. If the vehicle's arrival time is later than the arrival time of any dynamic obstacle, and the time difference is less than the preset safe meeting time threshold, or the traffic rules clearly state that this vehicle must yield, then the vehicle's priority is determined to be yielding. If the arrival time difference between this vehicle and multiple dynamic obstacles is less than the preset safe meeting time threshold... If the obstacle risk level is high and there is a risk of continuous intersection, the vehicle's priority is determined to be waiting. If the priority is priority, no additional processing is performed. If the priority is yield, a constraint to reduce speed to a preset yield speed, such as 10 km / h, is added before the conflict area. If the priority is waiting, a stopping point is added to obtain a corrected baseline speed sequence. The global driving path and the corrected baseline speed sequence are used as the initial optimal global driving path. The initial optimal global driving path is dynamically adjusted according to the obstacle risk level and road conditions to generate the final planned real-time driving trajectory.

[0031] The correspondence between road condition complexity levels and planning strategies includes: Level 1 complex road conditions are matched with a normal driving strategy, and the weight of the fourth cost term (distance to the lane centerline) is set normally during global path search; Level 2 complex road conditions are matched with a congestion adaptive strategy, and the weight of the third cost term (reciprocal of obstacle distance) is increased during global path search; Level 3 complex road conditions are matched with a low-speed cautious driving strategy, and the upper limit of comfort acceleration is reduced to a preset lower value during global path search, and the number of surviving nodes is reduced; Level 4 complex road conditions are matched with an unstructured road condition strategy, and the fourth cost term (lane centerline constraint) is turned off during global path search, and path search no longer relies on the lane centerline.

[0032] In this embodiment, a globally optimal path is generated through multi-layer node search and multi-dimensional cost weighted filtering. A reasonable vehicle speed is matched with road speed limits and curve curvature. Then, the vehicle priority is determined based on intersection traffic conflicts and traffic rules, and the vehicle speed sequence is corrected to obtain the initial optimal driving path. The generated path satisfies the mechanical constraints of vehicle driving, and the vehicle speed curve is smooth and continuous. At the same time, it takes into account traffic rules and traffic flow interaction, effectively avoiding traffic conflicts in complex areas such as intersections and merging areas, and improving the compliance, comfort and traffic safety of autonomous driving.

[0033] Preferably, based on the obstacle risk level and road condition scenario, the initial optimal global driving path is dynamically adjusted to generate the final planned real-time driving trajectory. Specifically, this includes: dynamically generating a preliminary obstacle avoidance driving trajectory based on the obstacle risk level and road condition scenario. This preliminary obstacle avoidance driving trajectory includes a first obstacle avoidance trajectory, a second obstacle avoidance trajectory, a third obstacle avoidance trajectory, a fourth obstacle avoidance trajectory, and a fifth obstacle avoidance trajectory; then performing deviation regression fusion between the preliminary obstacle avoidance driving trajectory and the initial optimal global driving path to generate the final planned real-time driving trajectory. Specifically, dynamically generating the preliminary obstacle avoidance driving trajectory based on the obstacle risk level and road condition scenario includes: when a low-risk obstacle is identified, using the path point connection of the initial optimal global driving path as a reference line, taking the safety boundary point on the side of the obstacle closest to the vehicle as the lateral offset target point, selecting the vehicle's current position as the starting point, and a predetermined regression distance forward along the reference line (e.g., 20m) as the ending point, and constructing a fifth-order polynomial. The lateral offset curve starts at the vehicle's current position and ends at a predetermined regression distance along the reference line. The lateral offset is defined as a fifth-order polynomial function of time, where time ranges from 0 to a predetermined detour duration. At the starting point, the lateral offset is 0, and both its first and second derivatives with respect to time are 0. At the ending point, the lateral offset equals the lateral distance between the obstacle safety boundary point and the reference line, and both its first and second derivatives with respect to time are 0. Substituting these six boundary conditions into the fifth-order polynomial and its first and second derivative expressions, the coefficients of the six polynomials are solved to obtain the lateral offset curve. The lateral offset curve satisfies the condition that both the first and second derivatives at the starting and ending points are continuous. The vehicle travels along the lateral offset curve to complete a small, smooth detour, maintaining the corresponding reference speed in the reference speed sequence throughout, thus forming the first obstacle avoidance trajectory.

[0034] When a medium-risk obstacle is identified, the reference speed corresponding to the current position in the reference speed sequence is multiplied by a preset deceleration ratio of 0.7 to obtain the target speed. A trajectory solving problem is constructed with lateral offset and longitudinal speed change as variables. The constraints are that the safe distance between the trajectory and the obstacle is greater than a preset safe distance threshold of 1m, the maximum lateral acceleration does not exceed 2m / s, and the maximum longitudinal deceleration does not exceed 3m / s. The model predictive control (MPC) algorithm is used to solve the problem and obtain the fine-tuned trajectory with lateral and longitudinal coupling as the second obstacle avoidance trajectory.

[0035] When a high-risk or extremely high-risk obstacle is detected, the system uses the vehicle's current position and speed as the initial state, and calculates forward at a preset maximum comfortable deceleration of 5 m / s. It determines if there exists a point where the shortest distance between the vehicle and the obstacle-occupied area is greater than zero before the vehicle comes to a complete stop. If such a point exists, a straight emergency braking trajectory along the current lane centerline is generated, ending at zero speed, serving as the third obstacle avoidance trajectory. If no suitable stopping point exists, the system checks if there is a continuous unoccupied space within a preset detection distance of 50 m in front of the obstacle in the adjacent lane, with a length greater than 1.5 times the vehicle's length. It also checks if the relative distance and speed of vehicles approaching from behind in the adjacent lane meet preset safe lane-changing conditions. These conditions are: a longitudinal relative distance of more than 30 m between the vehicle and the approaching vehicle, and a relative speed less than or equal to 10 m / s. If both conditions are met, the system proceeds as planned. Starting from the center line of the preceding lane and targeting the center line of the adjacent lane, a smooth lateral transition curve is generated using a sine function. Starting from the center line of the current lane and targeting the center line of the adjacent lane, the total lateral offset is equal to the lane width. A lane-changing time is set and divided into several time steps. At each time step, the lateral offset is equal to the total lateral offset multiplied by the ratio of the corresponding time step to the lane-changing time, minus the sine of that ratio divided by two, to obtain the smooth transition lateral offset. The longitudinal displacement is calculated based on the current vehicle speed and the applied preset maximum comfort deceleration according to the uniform deceleration motion law. The lane-changing time is not less than 3 seconds. At the same time, the preset maximum comfort deceleration is applied to form an emergency lane-changing detour trajectory, which serves as the fourth obstacle avoidance trajectory. If any condition is not met, the vehicle continues to stop along the current lane at the preset maximum comfort deceleration, and triggers external audible and visual warning signals, forming the fifth obstacle avoidance trajectory.

[0036] In this embodiment, five differentiated obstacle avoidance trajectories are designed for four levels of obstacle risk: minor detour, deceleration and fine adjustment, emergency braking, emergency lane change, and warning stop. This enables refined obstacle avoidance operations corresponding to risk levels. Low and medium risk scenarios ensure smooth driving, while high and extremely high risk scenarios prioritize driving safety. The design balances driving experience with the reliability of obstacle avoidance in extreme conditions, comprehensively covering all types of obstacle avoidance scenarios, including both routine and emergency scenarios, and significantly improving the vehicle's emergency obstacle avoidance capabilities in complex and high-risk road conditions.

[0037] Preferably, a deviation regression fusion is performed on the initial obstacle avoidance trajectory and the initial optimal global driving path to generate the final planned real-time driving trajectory. Specifically, this includes: acquiring the vehicle's current positioning information in real time; calculating the vertical distance from the vehicle's position to the initial optimal global driving path as the lateral position deviation; simultaneously calculating the difference between the vehicle's current heading angle and the tangential angle of the corresponding point on the initial optimal global driving path as the heading angle deviation; on the initial optimal global driving path, a preset forward aiming distance of 15m is intercepted from the vehicle's current position along the path direction to obtain the coordinates of this forward aiming point; the lateral offset between the vehicle's current heading extension line and this forward aiming point is calculated as the forward lateral deviation; based on the forward lateral deviation and the heading angle deviation, combined with a preset trajectory tracking gain of 0.8, a pure tracking algorithm is used to generate... The smooth regression trajectory for the initial optimal global driving path regression is as follows: Calculate the target steering curvature, which is equal to 2 times the forward lateral deviation divided by the square of the aiming distance. Calculate the heading angle correction curvature, which is equal to the heading angle deviation multiplied by the preset trajectory tracking gain divided by the aiming distance. Add the target steering curvature and the heading angle correction curvature to obtain the comprehensive target curvature. Generate an arc path from the vehicle's current position to the forward point using the comprehensive target curvature. This arc path is the smooth regression trajectory. Check the lateral acceleration at each point on the smooth regression trajectory. The lateral acceleration is equal to the curvature at that point multiplied by the square of the vehicle speed. If it exceeds the preset comfort upper limit, the curvature is reduced proportionally until the constraint is met. The smooth regression trajectory satisfies the condition of continuous curvature and lateral acceleration not exceeding the preset comfort upper limit of 2 m / s.

[0038] The smooth regression trajectory and the initial obstacle avoidance trajectory are compared point-by-point in time and space. When there is no conflict between the two, the smooth regression trajectory and the initial obstacle avoidance trajectory are weighted and fused to obtain the fused trajectory. The weighting is dynamically adjusted according to the obstacle risk level: for low risk, the regression trajectory weight is 0.7 and the obstacle avoidance trajectory weight is 0.3; for medium risk, the weights of the two are each 0.5. When there is a conflict between the two, the compromise trajectory that can simultaneously meet the obstacle safety distance constraint and the global path regression requirement is selected as the fused trajectory. If there is no feasible compromise trajectory, the initial obstacle avoidance trajectory is followed completely until the corresponding obstacle risk level is reduced to below medium risk, and then the smooth regression trajectory is switched. The fused trajectory or the final switched trajectory is used as the final planned real-time driving trajectory and output for use by the vehicle control actuator.

[0039] In this embodiment, by detecting multi-dimensional deviations in the vehicle's lateral, heading, and forward directions, a smooth trajectory that reverts to the global path is generated. This trajectory is then fused with the obstacle avoidance trajectory, conflict determination is performed, and adaptive switching is applied to output the final driving trajectory. This achieves a dynamic balance between safe obstacle avoidance and path regression, ensuring continuous curvature and smooth driving throughout the trajectory. It avoids vehicle jerking and sideslip caused by sudden trajectory changes. Furthermore, after the obstacle risk is eliminated, the vehicle can smoothly return to the preset driving path, balancing obstacle avoidance safety and path tracking accuracy.

[0040] Preferably, the update frequency of the high-precision real-time environment model is dynamically adjusted according to the road condition complexity level and obstacle risk level. Specifically, this includes: when the road condition complexity level is level four, or when there are obstacles of extremely high risk, the frame rate of the LiDAR and HD camera is forcibly increased to the maximum hardware-supported frame rate (typically 20Hz for LiDAR and 30 frames / second for the camera), and the update frequency of the high-precision real-time environment model is set to a preset upper limit value, which is the larger of the maximum hardware-supported frame rate of the LiDAR and the maximum frame rate of the HD camera; otherwise, the corresponding baseline update frequency is queried from a preset basic update frequency mapping table according to the current road condition complexity level. The basic update frequency mapping table stores the baseline update frequency values ​​corresponding to the four levels of complex road conditions: level one, level two, level three, and level four. The baseline frequency for level one is 10Hz, for level two it is 15Hz to 20Hz, for level three it is 20Hz to 30Hz, and for level four it is 30Hz to 5Hz. The baseline update frequency is 0Hz. The higher the road condition complexity level, the higher the baseline update frequency. Based on the highest risk level among all obstacles, the corresponding update frequency adjustment coefficient is retrieved from the preset frequency adjustment coefficient mapping table. The baseline update frequency is multiplied by the update frequency adjustment coefficient to obtain the target update frequency. The obstacle risk level and frequency adjustment coefficient mapping table stores the frequency adjustment coefficient values ​​corresponding to four levels: low risk, medium risk, high risk, and extremely high risk. The coefficient for low risk is 1 times the baseline value, the coefficient for medium risk is 1.2 to 1.5 times, the coefficient for high risk is 1.5 to 2.0 times, and the coefficient for extremely high risk is 2.0 to 3.0 times. The update frequency of the high-precision real-time environment model is dynamically adjusted based on the target update frequency.

[0041] In this embodiment, the update frequency of the environmental model is adjusted by matching the road condition complexity level with the benchmark update frequency and dynamically superimposing the frequency adjustment coefficient based on the highest obstacle risk level. The frequency range and forced upgrade rules corresponding to each level are clearly defined. This ensures that the more complex the road conditions and the higher the obstacle risk, the more timely the perception model update. This reduces the hardware computing power consumption and improves the operating efficiency under simple road conditions, while ensuring the ultra-high real-time performance of environmental perception in high-risk and complex scenarios. It also eliminates obstacle avoidance errors caused by perception delays and achieves the optimal match between computing power resources and driving safety.

[0042] This invention also provides an autonomous driving planning system for adaptive obstacle avoidance in complex road conditions, such as... Figure 2 The diagram shows the structure of an autonomous driving planning system for adaptive obstacle avoidance in complex road conditions. This system includes: a multi-source fusion modeling module, used to collect real-time multi-source environmental data around the vehicle through onboard sensors, and to fuse and process the multi-source sensor data to construct a high-precision real-time environmental model. The multi-source environmental data includes road structure information, obstacle information, vehicle status information, and road condition and traffic flow information; onboard sensors include LiDAR, high-definition cameras, millimeter-wave radar, body sensors, and an onboard positioning module; and a road condition complexity classification module, used to intelligently classify and identify the current road condition based on the high-precision real-time environmental model and a preset road condition feature database, determining the road condition complexity level; the road condition feature database is used to characterize different types of road conditions. The system includes a multi-dimensional feature parameter set for the complexity of road conditions in various driving scenarios; an obstacle risk assessment module for classifying, labeling, and predicting the situation of obstacles in a high-precision real-time environment model, and classifying the risk level of obstacles based on the classification and prediction results; and an adaptive obstacle avoidance control module for automatically matching corresponding obstacle avoidance strategies according to the level of road condition complexity, generating real-time driving trajectories based on the planned obstacle avoidance strategies and obstacle risk levels, outputting vehicle control commands based on the real-time driving trajectory to drive the vehicle to perform obstacle avoidance driving, and dynamically adjusting the update frequency of the high-precision real-time environment model according to the level of road condition complexity and obstacle risk levels to achieve dynamic updating and adaptive adjustment of the real-time driving trajectory.

[0043] Example 1: Complex scenario of mixed pedestrian and vehicle traffic on urban roads The driverless delivery vehicle travels on urban auxiliary roads, collecting multi-source environmental data in real time through LiDAR, high-definition cameras, millimeter-wave radar, vehicle sensors, and onboard positioning modules. After spatiotemporal alignment and feature fusion, a high-precision real-time environmental model is constructed, including a road structure layer, an obstacle layer, a vehicle status layer, and a road condition and traffic flow statistics layer. The onboard sensors and their data processing unit detect in real time that there are illegally parked vehicles in the lane ahead, and at the same time, a pedestrian suddenly crosses the road on the right-hand sidewalk.

[0044] The system extracts lane line integrity, static obstacle coverage, dynamic obstacle density, average traffic flow speed and speed variance from a high-precision real-time environment model to form a multi-dimensional road condition feature vector. It then calculates the weighted Euclidean distance between the vector and the standard feature vector of a level four complex scenario in the preset road condition feature database. If the minimum weighted Euclidean distance is less than the preset maximum similarity threshold, the system identifies the current road condition as level three complex and automatically matches the corresponding low-speed cautious driving planning obstacle avoidance strategy.

[0045] The obstacle module acquires the average speed of each obstacle in consecutive time frames. Vehicles with average speeds below a preset static speed threshold are classified as static obstacles; pedestrians with average speeds above the threshold are classified as dynamic obstacles. For static illegally parked vehicles, the collision probability is calculated based on the shortest relative distance between the vehicle's current planned path and its occupied area, its current speed, and size. For dynamic pedestrians, the module predicts their forward trajectory based on their current position, speed, and direction of movement, obtaining a spatiotemporal position sequence within 2 seconds of the pedestrian entering the vehicle's path. The relative distance between the vehicle and the pedestrian at each predicted time step is calculated to obtain the minimum predicted collision time. This minimum predicted collision time is input into a collision probability mapping function, which outputs a high collision probability. Combined with the predicted collision time, the pedestrian obstacle risk level is determined to be high risk.

[0046] The adaptive obstacle avoidance control module uses the vehicle's current position as the starting point and the destination as the ending point. In the grid map, it marks the grids occupied by illegally parked vehicles as impassable areas. It generates an initial optimal global driving path through a multi-cost weighted search and constructs a baseline speed sequence. Based on high-risk obstacles, it determines that a constant deceleration at a preset maximum comfort deceleration is insufficient to completely avoid the obstacle before stopping. If the adjacent lane on the left or a passable area meets the safe lane-changing conditions, it generates an emergency lane-changing detour trajectory as the fourth obstacle avoidance trajectory: starting from the current lane centerline and using the passable area on the left as the lateral offset target, it generates a smooth lateral transition curve and applies a preset maximum comfort deceleration to form a deceleration detour trajectory; it controls the vehicle to slow down and smoothly detour slightly to the left to avoid pedestrians and illegally parked vehicles.

[0047] During obstacle avoidance, the vertical distance and heading angle deviation from the vehicle's position to the initial optimal global driving path are calculated in real time, and the forward lateral deviation at the look-ahead point is obtained. Based on the trajectory tracking gain, a smooth regression trajectory that satisfies the curvature continuity and lateral acceleration comfort upper limit is generated. After the obstacle is completely removed from the driving area, the smooth regression trajectory is compared spatiotemporally with the fourth obstacle avoidance trajectory. When the shortest distance between the vehicle's current position and the area occupied by the obstacle in the high-precision real-time environment model is greater than a preset safe distance threshold (e.g., twice the vehicle length), and the obstacle is not on the vehicle's planned driving path for the next 3 seconds, it is determined that the obstacle has left the driving area. In the absence of conflict, weighted fusion is performed to guide the vehicle to smoothly return to the original global driving path and resume normal driving speed. At the same time, due to the presence of high-risk obstacles and the road condition complexity level of level three, the frame rate of the LiDAR and high-definition camera is increased according to the update frequency adjustment strategy, and the update frequency of the high-precision real-time environment model is set to a preset upper limit value to ensure real-time adaptive adjustment throughout the process.

[0048] Example 2: Rural unstructured road conditions without lane markings The driverless sanitation vehicle travels on rural roads with no lane markings and uneven road widths. It uses onboard sensors to build a high-precision real-time environmental model, and there are scattered obstacles such as weeds, stones, and temporary piles of materials on both sides of the road.

[0049] The system extracts multi-dimensional road condition feature vectors of the current scene from a high-precision real-time environment model. It then calculates weighted Euclidean distances between these vectors and the standard feature vectors of Level 4 complex scenes in a pre-set road condition feature database. If the minimum weighted Euclidean distance is greater than a pre-set maximum similarity threshold, the road condition is directly classified as Level 4 unstructured complex road condition. The system automatically matches unstructured road condition planning strategies, disables the traditional lane constraint planning mode, and no longer relies on lane centerlines for path search.

[0050] In the road structure layer of the raster map, grids occupied by static obstacles such as weeds, rocks, and temporary piles of materials, identified by LiDAR point clouds and camera images, are marked as impassable areas, while the remaining areas are marked as passable areas. Starting from the vehicle's current position and guided by the destination direction, within the minimum turning radius and allowable turning angle range, candidate path points are expanded layer by layer according to a preset angle resolution. The comprehensive cost of each candidate path point is obtained by weighted summation of the cumulative path length, the change in path curvature, the reciprocal of the distance to the nearest obstacle grid, and the distance to the center trend line of the passable area. An optimal travel trajectory, unconstrained by lanes, is dynamically fitted. The center trend line of the passable area is the midline between the two boundary lines after curve fitting of the boundaries of the passable area. This midline is obtained by fitting the average coordinates of the corresponding points on the left and right boundaries of the passable area.

[0051] The vehicle adaptively travels at low speed along the optimal travel trajectory. For scattered small obstacles, when the risk is low, a lateral offset curve is generated using the optimal travel trajectory as a reference line to form the first obstacle avoidance trajectory for fine-tuning. When the risk is medium, a lateral and longitudinal coupled fine-tuning trajectory that satisfies the safety distance and comfortable acceleration constraints is solved to form the second obstacle avoidance trajectory. A vehicle kinematic model is built, with state variables including the vehicle's lateral position, longitudinal position, heading angle, and longitudinal velocity, and control variables including the front wheel angle and longitudinal acceleration. A prediction time domain and sampling period are set, with the current vehicle state as the initial state. An objective function is constructed, which is obtained by weighted summation of the deviation of the lateral position from the reference line, the deviation of the longitudinal velocity from the target vehicle speed, and the rate of change of the control quantity. Constraints are set, including a safe distance between the trajectory and the obstacle greater than a preset threshold, lateral acceleration not exceeding a preset upper limit, longitudinal deceleration not exceeding a preset upper limit, and front wheel angle not exceeding the vehicle's maximum steering angle. The control sequence that minimizes the objective function is solved in the prediction time domain, and the first control quantity is taken as the control command at the current moment to generate the fine-tuning trajectory. For large-area obstacles, the system first determines whether a safe detour is allowed within the passable area. If a continuous space meeting the conditions exists, a lane-changing detour trajectory is generated, forming the fourth obstacle avoidance trajectory. If no detour is possible, the system stops at a preset maximum comfortable deceleration and triggers an audible and visual warning, forming the fifth obstacle avoidance trajectory. The entire driving process is lane-independent, ensuring stable operation. The system also calculates the deviation from the optimal travel trajectory in real time for smooth regression and dynamically adjusts the environmental model update frequency to a preset upper limit based on the four levels of complex road conditions.

[0052] Example 3: Complex Scenario of Road Construction and Congestion The road section is subject to road construction barriers and vehicles operating in the road, with dense traffic and frequent lane changes and weaving. Onboard sensors construct a high-precision real-time environmental model to acquire information on road structure, dynamic obstacle density, and traffic flow.

[0053] The system extracts multi-dimensional road condition feature vectors of the current scene, calculates the weighted Euclidean distance with the standard feature vectors, and identifies the standard scene corresponding to the minimum weighted Euclidean distance as a level-two complex road condition. The system then identifies the current scene as a level-two complex congested road condition and activates a congestion adaptive planning strategy. This strategy reduces the number of surviving nodes in the global path search from 20 to a preset lower value (e.g., 10), increases the angular resolution from 5 to a preset higher value (e.g., 10), and reduces search complexity. Simultaneously, it increases the weight of the third cost term (inverse distance to obstacles) from 0.3 to a preset higher value (e.g., 0.4) to enhance sensitivity to surrounding obstacles. The system switches the planning mode to a dense dynamic interaction mode and queries the corresponding baseline update frequency from the basic update frequency mapping table based on the level-two complex road condition.

[0054] The obstacle module tracks and predicts the situation of surrounding dynamic vehicles. Based on the position and speed sequences of continuous time frames, it predicts the lane-changing and cutting-in intentions of each vehicle. By analyzing the lateral position change rate of dynamic vehicles in continuous time frames, when the lateral position change rate exceeds a preset lane-changing threshold and the direction is pointing towards the lane line, it determines that the vehicle intends to change lanes. By analyzing the longitudinal distance change rate between dynamic vehicles and the vehicle itself, when the longitudinal distance decreases rapidly and the vehicle is in the same lane ahead of the vehicle, it determines that the vehicle intends to cut in and generates the corresponding spatiotemporal trajectory. It calculates the minimum predicted collision time and collision probability between the vehicle and each dynamic vehicle and dynamically assesses the risk level. According to the congestion adaptive strategy, for medium and high-risk dynamic vehicles, it generates a lateral and longitudinal coupled fine-tuning trajectory in real time with lateral offset and longitudinal speed change as variables, dynamically adjusting the following distance and driving trajectory to achieve proactive yielding or preventive avoidance.

[0055] The planning module simultaneously marks static obstacles such as construction barriers and vehicles occupying the road as impassable areas on the grid map. It generates a baseline trajectory based on global path search and overlays the predicted dynamic traffic flow onto this trajectory. While avoiding static construction obstacles, it generates a fused trajectory in advance based on the predicted dynamic traffic flow trend through minor adjustments such as deceleration or slight lateral offsets, ensuring a smooth transition in driving posture and avoiding collisions. Throughout the process, a frequency adjustment coefficient is determined based on the current highest risk level, multiplied by the secondary road condition baseline update frequency to obtain the target update frequency. This dynamically adjusts the update frequency of the high-precision real-time environmental model to ensure the real-time nature of the planning response, ultimately allowing vehicles to smoothly pass through congested construction sections.

[0056] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0057] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A driving planning method for autonomous driving with adaptive obstacle avoidance in complex road conditions, characterized in that, The method includes: The vehicle collects multi-source environmental data around the vehicle in real time using onboard sensors, and fuses the multi-source sensor data to construct a high-precision real-time environmental model. The multi-source environmental data includes road structure information, obstacle information, vehicle status information, and road condition and traffic flow information. The onboard sensors include lidar, high-definition cameras, millimeter-wave radar, body sensors, and onboard positioning modules. Based on the high-precision real-time environment model and the preset road condition feature database, the current driving road condition is intelligently classified and identified to determine the road condition complexity level. The road condition feature database is a set of multi-dimensional feature parameters that characterize the road condition complexity under different typical driving scenarios. The obstacles in the high-precision real-time environment model are classified, labeled, and their situations are predicted. Based on the classification, labeling, and situation prediction results, the obstacles are classified into risk levels to determine the obstacle risk level. The system automatically matches the corresponding obstacle avoidance strategy based on the road condition complexity level, generates a real-time driving trajectory based on the planned obstacle avoidance strategy and the obstacle risk level, outputs vehicle control commands based on the real-time driving trajectory to drive the vehicle to perform obstacle avoidance driving, and dynamically adjusts the update frequency of the high-precision real-time environment model based on the road condition complexity level and the obstacle risk level to achieve dynamic updating and adaptive adjustment of the real-time driving trajectory.

2. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 1, characterized in that, The process of fusing multi-source sensor data to construct a high-precision real-time environment model specifically includes: Spatiotemporal alignment is performed on the raw data collected by lidar, high-definition cameras, millimeter-wave radar, vehicle body sensors, and vehicle positioning modules: the data from each sensor are timestamped and unified to the same time reference; the detection results of each sensor are transformed from their respective sensor coordinate systems to a unified vehicle body coordinate system, and combined with the vehicle positioning data, they are transformed to the global world coordinate system. Obstacle feature fusion is performed based on spatiotemporally aligned multi-source data: the position and three-dimensional size information of obstacles are obtained through lidar point cloud detection, the speed and direction of movement of obstacles are obtained through millimeter-wave radar detection, and the outline boundary of obstacles is corrected through high-definition camera visual detection; the multi-source detection results of the same obstacle are correlated and matched to output unified obstacle attributes and obtain obstacle information, which includes obstacle position, size, speed and direction of movement; Road structure fusion is performed based on high-definition camera image data after spatiotemporal alignment and high-precision map data matched by vehicle positioning module: the lane lines and road boundaries of the current lane are identified by the high-definition camera, and the road reference data in the high-precision map are combined for correction and fusion to output road structure information, which includes lane lines and road boundaries; By integrating wheel speed detection data from vehicle body sensors with attitude determination data from onboard positioning modules, the vehicle's own state information is calculated, including vehicle speed and heading angle. Based on the fused dynamic obstacle information, the dynamic obstacle density, average traffic flow speed and speed variance within the current preset sensing range are statistically analyzed to obtain road condition and traffic flow information. Centered on the vehicle's current location, the road structure information, obstacle information, vehicle status information, and traffic flow information are mapped layer by layer onto a grid map of a preset resolution, thereby constructing a high-precision real-time environment model that includes a road structure layer, an obstacle layer, a vehicle status layer, and a traffic flow statistics layer.

3. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 2, characterized in that, The process of intelligently classifying and identifying the current road conditions based on the high-precision real-time environment model and the preset road condition feature database, and determining the level of road condition complexity, specifically includes: The high-precision real-time environment model extracts a multi-dimensional road condition feature vector for the current scene. The multi-dimensional road condition feature vector includes lane line integrity, static obstacle coverage, dynamic obstacle density, average traffic flow speed, and speed variance. For each feature component, a corresponding weight coefficient is preset. For each standard feature vector of each complex scenario in the preset road condition feature database, the difference between each component of the current feature vector and the corresponding component of the standard feature vector is calculated. The complex scenarios include level 1 complex road conditions, level 2 complex road conditions, level 3 complex road conditions, and level 4 complex road conditions. After squaring each difference, it is multiplied by the weight coefficient corresponding to that component. After summing all the weighted square values, the square root is taken to obtain the weighted Euclidean distance between the current scenario and the standard scenario. If the minimum weighted Euclidean distance is less than or equal to the preset maximum similarity threshold, the complexity level of the standard feature vector corresponding to the minimum weighted Euclidean distance is determined as the current road condition complexity level; if the minimum weighted Euclidean distance is greater than the preset maximum similarity threshold, it is directly determined as a level four complex road condition.

4. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 2, characterized in that, The process of classifying, labeling, and predicting the situation of obstacles in the high-precision real-time environment model specifically includes: The average motion speed of each obstacle is obtained in multiple consecutive time frames. If the average motion speed is lower than a preset static speed threshold, the obstacle is classified as a static obstacle; otherwise, it is classified as a dynamic obstacle. For static obstacles, the collision probability of each static obstacle is calculated based on the shortest relative distance between the vehicle's current planned driving path and the area occupied by the obstacle, the vehicle's current speed, and the size of the obstacle. For dynamic obstacles, based on the current position, speed and direction of motion tracked in the high-precision real-time environment model, the movement trajectory within a certain period of time is predicted to obtain the spatiotemporal position sequence of the dynamic obstacle; the spatiotemporal position sequence of the vehicle's planned driving path for a future period of time is obtained. At each prediction time step, the relative distance between the vehicle position and the predicted position of the obstacle is calculated, and the minimum relative distance and the corresponding relative speed are recorded throughout the entire prediction period. The minimum predicted collision time between the vehicle and the dynamic obstacle is calculated based on the minimum relative distance and relative speed. If the relative speed is close to zero or the trajectories of the two do not intersect, the predicted collision time is set to infinity. The minimum predicted collision time is input into a preset collision probability mapping function. The function outputs a collision probability that monotonically increases as the predicted collision time decreases, thereby generating the collision probability of the dynamic obstacle. The shortest relative distance and the minimum predicted collision time are output as the basis for calculating the predicted collision time.

5. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 4, characterized in that, The step of classifying the obstacles based on classification labels and situation prediction results to determine the obstacle risk level specifically includes: The obstacle risk levels include low, medium, high, and very high risk levels; Obtain the collision probability of obstacles; wherein the predicted collision time of static obstacles is calculated based on the shortest relative distance, the current speed of the vehicle and the preset emergency braking deceleration, according to the uniform deceleration motion model, and the predicted collision time of dynamic obstacles is the minimum predicted collision time; When the collision probability is greater than or equal to a preset high collision probability threshold and the predicted collision time is less than a preset emergency collision time threshold, the obstacle risk level is determined to be extremely high risk. When the collision probability is greater than or equal to a preset medium collision probability threshold and the predicted collision time is less than a preset safe collision time threshold, or when the collision probability is greater than or equal to a preset high collision probability threshold and the predicted collision time is greater than or equal to a preset emergency collision time threshold, the obstacle risk level is determined to be high risk. When the collision probability is greater than or equal to a preset low collision probability threshold and less than a preset medium collision probability threshold, and the predicted collision time is greater than or equal to a preset safe collision time threshold, the obstacle risk level is determined to be medium risk. When the collision probability is less than a preset low collision probability threshold, the obstacle risk level is determined to be low risk.

6. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 4, characterized in that, The automatic matching of corresponding obstacle avoidance strategies based on the road condition complexity level, and the generation of real-time driving trajectories based on the obstacle avoidance strategies and obstacle risk levels, specifically includes: Taking the vehicle's current location as the starting point and the destination as the ending point, on the high-precision real-time environment model grid map, the grids occupied by static obstacles are marked as impassable areas, and the remaining grids are marked as passable areas; Within the passable area, the search is expanded outward layer by layer from the starting point. During the expansion, the current heading of the vehicle is used as a reference. Within the turning angle range allowed by the minimum turning radius of the vehicle, multiple candidate path points for the next layer are generated according to a preset angle resolution. For each candidate path point, a comprehensive cost value is calculated. The comprehensive cost value is obtained by weighted summation of the cumulative path length, the change in path curvature, the reciprocal of the distance to the nearest static obstacle grid, and the perpendicular distance to the center line of the lane in the same direction. Each layer retains only a preset number of candidate path points with the lowest comprehensive cost value as surviving nodes, and continues to expand to the next layer until a candidate path point enters the preset reach range of the destination. The global driving path is obtained by tracing back the surviving nodes of each layer. Along the global driving path, the radius of curvature of each path point is calculated, and the safe speed for passing through curves is obtained by combining the preset comfort lateral acceleration upper limit. The smaller value between this speed and the road speed limit is taken as the target speed upper limit, and the speed of adjacent points is smoothed with the preset comfort acceleration upper limit as a constraint to generate a reference vehicle speed sequence. The system obtains the traffic rules for intersections, ramps, and merging areas, and determines the traffic priority based on the order in which the vehicle and other vehicles arrive at the conflict area. The priority is divided into priority passage, yielding, and waiting passage. If yielding is required, a constraint to decelerate to a preset yielding speed is applied. If waiting is required, a stopping waiting point is added to obtain a corrected base speed sequence. The global driving path and the corrected baseline vehicle speed sequence are used as the initial optimal global driving path; and the initial optimal global driving path is dynamically adjusted according to the obstacle risk level and road condition scenario to generate the final planned real-time driving trajectory.

7. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 6, characterized in that, The step of dynamically adjusting the initial optimal global driving path based on obstacle risk levels and road conditions to generate the final planned real-time driving trajectory specifically includes: Based on the obstacle risk level and road conditions, a preliminary obstacle avoidance trajectory is dynamically generated: When a low-risk obstacle is detected, the vehicle is laterally offset to the obstacle safety boundary point while maintaining the baseline speed, using the initial optimal global driving path as a reference line, to form the first obstacle avoidance trajectory. When a medium-risk obstacle is detected, the base vehicle speed is multiplied by a preset deceleration ratio to obtain the target vehicle speed. The lateral and longitudinal coupling fine-tuning trajectory that satisfies the safety distance and comfort acceleration constraints is then solved as the second obstacle avoidance trajectory. When a high-risk or extremely high-risk obstacle is detected, the system will decelerate at a preset maximum comfort deceleration to determine whether there is a collision-free stopping point before stopping. If there is, an emergency braking stopping trajectory along the center line of the current lane will be generated as the third obstacle avoidance trajectory. If the parking spot does not exist, check if there is an unoccupied continuous space with a length greater than the length of the vehicle within a preset distance in front of the adjacent lane, and check whether the relative distance and relative speed of the vehicles approaching from behind in the adjacent lane meet the preset safe lane-changing conditions; if both are met, generate an emergency lane-changing detour trajectory and apply the preset maximum comfort deceleration as the fourth obstacle avoidance trajectory; otherwise, continue to stop along the current lane with the preset maximum comfort deceleration and trigger an audible and visual warning to form the fifth obstacle avoidance trajectory; The initial obstacle avoidance trajectory and the initial optimal global driving path are fused by deviation regression to generate the final planned real-time driving trajectory.

8. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 7, characterized in that, The step of performing deviation regression fusion on the preliminary obstacle avoidance trajectory and the initial optimal global driving path to generate the final planned real-time driving trajectory specifically includes: The vehicle's current location information is acquired in real time. The vertical distance from the vehicle's location point to the initial optimal global driving path is calculated as the lateral position deviation. At the same time, the difference between the vehicle's current heading angle and the tangential angle of the corresponding point on the initial optimal global driving path is calculated as the heading angle deviation. On the initial optimal global driving path, a preset forward aiming distance is intercepted from the vehicle's current position along the path direction to obtain the coordinates of the forward point. The lateral offset between the vehicle's current heading extension line and the forward point is calculated as the forward lateral deviation. Based on the forward lateral deviation and heading angle deviation, combined with a preset trajectory tracking gain, a smooth regression trajectory is generated to regress to the initial optimal global driving path. The smooth regression trajectory satisfies the condition of continuous curvature and lateral acceleration not exceeding a preset comfort upper limit throughout. The smooth regression trajectory is compared point by point in time and space with the initial obstacle avoidance driving trajectory. If there is no conflict, the two are weighted and fused to obtain the fused trajectory. If there is a conflict, a compromise trajectory that simultaneously satisfies the obstacle safety distance constraint and the global path regression requirement is selected as the fused trajectory. If there is no feasible compromise trajectory, the initial obstacle avoidance driving trajectory is followed completely until the corresponding obstacle risk level is reduced to below medium risk, and then the smooth regression trajectory is switched. The fused trajectory or the final switched trajectory is used as the final planned real-time driving trajectory and output for use by the vehicle control actuator.

9. The autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in claim 4, characterized in that, The step of dynamically adjusting the update frequency of the high-precision real-time environment model based on the road condition complexity level and obstacle risk level specifically includes: When the road condition complexity level is level four complex road condition, or there are obstacles of extremely high risk level, the acquisition frame rate of the LiDAR and high-definition camera will be increased to the maximum hardware support value, and the update frequency of the high-precision real-time environment model will be set to the preset upper limit value. Otherwise, based on the current road condition complexity level, the corresponding baseline update frequency is queried from the preset baseline update frequency mapping table; the baseline update frequency mapping table stores the baseline update frequency values ​​corresponding to the four levels of complex road conditions: Level 1, Level 2, Level 3, and Level 4. Based on the highest risk level among all obstacles, the corresponding update frequency adjustment coefficient is queried from the preset frequency adjustment coefficient mapping table. The benchmark update frequency is multiplied by the update frequency adjustment coefficient to obtain the target update frequency. The obstacle risk level and frequency adjustment coefficient mapping table stores the frequency adjustment coefficient values ​​corresponding to the four levels of low risk, medium risk, high risk and extremely high risk. The update frequency of the high-precision real-time environment model is dynamically adjusted based on the target update frequency.

10. An autonomous driving planning system for adaptive obstacle avoidance in complex road conditions, applied in the autonomous driving planning method for adaptive obstacle avoidance in complex road conditions as described in any one of claims 1-9, characterized in that, include: The multi-source fusion modeling module is used to collect multi-source environmental data around the vehicle in real time through vehicle-mounted sensors, and to fuse and process the multi-source sensor data to build a high-precision real-time environmental model. The multi-source environmental data includes road structure information, obstacle information, vehicle status information, and road condition and traffic flow information. The vehicle-mounted sensors include lidar, high-definition camera, millimeter-wave radar, body sensors, and vehicle positioning module. The road condition complexity classification module is used to intelligently classify and identify the current driving road conditions based on the high-precision real-time environment model and the preset road condition feature database, and determine the road condition complexity level; the road condition feature database is a set of multi-dimensional feature parameters that characterize the road condition complexity under different typical driving scenarios. The obstacle risk assessment module is used to classify, label, and predict the situation of obstacles in the high-precision real-time environment model, and to classify the risk level of the obstacles based on the classification, labeling, and situation prediction results, thereby determining the obstacle risk level. The adaptive obstacle avoidance control module is used to automatically match the corresponding planned obstacle avoidance strategy according to the road condition complexity level, generate a real-time driving trajectory based on the planned obstacle avoidance strategy and obstacle risk level, output vehicle control commands according to the real-time driving trajectory to drive the vehicle to perform obstacle avoidance driving, and dynamically adjust the update frequency of the high-precision real-time environment model according to the road condition complexity level and obstacle risk level to realize the dynamic update and adaptive adjustment of the real-time driving trajectory.

Images