Automatic driving vehicle obstacle avoidance method and device based on road condition perception

Abstract

The invention discloses an automatic driving vehicle obstacle avoidance method and device based on road condition perception, and relates to the technical field of automatic driving obstacle avoidance, and the method comprises the steps: collecting the multi-dimensional perception data of a vehicle driving environment in real time, and generating a road obstacle feature information set; constructing an obstacle movement track prediction model, and carrying out collision risk quantitative evaluation to obtain a dynamic risk distribution map; a multi-mode obstacle avoidance decision tree is generated, and strategy feasibility verification is carried out; and generating a control instruction set based on the verified obstacle avoidance strategy, driving the vehicle to execute an obstacle avoidance action and feeding back execution state data in real time. The technical problem that the obstacle avoidance safety and reliability are low due to the fact that an automatic driving vehicle incomprehensively senses the road environment under the complex road condition in the prior art is solved, and the technical effects that efficient obstacle avoidance of the automatic driving vehicle under the complex road condition is achieved, and the driving safety and reliability are improved are achieved.

Description

Technical Field

[0001] This invention relates to the field of obstacle avoidance technology for autonomous driving, and more specifically to an obstacle avoidance method and apparatus for autonomous vehicles based on road condition perception. Background Technology

[0002] In the field of autonomous driving technology, vehicles need to perceive their surroundings and make obstacle avoidance decisions in a timely and accurate manner when driving in complex road environments. Existing technologies for obstacle avoidance in some autonomous vehicles suffer from problems such as a single dimension of road environment perception, insufficient accuracy in predicting obstacle trajectories, a lack of multi-objective optimization in obstacle avoidance strategy generation, and imperfect feasibility verification mechanisms. These issues can lead to untimely obstacle avoidance, unreasonable strategies, or execution deviations when vehicles encounter sudden obstacles or complex traffic scenarios, affecting the safety and reliability of autonomous driving.

[0003] Existing technologies suffer from the technical problem that autonomous vehicles do not have a comprehensive perception of the road environment in complex road conditions, resulting in low obstacle avoidance safety and reliability. Summary of the Invention

[0004] This application provides an obstacle avoidance method and apparatus for autonomous vehicles based on road condition perception, which is used to address the technical problem in the prior art that autonomous vehicles do not have comprehensive perception of the road environment in complex road conditions, resulting in low obstacle avoidance safety and reliability.

[0005] In view of the above problems, this application provides an obstacle avoidance method and device for autonomous vehicles based on road condition perception.

[0006] The first aspect of this application provides an obstacle avoidance method for autonomous vehicles based on road condition perception, the method comprising:

[0007] The system collects multi-dimensional perception data of the vehicle's driving environment in real time to generate a set of road obstacle feature information; it constructs an obstacle motion trajectory prediction model, performs a collision risk quantification assessment based on the road obstacle feature information set, and obtains a dynamic risk distribution map; it generates a multimodal obstacle avoidance decision tree based on the dynamic risk distribution map, and verifies the feasibility of the strategy by combining vehicle dynamic parameters; it generates a control command set based on the verified obstacle avoidance strategy, drives the vehicle to perform obstacle avoidance actions, and provides real-time feedback on the execution status data.

[0008] A second aspect of this application provides an obstacle avoidance device for autonomous vehicles based on road condition perception, the device comprising:

[0009] The system includes a multi-dimensional data acquisition module for real-time acquisition of multi-dimensional perception data of the vehicle's driving environment and generation of a road obstacle feature information set; a dynamic risk distribution map acquisition module for constructing an obstacle motion trajectory prediction model, performing a collision risk quantification assessment based on the road obstacle feature information set, and obtaining a dynamic risk distribution map; a feasibility verification module for generating a multi-modal obstacle avoidance decision tree based on the dynamic risk distribution map and verifying the feasibility of the strategy in conjunction with vehicle dynamic parameters; and a control command set generation module for generating a control command set based on the verified obstacle avoidance strategy, driving the vehicle to perform obstacle avoidance actions and providing real-time feedback on execution status data.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] This system collects multi-dimensional perception data of the vehicle's driving environment in real time to generate a road obstacle feature information set; constructs an obstacle motion trajectory prediction model; performs a collision risk quantification assessment based on the road obstacle feature information set to obtain a dynamic risk distribution map; generates a multimodal obstacle avoidance decision tree based on the dynamic risk distribution map; and verifies the feasibility of the strategy by combining vehicle dynamic parameters; based on the verified obstacle avoidance strategy, generates a control command set to drive the vehicle to execute obstacle avoidance actions and provides real-time feedback on execution status data. This achieves the technical effect of enabling autonomous vehicles to efficiently avoid obstacles in complex road conditions, improving driving safety and reliability. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of the obstacle avoidance method for autonomous vehicles based on road condition perception provided in an embodiment of this application;

[0014] Figure 2 This is a schematic diagram of the obstacle avoidance device for autonomous vehicles based on road condition perception, provided in an embodiment of this application.

[0015] Figure labeling: Multidimensional data acquisition module 10, dynamic risk distribution map acquisition module 20, feasibility verification module 30, control instruction set generation module 40. Detailed Implementation

[0016] This application provides an obstacle avoidance method and device for autonomous vehicles based on road condition perception, which addresses the technical problem in the prior art where autonomous vehicles have incomplete perception of the road environment in complex road conditions, resulting in low obstacle avoidance safety and reliability.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides an obstacle avoidance method for autonomous vehicles based on road condition perception, the method comprising:

[0019] Step S100: Collect multi-dimensional perception data of the vehicle's driving environment in real time and generate a set of road obstacle feature information.

[0020] Specifically, LiDAR point cloud data, camera visual data, and millimeter-wave radar detection data are acquired simultaneously to form a raw perception dataset. This raw perception dataset undergoes spatiotemporal alignment processing to generate a 3D environment fusion matrix. Obstacle contour features, motion state features, and material reflection features are extracted from the 3D environment fusion matrix to construct a road obstacle feature information set.

[0021] Step S200: Construct an obstacle motion trajectory prediction model, perform a collision risk quantification assessment based on the road obstacle feature information set, and obtain a dynamic risk distribution map.

[0022] Specifically, a long short-term memory network model is first established and historical trajectory sequences and environmental context features are input. Then, a set of virtual obstacle motion scenarios is constructed using an adversarial generative network to pre-train the model. After that, the model parameters are corrected online using real road test data to generate an obstacle motion trajectory prediction model. Based on this model, the collision risk is quantitatively assessed by combining the obstacle outline, motion state, and material reflection features in the road obstacle feature information set, and finally, a dynamic risk distribution map is obtained.

[0023] Step S300: Generate a multimodal obstacle avoidance decision tree based on the dynamic risk distribution map, and verify the feasibility of the strategy by combining vehicle dynamics parameters.

[0024] Specifically, the dynamic risk distribution map is divided into three response zones: an emergency avoidance zone, a strategy adjustment zone, and a safety monitoring zone. In the emergency avoidance zone, a path replanning module is activated to generate an emergency vehicle control strategy. In the strategy adjustment zone, a trajectory optimization module is initiated to optimize driving parameters. In the safety monitoring zone, a continuous monitoring mode is activated to maintain the current driving state, thus generating a multimodal obstacle avoidance decision tree. The obstacle avoidance strategies generated by the decision tree (such as steering angle, braking intensity, and acceleration) are then input into a vehicle dynamics model. This model includes parameters such as vehicle mass, tire characteristics, and suspension system. The vehicle's motion state during strategy execution is simulated by solving differential equations. Simultaneously, physical constraints are set, such as the steering angle not exceeding the vehicle's maximum steering angle, braking deceleration within the allowable range of tire-road adhesion coefficient, and acceleration change rate not exceeding the passenger comfort threshold. The simulation results are then validated. If the vehicle's motion state exceeds the physical constraints or fails to meet obstacle avoidance safety requirements after strategy execution, the strategy is deemed infeasible, and the process returns to the decision tree to regenerate a new strategy. If all constraints are met and obstacle avoidance is successful, the validation is passed, ensuring the generated obstacle avoidance strategy is feasible at the actual vehicle dynamics level.

[0025] Step S400: Generate a control instruction set based on the verified obstacle avoidance strategy, drive the vehicle to perform obstacle avoidance actions and provide real-time feedback on the execution status data.

[0026] Specifically, a control command set is generated based on the validated obstacle avoidance strategy. This command set is used to drive the vehicle to perform specific obstacle avoidance actions. Simultaneously, the status data of the vehicle's obstacle avoidance actions is collected and fed back in real time to form a closed-loop control. In this process, a closed-loop verification mechanism for the obstacle avoidance control commands is also established. The response delay data and steering torque deviation rate of the vehicle chassis actuators are acquired in real time. These data are compared with a first preset threshold and a second preset threshold, respectively. If the response delay data exceeds the first preset threshold or the steering torque deviation rate exceeds the second preset threshold, the redundant control module is activated to generate compensation control commands. The compensation results are then fed back to the closed-loop verification mechanism for dynamic correction, ensuring the accuracy and reliability of the obstacle avoidance actions.

[0027] In one possible implementation, step S100 further includes:

[0028] Step S110: Simultaneously acquire lidar point cloud data, camera visual data, and millimeter-wave radar detection data to form the original perception dataset.

[0029] Step S120: Perform spatiotemporal alignment processing on the original perception dataset to generate a three-dimensional environment fusion matrix.

[0030] Step S130: Extract obstacle contour features, motion state features, and material reflection features from the three-dimensional environment fusion matrix to construct the road obstacle feature information set.

[0031] Specifically, vehicle driving environment data is collected synchronously by sensors such as LiDAR, cameras, and millimeter-wave radar. LiDAR acquires environmental point cloud data to present a three-dimensional spatial structure, cameras collect visual images to identify target appearance features, and millimeter-wave radar detects target distance, speed, and other information. These multi-source data together constitute a raw perception dataset containing multi-dimensional environmental information, providing a foundation for subsequent obstacle feature extraction.

[0032] Because the time reference and spatial coordinate system of data collected by sensors such as LiDAR, cameras, and millimeter-wave radar differ, spatiotemporal alignment processing is required for the original perception dataset. First, a time synchronization algorithm is used to eliminate timestamp discrepancies in the data from each sensor, ensuring data consistency in the time dimension. Then, a coordinate transformation matrix is ​​used to uniformly transform the spatial coordinate systems of different sensors to the vehicle's global coordinate system. Finally, the calibrated multi-source data is fused to generate a 3D environment fusion matrix. This matrix integrates point cloud, visual, and radar detection information from the same spatiotemporal context, providing accurate environmental representation for subsequent obstacle feature extraction.

[0033] From the 3D environment fusion matrix, feature extraction algorithms are used to identify and extract the contour features of obstacles, such as shape and size, to clarify their geometric form. Simultaneously, motion state features, including velocity, acceleration, and direction of motion, are acquired to understand the obstacle's movement trend. Furthermore, material reflection features are extracted, such as lidar reflectivity, color and texture features from camera images, and echo intensity from millimeter-wave radar, to determine the obstacle's material properties. These obstacle contour features, motion state features, and material reflection features extracted from the 3D environment fusion matrix are integrated to construct a road obstacle feature information set, providing crucial data support for subsequent obstacle trajectory prediction and collision risk assessment.

[0034] In one possible implementation, step S200 further includes:

[0035] Step S210: Establish a long short-term memory network model by inputting historical trajectory sequences and environmental context features.

[0036] Step S220: Pre-train the Long Short-Term Memory Network model by constructing a set of virtual obstacle motion scenarios through a Generative Adversarial Network.

[0037] Step S230: Use real road test data to correct the parameters of the long short-term memory network model online, and generate the obstacle motion trajectory prediction model.

[0038] Specifically, a Long Short-Term Memory (LSTM) network model is constructed. This model can handle long-term dependencies in time-series data. It takes the historical trajectory sequence of obstacles, such as their position coordinates, velocity, and acceleration over a historical time period, as well as environmental context features such as surrounding road environment characteristics, traffic rule constraints, and the states of other traffic participants as inputs, providing a model foundation for subsequent prediction of obstacle movement trajectories.

[0039] By leveraging the game-like mechanism of generators and discriminators in Generative Adversarial Networks (GANs), a set of virtual obstacle motion scenarios is constructed, encompassing diverse traffic scenarios (such as urban roads, highways, and complex intersections), different weather conditions (rain, fog, and sunshine), various obstacle types (pedestrians, motor vehicles, and non-motorized vehicles), and multiple motion patterns (uniform speed, acceleration, turning, and lane changing). This set of virtual scenarios can simulate rare or dangerous traffic conditions in reality. By pre-training a Long Short-Term Memory (LSTM) network model with these scenarios, the model can learn the motion patterns and characteristics of obstacles in different situations before encountering real-world complex scenarios, thereby improving the model's adaptability to complex environments and its generalization ability for trajectory prediction.

[0040] In real-world road testing scenarios, real-time road test data containing the actual motion trajectories of obstacles is collected. This data encompasses dynamic information such as the position, speed, and acceleration of obstacles under different road conditions, traffic flow, and weather conditions. This real-world data is input into a Long Short-Term Memory (LSTM) network model pre-trained in a virtual scenario. An online learning mechanism is used to correct the model parameters in real time. By comparing the deviation between the model's predicted trajectory and the actual trajectory, the network weights are optimized based on the backpropagation algorithm. This allows the model to gradually adapt to the uncertainty and complexity of obstacle motion in real-world roads, ultimately generating an obstacle motion trajectory prediction model with high robustness and accurate prediction capabilities.

[0041] In one possible implementation, step S300 further includes:

[0042] Step S310: Divide the dynamic risk distribution map into three levels of response areas: emergency avoidance zone, strategy adjustment zone, and safety monitoring zone.

[0043] Step S320: Activate the path replanning module for the emergency avoidance zone to generate a vehicle emergency control strategy.

[0044] Step S330: Activate the trajectory optimization module in the strategy adjustment area to perform driving parameter optimization and adjustment.

[0045] Step S340: Enable continuous monitoring mode for the safety monitoring zone to maintain the current driving status.

[0046] Specifically, based on the level and urgency of collision risk, the dynamic risk distribution map is divided into three response zones: an emergency avoidance zone, a strategy adjustment zone, and a safety monitoring zone. The emergency avoidance zone is a high-risk area; when an obstacle is in this zone, the risk of collision between the vehicle and the obstacle is extremely high, requiring immediate avoidance measures. The strategy adjustment zone is a medium-risk area; in this zone, there is a potential collision risk between the obstacle and the vehicle, requiring optimization and adjustment of driving parameters to prevent the risk. The safety monitoring zone is a low-risk area; obstacles in this zone pose a relatively small threat to vehicle safety, and their status only needs continuous monitoring. This hierarchical division allows for differentiated responses to obstacles of different risk levels, improving the targeting and effectiveness of obstacle avoidance strategies.

[0047] When an obstacle in the dynamic risk distribution map is located in the emergency avoidance zone, the path replanning module is immediately activated. This module, based on the vehicle's current dynamic parameters (such as speed and steering angle), obstacle contours and trajectory data from the road obstacle feature information set, and the surrounding vehicle status obtained through real-time V2X communication, replans a safe obstacle avoidance path using a global path search algorithm (such as the A algorithm or the RRT algorithm). Simultaneously, the path replanning module generates a vehicle emergency control strategy that includes actions such as emergency braking and extreme steering. This strategy must be verified by vehicle dynamic parameters (such as whether the steering torque exceeds the vehicle's limits and whether the braking distance meets the safety threshold) to ensure the physical feasibility of the emergency control strategy, ultimately driving the vehicle to perform rapid obstacle avoidance maneuvers.

[0048] When an obstacle is within the strategy adjustment zone, the trajectory optimization module is activated, employing either Model Predictive Control (MPC) or Quadratic Programming (QP) algorithms to optimize the driving trajectory and parameters. Using the vehicle's current dynamic state (position, velocity, acceleration) and obstacle trajectory prediction data as input, an optimization objective function is constructed within a finite prediction time domain. This function includes obstacle avoidance safety constraints (such as minimum safe distance), comfort constraints (such as maximum jerk), and energy consumption constraints (such as motor power loss). Through rolling optimization, the optimal control quantity for the current moment is solved, enabling dynamic adjustment of driving parameters such as speed, acceleration, and steering angle. This allows the vehicle to avoid obstacles while controlling longitudinal acceleration / deceleration fluctuations within 0.5 m / s². 2 The rate of change of internal and lateral acceleration is controlled within 0.3 m / s². 3 Within this range, a balance between passenger comfort and energy efficiency is ensured.

[0049] When an obstacle is within the safe monitoring zone, a continuous monitoring mode is activated. Multiple sensors, including LiDAR, cameras, and millimeter-wave radar, collect real-time data on the obstacle's motion (such as position, speed, and acceleration) and environmental context (such as road curvature and surrounding vehicle dynamics). This data, combined with traffic flow data obtained through V2X communication, continuously tracks and analyzes the relative positional relationship and motion trends between the obstacle and the vehicle. Provided the obstacle poses no direct threat to vehicle safety, the vehicle's current speed, steering angle, and other state parameters remain unchanged. Simultaneously, the monitored real-time data is input into a dynamic risk distribution map for periodic updates. If the obstacle's risk level rises to the strategy adjustment zone or emergency avoidance zone, the corresponding obstacle avoidance response mechanism is immediately triggered, ensuring continuous safety monitoring even in low-risk scenarios.

[0050] In one possible implementation, step S300 further includes:

[0051] Step S350: Construct a reinforcement learning reward function and build a multi-objective evaluation system based on obstacle avoidance success rate, passenger comfort, and energy efficiency.

[0052] Step S360: Based on the multi-objective evaluation system, iteratively train the decision tree parameters in a virtual simulation environment to generate an optimized multimodal obstacle avoidance decision tree model.

[0053] Step S370: Based on the verification results of the multimodal obstacle avoidance decision tree model in the simulation environment, establish a strategy effect mapping relationship library.

[0054] Step S380: Dynamically update the weight parameters of the multimodal obstacle avoidance decision tree model based on the deviation between the strategy execution effect in the actual obstacle avoidance process and the strategy effect mapping relationship library.

[0055] Specifically, when constructing the reinforcement learning reward function, a multi-objective weighted summation algorithm is used to implement a multi-objective evaluation system based on obstacle avoidance success rate, passenger comfort, and energy efficiency. Specifically, it is constructed as follows: First, a reward term for obstacle avoidance success rate is defined. When the minimum distance between the vehicle and the obstacle is greater than a safety threshold and no collision occurs, a positive reward value R is given. safe Conversely, negative rewards are given for positive results; secondly, reward items for passenger comfort are designed, and the longitudinal acceleration 'a' of the vehicle during operation is collected. x and lateral acceleration a y Set the comfort threshold range [a] x-min , a x-max ] and [a y-min , a y-max The penalty value R is calculated based on the degree to which the acceleration deviates from the threshold. comfortThe greater the deviation, the greater the penalty; then, determine the reward for energy efficiency, using energy consumption E per unit distance as the indicator, with lower energy consumption resulting in higher rewards, and set the reward function R. energy The energy consumption is a decreasing function; finally, the three reward terms are linearly combined using weighted coefficients W1, W2, and W3 (W1 + W2 + W3 = 1) to form the final reinforcement learning reward function R = W1R. safe +W2R comfort +W3R energy This enables multi-objective quantitative evaluation of obstacle avoidance strategies.

[0056] Based on the established multi-objective evaluation system, a multimodal obstacle avoidance decision tree is iteratively trained in a virtual simulation environment. Virtual environments with diverse scenarios, including urban roads, highways, and complex intersections, are constructed using simulation platforms such as Unity or CARLA to simulate different weather conditions, traffic flow, and the movement of obstacle types (e.g., pedestrians, motor vehicles, and non-motorized vehicles). The multimodal obstacle avoidance decision tree model is integrated into the simulation environment. For each simulated scenario, the decision tree outputs an obstacle avoidance strategy, and the evaluation system calculates the reward value based on indicators such as obstacle avoidance success rate, passenger comfort, and energy efficiency. The parameters of the decision tree are iteratively optimized using reinforcement learning algorithms (e.g., PPO, DDPG), continuously adjusting the conditional thresholds and branch weights of decision nodes. This allows the model to learn the optimal obstacle avoidance strategy through multiple training iterations, ultimately generating an optimized multimodal obstacle avoidance decision tree model that performs better under multi-objective balance.

[0057] After fully validating the optimized multimodal obstacle avoidance decision tree model in a virtual simulation environment, decision strategies and corresponding multi-objective evaluation results were collected under different test scenarios to establish a strategy-effect mapping relationship library. For various scenarios such as urban roads, highways, and complex intersections, and under different weather conditions such as sunny, rainy, and foggy days, the obstacle avoidance strategies output by the decision tree (such as emergency braking, steering avoidance, and trajectory optimization) were recorded. Simultaneously, evaluation results such as obstacle avoidance success rate, passenger comfort indicators (such as acceleration changes), and energy efficiency data (such as energy consumption per unit distance) were collected. The strategies and multi-objective effect data were associated and stored to form a strategy-effect mapping relationship library. This library serves as a reference benchmark for subsequent strategy optimization, comparing the strategy execution effects in actual obstacle avoidance processes and providing data support for dynamically updating the weight parameters of the decision tree model.

[0058] During actual road obstacle avoidance, real-time data on the obstacle avoidance strategies executed by vehicles and their corresponding actual effects are collected, including obstacle avoidance success rate, passenger comfort index, and energy efficiency. These actual performance results are compared with the expected results in the strategy effect mapping database, and the deviation is calculated, quantified using methods such as mean squared error or absolute error. Based on the magnitude of the deviation, a parameter update algorithm from reinforcement learning (such as gradient descent) is used to dynamically adjust the weight parameters of each objective function in the multimodal obstacle avoidance decision tree model. If the actual obstacle avoidance success rate is lower than expected, the weight corresponding to the obstacle avoidance success rate is increased; if the passenger comfort index does not meet expectations, the comfort weight is adjusted, and so on, continuously optimizing the model in practical applications to improve the adaptability and effectiveness of the obstacle avoidance strategy.

[0059] In one possible implementation, step S310 further includes:

[0060] Step S311: Receive V2X communication data in real time and extract surrounding vehicle status information and traffic control information.

[0061] Step S312: The V2X communication data and vehicle perception data are fused with confidence weighting to generate global road state data.

[0062] Step S313: Correct the dynamic risk distribution map based on the global road status data.

[0063] Specifically, the vehicle-mounted V2X communication module receives real-time communication data from surrounding vehicles and infrastructure (such as traffic lights and roadside units), extracting status information such as the location, speed, acceleration, and direction of travel of surrounding vehicles, as well as traffic control information (such as speed limit instructions, restricted areas, and traffic incident warnings), providing external data support for subsequent road condition analysis.

[0064] When performing confidence-weighted fusion of V2X communication data and vehicle-mounted perception data, firstly, based on factors such as the reliability of the data source, transmission signal strength, and historical data accuracy, corresponding confidence weights are assigned to both V2X communication data (e.g., surrounding vehicle status, traffic control information) and vehicle-mounted perception data (LiDAR point clouds, camera vision, millimeter-wave radar detection data). Then, a weighted fusion algorithm is used to process the two types of data. For example, for the location information of the same obstacle, if the confidence weight of the V2X data is 0.6 and the confidence weight of the vehicle-mounted perception data is 0.4, the fused location information is obtained through weighted calculation. This eliminates data inconsistencies and integrates complementary information, ultimately generating global road state data that includes road obstacle distribution, traffic flow status, and environmental characteristics, providing comprehensive and accurate data support for subsequent correction of the dynamic risk distribution map.

[0065] When revising the dynamic risk distribution map based on the generated global road state data, the first step is to compare the real-time location, trajectory, type attributes (such as size, speed, and acceleration) of obstacles in the global road state data with the corresponding information in the original dynamic risk distribution map. For new obstacle locations and movement trends confirmed through the fusion of V2X communication data and vehicle-mounted perception data, their coordinates and risk levels in the map are updated. For traffic control information, such as speed limits ahead or the presence of temporary construction areas, the collision risk of the area is reassessed and the risk level is adjusted. In this way, the dynamic risk distribution map can more accurately reflect the actual risk situation of the current road, providing a more reliable basis for the subsequent generation of a multimodal obstacle avoidance decision tree.

[0066] In one possible implementation, step S400 further includes:

[0067] Step S410: Establish a closed-loop verification mechanism for obstacle avoidance control commands and obtain real-time response delay data and steering torque deviation rate of the vehicle chassis execution unit.

[0068] Step S420: Compare the response delay data with a first preset threshold, and at the same time compare the steering torque deviation rate with a second preset threshold.

[0069] Step S430: When the response delay data exceeds the first preset threshold or the steering torque deviation rate exceeds the second preset threshold, the redundant control module is activated to generate a compensation control command, and the compensation result is fed back to the closed-loop verification mechanism for dynamic correction.

[0070] Specifically, a closed-loop verification mechanism for obstacle avoidance control commands is established. This involves deploying high-precision sensors and data acquisition modules in the vehicle chassis actuators (such as braking, steering, and drive systems) to collect real-time data on the response delay of control commands and the steering torque deviation rate. The response delay data is recorded by a timer, tracking the time interval from the issuance of the obstacle avoidance control command to the start of the actuator's action, and is used to evaluate the real-time performance of the actuator system. The steering torque deviation rate is calculated by using a torque sensor to obtain the difference between the actual output torque and the torque required by the command, and then calculating the ratio to the commanded torque, reflecting the execution accuracy. This data is transmitted in real-time to the central processing unit via the vehicle network, providing accurate real-time data support for subsequent verification of the effectiveness of control commands and dynamic correction.

[0071] The real-time acquired response delay data of the vehicle chassis actuator is compared with a pre-set first threshold, and the steering torque deviation rate is compared with a second preset threshold. The first preset threshold is set according to the real-time requirements of the vehicle control system and is used to determine whether the actuator's response to obstacle avoidance control commands is timely; the second preset threshold is determined based on the precision standards of the vehicle steering system and is used to measure whether the deviation between the actual steering torque and the command requirement is within the allowable range. By synchronously comparing these two data points with their corresponding thresholds, the normal operating status of the chassis actuator can be accurately determined, providing a crucial basis for whether to trigger the redundant control module subsequently.

[0072] When the response delay data exceeds a first preset threshold or the steering torque deviation rate exceeds a second preset threshold, the redundant control module is activated and a compensation control command is generated through the following means: First, after the central control unit detects abnormal data, it immediately triggers the hardware or software redundancy mechanism of the redundant control module, such as activating the backup controller or switching to the backup control algorithm. The redundant control module calculates the corresponding compensation amount based on a preset compensation model (a compensation algorithm based on PID control), combined with the type and degree of the current anomaly (such as response delay duration and torque deviation ratio), and generates a compensation control command containing correction parameters. This command is sent to the chassis execution unit through a redundant communication link independent of the main control channel, forcibly adjusting the timing of the execution action or the torque output value. Simultaneously, after compensation, the sensors collect new response delay data and steering torque deviation rate in real time. These are compared with the abnormal data before compensation through a closed-loop verification mechanism. If a deviation still exists, the compensation parameters are iteratively optimized until the data returns to the threshold range, ensuring the execution accuracy and real-time performance of the obstacle avoidance control command.

[0073] Example 2 is based on the same inventive concept as the obstacle avoidance method for autonomous vehicles based on road condition perception in the previous examples, such as... Figure 2 As shown, this application provides an obstacle avoidance device for autonomous vehicles based on road condition perception. The device and method embodiments in this application are based on the same inventive concept. The device includes:

[0074] The multi-dimensional data acquisition module 10 is used to collect multi-dimensional perception data of the vehicle's driving environment in real time and generate a set of road obstacle feature information.

[0075] The dynamic risk distribution map acquisition module 20 is used to construct an obstacle motion trajectory prediction model, perform a collision risk quantification assessment based on the road obstacle feature information set, and obtain a dynamic risk distribution map.

[0076] The feasibility verification module 30 is used to generate a multimodal obstacle avoidance decision tree based on the dynamic risk distribution map and to verify the feasibility of the strategy in combination with vehicle dynamic parameters.

[0077] The control instruction set generation module 40 is used to generate a control instruction set based on the verified obstacle avoidance strategy, drive the vehicle to perform obstacle avoidance actions and provide real-time feedback on the execution status data.

[0078] Furthermore, the device is also used to perform the following functions:

[0079] Simultaneously acquire lidar point cloud data, camera visual data, and millimeter-wave radar detection data to form an original perception dataset; perform spatiotemporal alignment processing on the original perception dataset to generate a three-dimensional environment fusion matrix; extract obstacle contour features, motion state features, and material reflection features from the three-dimensional environment fusion matrix to construct the road obstacle feature information set.

[0080] Furthermore, the device is also used to perform the following functions:

[0081] A long short-term memory network model is established, and historical trajectory sequences and environmental context features are input. A virtual obstacle motion scene set is constructed through an adversarial generative network to pre-train the long short-term memory network model. The parameters of the long short-term memory network model are corrected online using real road test data to generate the obstacle motion trajectory prediction model.

[0082] Furthermore, the device is also used to perform the following functions:

[0083] The dynamic risk distribution map is divided into three response zones: an emergency avoidance zone, a strategy adjustment zone, and a safety monitoring zone. A path replanning module is activated in the emergency avoidance zone to generate an emergency vehicle control strategy. A trajectory optimization module is initiated in the strategy adjustment zone to optimize and adjust driving parameters. A continuous monitoring mode is enabled in the safety monitoring zone to maintain the current driving state.

[0084] Furthermore, the device is also used to perform the following functions:

[0085] A reinforcement learning reward function is constructed, and a multi-objective evaluation system is built based on obstacle avoidance success rate, passenger comfort, and energy efficiency. Based on the multi-objective evaluation system, the decision tree parameters are iteratively trained in a virtual simulation environment to generate an optimized multimodal obstacle avoidance decision tree model. Based on the verification results of the multimodal obstacle avoidance decision tree model in the simulation environment, a strategy effect mapping relationship library is established. According to the deviation between the strategy execution effect in the actual obstacle avoidance process and the strategy effect mapping relationship library, the weight parameters of the multimodal obstacle avoidance decision tree model are dynamically updated.

[0086] Furthermore, the device is also used to perform the following functions:

[0087] Real-time V2X communication data is received to extract surrounding vehicle status information and traffic control information; the V2X communication data and vehicle-mounted perception data are fused with confidence weighting to generate global road status data; and the dynamic risk distribution map is corrected based on the global road status data.

[0088] Furthermore, the device is also used to perform the following functions:

[0089] A closed-loop verification mechanism for obstacle avoidance control commands is established to acquire real-time response delay data and steering torque deviation rate of the vehicle chassis execution unit; the response delay data is compared with a first preset threshold, and the steering torque deviation rate is compared with a second preset threshold; when the response delay data exceeds the first preset threshold or the steering torque deviation rate exceeds the second preset threshold, the redundant control module is activated to generate compensation control commands, and the compensation results are fed back to the closed-loop verification mechanism for dynamic correction.

[0090] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0091] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0092] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. An obstacle avoidance method for autonomous vehicles based on road condition perception, characterized in that, The method comprises: Real-time collection of multi-dimensional perception data of the vehicle's driving environment to generate a set of road obstacle feature information; Construct an obstacle motion trajectory prediction model, perform a collision risk quantification assessment based on the road obstacle feature information set, and obtain a dynamic risk distribution map; A multimodal obstacle avoidance decision tree is generated based on the dynamic risk distribution map, and the feasibility of the strategy is verified by combining vehicle dynamics parameters. Based on the verified obstacle avoidance strategy, a set of control instructions is generated to drive the vehicle to perform obstacle avoidance actions and provide real-time feedback on the execution status data.

2. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 1, characterized in that, Real-time acquisition of multi-dimensional perception data of the vehicle's driving environment generates a set of road obstacle feature information, including: Simultaneously acquire lidar point cloud data, camera visual data, and millimeter-wave radar detection data to form the original perception dataset; The original sensor dataset is spatiotemporally aligned to generate a 3D environment fusion matrix; The obstacle contour features, motion state features, and material reflection features are extracted from the three-dimensional environment fusion matrix to construct the road obstacle feature information set.

3. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 1, characterized in that, Constructing an obstacle motion trajectory prediction model, including: Establish a long short-term memory network model by inputting historical trajectory sequences and environmental context features; The Long Short-Term Memory Network model is pre-trained by constructing a set of virtual obstacle motion scenarios using a Generative Adversarial Network. The parameters of the Long Short-Term Memory Network model are corrected online using real road test data to generate the obstacle motion trajectory prediction model.

4. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 1, characterized in that, A multimodal obstacle avoidance decision tree is generated based on the dynamic risk distribution map, including: The dynamic risk distribution map is divided into three levels of response areas: emergency avoidance zone, strategy adjustment zone, and safety monitoring zone. The path replanning module is activated for the emergency avoidance zone to generate a vehicle emergency control strategy; The trajectory optimization module is activated in the strategy adjustment area to optimize and adjust driving parameters. Continuous monitoring mode is enabled for the safety monitoring zone to maintain the current driving status.

5. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 4, characterized in that, The method further includes: Construct a reinforcement learning reward function and build a multi-objective evaluation system based on obstacle avoidance success rate, passenger comfort, and energy efficiency; Based on the multi-objective evaluation system, the decision tree parameters are iteratively trained in a virtual simulation environment to generate an optimized multimodal obstacle avoidance decision tree model. Based on the verification results of the multimodal obstacle avoidance decision tree model in the simulation environment, a strategy effect mapping relationship library is established. The weight parameters of the multimodal obstacle avoidance decision tree model are dynamically updated based on the deviation between the actual strategy execution effect and the strategy effect mapping relationship library during the obstacle avoidance process.

6. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 1, characterized in that, The method comprises: It receives V2X communication data in real time and extracts information on the status of surrounding vehicles and traffic control information. The V2X communication data and vehicle perception data are fused with confidence weighting to generate global road state data. The dynamic risk distribution map is corrected based on the global road condition data.

7. The obstacle avoidance method for autonomous vehicles based on road condition perception as described in claim 1, characterized in that, The method comprises: Establish a closed-loop verification mechanism for obstacle avoidance control commands and obtain real-time response delay data and steering torque deviation rate of the vehicle chassis execution unit; The response delay data is compared with a first preset threshold, and the steering torque deviation rate is compared with a second preset threshold. When the response delay data exceeds the first preset threshold or the steering torque deviation rate exceeds the second preset threshold, the redundant control module is activated to generate a compensation control command, and the compensation result is fed back to the closed-loop verification mechanism for dynamic correction.

8. An obstacle avoidance device for autonomous vehicles based on road condition perception, characterized in that, The apparatus is used to implement the obstacle avoidance method for autonomous vehicles based on road condition perception as described in any one of claims 1-7, the apparatus comprising: The multi-dimensional data acquisition module is used to collect multi-dimensional perception data of the vehicle's driving environment in real time and generate a set of road obstacle feature information. The dynamic risk distribution map acquisition module is used to construct an obstacle motion trajectory prediction model, perform a collision risk quantification assessment based on the road obstacle feature information set, and obtain a dynamic risk distribution map. The feasibility verification module is used to generate a multimodal obstacle avoidance decision tree based on the dynamic risk distribution map and to verify the feasibility of the strategy in combination with vehicle dynamics parameters. The control instruction set generation module is used to generate control instruction sets based on the verified obstacle avoidance strategy, drive the vehicle to perform obstacle avoidance actions, and provide real-time feedback on execution status data.

Images