Autonomous obstacle avoidance method and system for low-speed working vehicle

Abstract

The invention provides an autonomous obstacle avoidance method and system for a low-speed working vehicle, and relates to the technical field of vehicle obstacle avoidance, and the method comprises the steps: constructing a transient environment sensing field domain; obtaining environment sensing data; performing multi-target dynamic obstacle detection tracking; performing space-time motion state fitting to obtain a plurality of obstacle prediction trajectories; performing multi-constraint dynamic obstacle avoidance fitting to obtain a plurality of obstacle avoidance bypassing tracks; performing path regression feasibility analysis to generate a plurality of obstacle avoidance regression tracks; performing control complexity joint evaluation, and outputting a plurality of trajectory optimization evaluation parameters; and screening and outputting an autonomous obstacle avoidance linkage trajectory, and after trajectory tracking control is executed, performing obstacle avoidance decision control closed-loop iteration. According to the invention, a technical problem that the accuracy of an obstacle avoidance decision is reduced due to the fact that vehicle obstacle avoidance in the prior art usually depends on a single type of sensor to carry out environment perception and accurate environment data cannot be acquired in time and thus an obstacle cannot be accurately identified and predicted is solved.

Description

Technical Field

[0001] This invention relates to the field of vehicle obstacle avoidance technology, and specifically to an autonomous obstacle avoidance method and system for low-speed operating vehicles. Background Technology

[0002] Low-speed work vehicles are widely used in logistics, agriculture, mining and other fields, especially in environments where human operation is inconvenient or dangerous. The typical characteristics of these vehicles are their low speed and the need for autonomous navigation and obstacle avoidance in complex and dynamic environments. For example, in agriculture, low-speed work vehicles need to avoid obstacles such as mounds of soil, shrubs, and other vehicles in farmland; in mines or warehouses, low-speed work vehicles need to cope with irregular terrain, obstacles and narrow working spaces.

[0003] However, existing low-speed operating vehicles typically rely on a single type of sensor for environmental perception, which cannot comprehensively and accurately acquire all information in complex environments. For example, lidar performs poorly in rainy or foggy weather, and cameras perform poorly in low-light environments. Therefore, this single perception method prevents vehicles from providing continuous and stable perception data in various weather conditions and complex environments, thus failing to accurately identify and predict obstacles, resulting in a decrease in the accuracy and responsiveness of obstacle avoidance decisions. Summary of the Invention

[0004] This application provides an autonomous obstacle avoidance method and system for low-speed operating vehicles, aiming to solve the technical problem that existing vehicle obstacle avoidance technologies typically rely on a single type of sensor for environmental perception, which cannot obtain accurate environmental data in a timely manner, thus failing to accurately identify and predict obstacles, leading to a decrease in the accuracy of obstacle avoidance decisions.

[0005] The first aspect disclosed in this application provides an autonomous obstacle avoidance method for low-speed working vehicles. The method includes: constructing a transient environmental perception field based on the driving direction of the low-speed working vehicle; controlling a multimodal sensor array to perform environmental perception within the transient environmental perception field to obtain environmental perception data; constructing a locally occupied grid map based on the environmental perception data, performing multi-target dynamic obstacle detection and tracking to obtain multiple real-time motion trends of multiple real-time obstacles; fitting the spatiotemporal motion state based on the multiple real-time motion trends to obtain multiple obstacle prediction trajectories; and determining the predicted trajectories of the low-speed working vehicle based on its predetermined driving trajectory and the multiple obstacles. The system predicts the object's trajectory and performs multi-constraint dynamic obstacle avoidance fitting to obtain multiple obstacle avoidance trajectories. Based on these multiple obstacle avoidance trajectories and the predetermined driving trajectory, it performs path regression feasibility analysis to generate multiple obstacle avoidance regression trajectories. It then performs a joint evaluation of the control complexity of the multiple obstacle avoidance trajectories and the multiple obstacle avoidance regression trajectories, outputting multiple trajectory optimization evaluation parameters. Based on the descending order of the multiple trajectory optimization evaluation parameters, it selects and outputs an autonomous obstacle avoidance linkage trajectory from the multiple obstacle avoidance trajectories and the multiple obstacle avoidance regression trajectories. After executing the trajectory tracking control of the low-speed operating vehicle, it performs closed-loop iteration of obstacle avoidance decision control based on the transient environment update field.

[0006] The second aspect of this application discloses an autonomous obstacle avoidance system for low-speed working vehicles. The system is used in the aforementioned autonomous obstacle avoidance method for low-speed working vehicles. The system includes: a perception field construction module for constructing a transient environmental perception field based on the driving direction of the low-speed working vehicle; an environment perception module for controlling a multimodal sensor array to perform environmental perception within the transient environmental perception field and obtain environmental perception data; a detection and tracking module for constructing a locally occupied grid map based on the environmental perception data, performing multi-target dynamic obstacle detection and tracking, and obtaining multiple real-time motion trends of multiple real-time obstacles; a state fitting module for performing spatiotemporal motion state fitting based on the multiple real-time motion trends to obtain multiple obstacle prediction trajectories; and an obstacle avoidance fitting module for... The predetermined driving trajectory of the low-speed working vehicle and the predicted trajectories of multiple obstacles are used to perform multi-constraint dynamic obstacle avoidance fitting to obtain multiple obstacle avoidance detour trajectories; a feasibility analysis module is used to perform path regression feasibility analysis based on the multiple obstacle avoidance detour trajectories and the predetermined driving trajectory to generate multiple obstacle avoidance regression trajectories; a joint evaluation module is used to perform joint evaluation of the control complexity of the multiple obstacle avoidance detour trajectories and multiple obstacle avoidance regression trajectories, and output multiple trajectory optimization evaluation parameters; a closed-loop iteration module is used to select and output autonomous obstacle avoidance linkage trajectories from the multiple obstacle avoidance detour trajectories and multiple obstacle avoidance regression trajectories according to the descending sorting results of the multiple trajectory optimization evaluation parameters, and after executing the trajectory tracking control of the low-speed working vehicle, the obstacle avoidance decision control closed-loop iteration is performed according to the transient environment update field.

[0007] One or more technical solutions provided in this application have at least the following beneficial effects:

[0008] By constructing a transient environmental perception field based on the driving direction of low-speed operating vehicles, real-time perception of the road and environment ahead can be ensured. Controlling a multimodal sensor array to collect environmental data within the transient environmental perception field allows for the acquisition of multidimensional data on obstacles, road conditions, and other environmental information, enhancing the vehicle's adaptability to dynamic environments. Rasterizing the perception data results in a locally occupied grid map that subdivides the environment into smaller blocks. Combined with multi-target dynamic obstacle detection, this accurately tracks obstacle movement trends, enabling obstacle avoidance decisions based on precise dynamic information. Spatiotemporal motion state fitting analyzes the movement trends of multiple obstacles to predict their future trajectories, effectively preventing collision risks caused by sudden obstacles. Dynamic interaction between the predetermined driving trajectory and the predicted obstacle trajectory allows for multi-constraint obstacle avoidance fitting, generating multiple obstacle avoidance trajectories. This ensures that the detour trajectories not only avoid obstacles but also maintain vehicle stability, efficiency, and safety. Path regression feasibility analysis ensures the vehicle... The system can smoothly return to the predetermined trajectory after obstacle avoidance. Through this analysis, the generated obstacle avoidance and return trajectory restores the vehicle's original driving route as much as possible while avoiding obstacles. This not only improves the continuity of the obstacle avoidance process but also reduces the possibility of the vehicle frequently changing direction in a dynamic environment, improving the smoothness and stability of the path. Through joint evaluation of control complexity, multiple obstacle avoidance and return trajectories are comprehensively evaluated. By quantifying the control complexity and response characteristics of each trajectory, the optimal trajectory is selected. This optimization evaluation can balance multiple factors to ensure that the vehicle is both efficient and safe during obstacle avoidance. By sorting the optimization evaluation parameters of multiple trajectories in descending order, the most suitable obstacle avoidance trajectory is selected, thereby ensuring the vehicle's optimal obstacle avoidance behavior in complex environments. Finally, after the vehicle executes the selected trajectory, trajectory tracking control ensures that the vehicle travels precisely along the trajectory. At the same time, the closed-loop iteration of the transient environment update field ensures that the system dynamically adjusts according to real-time changes in the environment, enabling the vehicle to continuously make obstacle avoidance decisions and ensuring the continuity and safety of driving.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of an autonomous obstacle avoidance method for low-speed operating vehicles provided in an embodiment of this application.

[0011] Figure 2This is a schematic diagram of an autonomous obstacle avoidance system for low-speed operating vehicles provided in an embodiment of this application.

[0012] Figure labeling: Perception field construction module 10, environmental perception module 20, detection and tracking module 30, state fitting module 40, obstacle avoidance fitting module 50, feasibility analysis module 60, joint evaluation module 70, closed-loop iteration module 80. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] Example 1, as Figure 1 As shown in the figure, this application provides an autonomous obstacle avoidance method for low-speed operating vehicles, the method comprising:

[0015] A transient environmental perception field is constructed based on the travel direction of low-speed operating vehicles.

[0016] The system acquires the real-time driving speed, maximum braking acceleration, and control response delay of low-speed operating vehicles. This information is used to calculate the minimum safe perception distance for these vehicles. Using the vehicle's current position as input, the system extracts the real-time road network width. These parameters determine the range of the vehicle's perception in the environment. By using an environmental topology feature library, the minimum safe perception distance is matched with the actual road network width, and boundary constraints are corrected to generate a dynamic perception area that adapts to environmental changes—the transient environmental perception field.

[0017] The multimodal sensing array is controlled to perform environmental sensing within the transient environmental sensing field and obtain environmental sensing data.

[0018] Multiple field size ranges and sensor activation modes are predefined. These modes determine the activation method of sensors under different environments; for example, different sensors are selected based on factors such as distance and obstacle type. Based on the constructed transient environmental perception field, the activation mode that best matches the current environment is selected from the sensor scheduling decision library. After activating the selected mode, a multimodal sensor array, such as radar, lidar, and cameras, is controlled to perform environmental perception within the transient environmental perception field, acquiring data from different sensors, such as distance, velocity, and angle. The single-modal data obtained from different sensors are spatiotemporally aligned to ensure consistency in time and space, resulting in complete environmental perception data.

[0019] Based on the environmental perception data, a local occupancy grid map is constructed, and multi-target dynamic obstacle detection and tracking are performed to obtain multiple real-time movement trends of multiple real-time obstacles.

[0020] Environmental perception data is rasterized, dividing the environment into multiple grids for more detailed analysis. Based on the spatial confidence weights of each sensor, multiple single-modal grid data are fused to generate a locally occupied grid map, indicating whether each grid is occupied by an obstacle. Euclidean distance clustering is used to cluster the locally occupied grid map, identifying multiple obstacle clusters, each representing an obstacle. For each obstacle cluster, its geometric features (such as size and shape) and motion features (such as velocity and acceleration) are extracted to obtain multiple real-time motion trends. These features aid in subsequent obstacle type identification and motion trend analysis.

[0021] Based on the multiple real-time motion trends, spatiotemporal motion states are fitted to obtain multiple obstacle prediction trajectories.

[0022] Multi-order time-domain analysis is performed on the real-time motion trend of each obstacle to extract information such as real-time position, velocity, acceleration, rate of change of motion direction, and radius of curvature. This data helps to understand the obstacle's motion pattern. Based on the rate of change of motion direction and radius of curvature, the obstacle's motion pattern is identified, such as linear motion, circular motion, or other complex trajectories. According to the obstacle's motion pattern, a motion model based on physical laws is selected, such as a uniform linear motion model or a uniformly accelerated motion model. Using the selected motion model, the obstacle's real-time position, velocity, and acceleration are used as initial states to predict the trajectory under spatiotemporal perturbation assumptions. Multiple candidate trajectories are generated through multiple perturbation assumptions, representing the obstacle's possible future position. Based on multi-dimensional probability assessments, such as motion continuity, environmental rationality, and interaction safety, the multiple candidate trajectories are fused to obtain the final predicted obstacle trajectory.

[0023] Based on the predetermined driving trajectory of the low-speed operating vehicle and the predicted trajectories of the multiple obstacles, multi-constraint dynamic obstacle avoidance fitting is performed to obtain multiple obstacle avoidance and detour trajectories.

[0024] By combining the predetermined travel trajectory of a low-speed operating vehicle with multiple predicted obstacle trajectories, multi-constraint dynamic obstacle avoidance fitting is performed based on factors such as vehicle size, kinematic constraints, and surrounding environmental limitations. These constraints include, but are not limited to, the vehicle's turning radius, speed limits, obstacle avoidance area, and road network topology. Using an obstacle avoidance algorithm, multiple obstacle avoidance trajectories are generated based on these constraints. These trajectories represent how the vehicle can bypass obstacles and continue moving forward in the presence of multiple obstacles.

[0025] Based on the multiple obstacle avoidance and detour trajectories and the predetermined driving trajectory, a path regression feasibility analysis is performed to generate multiple obstacle avoidance regression trajectories.

[0026] For each obstacle avoidance trajectory, its degree of matching with the predetermined driving trajectory is analyzed, and it is determined whether the post-obstacle avoidance trajectory meets actual driving requirements. Specifically, regression feasibility analysis includes analyzing whether the trajectory can be smoothly connected to the original planned path and whether it meets safe driving requirements. Based on the regression feasibility analysis, path planning algorithms, such as Clothoid curve smoothing technology, are used to smoothly connect the detour trajectories, generating multiple obstacle avoidance regression trajectories. These trajectories allow the vehicle to avoid obstacles while returning to the original driving path as much as possible.

[0027] The control complexity of the multiple obstacle avoidance trajectories and multiple obstacle avoidance return trajectories is jointly evaluated, and multiple trajectory optimization evaluation parameters are output.

[0028] For each trajectory, including obstacle avoidance and rerouting trajectories, control costs are quantified. The control requirements of the vehicle when executing these trajectories are analyzed, such as braking frequency, steering angle changes, and drive motor acceleration changes. The control response characteristics of the trajectory are quantified, including lateral tracking error, steering command correction frequency, and speed lag time. These indicators measure the vehicle's accuracy and response speed when executing the trajectory. A comprehensive evaluation is performed by combining control costs and control response characteristics. These parameters are then fused through hierarchical weighting to obtain trajectory optimization evaluation parameters, ultimately providing a basis for selecting the optimal trajectory.

[0029] Based on the descending order of the multiple trajectory optimization evaluation parameters, an autonomous obstacle avoidance linkage trajectory is selected and output from the multiple obstacle avoidance detour trajectories and multiple obstacle avoidance return trajectories. After executing the trajectory tracking control of the low-speed working vehicle, the obstacle avoidance decision control closed-loop iteration is performed according to the transient environment update field.

[0030] Based on the descending order of multiple trajectory optimization evaluation parameters, the optimal autonomous obstacle avoidance linkage trajectory is selected from multiple obstacle avoidance detour trajectories and obstacle avoidance return trajectories. The low-speed operating vehicle is then controlled to track the selected autonomous obstacle avoidance linkage trajectory. At this time, the vehicle control system performs closed-loop control to ensure that the vehicle travels accurately along the selected trajectory. As the vehicle moves, the surrounding environment, such as the position of obstacles and the vehicle's driving state, will constantly change. Therefore, the transient environmental perception field will be updated in real time. Based on the new environmental information, iterative updates will be performed to adjust the vehicle's driving trajectory and ensure the continuous safety of the obstacle avoidance process.

[0031] Furthermore, the method for constructing a transient environmental perception field based on the travel direction of low-speed operating vehicles includes:

[0032] The system interacts with the real-time driving speed, maximum braking acceleration, and control response delay of the low-speed operating vehicle; calculates the minimum safe perception distance based on the real-time driving speed, maximum braking acceleration, and control response delay; extracts the real-time traffic road network width based on the real-time road network coordinates of the low-speed operating vehicle; after matching the standard environmental perception field with the environmental topology feature library using the minimum safe perception distance, it uses the real-time traffic road network width to correct the field boundary constraints and outputs the transient environmental perception field.

[0033] Real-time driving speed is the vehicle's current speed, which can be obtained through a vehicle speed sensor; maximum braking acceleration is the acceleration required for the vehicle to come to a complete stop from its current speed, determined by the performance of the vehicle's braking system. Control response delay is the time delay required from receiving a command from the vehicle's control system to executing that command, including the system's processing time and the sensor's data processing time.

[0034] The minimum safe perception distance is calculated using real-time driving speed, maximum braking acceleration, and control response delay. This distance refers to the minimum distance required for the vehicle to safely stop at its current speed, based on braking capacity and control response time. The minimum safe perception distance is calculated using the following formula: ,in, It is the minimum safe sensing distance. It is the real-time driving speed. It controls response latency. It is the maximum braking acceleration.

[0035] Real-time road network coordinates, i.e., the vehicle's specific location, are obtained through the vehicle's positioning system. Road network data, provided by high-precision maps or road databases, is then used to extract the width information of the currently located road. This width information is obtained by matching the vehicle's positioning coordinates with pre-stored road network data. Road network width is a crucial parameter when constructing the perception field, affecting the size of the vehicle's perception area and obstacle avoidance strategies. Especially in narrow roads or congested environments, road network width directly impacts the vehicle's safe perception range.

[0036] The environmental topology feature library contains standard environmental perception fields for different types of roads. These fields are defined based on different environments (e.g., open roads, urban roads, narrow passages) and vehicle parameters (e.g., vehicle width, vehicle length), and have preset perception ranges and obstacle detection areas. By matching the minimum safe perception distance with the standard fields in the environmental topology feature library, a suitable perception field is selected to ensure that the perception field can cover all potential danger areas and that the vehicle can perceive all possible obstacles in real time. The boundaries of the perception field are adjusted according to the real-time road network width. For example, on narrower roads, the boundaries of the perception field shrink to adapt to the current road environment; while on wider roads, the boundaries expand. Through matching and adjustment, a transient environmental perception field adapted to the current environment is finally obtained. This perception field dynamically adjusts according to the vehicle's real-time driving status and changes in road conditions.

[0037] Furthermore, the method of controlling a multimodal sensing array to perform environmental sensing within the transient environmental sensing field and obtain environmental sensing data includes:

[0038] Multiple sensing modal activation modes with predefined field size ranges are used to construct a sensor scheduling decision library. The transient environmental sensing field is used to match the transient activation mode in the sensor scheduling decision library. The transient activation mode is used to control the multimodal sensing array to perform environmental sensing in the transient environmental sensing field, and M single-modal sensing data are obtained. The M single-modal sensing data are spatiotemporally aligned to obtain the environmental sensing data.

[0039] Different environmental conditions, such as varying obstacle densities, field of view ranges, and speed requirements, are divided into multiple field size intervals. Each interval represents a different environmental scenario, such as open roads, complex intersections, and narrow passages, where vehicle perception requirements differ. Based on different field sizes and environmental needs, multiple sensor activation modes are predefined. These modes determine which sensors should be activated, when to activate them, and how to adjust their operating states in different situations. For example, in relatively open areas, long-range sensors need to be activated, while in complex urban environments, multiple sensors need to be activated simultaneously for comprehensive perception. Based on these different fields and activation modes, a sensor scheduling decision library is constructed. This library stores the sensor combinations and configurations activated according to different environmental scenarios, enabling vehicles to make optimal sensor activation decisions under various conditions.

[0040] Based on the transient environmental perception field, the most matching sensor activation mode is searched in the sensor scheduling decision library. This process selects the most suitable sensor configuration by comparing the current environmental information with the predefined environmental conditions in the library.

[0041] Based on the matched transient activation mode, the multimodal sensor array is controlled to perform perception tasks within the current transient environmental perception field. These sensors collect information about the environment around the vehicle, such as the position, shape, and speed of obstacles. Each sensor works independently and provides single-modal perception data. For example, lidar provides distance information, cameras provide image data, and radar provides speed data. Each sensor generates independent perception data, which is called single-modal perception data.

[0042] Data from different sensors differ in time and space. For example, LiDAR and cameras operate at different frequencies, or they have different viewing angles and measurement ranges. Therefore, spatiotemporal alignment of M single-modal sensing data means calibrating the data acquired by different sensors according to the same time and space reference frame so that they can be accurately superimposed. Through spatiotemporal alignment, complete and comprehensive environmental perception data is obtained, which includes information from all sensors and can provide a comprehensive view of the surrounding environment.

[0043] Furthermore, based on the environmental perception data, a locally occupied grid map is constructed, and multi-target dynamic obstacle detection and tracking are performed to obtain multiple real-time movement trends of multiple real-time obstacles. The method includes:

[0044] The environmental perception data of M single-modal sensing data are processed into M single-modal grid layers by data rasterization; M spatial confidence weights of M single-modal sensors are matched according to the real-time operating environment of the transient environmental perception field; dynamic weight fusion of the M single-modal grid layers is performed based on the M spatial confidence weights to obtain the local occupancy grid map; Euclidean distance clustering is performed on the local occupancy grid map to obtain multiple obstacle clusters; multiple geometric features and multiple motion features of the multiple obstacle clusters are extracted; obstacle type identification is performed based on the multiple geometric features to obtain the multiple real-time obstacles; and multi-level state prediction of motion trends is performed based on the multiple motion features to obtain the multiple real-time motion trends.

[0045] Data obtained from different sensors is called unimodal sensing data. This unimodal sensing data is mapped onto a fixed spatial grid, where each grid represents a small region of the environment. Each grid stores a value indicating whether the region is occupied by an obstacle, or other environmental features such as the distance or speed of obstacles. This results in M ​​unimodal grid layers, each containing information about different dimensions of the environment.

[0046] Based on the different characteristics of the environment in the transient environmental sensing field, different spatial confidence weights are assigned to the output data of each sensor. The confidence weight represents the reliability or importance of a sensor in a specific environment. For example, if obstacles in a certain area are mainly detected by LiDAR, and the LiDAR data is particularly accurate in that area, then the LiDAR can be given a higher weight; conversely, if the signal quality in that area is poor, then the LiDAR's weight is reduced. The spatial confidence weights assigned to the sensing data of all sensors can be dynamically adjusted based on factors such as sensor accuracy, operating range, and environmental changes.

[0047] By utilizing the spatial confidence weights of each sensor, data from M single-modal grid layers are weighted and fused. The data from each grid layer is adjusted according to its corresponding confidence weight; sensor data with higher weights have a greater impact on the final result, while data with lower weights have a smaller impact. Through weighted fusion, a comprehensive local occupancy grid map is obtained. This map can display the occupancy status of the vehicle's surrounding environment, such as which areas are occupied by obstacles and which areas are open.

[0048] Clustering algorithms based on Euclidean distance, such as K-means clustering or DBSCAN, are used to cluster obstacle data in locally occupied grid maps. Euclidean distance refers to the straight-line distance between two points in space. The clustering algorithm will group grid data that are close to each other into the same cluster. After clustering, obstacles in locally occupied grid maps will be divided into multiple clusters, and each cluster represents a group of close obstacles.

[0049] Geometric features include the shape, size, boundaries, and center position of obstacles, which help determine the specific physical properties of obstacles. Motion features reflect the motion state of obstacles, including velocity, acceleration, direction of motion, and rate of change of trajectory. These geometric and motion features will be used for obstacle identification, classification, and dynamic prediction.

[0050] By analyzing the geometric features of multiple obstacle clusters and using pattern recognition techniques, such as machine learning algorithms, deep learning models, or rule-based classification methods, different types of obstacles can be identified. For example, obstacles can be classified into static obstacles, dynamic obstacles, moving objects, etc. Identifying obstacle types helps predict their movement patterns and interactions.

[0051] Based on the motion characteristics of obstacles, multi-level state prediction is performed. The motion state of an obstacle involves multiple stages, such as current velocity, acceleration, and future acceleration changes. Multi-level state prediction can obtain the motion trend of the obstacle over a future period. This can be achieved through kinematic models or more complex methods such as Kalman filtering and particle filtering. The prediction results include the obstacle's future position, velocity, and acceleration, thus providing a basis for subsequent trajectory planning and obstacle avoidance decisions.

[0052] Furthermore, by fitting the spatiotemporal motion state based on the multiple real-time motion trends, multiple obstacle prediction trajectories are obtained. The method includes:

[0053] A multi-order time-domain analysis is performed on the first real-time motion trend to extract the first real-time position, first real-time velocity, first real-time acceleration, first motion direction change rate, and first radius of curvature. A first motion pattern is identified based on the first motion direction change rate and first radius of curvature. A motion model is selected based on the first real-time obstacle and the first motion pattern, and the first motion model is retrieved. The first real-time position, first real-time velocity, and first real-time acceleration are used as the initial state of the first motion and input into the first motion model for spatiotemporal multi-perturbation hypothetical trajectory prediction, resulting in multiple candidate predicted trajectories. Based on multi-dimensional probability evaluation, the trajectories of the multiple candidate predicted trajectories are fused to output the first obstacle predicted trajectory.

[0054] Through multi-order time-domain analysis, the first real-time motion trend is analyzed in multiple orders. Multi-order means analyzing the state changes over multiple time periods and extracting different motion state information. Among them, the first real-time position is the current position of the obstacle, the first real-time velocity is the current velocity of the obstacle, which can be calculated by the rate of change of distance over time, the first real-time acceleration is the acceleration of the obstacle, which can be calculated by the rate of change of velocity over time, the first rate of change of motion direction is the rate of change of the obstacle's motion direction, reflecting the obstacle's turning speed, and the first radius of curvature is the radius of curvature of the obstacle's motion trajectory, reflecting the curvature of the obstacle's trajectory and used to describe whether the obstacle is moving along a curve.

[0055] If the rate of change of the obstacle's first direction of motion is large, it indicates that the obstacle is turning or changing trajectory, and such objects have more complex motion patterns. The size of the first radius of curvature directly affects the obstacle's first motion pattern. A larger first radius of curvature means that the obstacle is making a gentle turn, while a smaller first radius of curvature means that the obstacle is making a sharp turn. The first radius of curvature can help distinguish vehicles from other types of obstacles, such as pedestrians and stationary objects. By analyzing the rate of change of the obstacle's first direction of motion and the first radius of curvature, motion pattern recognition can be performed on the obstacle. Common motion patterns include: linear motion, where the obstacle's velocity direction hardly changes and the radius of curvature is large; turning motion, where the obstacle's velocity direction changes significantly and the radius of curvature is small; and sudden stopping or acceleration motion, where the obstacle's acceleration changes significantly.

[0056] Based on the first real-time obstacle and the first motion pattern, a suitable motion model is selected for obstacle trajectory prediction. The motion models include: a uniform linear motion model for obstacles moving at a constant speed along a straight line; a uniformly accelerated motion model for obstacles with constant speed changes; and a curvilinear motion model for obstacles traveling along a curved path, which is based on a more complex dynamic model. The appropriate motion model is selected according to the obstacle's motion pattern; for example, if the obstacle is moving in a straight line, a uniform linear motion model is selected; if the obstacle is turning, a curvilinear motion model is selected. After retrieving the corresponding motion model, it can be used for spatiotemporal trajectory prediction.

[0057] The first real-time position, first real-time velocity, and first real-time acceleration are used as the initial motion state, reflecting the current motion state of the obstacle. These data are input into the first motion model. By adding perturbations to the first motion model—that is, changes in some assumptions—multiple candidate predicted trajectories are generated. The perturbations include: velocity perturbation, assuming a change in the obstacle's velocity; acceleration perturbation, assuming a change in the obstacle's acceleration; and direction perturbation, assuming a change in the obstacle's direction of travel. The multiple candidate predicted trajectories generated through these perturbation assumptions can cover various possible motion scenarios of the obstacle, thereby improving the accuracy and reliability of the prediction.

[0058] The multi-dimensional probability assessment includes evaluation items such as motion continuity, environmental rationality, and interaction safety. By weighting multiple evaluation dimensions, the comprehensive confidence probability of each trajectory is calculated. Each trajectory receives a confidence value based on its evaluation results, with higher confidence values ​​indicating a greater likelihood of occurrence. Multiple candidate predicted trajectories are then fused based on the confidence values, and the most likely trajectory is selected as the first obstacle prediction trajectory. This trajectory serves as the basis for subsequent obstacle avoidance decisions, helping low-speed operating vehicles prepare for obstacle avoidance in advance.

[0059] Furthermore, the method involves fusing the multiple candidate predicted trajectories based on multi-dimensional probability evaluation to output a first obstacle predicted trajectory, the method comprising:

[0060] Environmental semantic constraints are corrected and interaction forces between obstacles are compensated for on the multiple candidate predicted trajectories to obtain multiple corrected predicted trajectories; multidimensional probability evaluation is performed on the multiple corrected predicted trajectories to obtain multiple trajectory comprehensive confidence probabilities, wherein the evaluation items of the multidimensional probability evaluation include motion continuity, environmental rationality and interaction safety; based on the multiple trajectory comprehensive confidence probabilities, the multiple corrected predicted trajectories are fused to output the first obstacle predicted trajectory.

[0061] Candidate predicted trajectories may not perfectly conform to actual environmental constraints, such as the actual width of the road, speed limits, and physical limitations of obstacles. Therefore, constraint corrections are needed based on semantic information of the environment. For example, in certain specific areas, such as narrow passages or intersections, candidate predicted trajectories need to be adjusted to avoid static or dynamic obstacles. For instance, if a candidate trajectory traverses a narrow passage or a restricted area, the trajectory can be adjusted to ensure that the vehicle always remains within a feasible driving range.

[0062] Obstacles can interact with each other, especially on narrow roads, where the movement of multiple obstacles can influence each other. For example, two moving obstacles may avoid each other. Therefore, the effects of these interactions need to be considered in trajectory prediction. For instance, if obstacle A and obstacle B influence each other, their trajectories will change. Appropriate trajectory corrections are needed based on these interactions to ensure that the interactions between obstacles are not ignored.

[0063] By correcting environmental semantic constraints and compensating for interaction forces, multiple corrected predicted trajectories are obtained, which are more in line with the actual environment and the relationship between obstacles.

[0064] The evaluation items for multidimensional probability assessment include motion continuity, environmental rationality, and interaction safety. Among them, motion continuity assesses the smoothness and coherence of the trajectory, judging whether the trajectory conforms to the actual motion law. For example, whether the trajectory has unreasonable sharp turns or discontinuous jumps. The smoothness of the trajectory is closely related to the physical motion characteristics of obstacles. Environmental rationality assesses whether the trajectory conforms to the actual constraints of the environment, including road conditions, obstacle distribution, and drivable areas. If the trajectory crosses prohibited areas or violates the physical constraints of the road, the rationality of the trajectory is poor. Interaction safety assesses the safety of the trajectory when interacting with other obstacles or dynamic objects. For example, if the trajectory path collides with other obstacles or is too close, its interaction safety is low.

[0065] By combining multiple evaluation items such as motion continuity, environmental rationality, and interaction safety, the comprehensive confidence probability of each trajectory is calculated. Each evaluation item is assigned a certain weight according to the specific situation, thereby reflecting the overall credibility of the trajectory.

[0066] Based on the comprehensive confidence probability of each corrected predicted trajectory, all candidate trajectories are weighted and fused. Trajectories with high confidence have a greater impact on the final prediction result, while trajectories with low confidence have their weight reduced or are excluded. Fusion methods can use weighted average or other fusion algorithms. Through trajectory fusion, the final first obstacle prediction trajectory is obtained, which represents the most likely future movement path of the obstacle. This predicted trajectory will serve as the basis for the vehicle's obstacle avoidance decision-making, used to plan the next obstacle avoidance action.

[0067] Furthermore, based on the predetermined travel trajectory of the low-speed operating vehicle and the predicted trajectories of the multiple obstacles, multi-constraint dynamic obstacle avoidance fitting is performed to obtain multiple obstacle avoidance trajectories. The method includes:

[0068] Based on the vehicle dimensions of the low-speed operating vehicle and the predetermined driving trajectory, a spatiotemporal safety space channel is constructed in a four-dimensional spatiotemporal grid space; the multiple obstacle avoidance detour trajectories are mapped to the four-dimensional spatiotemporal grid space, and multi-resolution conflict detection is performed with the spatiotemporal safety space channel to locate distributed conflict regions; based on the distributed conflict regions, detour routes are fitted in the spatiotemporal safety space channel based on vehicle kinematic constraints and road network topology constraints to obtain the multiple obstacle avoidance detour trajectories.

[0069] Vehicle dimensions include the vehicle's length, width, height, and turning radius. This dimensional information is used to define the safe space required for the vehicle during movement. The planned travel trajectory is the path the vehicle will follow in the future, providing guidance for the vehicle's direction of travel.

[0070] The four-dimensional spatiotemporal grid space is a space containing spatial and temporal information, composed of X, Y (planar coordinates), time dimension, and the vehicle's longitudinal or lateral velocity. This space is divided according to the vehicle's dynamic characteristics, driving trajectory, and time changes. Using the vehicle's size information and predetermined driving trajectory, a spatiotemporal safety space channel is constructed within the four-dimensional spatiotemporal grid space. This channel ensures that the vehicle can travel within the predetermined trajectory range while avoiding collisions with obstacles. This space channel takes into account factors such as the vehicle's physical limitations, turning radius, and speed limits, providing a safe driving space for the vehicle.

[0071] Multiple obstacle avoidance trajectories are mapped onto a four-dimensional spatiotemporal grid space. Each trajectory occupies a certain area both spatially and temporally, forming a series of trajectory points. Multi-resolution collision detection is performed on these trajectories. Multi-resolution detection means performing collision detection at different resolution levels, allowing for checks at both coarser (large-scale) and finer (small-scale) scales. This method effectively balances computational efficiency and accuracy: detecting large potential conflict areas at lower resolutions and checking more detailed local areas at higher resolutions to ensure the accuracy of the detour path. Through collision detection, conflict areas between the obstacle avoidance trajectory and the spatiotemporal safety space are located. These areas represent parts where the vehicle may collide with obstacles and require further adjustment and optimization.

[0072] A distributed conflict region is a collection of multiple potential conflict regions, representing several areas that a vehicle needs to traverse during obstacle avoidance. Each conflict region involves one or more obstacles. Vehicle kinematic constraints refer to the physical and dynamic limitations that a vehicle must follow when traversing, such as the minimum turning radius, speed limits, and acceleration limits. The detour path must be adjusted under these constraints to ensure that the vehicle can complete the detour smoothly and safely. Road network topology constraints consider the road layout, traffic rules, and vehicle travel paths. For example, vehicles cannot cross restricted areas, and the detour path must follow the actual road layout and traffic signs. Combining vehicle kinematic constraints and road network topology constraints, detour routes are fitted within the distributed conflict regions. This process ensures that the detour path not only avoids obstacles but also conforms to the vehicle's dynamic characteristics and road constraints. Ultimately, multiple optimized obstacle avoidance trajectories are generated.

[0073] Furthermore, based on the multiple obstacle avoidance trajectories and the predetermined driving trajectory, a path regression feasibility analysis is performed to generate multiple obstacle avoidance regression trajectories. The method includes:

[0074] Calculate the first obstacle avoidance detour trajectory and the first feasible regression point of the predetermined driving trajectory; use a Clothoid curve to smoothly connect the trajectory endpoint of the first obstacle avoidance detour trajectory and the first feasible regression point to generate the first obstacle avoidance regression trajectory.

[0075] The first obstacle avoidance trajectory is the path generated after encountering an obstacle and detouring around it. The predetermined driving trajectory is the driving path initially planned by the vehicle. The selection of the return point is crucial to ensuring that the vehicle smoothly returns from the obstacle avoidance trajectory to the predetermined driving trajectory. When calculating the return point, based on factors such as the vehicle's current position, speed, and the geometry of the road ahead, the return point is located near the end of the obstacle avoidance trajectory and meets the following conditions: the return point is the point closest to the predetermined trajectory, allowing the vehicle to easily adjust its direction and return; the return point should conform to the road topology and safety restrictions, such as road width and traffic signs; the location of the return point should ensure that the vehicle can transition smoothly, avoiding excessive turning angles or sudden acceleration or deceleration.

[0076] A Clothoid curve (also known as a Euler curve or gradient curve) is a curve used for smooth connections between roads or tracks. It smoothly transitions between changes in track curvature. During obstacle avoidance, the Clothoid curve effectively avoids abrupt steering changes, ensuring vehicle stability and safety during turns. When connecting the endpoint of the first obstacle avoidance trajectory to the first feasible return point, the Clothoid curve smoothly transitions the geometry of the two tracks, allowing the vehicle to naturally return to the predetermined trajectory after the detour. By adjusting the curve parameters, the turning radius of the vehicle within this track segment gradually changes without generating abrupt steering forces. The resulting first obstacle avoidance return trajectory, obtained through a smooth connection using the Clothoid curve, connects the endpoint of the obstacle avoidance trajectory to the return point, ensuring the vehicle smoothly returns to its original driving trajectory and guaranteeing safe driving.

[0077] Furthermore, the method involves jointly evaluating the control complexity of the multiple obstacle avoidance trajectories and multiple obstacle avoidance return trajectories, and outputting multiple trajectory optimization evaluation parameters.

[0078] The control cost of the first obstacle avoidance trajectory and the first obstacle avoidance return trajectory is quantified to obtain the cumulative frequency of the first braking action, the cumulative angle change of the first steering motor, the rate of change of acceleration of the first drive motor, and the first cumulative obstacle avoidance energy consumption. The control response characteristics of the first obstacle avoidance trajectory and the first obstacle avoidance return trajectory are quantified to obtain the first maximum lateral tracking error, the first steering command correction frequency, and the first speed lag time. The first braking action cumulative frequency, the cumulative angle change of the first steering motor, the rate of change of acceleration of the first drive motor, the first cumulative obstacle avoidance energy consumption, the first maximum lateral tracking error, the first steering command correction frequency, and the first speed lag time are hierarchically weighted and fused to output the first trajectory optimization evaluation parameters.

[0079] During obstacle avoidance, the vehicle's braking system will activate and deactivate multiple times. The cumulative frequency of the first braking action refers to the frequency of braking operations during the detour or return trajectory. Frequent braking implies higher energy consumption and poorer control efficiency. The steering motor will make multiple adjustments during the detour trajectory. The cumulative change in steering angle of the first steering motor measures the steering adjustment range throughout the obstacle avoidance process. Frequent or excessive steering increases control difficulty and mechanical wear. The rate of change in acceleration of the first drive motor reflects the vehicle's acceleration and deceleration during the detour and return trajectory. A large rate of change in acceleration may lead to vehicle instability or an unsmooth control response. The energy consumption throughout the obstacle avoidance process can be quantified by evaluating the energy consumption of the vehicle when executing the detour and return trajectory. This includes braking energy consumption, acceleration energy consumption, and control system energy consumption, reflecting the overall efficiency of the trajectory.

[0080] The first maximum lateral tracking error refers to the maximum distance the vehicle deviates from the predetermined trajectory. A larger lateral error means the vehicle deviates far from the trajectory, which may affect the obstacle avoidance effect. By quantifying the maximum lateral error, we can understand the accuracy of the vehicle's control during obstacle avoidance. The first steering command correction frequency indicates how many steering command corrections the vehicle needs to make during the detour. Frequent corrections mean the instability of trajectory planning and increase the complexity of control. The first speed lag time refers to the time delay from when the vehicle receives a speed adjustment command to when the actual speed changes. A larger lag time may cause the vehicle to react too slowly, especially in complex obstacle avoidance scenarios, where the delay increases the risk of collision.

[0081] The cumulative frequency of the first braking action, the cumulative change in the first steering motor angle, the rate of change in the first drive motor acceleration, the cumulative energy consumption for first obstacle avoidance, the first maximum lateral tracking error, the frequency of the first steering command correction, and the first speed lag time are weighted and fused. Each indicator is assigned a different weight according to its degree of influence on trajectory execution. More critical indicators, such as energy consumption and tracking error, are given higher weights. A hierarchical weighting method is used to weight the indicators at different levels to balance the influence of various aspects. Based on the weighted fusion result, a comprehensive first trajectory optimization evaluation parameter is output, which can reflect the quality of the trajectory.

[0082] Example 2, based on the same inventive concept as the autonomous obstacle avoidance method for low-speed operating vehicles in the previous examples, such as... Figure 2 As shown in the figure, this application provides an autonomous obstacle avoidance system for low-speed working vehicles, the system comprising:

[0083] The system comprises the following modules: a perception field construction module 10, used to construct a transient environmental perception field based on the driving direction of the low-speed working vehicle; an environmental perception module 20, used to control a multimodal sensor array to perform environmental perception within the transient environmental perception field and obtain environmental perception data; a detection and tracking module 30, used to construct a locally occupied grid map based on the environmental perception data, perform multi-target dynamic obstacle detection and tracking, and obtain multiple real-time motion trends of multiple real-time obstacles; a state fitting module 40, used to perform spatiotemporal motion state fitting based on the multiple real-time motion trends, and obtain multiple obstacle prediction trajectories; and an obstacle avoidance fitting module 50, used to perform obstacle avoidance fitting based on the predetermined driving trajectory of the low-speed working vehicle and the multiple obstacle prediction trajectories. The system performs multi-constraint dynamic obstacle avoidance fitting to obtain multiple obstacle avoidance trajectories; a feasibility analysis module 60 is used to perform path regression feasibility analysis based on the multiple obstacle avoidance trajectories and the predetermined driving trajectory, generating multiple obstacle avoidance regression trajectories; a joint evaluation module 70 is used to perform joint evaluation of the control complexity of the multiple obstacle avoidance trajectories and the multiple obstacle avoidance regression trajectories, outputting multiple trajectory optimization evaluation parameters; a closed-loop iteration module 80 is used to select and output autonomous obstacle avoidance linkage trajectories from the multiple obstacle avoidance trajectories and the multiple obstacle avoidance regression trajectories according to the descending order of the multiple trajectory optimization evaluation parameters, and after executing the trajectory tracking control of the low-speed operating vehicle, it performs closed-loop iteration of obstacle avoidance decision control based on the transient environment update field.

[0084] Furthermore, the sensing field construction module 10 is used to perform the following operation steps:

[0085] The system interacts with the real-time driving speed, maximum braking acceleration, and control response delay of the low-speed operating vehicle; calculates the minimum safe perception distance based on the real-time driving speed, maximum braking acceleration, and control response delay; extracts the real-time traffic road network width based on the real-time road network coordinates of the low-speed operating vehicle; after matching the standard environmental perception field with the environmental topology feature library using the minimum safe perception distance, it uses the real-time traffic road network width to correct the field boundary constraints and outputs the transient environmental perception field.

[0086] Furthermore, the environment perception module 20 is used to perform the following operation steps:

[0087] Multiple sensing modal activation modes with predefined field size ranges are used to construct a sensor scheduling decision library. The transient environmental sensing field is used to match the transient activation mode in the sensor scheduling decision library. The transient activation mode is used to control the multimodal sensing array to perform environmental sensing in the transient environmental sensing field, and M single-modal sensing data are obtained. The M single-modal sensing data are spatiotemporally aligned to obtain the environmental sensing data.

[0088] Furthermore, the detection and tracking module 30 is used to perform the following operation steps:

[0089] The environmental perception data of M single-modal sensing data are processed into M single-modal grid layers by data rasterization; M spatial confidence weights of M single-modal sensors are matched according to the real-time operating environment of the transient environmental perception field; dynamic weight fusion of the M single-modal grid layers is performed based on the M spatial confidence weights to obtain the local occupancy grid map; Euclidean distance clustering is performed on the local occupancy grid map to obtain multiple obstacle clusters; multiple geometric features and multiple motion features of the multiple obstacle clusters are extracted; obstacle type identification is performed based on the multiple geometric features to obtain the multiple real-time obstacles; and multi-level state prediction of motion trends is performed based on the multiple motion features to obtain the multiple real-time motion trends.

[0090] Furthermore, the state fitting module 40 is used to perform the following operation steps:

[0091] A multi-order time-domain analysis is performed on the first real-time motion trend to extract the first real-time position, first real-time velocity, first real-time acceleration, first motion direction change rate, and first radius of curvature. A first motion pattern is identified based on the first motion direction change rate and first radius of curvature. A motion model is selected based on the first real-time obstacle and the first motion pattern, and the first motion model is retrieved. The first real-time position, first real-time velocity, and first real-time acceleration are used as the initial state of the first motion and input into the first motion model for spatiotemporal multi-perturbation hypothetical trajectory prediction, resulting in multiple candidate predicted trajectories. Based on multi-dimensional probability evaluation, the trajectories of the multiple candidate predicted trajectories are fused to output the first obstacle predicted trajectory.

[0092] Furthermore, the state fitting module 40 is used to perform the following operation steps:

[0093] Environmental semantic constraints are corrected and interaction forces between obstacles are compensated for on the multiple candidate predicted trajectories to obtain multiple corrected predicted trajectories; multidimensional probability evaluation is performed on the multiple corrected predicted trajectories to obtain multiple trajectory comprehensive confidence probabilities, wherein the evaluation items of the multidimensional probability evaluation include motion continuity, environmental rationality and interaction safety; based on the multiple trajectory comprehensive confidence probabilities, the multiple corrected predicted trajectories are fused to output the first obstacle predicted trajectory.

[0094] Furthermore, the obstacle avoidance fitting module 50 is used to perform the following operation steps:

[0095] Based on the vehicle dimensions of the low-speed operating vehicle and the predetermined driving trajectory, a spatiotemporal safety space channel is constructed in a four-dimensional spatiotemporal grid space; the multiple obstacle avoidance detour trajectories are mapped to the four-dimensional spatiotemporal grid space, and multi-resolution conflict detection is performed with the spatiotemporal safety space channel to locate distributed conflict regions; based on the distributed conflict regions, detour routes are fitted in the spatiotemporal safety space channel based on vehicle kinematic constraints and road network topology constraints to obtain the multiple obstacle avoidance detour trajectories.

[0096] Furthermore, the feasibility analysis module 60 is used to perform the following steps:

[0097] Calculate the first obstacle avoidance detour trajectory and the first feasible regression point of the predetermined driving trajectory; use a Clothoid curve to smoothly connect the trajectory endpoint of the first obstacle avoidance detour trajectory and the first feasible regression point to generate the first obstacle avoidance regression trajectory.

[0098] Furthermore, the joint evaluation module 70 is used to perform the following operational steps:

[0099] The control cost of the first obstacle avoidance trajectory and the first obstacle avoidance return trajectory is quantified to obtain the cumulative frequency of the first braking action, the cumulative angle change of the first steering motor, the rate of change of acceleration of the first drive motor, and the first cumulative obstacle avoidance energy consumption. The control response characteristics of the first obstacle avoidance trajectory and the first obstacle avoidance return trajectory are quantified to obtain the first maximum lateral tracking error, the first steering command correction frequency, and the first speed lag time. The first braking action cumulative frequency, the cumulative angle change of the first steering motor, the rate of change of acceleration of the first drive motor, the first cumulative obstacle avoidance energy consumption, the first maximum lateral tracking error, the first steering command correction frequency, and the first speed lag time are hierarchically weighted and fused to output the first trajectory optimization evaluation parameters.

[0100] Through the foregoing detailed description of the autonomous obstacle avoidance method for low-speed operating vehicles, those skilled in the art can clearly understand the autonomous obstacle avoidance system for low-speed operating vehicles in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for autonomous obstacle avoidance for a low speed work vehicle, characterized in that, The method comprises: constructing a transient environment perception field according to the driving direction of the low-speed working vehicle; controlling a multi-modal sensor array to perform environment perception in the transient environment perception field to obtain environment perception data; constructing a local occupancy grid map based on the environment perception data, performing multi-target dynamic obstacle detection and tracking, and obtaining multiple real-time motion trends of multiple real-time obstacles; performing spatio-temporal motion state fitting according to the multiple real-time motion trends to obtain multiple obstacle prediction trajectories; performing multi-constraint dynamic obstacle avoidance fitting according to the predetermined driving trajectory of the low-speed working vehicle and the multiple obstacle prediction trajectories to obtain multiple obstacle avoidance detour trajectories; performing path regression feasibility analysis based on the multiple obstacle avoidance detour trajectories and the predetermined driving trajectory to generate multiple obstacle avoidance regression trajectories; performing control complexity joint evaluation on the multiple obstacle avoidance detour trajectories and the multiple obstacle avoidance regression trajectories to output multiple trajectory optimization evaluation parameters; According to the descending order arrangement result of the multiple trajectory optimization evaluation parameters, the autonomous obstacle avoidance linkage trajectory is output from the multiple obstacle avoidance detour trajectories and the multiple obstacle avoidance regression trajectories, and after the trajectory tracking control of the low-speed working vehicle is performed, the obstacle avoidance decision control closed-loop iteration is performed according to the transient environment update field.

2. The autonomous obstacle avoidance method for a low-speed work vehicle according to claim 1, characterized by, According to the driving direction of the low-speed working vehicle, a transient environment perception field is constructed, and the method comprises: interacting the real-time driving speed, the maximum braking acceleration and the control response time delay of the low-speed working vehicle; calculating the minimum safe perception distance according to the real-time driving speed, the maximum braking acceleration and the control response time delay; extracting the real-time passing road network width according to the real-time road network coordinates of the low-speed working vehicle; after the minimum safe perception distance is matched with the standard environment perception field in the environment topology feature library, the real-time passing road network width is used for field boundary constraint correction, and the transient environment perception field is output.

3. The autonomous obstacle avoidance method for a low-speed work vehicle according to claim 1, wherein, controlling a multi-modal sensor array to perform environment perception in the transient environment perception field to obtain environment perception data, the method comprising: predefining multiple sensor modal activation modes of multiple field size intervals to construct a sensor scheduling decision library; matching the transient activation mode in the sensor scheduling decision library by using the transient environment perception field; controlling the multi-modal sensor array in the transient environment perception field to perform environment perception by using the transient activation mode to obtain M single-modal perception data; spatio-temporal aligning the M single-modal perception data to obtain the environment perception data.

4. The autonomous obstacle avoidance method for a low-speed work vehicle according to claim 1, characterized by, based on the environment perception data, constructing a local occupancy grid map, performing multi-target dynamic obstacle detection and tracking, and obtaining multiple real-time motion trends of multiple real-time obstacles, the method comprising: performing data gridding processing on the M single-modal perception data in the environment perception data to obtain M single-modal grid layers; matching M spatial confidence weights of M single-modal sensors according to the real-time working environment of the transient environment perception field; performing dynamic weight fusion of the M single-modal grid layers according to the M spatial confidence weights to obtain the local occupancy grid map; performing Euclidean distance clustering on the local occupancy grid map to obtain multiple obstacle clusters; extracting a plurality of geometric features and a plurality of motion features of the plurality of obstacle clusters; performing obstacle type recognition based on the plurality of geometric features to obtain the plurality of real-time obstacles, and performing motion trend multi-order state prediction based on the plurality of motion features to obtain the plurality of real-time motion trends.

5. The autonomous obstacle avoidance method for a low-speed work vehicle according to claim 1, wherein, According to the plurality of real-time motion trends, a spatio-temporal motion state fitting is performed to obtain a plurality of obstacle prediction trajectories, and the method comprises: performing multi-order time domain analysis on the first real-time motion trend, and extracting a first real-time position, a first real-time speed, a first real-time acceleration, a first motion direction change rate and a first curvature radius; identifying a first motion pattern based on the first motion direction change rate and the first curvature radius; performing motion model selection according to the first real-time obstacle and the first motion pattern, and calling a first motion model; inputting the first real-time position, the first real-time speed and the first real-time acceleration as a first motion initial state into the first motion model to perform spatio-temporal multi-disturbance hypothesis trajectory prediction, and obtaining a plurality of candidate prediction trajectories; performing trajectory fusion of the plurality of candidate prediction trajectories based on multi-dimensional probability evaluation, and outputting a first obstacle prediction trajectory.

6. The autonomous obstacle avoidance method for a low-speed work vehicle according to claim 5, wherein, performing trajectory fusion of the plurality of candidate prediction trajectories based on multi-dimensional probability evaluation, and outputting a first obstacle prediction trajectory, and the method comprises: performing environment semantic constraint correction and obstacle interaction force compensation on the plurality of candidate prediction trajectories to obtain a plurality of modified prediction trajectories; performing multi-dimensional probability evaluation on the plurality of modified prediction trajectories to obtain a plurality of trajectory comprehensive confidence probabilities, wherein the evaluation items of the multi-dimensional probability evaluation include motion continuity, environmental rationality and interaction safety; performing trajectory fusion of the plurality of modified prediction trajectories based on the plurality of trajectory comprehensive confidence probabilities, and outputting the first obstacle prediction trajectory.

7. The autonomous obstacle avoidance method for low speed work vehicle of claim 1, wherein, According to the predetermined driving trajectory of the low-speed operation vehicle and the plurality of obstacle prediction trajectories, a multi-constraint dynamic obstacle avoidance fitting is performed to obtain a plurality of obstacle avoidance detour trajectories, and the method comprises: constructing a spatio-temporal safety space channel in a four-dimensional spatio-temporal grid space according to the vehicle size of the low-speed operation vehicle and the predetermined driving trajectory; mapping the plurality of obstacle avoidance detour trajectories to the four-dimensional spatio-temporal grid space, and performing multi-resolution conflict detection with the spatio-temporal safety space channel to locate a distributed conflict area; According to the distributed conflict area, a detour route fitting based on vehicle kinematics constraints and road network topological structure constraints is performed in the spatio-temporal safety space channel to obtain the plurality of obstacle avoidance detour trajectories.

8. The autonomous obstacle avoidance method for a low-speed work vehicle of claim 1, wherein, Based on the plurality of obstacle avoidance detour trajectories and the predetermined driving trajectory, a path regression feasibility analysis is performed to generate a plurality of obstacle avoidance regression trajectories, and the method comprises: calculating a first feasible regression point of the first obstacle avoidance detour trajectory and the predetermined driving trajectory; adopting a Clothoid curve to smoothly connect the trajectory endpoint of the first obstacle avoidance detour trajectory and the first feasible regression point to generate a first obstacle avoidance regression trajectory.

9. The autonomous obstacle avoidance method for low speed operation oriented vehicle according to claim 1, wherein, performing control complexity joint evaluation on the plurality of obstacle avoidance detour trajectories and the plurality of obstacle avoidance regression trajectories to output a plurality of trajectory optimization evaluation parameters, and the method comprises: The first obstacle avoidance detour trajectory and the first obstacle avoidance return trajectory are subjected to control cost quantification, to obtain a first brake action cumulative frequency, a first steering motor cumulative rotation angle change, a first drive motor acceleration change rate, and a first cumulative obstacle avoidance energy consumption; The first obstacle avoidance detour trajectory and the first obstacle avoidance return trajectory are subjected to control response characteristic quantification, to obtain a first maximum lateral tracking error, a first steering instruction correction frequency, and a first speed lag time; The first brake action cumulative frequency, the first steering motor cumulative rotation angle change, the first drive motor acceleration change rate, the first cumulative obstacle avoidance energy consumption, the first maximum lateral tracking error, the first steering instruction correction frequency, and the first speed lag time are hierarchically weighted and fused to output a first trajectory optimization evaluation parameter.

10. An autonomous obstacle avoidance system for a low speed work vehicle, characterized by The system for implementing the autonomous obstacle avoidance method for the low-speed operation vehicle according to any one of claims 1-9, the system comprising: a perception field construction module configured to construct a transient environment perception field according to a driving direction of the low-speed operation vehicle; an environment perception module configured to control a multi-modal sensor array to perform environment perception in the transient environment perception field, to obtain environment perception data; a detection and tracking module configured to construct a local occupancy grid map based on the environment perception data, to perform multi-target dynamic obstacle detection and tracking, to obtain a plurality of real-time motion trends of a plurality of real-time obstacles; a state fitting module configured to perform spatiotemporal motion state fitting according to the plurality of real-time motion trends, to obtain a plurality of obstacle prediction trajectories; an obstacle avoidance fitting module configured to perform multi-constraint dynamic obstacle avoidance fitting according to a predetermined driving trajectory of the low-speed operation vehicle and the plurality of obstacle prediction trajectories, to obtain a plurality of obstacle avoidance detour trajectories; a feasibility analysis module configured to perform path return feasibility analysis based on the plurality of obstacle avoidance detour trajectories and the predetermined driving trajectory, to generate a plurality of obstacle avoidance return trajectories; a joint evaluation module configured to perform control complexity joint evaluation on the plurality of obstacle avoidance detour trajectories and the plurality of obstacle avoidance return trajectories, to output a plurality of trajectory optimization evaluation parameters; a closed-loop iteration module configured to, according to a descending order arrangement result of the plurality of trajectory optimization evaluation parameters, select an autonomous obstacle avoidance linkage trajectory from the plurality of obstacle avoidance detour trajectories and the plurality of obstacle avoidance return trajectories, and perform trajectory tracking control of the low-speed operation vehicle, and then perform obstacle avoidance decision control closed-loop iteration according to a transient environment update field.

Images