Method for selecting transverse motion control object of autonomous vehicle

Abstract

The invention relates to the technical field of automatic driving vehicle control, and discloses an automatic driving vehicle transverse motion control object selection method. According to the method, real-time sensing data of the surrounding environment of the automatic driving vehicle is collected in an omnibearing mode through a sensing device, the data covers various kinds of information around the vehicle, and a rich basis is provided for follow-up decision making. A plurality of control object selection points are generated based on the vehicle state parameters, and a control scheduling sequence is formed according to the points. A control scheduling command is generated according to real-time sensing data, the transverse movement of the vehicle is adjusted by means of the command, the vehicle can be ensured to run according to a preset strategy, meanwhile, a control metric value corresponding to each control object selection point is obtained, and the metric values reflect the effect of each selection point in the control process. And aggregating the control metric values of all the control object selection points, generating a comprehensive control analysis result, and outputting a control termination signal or a control retry signal based on the control analysis result so as to optimize a control strategy.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle control technology, specifically to a method for selecting a control object for lateral movement of an autonomous vehicle. Background Technology

[0002] With the rapid development of technology, autonomous driving has become a research hotspot and development direction in the global transportation field. From the initial driver assistance systems to the gradual advancement towards high automation and even fully autonomous driving, autonomous driving technology has continuously progressed, gradually realizing core functions such as vehicle environmental perception, route planning, and control execution. Major automakers and technology companies are actively investing in research and development, pushing autonomous driving technology towards higher levels of automation. This technology is of great significance for improving traffic safety, alleviating traffic congestion, and enhancing the travel experience. In terms of improving traffic safety, autonomous driving systems can quickly and accurately perceive road conditions and react in a timely manner, avoiding traffic accidents caused by human negligence or fatigue driving. In terms of alleviating traffic congestion, intelligent route planning and reasonable speed control can make traffic flow smoother. At the same time, autonomous driving can also bring passengers a more comfortable and convenient travel experience, allowing them to rest, work, or entertain themselves freely during the journey.

[0003] In autonomous driving systems, lateral motion control is a crucial component, responsible for controlling the vehicle's direction of travel and ensuring it follows a predetermined path. Whether maintaining a stable lane on a highway or navigating complex road conditions such as turns and lane changes in urban areas, precise lateral motion control plays a decisive role. For example, on highways, where vehicles need to stay within their lanes, lateral motion control precisely adjusts the steering wheel angle to ensure the vehicle maintains a suitable distance from the lane markings. When turning in urban areas, it accurately controls the vehicle's steering based on factors such as the curve radius and vehicle speed, allowing the vehicle to smoothly navigate the curve. In short, the accuracy and reliability of lateral motion control directly impact the safety and stability of autonomous driving.

[0004] In the field of autonomous driving, the input source for lateral motion control largely relies on lane line recognition and definition. The principle is that after lane line recognition, the system determines the relative position of the autonomous vehicle to the lane lines, thereby controlling the vehicle's lateral movement to keep it within its lane and prevent it from erroneously veering out of the lane. For example, a common lane-keeping assist system uses a camera to detect lane boundary lines. When the vehicle approaches the boundary, the system vibrates the steering wheel to alert the driver. If the driver does not explicitly change lanes and the vehicle does not leave the lane, the system will continuously vibrate to remind the driver and may even automatically fine-tune the steering wheel to bring the vehicle back to the center of the lane.

[0005] However, in actual driving conditions, unclear or missing lane markings are commonplace. On the one hand, due to long-term use, lane markings are easily worn and faded due to frequent vehicle traffic and environmental erosion. In older urban roads, due to their age and high traffic volume, lane markings become blurred after years of tire friction; on some mountain roads, the paint gradually peels off due to long-term rain and sun exposure. On the other hand, sudden changes in road direction or temporary road closures can also prevent lane markings from being recognized by relevant sensors. For example, during road construction, existing lane markings may be covered or replaced by new temporary markings, which may not be clear or standardized; during traffic control measures following an accident, lane settings may be temporarily altered without timely and clear lane markings. Summary of the Invention

[0006] The purpose of this invention is to provide a method for selecting the lateral motion control object of an autonomous vehicle, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for selecting a lateral motion control object for an autonomous vehicle, the method comprising: Real-time perception data of the surrounding environment of autonomous vehicles are collected using sensing devices. Multiple control object selection points are generated based on vehicle state parameters, and a control scheduling sequence is formed based on the control object selection points; Based on real-time sensing data, control and scheduling commands are generated, and the lateral movement of the vehicle is adjusted according to the control and scheduling commands. The control metric value corresponding to the selected point of each controlled object is also obtained. Aggregate the control metrics of all selected points of the controlled objects, generate control analysis results, and output control termination signals or control retry signals.

[0008] Preferably, the step of generating multiple control object selection points based on vehicle state parameters includes: acquiring three-dimensional point cloud data of the vehicle's surrounding environment through an onboard perception system, and extracting lane line features and obstacle contours using point cloud data processing algorithms; calculating lateral motion constraints based on the vehicle dynamics model to generate an initial set of control candidate points; grouping the initial set of control candidate points using a clustering analysis algorithm and removing outliers; calculating the center point of each cluster as a control object selection point based on the grouping results; and dynamically adjusting the density of control object selection points based on real-time traffic flow data, increasing the number of selection points in densely trafficked areas.

[0009] Preferably, the step of forming a control scheduling sequence based on the selected points of the controlled object includes: assigning a priority score to each selected point of the controlled object, the score being calculated based on the current distance between the point and the vehicle, the curvature of the lane where the point is located, and the density of obstacles around the point; sorting the selected points of the controlled object according to the priority scores from high to low; introducing a time window mechanism to divide the sorted points into multiple time segments; within each time segment, calculating the reachability of the points based on the expected speed of the vehicle and adjusting the order of the points; and finally connecting the adjusted point sequence into a control scheduling sequence.

[0010] Preferably, the starting control object selection point is determined by: real-time monitoring of the offset of the vehicle's lateral position from the lane centerline; calculating the Euclidean distance between each control object selection point and the vehicle's current position; and combining the offset and distance with a weighted sum, selecting the point with the smallest weighted value as the starting control object selection point.

[0011] Preferably, the step of generating control scheduling commands based on real-time sensing data includes: continuously collecting vehicle lateral acceleration and yaw rate data; constructing a data buffer based on a sliding window to store sensor readings over a recent period; smoothing the buffer data using a digital filter; calculating the deviation between the processed data and the expected control target; if the deviation is less than the tolerance range, generating a command to maintain the current scheduling; if the deviation exceeds the tolerance range, generating an acceleration or deceleration scheduling command based on the direction of the deviation.

[0012] Preferably, the adjustment of the vehicle's lateral movement by means of control scheduling commands includes: parsing the type of control scheduling command; if it is a maintenance command, maintaining the current lateral control output; if it is an acceleration command, increasing the steering motor torque output; if it is a deceleration command, decreasing the torque output and activating brake assist; and monitoring the vehicle's attitude in real time during the adjustment process, and fine-tuning the control quantity through feedback loops.

[0013] Preferably, the process of aggregating the control metric values ​​of all selected control object points to generate control analysis results includes: collecting multiple sample values ​​of control metrics for each selected control object point; calculating the mean and standard deviation of the sample values; comparing the metric distributions of different points using hypothesis testing methods; if the distribution consistency passes the test, merging the data to generate an overall characteristic curve; otherwise, retaining the differences between points to mark abnormal areas; and finally outputting control analysis results with confidence intervals.

[0014] Preferably, the step of selecting the point with the smallest weighted value as the starting control object selection point includes: real-time acquisition of lane line recognition confidence data, calculation of dynamic weight coefficients in combination with vehicle lateral position offset; adjustment of the sensitivity of the weight coefficients according to weather conditions, optimization of the weight calculation process by integrating historical starting point selection success rate data; and finally, selection of the point with the smallest comprehensive weighted value as the starting control object selection point.

[0015] Preferably, the application of cluster analysis algorithm to group the initial control candidate point set includes: using density clustering algorithm to identify high-density regions in the point set; calculating the local density and minimum distance of each point; determining the cluster center based on the density peak and assigning the point to the nearest cluster center; removing outliers with a density below a threshold; and dynamically adjusting the cluster radius to adapt to different traffic scenarios.

[0016] Preferably, the step of assigning priority scores to the selected points for each controlled object includes: extracting the position coordinates of the point and the current pose of the vehicle; calculating the consistency between the point and the vehicle's heading angle; evaluating the visibility and lighting conditions of the area where the point is located; and integrating historical control success rate data to calculate a comprehensive score using a weighted summation formula.

[0017] Compared with the prior art, the beneficial effects of the present invention are: One of the greatest advantages of the lateral motion control object selection method for autonomous vehicles proposed in this invention is its excellent adaptability to complex road conditions. Traditional lateral motion control methods rely heavily on lane line recognition. Once lane lines become unclear or missing for various reasons, they become difficult to function properly, seriously threatening the safety and stability of vehicle driving. This method overcomes this limitation by not using lane lines as the sole control basis. In road construction scenarios, lane lines are often partially covered or obscured due to the accumulation of construction materials and the placement of warning signs, making them difficult to identify. In such cases, this method collects real-time perception data of the vehicle's surrounding environment through sensing devices. This data covers various information about the vehicle's surroundings, such as the position, speed, and direction of other vehicles, obstacles on the road, and roadside landmarks. Based on this rich perception data, the vehicle can generate multiple control object selection points and form a control scheduling sequence based on these points. This means that the vehicle no longer relies solely on lane lines to determine its driving direction but comprehensively considers various factors in the surrounding environment to autonomously find a suitable driving path. Even if the lane markings completely disappear, the vehicle can still drive stably, effectively avoiding deviating from the lane and greatly reducing the risk of traffic accidents.

[0018] This method also demonstrates significant advantages in situations where natural disasters, such as torrential rain, floods, and mudslides, damage roads and disrupt lane markings. Heavy rain can cause severe flooding, washing away the lane marking paint and making them illegible; floods and mudslides can directly destroy roads, altering their original shape and completely eliminating lane markings. In these extreme situations, lane-based control methods often struggle, while this method can flexibly adjust vehicle direction by generating real-time sensing data and control object selection points. Vehicles can use the trajectories of other vehicles as a reference, or rely on undamaged buildings, trees, and other landmarks along the roadside to determine their position and direction, thus ensuring safe driving in adverse road conditions and protecting the lives of passengers.

[0019] This method acquires real-time perception data of the surrounding environment of autonomous vehicles from all directions and multiple dimensions through sensing devices. These sensing devices include, but are not limited to, cameras, lidar, and millimeter-wave radar, each with its unique advantages, capable of acquiring information about the vehicle's surroundings from different angles. Cameras can capture visual images of the vehicle's surroundings, providing rich texture and color information to help identify various signs, markings, and features of other vehicles and pedestrians on the road; lidar, by emitting laser beams and receiving reflected signals, accurately measures the distance between the vehicle and surrounding objects, constructing a three-dimensional model of the surrounding environment, which is highly effective for detecting the position and shape of obstacles; millimeter-wave radar uses electromagnetic waves in the millimeter-wave band to detect the distance, speed, and angle of target objects, has strong penetration capabilities, and can operate normally even in adverse weather conditions. By fusing this multi-source data, the vehicle can obtain comprehensive and accurate information about its surrounding environment. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the method for selecting lateral motion control objects for autonomous vehicles according to the present invention. Figure 2 A flowchart for generating multiple control object selection points based on vehicle state parameters; Figure 3 This is a flowchart for forming a control scheduling sequence based on the points selected by the controlled object. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a method for selecting a lateral motion control object for an autonomous vehicle, the method comprising: Precise control of vehicle lateral movement is achieved through a series of coordinated processes. Real-time perception data of the surrounding environment of the autonomous vehicle is collected using sensors, including but not limited to camera images, LiDAR point clouds, and millimeter-wave radar signals, providing foundational information for subsequent processing. Multiple control object selection points are generated based on vehicle state parameters, encompassing dynamic indicators such as vehicle speed, steering angle, and yaw rate. The generation of control object selection points depends on environmental characteristics and vehicle motion constraints. A control scheduling sequence is formed based on the control object selection points, considering the spatial distribution and temporal reachability of the points. Control scheduling commands are generated based on real-time perception data, with command generation based on deviation analysis between sensor readings and expected targets. The lateral movement of the vehicle is adjusted using these commands, involving torque control of the steering system. Control metrics corresponding to each control object selection point are obtained, reflecting control effectiveness such as path tracking error. The control metrics of all control object selection points are aggregated to generate control analysis results. The analysis results integrate multi-point data using statistical methods and output control termination or retry signals, the output of which depends on the confidence level of the analysis results.

[0023] Example 1: See Figure 2 The process of generating multiple control object selection points begins with the vehicle-mounted perception system acquiring 3D point cloud data of the vehicle's surrounding environment. This 3D point cloud data is generated through solid-state LiDAR scanning and stereo matching fusion using dual forward-looking cameras. The LiDAR scanning frequency is 20 Hz, and the camera acquisition frame rate is 30 frames per second. The two data streams are aligned in time using a hardware synchronization signal. The point cloud data processing algorithm extracts lane line features and obstacle contours. First, the algorithm downsamples the original 3D point cloud using a voxel mesh to reduce the amount of data. The downsampled point cloud is then fed into the ground point extraction module, which uses a filtering algorithm based on radial slope constraints to distinguish between ground points and non-ground points. Non-ground point clusters are grouped using a Euclidean clustering algorithm. Each point cluster is fed into a feature calculation unit to calculate its 2D bounding box and main direction. Lane line features are identified from the ground point cloud using a region growing algorithm based on intensity values. Obstacle contours are obtained by calculating the convex hull polygons of the non-ground point clusters.

[0024] The vehicle dynamics model calculates lateral motion constraints. The vehicle dynamics model is a three-degree-of-freedom bicycle model including lateral, longitudinal, and yaw motions. Model parameters include vehicle mass, moment of inertia about the Z-axis, front wheel lateral stiffness, and rear wheel lateral stiffness. Model inputs are steering wheel angle and vehicle speed. Model state variables are center-of-gravity sideslip angle and yaw rate. The model solves differential equations using numerical integration to obtain the vehicle's trajectory set over a future period. Lateral motion constraints are defined as boundary conditions that maintain a minimum safe distance between the vehicle trajectory and the lane boundary. An initial control candidate point set is generated, using the road centerline as a baseline. A lateral offset point is set every meter along the baseline, with the lateral offset range determined by the lane width minus the vehicle width. Each point is appended with a timestamp calculated from the vehicle's current speed.

[0025] Clustering analysis algorithms are applied to group the initial control candidate point set. The algorithm employs a density-based noise-based spatial clustering algorithm, which sets two core parameters: the neighborhood search radius and the minimum number of points. The initial neighborhood search radius is set to five meters, and the minimum number of points is set to four. The density clustering algorithm identifies high-density regions within the point set, defined as connected regions where the number of points in the neighborhood exceeds the minimum number of points. The local density and minimum distance are calculated for each point. Local density is defined as the number of other points contained within a sphere centered on the point and with the neighborhood search radius as its radius. The minimum distance is defined as the minimum Euclidean distance between the point and any point with a higher local density. Cluster centers are determined based on density peaks. A point with a density peak satisfies the condition that its local density is greater than the local density of all its neighbors and its minimum distance is greater than a set threshold. The point is then assigned to the nearest cluster center, and the assignment process uses a breadth-first search algorithm to traverse the density-reachable point set. Outliers with a density below a threshold are removed. Outlier criteria are that the local density of a point is less than two and the minimum distance is greater than ten meters. The clustering radius is dynamically adjusted to adapt to different traffic scenarios. In highway scenarios, the clustering radius is increased to eight meters to cover a wider area, while in urban road scenarios, the clustering radius is decreased to three meters to improve resolution.

[0026] Based on the grouping results, the center point of each cluster is calculated as the control object selection point. The coordinates of the cluster center point are obtained by calculating the geometric mean of all points within the cluster, and the timestamp of the center point is the median of the timestamps of the points within the cluster. The density of the control object selection points is dynamically adjusted according to real-time traffic flow data. Real-time traffic flow data is obtained by tracking surrounding vehicles through vehicle-mounted radar. The traffic flow data includes the number, speed, and position of surrounding vehicles. The number of selection points is increased in dense traffic areas. Dense traffic areas are defined as road sections with more than five vehicles within a 100-meter radius. The increase in point density is achieved by reducing the longitudinal sampling interval to 0.5 meters.

[0027] The process of grouping the initial control candidate point set using clustering analysis algorithms is further refined. Density clustering algorithms identify high-density regions within the point set, corresponding to areas with clustered obstacles or clearly defined lane lines. The local density and minimum distance for each point are calculated. Local density is calculated using a Gaussian kernel density estimation method, with the bandwidth of the Gaussian kernel adaptively adjusted based on the point cloud density. The minimum distance is used to distinguish core points from noise points. Cluster centers are determined based on density peaks, identified by calculating the product of each point's local density and minimum distance and finding local maxima. Points are assigned to the nearest cluster centers using a KD-tree-based nearest neighbor search algorithm. The KD-tree is constructed using the point's 3D coordinates and timestamp. Outliers with densities below a threshold are removed. The threshold setting considers the ranging error of the LiDAR and the point cloud registration error. The cluster radius is dynamically adjusted to adapt to different traffic scenarios, based on historical traffic pattern data downloaded from a cloud-based traffic management platform.

[0028] The point cloud data processing algorithm employs a multi-stage processing architecture. The first stage performs coordinate system unification, transforming the LiDAR point cloud from the sensor coordinate system to the vehicle coordinate system, and mapping the visual detection results from the image coordinate system to the vehicle coordinate system through inverse perspective transformation. The second stage performs timestamp alignment, adding a GPS timestamp to each point cloud frame and image frame, and compensating for the acquisition time difference between sensors using linear interpolation. The third stage performs data fusion, using a Kalman filter to estimate the state of the LiDAR point cloud and visual bounding box of the same object, generating a fused target list with velocity vectors.

[0029] The vehicle dynamics model is solved using the fourth-order Runge-Kutta method for numerical integration, with an integration step size of 0.1 seconds and a prediction time domain length of 3 seconds. Lateral motion constraints are generated based on lane boundary information provided by a high-precision map. The distance between the vehicle trajectory and the lane boundary is obtained by calculating the perpendicular distance from the trajectory point to the lane boundary line segment. The minimum safe distance is set to half the vehicle width plus a 0.3-meter margin. The parameter adjustment strategy for the density clustering algorithm is based on real-time point cloud density, obtained by counting the number of points per unit volume. When the point cloud density exceeds ten points per cubic meter, the neighborhood search radius is automatically reduced by 20% to improve clustering sensitivity. The outlier removal module adds a secondary verification mechanism, performing a motion consistency check on initially identified outliers. This check compares the instantaneous velocity of the point with the average velocity of surrounding points; points with a velocity difference exceeding two meters per second are reclassified as valid points. The density adjustment module for the selected control points integrates V2X communication data, which provides traffic light status and road construction information. When a road construction warning is received ahead, the point density within a 100-meter range upstream of the construction area is doubled to improve control accuracy. The ground point extraction module in the point cloud data processing algorithm incorporates rain-adaptive processing. This rain-adaptive processing identifies noise points caused by slippery road surfaces by analyzing the point cloud reflection intensity histogram. Points with reflection intensity below a threshold are marked as raindrop noise and filtered out.

[0030] Example 2: See Figure 3 The steps involved in forming the control scheduling sequence include assigning a priority score to each selected point of the controlled object. The priority score is calculated based on a combination of the current distance between the point and the vehicle, the curvature of the lane where the point is located, and the density of obstacles around the point. The current distance between the point and the vehicle is obtained by fusing data from the Global Positioning System (GPS) and Inertial Measurement Unit (INS), and the distance is calculated using the two-dimensional Euclidean distance formula from the center point of the vehicle's rear axle to the selected point of the controlled object. The curvature of the lane where the point is located is extracted from the lane centerline of a high-precision map, and the curvature is calculated using the three-point circle method, taking the reciprocal of the radius of curvature for three consecutive path points. The density of obstacles around the point is obtained by statistically analyzing the number of dynamic obstacles within a five-meter radius around the selected point of the controlled object. Dynamic obstacles include vehicles, pedestrians, and non-motorized vehicles.

[0031] The selected control points are sorted from highest to lowest priority score using a max-heap data structure. Each node in the max-heap stores the control point's ID and priority score. The heap adjustment algorithm employs a bottom-up filtering operation. A time window mechanism is introduced, with the window length adaptively calculated based on the vehicle's current speed. The time window length equals three seconds multiplied by the vehicle speed conversion factor. The sorted points are divided into multiple time segments, each corresponding to a control cycle with a length of 200 milliseconds. Within each time segment, the reachability of a point is calculated based on the vehicle's expected speed, derived from the speed curve output by the path planning module. The reachability calculation uses a vehicle kinematics model to predict the area the vehicle can reach in future time segments. The order of points is adjusted, considering both spatial uniformity and temporal continuity. Spatial uniformity is ensured by calculating the minimum distance between points, while temporal continuity requires a smooth transition between points in adjacent time segments. Finally, the adjusted point sequence is connected into a control scheduling sequence using a doubly linked list data structure. Each node contains the spatial coordinates, timestamp, and execution status flag of the selected control point.

[0032] The process of assigning priority scores to selected points for each controlled object is further detailed. The point's position coordinates and the vehicle's current pose are extracted. The vehicle's current pose includes eastward coordinates, northward coordinates, and heading angle; the coordinate data comes from the integrated navigation system. The consistency between the point and the vehicle's heading angle is calculated, obtained by calculating the cosine of the angle between the point's position vector relative to the vehicle and the vehicle's heading vector. Visibility and lighting conditions in the area where the point is located are evaluated. Visibility is assessed using the point cloud density and penetration rate of the LiDAR, while lighting conditions are quantified using the average grayscale value and contrast of the forward-looking camera image. Historical control success rate data is integrated; this data is extracted from the control system's operation log and records the execution results of each selected point for each controlled object over the past one hundred control cycles. A weighted summation formula is used to calculate a comprehensive score. This formula includes distance weight coefficients, curvature weight coefficients, obstacle density weight coefficients, visibility weight coefficients, and historical success rate weight coefficients; each weight coefficient is pre-set using expert experience.

[0033] The priority scoring process employs a multi-level weighted architecture. The first level calculates the base spatial score, which includes distance, curvature, and obstacle density components. The distance component uses a negative exponential function mapping, with higher scores for closer distances. The curvature component is processed using a piecewise function, receiving full marks when the curvature radius is greater than 200 meters and linearly decreasing when it is less than 50 meters. The obstacle density component is calculated using an inverse proportional function, receiving full marks when the density is zero and resetting the score to zero when the density exceeds 0.1 units per square meter.

[0034] The second level calculates the environmental adaptation score, which includes visibility and illumination components. The visibility component is calculated using the effective echo rate of the LiDAR point cloud; a score of full marks is awarded when the effective echo rate is above 90%, and zero when it is below 50%. The illumination component is calculated using the grayscale histogram distribution of the camera images, with illumination uniformity used to adjust the score. The third level calculates the historical performance score, based on the historical execution records of the selected points on the controlled object. Historical data for each point includes the number of successes, failures, and timeouts. The scoring formula uses a Bayesian probability model to calculate the prior probability of the next success. The scores from these three levels are weighted and fused to obtain the final priority score. The weights of each level are dynamically adjusted according to the real-time driving scenario; spatial scoring is given greater weight in highway scenarios, while environmental adaptation scoring is given greater weight in urban road scenarios.

[0035] The time window mechanism employs a sliding window method, with the window length proportional to vehicle speed and a fixed sliding step size of 50 milliseconds. Point sequence optimization within each time segment is solved using a genetic algorithm. The chromosome encoding in the genetic algorithm represents the point arrangement order, and the fitness function comprehensively considers point priority scores, spatial distribution uniformity, and temporal continuity. The population size is set to 50 individuals, the number of iterations is 100, the crossover probability is 0.8, and the mutation probability is 0.1. An integrity check mechanism is added to the connection operation of the control scheduling sequence. This integrity check includes timestamp continuity verification, spatial coordinate monotonicity testing, and control variable change rate limitation. Timestamp continuity verification requires that the time difference between adjacent points be within a reasonable range; spatial coordinate monotonicity testing ensures that the point sequence does not exhibit backtracking; and control variable change rate limitation ensures smooth transition of steering commands. Sequence storage uses a circular buffer structure, with the buffer size dynamically allocated based on the maximum prediction time domain. The calculation of the vehicle's expected speed integrates the navigation system's expected speed and the traffic environment's speed limits. The expected speed curve generation process considers a curve deceleration model and a following distance safety model. The cornering deceleration model calculates the safe passing speed based on lane curvature, while the following safety distance model calculates the appropriate speed based on the speed of the vehicle in front and the relative distance. Point accessibility analysis uses a vehicle kinematics model to predict the reachable area, considering maximum steering angular velocity and maximum lateral acceleration constraints. The reachable area is represented as a sector starting from the vehicle's current position. Historical control success rate data is updated using a sliding window averaging method, with the sliding window size being the last one hundred control records. The success rate statistics for the corresponding point are updated after each control execution. Adaptive adjustment of weight coefficients is implemented using a fuzzy logic controller. The inputs to the fuzzy logic controller are vehicle speed, road type, and weather conditions, and the output is the adjustment ratio of each weight coefficient. The final output format of the control scheduling sequence adopts the general autonomous driving interface standard. The sequence data includes the global coordinates, local coordinates, timestamp, priority score, and control command preview information for each point.

[0036] Example 3: The process of determining the initial control object selection point is based on the fusion processing of multi-source sensor data. Real-time monitoring of the vehicle's lateral position offset from the lane centerline is achieved through a forward-looking vision system. The forward-looking vision system uses a 1920×1080 resolution color CMOS sensor with a frame rate of 30 Hz. The lane line recognition algorithm is based on a deep convolutional neural network architecture. The neural network model uses an encoder-decoder structure. The encoder part uses a ResNet34 skeleton network to extract image features, and the decoder part gradually restores the spatial resolution through transposed convolutional layers, finally outputting a probability map of each pixel belonging to the lane line. The lane centerline calculation module refines the probability map and uses a skeleton extraction algorithm to obtain a lane centerline with a single pixel width. The vehicle's lateral position offset is obtained by calculating the vertical distance from the vehicle's center point to the lane centerline, with the sign of the distance value indicating the offset direction.

[0037] The Euclidean distance between each selected point of the controlled object and the current position of the vehicle is calculated using two-dimensional coordinates in the vehicle coordinate system. The origin of the vehicle coordinate system is located at the center of the rear axle, the X-axis points in the direction of vehicle movement, and the Y-axis points to the left. The coordinates of the selected points of the controlled objects are transformed from the global coordinate system to the vehicle coordinate system through coordinate transformation, and the coordinate transformation matrix includes translation and rotation components. The Euclidean distance is calculated as the square root of the sum of the squares of the straight-line distances between two points, and the calculation process uses a fast square root algorithm to ensure real-time performance.

[0038] By incorporating a dynamic weight adjustment mechanism into the process of weighted summation of offset and distance, the weighted summation formula is expressed as:

[0039] in: This represents the comprehensive score value of the selected point of the i-th control object. Let represent the Euclidean distance between the i-th point and the vehicle. This indicates the lateral offset of the vehicle's current position from the lane centerline. This represents the distance weighting coefficient. This represents the offset weighting coefficient. Distance weighting coefficient. The value range is from 0.6 to 0.8, and the offset weighting coefficient is... The value ranges from 0.2 to 0.4, and the weighting coefficient is dynamically adjusted according to real-time traffic conditions. The point with the smallest weighted value is selected as the initial control object selection point. The selection algorithm iterates through the comprehensive score of all control object selection points and uses a linear search method to find the point corresponding to the minimum value. Real-time acquisition of lane line recognition confidence data is obtained through the output of a neural network classifier. The confidence value ranges from 0 to 1, indicating the reliability of the recognition result. Dynamic weighting coefficients are calculated in conjunction with the vehicle's lateral position offset. An adaptive adjustment factor is introduced in the calculation of dynamic weighting coefficients, and the adjustment factor is updated based on the sliding window average of historical data.

[0040] The sensitivity of adjusting weighting coefficients based on weather conditions is achieved through an environmental perception module, which integrates data from rain, light, and temperature sensors. A weather condition classifier categorizes the current environment into sunny, rainy, and foggy conditions, each with a different weighting adjustment strategy. The weighting calculation process is optimized by fusing historical starting point selection success rate data. Historical data storage uses a circular buffer structure, recording the success or failure of the most recent 1000 selection operations. The success rate calculation employs an exponentially weighted moving average method, with recent data receiving higher weights. Acquiring lane line recognition confidence data involves multiple processing steps. The raw image is preprocessed before being input into the neural network model. Preprocessing operations include Gaussian filtering for noise reduction, contrast enhancement, and color space conversion. The neural network model outputs the probability value of each pixel belonging to a lane line. Confidence is calculated by statistically averaging the probability values ​​within the region of interest. Dynamic weighting coefficient adjustments consider sensor reliability indicators; when the camera is obstructed by dirt, the weight ratio of visual data is automatically reduced.

[0041] The measurement of vehicle lateral position offset employs a multi-sensor fusion strategy, incorporating GPS positioning data and inertial navigation system data in addition to the vision system. GPS positioning data provides absolute position reference, while the inertial navigation system provides continuous displacement measurements. The data fusion algorithm uses a Kalman filter to estimate the precise offset of the vehicle relative to the lane centerline. The offset value is updated every 10 milliseconds, with the update frequency consistent with the main cycle of the control system. The storage structure for historical starting point selection success rate data includes three fields: timestamp, scene features, and selection result. Scene features include parameters such as vehicle speed, road curvature, and weather conditions. The selection result records whether the selection successfully guided the vehicle to complete the lateral control task. Data retrieval is based on the nearest neighbor algorithm, finding the most similar historical records to the current scene, with similarity calculated using Euclidean distance. The optimization of the weight calculation process incorporates machine learning methods, training a weight coefficient prediction model through supervised learning. The training dataset contains a large number of historical operation records, with feature vectors including environmental parameters, vehicle status, and sensor readings, and labels representing manually labeled optimal weight coefficients. The prediction model uses a gradient boosting decision tree algorithm, and the model is retrained every 24 hours to adapt to long-term environmental changes. The adaptive adjustment mechanism for dynamic weighting coefficients includes a feedback control loop that monitors the control effect in real time after the initial point selection. The control effect is evaluated using the lateral error integral index. When the control effect is unsatisfactory, the weighting coefficient allocation ratio is automatically adjusted, with the adjustment magnitude calculated based on PID control principles. This closed-loop adjustment mechanism ensures that the weighting coefficients are always in an optimal state.

[0042] The decision-making process for selecting the point with the smallest weighted value as the initial control object adds a safety verification step. This verification includes point accessibility checks, collision risk assessments, and control feasibility analyses. The accessibility check, based on the vehicle's kinematics model, predicts whether the vehicle can reach the point within a specified time. The collision risk assessment calculates the minimum distance between the point and obstacles to ensure the safety margin meets requirements. The control feasibility analysis verifies whether the curvature of the point is within the vehicle's maximum steering capability.

[0043] The final selected starting control point must meet multiple constraints, including safe distance requirements from vehicles in front and behind, traffic regulations, and passenger comfort considerations. The safe distance is calculated using a two-second following distance principle, taking into account current speed and braking distance. Traffic regulations include not crossing solid lines and not using emergency lanes. Passenger comfort is controlled through lateral acceleration and jerk indicators to ensure smooth acceleration and deceleration. The entire starting point selection process is completed within 100 milliseconds, meeting the timing requirements of the real-time control system. The selection result is output to the control scheduling sequence generation module as the starting point for sequence planning. Simultaneously, intermediate data from the selection process is recorded in a black box memory for subsequent analysis and algorithm optimization. The system continuously monitors the selection effect, triggering a reselection mechanism when significant environmental changes occur to ensure the optimal starting control point is always used.

[0044] Example 4: The process of generating control and scheduling commands based on real-time sensing data begins with the continuous acquisition of sensor data. Vehicle lateral acceleration data is acquired through a microelectromechanical system (MEMS) accelerometer sensor installed at the vehicle's center of gravity. The sensor has a range of ±2g, a resolution of 0.001g, and a sampling frequency of 100 Hz. Yaw velocity data comes from a gyroscope sensor using fiber optic gyroscope technology, with a measurement range of ±100 degrees / second and a noise density of 0.01 degrees / second / √Hz. Sensor data is transmitted to the control unit via a CAN bus, conforming to the ISO11898 standard. The data frame includes a timestamp, sensor identifier, and checksum information. A circular buffer data structure is used to construct a sliding window-based data buffer. The buffer size is set to 1000 sampling points, corresponding to a 10-second time window. The data write pointer increments at a frequency of 100 Hz, automatically wrapping back to the starting position when it reaches the end of the buffer. The buffer management algorithm maintains the synchronization of the read and write pointers to ensure data integrity. It stores the raw data of sensor readings, including acceleration and angular velocity, over a recent period, with each data point accompanied by a microsecond-level timestamp of the acquisition time.

[0045] The buffer data is smoothed using a digital filter, specifically a fourth-order Butterworth low-pass filter, with the cutoff frequency dynamically adjusted based on the vehicle's motion. The filter design incorporates group delay compensation, employing a forward-backward filtering method to achieve zero phase distortion. Filter coefficients are calculated in real-time, using a bilinear transform to convert the analog filter to a digital one. The smoothed data is stored in a separate area of ​​the buffer, retaining both the original and filtered data for subsequent analysis. The deviation between the processed data and the expected control target, derived from the reference trajectory generated by the path planning module, is calculated. The deviation calculation includes two components: lateral position deviation and heading angle deviation. The lateral position deviation is obtained by comparing the vertical distance between the vehicle's current position and the reference path, while the heading angle deviation is calculated as the difference between the vehicle's actual heading and the tangential angle of the reference path. The deviation fusion algorithm uses a weighted summation method, with weight coefficients adaptively adjusted based on vehicle speed. Referring to Table 1, if the deviation value is less than the tolerance range, the tolerance range is determined using a lookup table, with the table index being vehicle speed and road curvature.

[0046] Table 1: Tolerance Range Data Table

[0047] This generates a command to maintain the current schedule. The maintenance command contains instructions to maintain the current steering angle and torque output. The command encoding uses a fixed-length binary format and includes a command type identifier, parameter fields, and a cyclic redundancy check code.

[0048] If the deviation exceeds the tolerance range, an acceleration or deceleration command is generated based on the deviation direction. The deviation direction is determined by comparing the current deviation value with historical deviation values, using a least-squares method to fit the deviation sequence of the most recent 10 sampling points. The acceleration command increases the steering motor torque, with the torque increment proportional to the deviation magnitude. The deceleration command reduces torque output and activates brake assist; the brake assist system intervenes in advance based on deviation rate prediction. The data acquisition module achieves multi-sensor time synchronization using a precision clock protocol to achieve microsecond-level synchronization accuracy. Each sensor data packet carries a GPS timestamp, and the data fusion unit performs interpolation alignment based on the timestamp. Data validity verification includes range checks, continuity checks, and consistency checks; abnormal data is marked and excluded from subsequent processing.

[0049] A sliding window buffer enables data compression, automatically increasing the sampling interval and dynamically adjusting the effective buffer length when the vehicle is in steady-state driving. A buffer data index establishes a fast query mechanism, supporting multi-dimensional retrieval by timestamp and sensor type. Data persistence employs a first-in, first-out (FIFO) strategy, triggering event recording before old data is overwritten by new data. Digital filter parameters are adaptively adjusted based on vehicle dynamics, increasing the cutoff frequency at low speeds to maintain response speed and decreasing it at high speeds to enhance anti-interference capabilities. Filter state variables are monitored in real-time, automatically switching to anti-aliasing filtering mode upon detecting transient impacts. Filter performance indicators are calculated online, including signal-to-noise ratio improvement and phase lag compensation. Deviation calculation incorporates a predictive compensation mechanism, using a Kalman filter to predict future deviation values ​​and compensate for inherent system delays. Deviation fusion weighting coefficients are adjusted based on road surface adhesion coefficients; for low-adhesion roads, the weight of heading angle deviation is reduced, while the weight of position deviation is increased. Reliability assessment of deviation values ​​considers sensor confidence and environmental interference factors; a conservative control strategy is employed for low-reliability deviation data.

[0050] The online update mechanism for the tolerance range table is optimized based on historical control performance data. For operating conditions with many successful control records, the tolerance range is appropriately increased, while for conditions requiring frequent adjustments, the tolerance range is decreased. Table lookups use bilinear interpolation to smooth abrupt changes at discrete parameter boundaries. Adaptive adjustment of the tolerance range considers passenger comfort indicators, dynamically optimizing parameters based on passenger feedback. The scheduling command generation module implements a multi-priority decision-making mechanism: safety-related deviations trigger the highest-priority response, while comfort-related deviations employ a gradual adjustment strategy. Command parameter calibration is based on real-vehicle test data, establishing a mapping relationship between deviation amounts and control parameters. The command output interface supports multiple bus protocols, including CAN, etc. It supports Ethernet to meet the compatibility requirements of different architecture platforms.

[0051] The torque increment calculation for acceleration dispatch commands incorporates a saturation limit to prevent actuator overload. The increment coefficient is dynamically adjusted based on motor temperature, reducing output capacity under high-temperature conditions. The braking assist intervention timing for deceleration dispatch commands is calculated based on a model predictive control algorithm, with the lead time taking into account braking system response delay and road conditions. Command safety verification includes coordination checks with longitudinal control to avoid conflicts between lateral and longitudinal control. The entire control dispatch command generation process is completed within a 10-millisecond control cycle, meeting real-time requirements. The system continuously monitors command execution effects, performing closed-loop optimization by comparing expected deviation changes with actual deviation changes. Historical command data is recorded for offline analysis and algorithm improvement, establishing a correlation model between control performance and environmental conditions.

[0052] Example 5: The process of adjusting the lateral movement of the vehicle using control scheduling commands begins with the parsing phase of the control scheduling commands. These commands are received via the controller local area network (LAN) bus, and the command frame contains a command type identifier, parameter set, and verification field. The command parsing module is implemented using a finite state machine, which includes idle, parsing, verification, and execution states. During parsing, the command format is rigorously checked, including identifier validity verification, parameter range checks, and checksum calculations. If it is a hold command, it contains instructions to maintain the current control output. The control system reads the current steering motor torque value and operating mode. The steering motor controller enters hold mode, which maintains the current torque output through closed-loop control. The torque feedback signal comes from the motor's built-in Hall sensor. The control system simultaneously monitors the duration of the hold command; when the duration exceeds a set threshold, a system self-test program is triggered.

[0053] For acceleration commands, the instruction includes increasing the torque output of the steering motor. The torque increment is calculated based on the deviation magnitude and vehicle speed. The torque control algorithm uses a proportional-integral-derivative (PID) controller, with the proportional coefficient dynamically adjusted according to vehicle speed. At high speeds, the proportional coefficient is decreased to ensure stability, while at low speeds, it is increased to improve response speed. Torque output is limited within the motor's safe operating range, and an overload protection circuit monitors motor current and temperature in real time. For deceleration commands, the instruction includes reducing torque output and activating brake assist. The torque reduction is calculated based on the deviation change rate. The brake assist system employs a gradual intervention strategy, with the initial brake pressure set at 50% of the base value, subsequently increasing linearly based on the deviation correction effect. The brake control system coordinates with the steering system to avoid braking intervention affecting steering accuracy.

[0054] The vehicle attitude is monitored in real time during adjustment, with attitude data sourced from the inertial measurement unit and wheel speed sensors. Attitude monitoring includes multiple parameters such as lateral acceleration, yaw rate, and wheel slip ratio. Monitoring data is updated at a frequency of 100 Hz, and each parameter has a set normal operating range and a danger threshold. When a parameter exceeds the danger threshold, the system triggers an emergency intervention procedure. Control quantities are fine-tuned through a feedback loop employing a multi-rate control structure, with an inner loop control frequency of 200 Hz and an outer loop control frequency of 100 Hz. The fine-tuning algorithm is based on model predictive control theory, using a vehicle dynamics model to predict the system state over a future period. Control quantity adjustments consider actuator response characteristics to avoid system oscillations caused by over-adjustment. Steering motor torque control enables soft start and soft stop functions, with the torque change slope limited to within 10% of the rated value per second. The motor controller uses space vector pulse width modulation technology to improve DC bus voltage utilization. Power device temperature is monitored in real time, and automatic derating is implemented when the temperature exceeds safety limits. The braking assist system works in conjunction with the vehicle stability control system, with brake pressure regulation employing a pressure-boosting-depressurization cyclic mode. The pressure sensor provides real-time feedback of the actual braking pressure, which is compared with the target pressure to form a closed-loop control. The anti-lock braking system (ABS) remains active during intervention to prevent wheel lock-up and instability.

[0055] Vehicle attitude monitoring incorporates a sensor data fusion algorithm, with a Kalman filter estimating the vehicle's actual motion state. The filter model includes a seven-DOF vehicle model, considering longitudinal, lateral, vertical, and yaw motions. The estimation results are compared with sensor measurements; when the difference exceeds a threshold, a sensor fault diagnosis procedure is triggered. The fine-tuning of the feedback loop is calculated based on the error integral and derivative terms. The integral term eliminates steady-state errors, while the derivative term suppresses overshoot. An anti-saturation mechanism is implemented in the integral term to limit integral growth when the error remains large. The derivative term uses an incomplete derivative form to reduce the impact of measurement noise. The control system implements feedforward compensation, with the feedforward amount calculated based on road curvature and vehicle speed. The compensation amount is injected into the control loop to improve the system's ability to suppress anticipated disturbances. The feedforward parameters are calibrated through real-vehicle testing, establishing a three-dimensional mapping table of curvature-vehicle speed-compensation amount. Actuator health monitoring is integrated throughout the entire adjustment process, monitoring indicators including motor current waveform, brake pressure response characteristics, and sensor signal quality. An anomaly detection algorithm analyzes the monitoring data to identify potential faults early. When performance degradation is detected, the system automatically adjusts control parameters or switches to a backup scheme.

[0056] The control output smoothing process employs a first-order low-pass filter, with the filter cutoff frequency adjusted according to the control mode. Smoothing ensures continuous control input changes, avoiding mechanical shocks caused by abrupt changes. Filter parameters are optimized online to balance response speed and stability requirements. The system establishes a comprehensive safety monitoring mechanism, including a watchdog timer, memory protection unit, and program flow monitoring. The watchdog timer ensures timely execution of the control cycle, the memory protection unit prevents unauthorized access, and program flow monitoring detects abnormal jumps. Multiple protection mechanisms ensure system safety degradation under abnormal conditions. The entire adjustment process implements online parameter self-tuning, with the self-tuning algorithm based on model reference adaptive control theory. The reference model describes the ideal system characteristics; the actual system output is compared with the reference model output, and the controller parameters are automatically adjusted to approximate the system characteristics to the reference model. The self-tuning process is activated during steady-state vehicle operation to avoid interfering with normal control.

[0057] The control system maintains real-time communication with other domain controllers in the vehicle, sharing control status and sensor information. The communication protocol uses automotive Ethernet, with a transmission latency of less than 1 millisecond. Information sharing enables coordinated optimization of lateral and longitudinal control, improving overall vehicle driving performance. The adjustment effect is evaluated based on a comprehensive assessment of multiple indicators, including path tracking accuracy, ride comfort, and energy consumption. Evaluation results are used for continuous optimization of the control system, establishing a correlation model between control parameters and performance indicators. Long-term operational data recording and analysis reveal control patterns and improve control strategies. Fault response strategies include multiple levels: minor faults employ adaptive parameter adjustment, moderate faults trigger control mode degradation, and severe faults execute minimum-risk operations. Each fault level corresponds to specific countermeasures to ensure safe system operation under various conditions. Detailed logs are recorded during fault handling to support post-event analysis and liability determination.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for selecting a lateral motion control object for an autonomous vehicle, characterized in that, The method is implemented through the following process: Real-time perception data of the surrounding environment of autonomous vehicles are collected using sensing devices. Multiple control object selection points are generated based on vehicle state parameters, and a control scheduling sequence is formed based on the control object selection points; Based on real-time sensing data, control and scheduling commands are generated, and the lateral movement of the vehicle is adjusted according to the control and scheduling commands. The control metric value corresponding to the selected point of each controlled object is also obtained. Aggregate the control metrics of all selected points of the controlled objects, generate control analysis results, and output control termination signals or control retry signals.

2. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 1, characterized in that, The process of generating multiple control object selection points based on vehicle state parameters includes: acquiring three-dimensional point cloud data of the vehicle's surrounding environment through an onboard perception system; extracting lane line features and obstacle contours using point cloud data processing algorithms; calculating lateral motion constraints based on a vehicle dynamics model to generate an initial set of control candidate points; grouping the initial set of control candidate points using a clustering analysis algorithm and removing outliers; calculating the center point of each cluster as a control object selection point based on the grouping results; and dynamically adjusting the density of control object selection points based on real-time traffic flow data, increasing the number of selection points in densely trafficked areas.

3. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 1, characterized in that, The process of forming a control scheduling sequence based on the selected points of the controlled object includes: assigning a priority score to each selected point of the controlled object, the score being calculated based on the current distance between the point and the vehicle, the curvature of the lane where the point is located, and the density of obstacles around the point; sorting the selected points of the controlled object according to the priority scores from high to low; introducing a time window mechanism to divide the sorted points into multiple time segments; within each time segment, calculating the reachability of the points based on the expected speed of the vehicle and adjusting the order of the points; and finally connecting the adjusted point sequence into a control scheduling sequence.

4. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 3, characterized in that, The starting control object selection point is determined by: real-time monitoring of the offset of the vehicle's lateral position from the lane centerline; calculating the Euclidean distance between each control object selection point and the vehicle's current position; and combining the offset and distance with a weighted sum, selecting the point with the smallest weighted value as the starting control object selection point.

5. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 1, characterized in that, The process of generating control and scheduling commands based on real-time sensing data includes: continuously collecting vehicle lateral acceleration and yaw rate data; constructing a data buffer based on a sliding window to store sensor readings over a recent period; smoothing the buffer data using a digital filter; calculating the deviation between the processed data and the expected control target; if the deviation is less than the tolerance range, generating a command to maintain the current scheduling; if the deviation exceeds the tolerance range, generating an acceleration or deceleration scheduling command based on the direction of the deviation.

6. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 1, characterized in that, The method of adjusting the lateral movement of the vehicle by means of control scheduling commands includes: parsing the type of control scheduling command; if it is a maintenance command, maintaining the current lateral control output; if it is an acceleration command, increasing the torque output of the steering motor; if it is a deceleration command, decreasing the torque output and activating brake assist; and monitoring the vehicle attitude in real time during the adjustment process and fine-tuning the control quantity through feedback loop.

7. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 1, characterized in that, The process of aggregating control metric values ​​from all selected control object points to generate control analysis results includes: collecting multiple sample values ​​of control metrics for each selected control object point; calculating the mean and standard deviation of the sample values; comparing the metric distributions of different points using hypothesis testing methods; if the distribution consistency passes the test, merging the data to generate an overall characteristic curve; otherwise, retaining the differences between points and marking outlier areas; and finally outputting control analysis results with confidence intervals.

8. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 4, characterized in that, The selection of the point with the smallest weighted value as the starting control object selection point includes: real-time acquisition of lane line recognition confidence data, calculation of dynamic weight coefficients based on vehicle lateral position offset; adjustment of the sensitivity of the weight coefficients according to weather conditions, optimization of the weight calculation process by integrating historical starting point selection success rate data; and finally selection of the point with the smallest comprehensive weighted value as the starting control object selection point.

9. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 2, characterized in that, The application of cluster analysis algorithm to group the initial control candidate point set includes: using density clustering algorithm to identify high-density regions in the point set; calculating the local density and minimum distance of each point; determining the cluster center based on the density peak and assigning the point to the nearest cluster center; removing outliers with density below the threshold; and dynamically adjusting the cluster radius to adapt to different traffic scenarios.

10. The method for selecting a lateral motion control object for an autonomous vehicle according to claim 3, characterized in that, The process of assigning priority scores to points selected for each controlled object includes: extracting the point's position coordinates and the vehicle's current pose; calculating the consistency between the point and the vehicle's heading angle; assessing the visibility and lighting conditions of the area where the point is located; and integrating historical control success rate data to calculate a comprehensive score using a weighted summation formula.

Images