VGA (Video Graphics Array) real-time path re-planning method for dynamic obstacles

Abstract

The invention discloses a VGA (Video Graphics Array) real-time path re-planning method for a dynamic obstacle, which comprises the following steps of: according to occupation state representation, acquiring time sequence data of a movement track of the obstacle, calculating a probability density function of a future position by adopting a Monte Carlo method, and determining a collision probability value of each candidate path; if the collision probability value exceeds a preset threshold value, Bayesian inference is introduced to update an obstacle threat degree evaluation function, the path confidence is judged, and an adjusted safety probability optimization result is obtained; according to a safety probability optimization result, a spatial clustering algorithm is applied to multiple obstacle scenes to aggregate obstacles which are close in distance and similar in motion direction into a group, and a composite obstacle represented by an envelope boundary model is obtained; if the aggregated obstacle area processing state shows that the groups are dispersed, the aggregated state is removed through real-time tracking data to recover independent avoidance processing of the single obstacle, and updated dynamic detection parameters are obtained.

Description

Technical Field

[0001] This invention belongs to the field of path planning technology, and in particular relates to a VGA real-time path replanning method for dynamic obstacles. Background Technology

[0002] In high-speed autonomous navigation scenarios for mobile robots or unmanned vehicles, dynamic obstacles in the environment, such as pedestrians suddenly crossing the road, vehicles rapidly changing lanes, or drones in flight, continuously change position and speed, causing pre-generated global paths to become invalid at any time. Therefore, real-time path replanning capability has become a core requirement for ensuring navigation safety and efficiency. This capability directly determines whether the system can output a passable and safe path within milliseconds, avoiding collisions.

[0003] While existing real-time path replanning methods can handle obstacle movement, they often assume that the future position of obstacles can be described by a single deterministic trajectory, or indirectly handle movement by simply increasing the static safety distance. This leads to either overly conservative paths with frequent pauses when obstacles experience sudden changes in velocity or dense clustering, or a high risk of residual collisions. It is difficult to simultaneously guarantee real-time performance and safety on embedded platforms with limited computing resources. Dynamic obstacles themselves possess high uncertainty; their positions and velocities change continuously over time and are difficult to predict precisely. This uncertainty makes traditional deterministic grid representations unable to accurately reflect the true occupancy situation at future moments, thus preventing path planning algorithms from providing reliable safety probabilities when evaluating candidate paths. When multiple dynamic obstacles appear simultaneously in a nearby area, their respective uncertainty regions overlap and intertwine, forming complex spatiotemporal threat regions. The uncertainty of a single obstacle is amplified into a complex uncertainty at the group level, forcing the planner to process a large number of interrelated probabilistic occupancy units simultaneously within a very short decision cycle, drastically increasing computational load and rapidly degrading real-time performance. For example, when three cars change lanes in different directions at an intersection at the same time, the possible trajectories of each car will form a fan-shaped uncertain area that expands over time. These areas quickly overlap to form a continuous danger zone. If the planner still calculates the probability of each obstacle separately, the number of probability updates that the grid nodes need to handle will increase exponentially, making it impossible to complete the path search within a control cycle of tens of milliseconds.

[0004] Therefore, how to achieve millisecond-level path safety probability assessment and complete replanning under the combined effects of dynamic obstacle motion uncertainty and multi-obstacle composite uncertainty has become a key issue restricting the reliable operation of the current system in complex real-world scenarios. Summary of the Invention

[0005] This invention proposes a VGA real-time path replanning method for dynamic obstacles to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a VGA real-time path replanning method for dynamic obstacles, comprising the following steps:

[0007] The raw data of dynamic obstacles is collected and converted into discrete raster values ​​to obtain the occupancy status representation of the obstacles in the VGA grid space;

[0008] Based on the occupancy status, the time series data of the obstacle's movement trajectory are obtained, the Monte Carlo method is used to predict the probability density function of the obstacle's future position, and the collision probability value of each candidate path is calculated.

[0009] If the collision probability value exceeds a preset threshold, Bayesian inference is introduced to update the obstacle threat assessment function, assess the path confidence, and output the safety probability optimization result.

[0010] Based on the safety probability optimization results, obstacles are aggregated into groups to construct composite obstacles represented by an envelope boundary model;

[0011] Based on the composite obstacles, a unified area is marked in the VGA grid, and a probabilistic occupancy grid map construction mechanism is used to simplify the number of detections for path search, and the processing status of the aggregated obstacle area is determined.

[0012] If the processing status shows that the group is scattered, the aggregation status is deactivated based on real-time tracking data, the independent avoidance processing of individual obstacles is restored, and updated dynamic detection parameters are obtained.

[0013] Based on the updated dynamic detection parameters, and combined with the threat weight coefficient adjustment mechanism, the security probability optimization results are integrated to generate the final path selection strategy.

[0014] Optionally, obtaining the occupancy state representation of the obstacle in the VGA mesh space includes:

[0015] Raw data of dynamic obstacles is collected using sensors to obtain point cloud sequences;

[0016] The displacement difference of the point cloud sequence is calculated using a vectorization method to obtain the velocity vector and acceleration vector.

[0017] The point cloud sequence is clustered based on location information, velocity vector, acceleration vector, volume attribute, and shape attribute to form multiple independent obstacle instances;

[0018] The state vector of each individual obstacle instance is tracked using a Kalman filter, and its position, velocity, and acceleration information are updated.

[0019] A fixed spatial region is divided using a voxel grid, and the tracked position information is mapped to the corresponding voxel grid.

[0020] Calculate the set of voxel grids covered by each independent obstacle instance based on its volume and shape attributes, and mark it as occupied.

[0021] The occupancy status of the voxel grid space at the current moment is determined by comparing the set of occupied voxel grids with the set of free voxel grids.

[0022] Optionally, calculating the collision probability value for each candidate path includes:

[0023] Based on the correlation data between the occupancy status and the movement trajectory, information on the positional changes of obstacles at different time points is obtained, and then processed into continuous trajectory data through time series analysis.

[0024] Based on historical trajectory data, predict the possible location range of obstacles in the next time period and generate a set of location distributions;

[0025] For the set of location distributions, the overlap between candidate paths and possible obstacle locations is analyzed. If the overlap area exceeds a preset threshold, the path is marked as a high-risk path.

[0026] For candidate paths that are not high-risk, the Monte Carlo method is used to perform probability assessment and calculate the collision risk value of each path.

[0027] Based on the collision risk value, paths below the safety threshold are selected to form a set of feasible paths.

[0028] Optionally, the evaluation of path confidence and output of safety probability optimization results includes:

[0029] If the collision probability value exceeds the preset threshold, the obstacle threat assessment function is updated using the Bayesian inference method to obtain the adjusted threat data.

[0030] Based on the updated threat data, assess the confidence level of the current path to obtain a preliminary security probability value;

[0031] The initial safety probability value is optimized by combining dynamic environmental information to obtain the optimized safety probability result;

[0032] If the optimized safety probability result is still lower than the safety standard, the path confidence is checked a second time and the path confidence data is updated.

[0033] Based on the updated path credibility data, the collision probability is recalculated to determine whether the threat assessment needs to be updated again, forming the basis for the final security decision.

[0034] Based on the reference data for generating path planning in the final security decision, the system outputs a security probability optimization result applicable to the current environment.

[0035] Optionally, the construction of the composite obstacle represented by the envelope boundary model includes:

[0036] Acquire the position and trajectory information of obstacles in a multi-obstacle scene to determine their distribution;

[0037] A spatial clustering algorithm is used to group obstacles that are close in location and move in the same direction, resulting in multiple obstacle groups;

[0038] For each obstacle group, extract its spatial extent data and construct an envelope boundary to determine the boundary extent of the group;

[0039] Geometric analysis methods are used to smooth the envelope boundary to obtain an optimized boundary description;

[0040] If the optimized boundary description deviates from the actual distribution, the boundary point positions are adjusted through comparative analysis to determine the final boundary model.

[0041] Based on the final boundary model and the motion consistency of the group, composite obstacles are generated, and their dynamic influence range in the scene is determined.

[0042] Optionally, determining the processing state of the aggregated obstacle region includes:

[0043] Based on the spatial distribution information of complex obstacles, an initial spatial data structure is constructed to form a preliminary environmental representation.

[0044] Spatial data is processed using a probabilistic occupancy method to generate a probabilistic occupancy raster map and determine the occupancy probability value of each raster cell.

[0045] The space is divided into multiple VGA grid cells to obtain grid structure data;

[0046] Regions that meet the unified labeling conditions are extracted from the grid structure data, and adjacent similar probability units are merged to determine the range of the initial labeled region;

[0047] An aggregation operation is performed on the initially marked area to integrate the scattered obstacle areas into a continuous whole unit, resulting in the aggregated obstacle area;

[0048] Analyze the state of the obstacle area after aggregation. If the probability of a certain area being occupied is higher than a preset threshold, it is marked as a high-risk area, and the final processing state is determined.

[0049] Optionally, obtaining the updated dynamic detection parameters includes:

[0050] The dispersion of the group is determined based on the processing status of the obstacle area after aggregation. If dispersion occurs, real-time tracking data is obtained from the sensor to identify the initial dispersion unit.

[0051] By comparing the occupancy values ​​of adjacent grids with real-time tracking data, the boundaries of independent obstacle units are determined, and the set of separated obstacle units is obtained.

[0052] The aggregation state is deactivated, and the avoidance trajectory is replanned for each individual obstacle unit to obtain updated avoidance trajectory data.

[0053] The velocity threshold and position offset information are extracted from the updated avoidance trajectory data to determine the dynamic detection parameters.

[0054] Optionally, the strategy for generating the final path selection includes:

[0055] Numerical inputs for path replanning are obtained based on dynamically detected parameters;

[0056] A threat weighting coefficient adjustment mechanism is adopted, and the weighted risk score of each feasible path is calculated in combination with the security probability optimization results;

[0057] Prioritize each path based on its weighted risk score;

[0058] If multiple paths have similar risk scores, a secondary analysis is performed using a probability optimization method to select the path with the lowest collision probability.

[0059] Generate the final path selection strategy and detailed planning data based on the selected path;

[0060] During path execution, environmental changes are monitored in real time. If the threat weight changes significantly, a replanning mechanism is triggered to update the optimal path.

[0061] Compared with the prior art, the present invention has the following advantages and technical effects:

[0062] This invention discloses an intelligent obstacle avoidance method based on dynamic obstacle perception and path optimization. Addressing the path planning challenges caused by the dynamic changes of multiple obstacles in complex environments, it integrates obstacle data vectorization processing, collision probability prediction, and path replanning to form a logically coherent solution. This invention converts obstacle physical attributes into discrete grid values, combines Monte Carlo methods to predict future location probabilities, and introduces Bayesian inference to update threat assessment, ensuring optimized path confidence. In multi-obstacle scenarios, spatial clustering algorithms are applied to aggregate similar obstacles, constructing an envelope boundary model to reduce the number of detections, while probabilistic occupancy of the grid map improves search efficiency. When groups disperse, real-time tracking data is de-aggregated, restoring independent avoidance processing, and finally, a threat weight coefficient adjustment mechanism determines the lowest collision path. The core innovation of this invention lies in the combination of dynamic perception and probabilistic optimization, significantly improving the safety and efficiency of path planning, and is suitable for intelligent navigation needs in complex dynamic environments. Attached Figure Description

[0063] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0067] Example 1

[0068] like Figure 1 As shown, this embodiment provides a VGA real-time path replanning method for dynamic obstacles, including the following steps:

[0069] The raw data of dynamic obstacles is collected and converted into discrete raster values ​​to obtain the occupancy status representation of the obstacles in the VGA grid space;

[0070] Based on the occupancy status representation, the time series data of the obstacle's movement trajectory are obtained. The Monte Carlo method is used to predict the probability density function of the obstacle's future position, and the collision probability value of each candidate path is calculated.

[0071] If the collision probability value exceeds the preset threshold, Bayesian inference is introduced to update the obstacle threat assessment function, assess the path confidence, and output the safety probability optimization result.

[0072] Based on the safety probability optimization results, obstacles are aggregated into groups to construct composite obstacles represented by an envelope boundary model;

[0073] Based on the composite obstacles, the same area is marked in the VGA grid, and a probabilistic occupancy grid map construction mechanism is used to simplify the number of detections for path search and determine the processing status of the aggregated obstacle area.

[0074] If the processing status shows that the group is scattered, the aggregation status is removed based on real-time tracking data, the independent avoidance processing of individual obstacles is restored, and the updated dynamic detection parameters are obtained.

[0075] Based on the updated dynamic detection parameters, and combined with the threat weight coefficient adjustment mechanism, the security probability optimization results are integrated to generate the final path selection strategy.

[0076] Specifically, the following steps are included:

[0077] Step S101: Collect raw data of dynamic obstacles through sensors, and use vectorization method to convert physical attributes such as position, velocity, acceleration, volume and shape into discrete grid values ​​to obtain the occupancy status representation of the obstacle in VGA grid space.

[0078] Specifically, raw data of dynamic obstacles is collected by sensors to obtain a point cloud sequence; a vectorization method is used to calculate the displacement difference of each point in the point cloud sequence to obtain the velocity vector and acceleration vector; the point cloud sequence is clustered into multiple independent obstacle instances based on position information, velocity vector, acceleration vector, volume attribute, and shape attribute; the state vector of each independent obstacle instance is tracked by Kalman filtering, and the position information, velocity vector, and acceleration vector are updated; a fixed spatial region is divided using a voxel grid, and the tracked position information is mapped to the corresponding voxel grid; based on the volume attribute and shape attribute, the set of voxel grids covered by each independent obstacle instance is calculated and marked as occupied; the occupancy status of the voxel grid space at the current moment is determined by comparing the occupied voxel grid set with the free voxel grids.

[0079] For example, within 50 meters in front of an autonomous vehicle, a dynamic obstacle point cloud sequence is collected using LiDAR, with approximately 100,000 points per frame.

[0080] In this embodiment, the displacement difference is first calculated after matching the points on the same object in adjacent frames to obtain the velocity vector of each point. For example, the velocity of a pedestrian's leg is 1.2 meters per second forward and 0.3 meters per second lateral. Then, the acceleration vector is obtained by second-order difference of the velocity, which is usually within 0.5 meters per second².

[0081] Specifically, clustering based on proximity, consistent velocity direction, similar acceleration amplitude, similar volume, and shape characteristics (such as aspect ratio) can group all points belonging to the same electric vehicle into one class, forming independent obstacle instances and avoiding the erroneous merging of two adjacent bicycles. After clustering, 5 to 15 independent dynamic instances can typically be stably obtained.

[0082] In this embodiment, a Kalman filter is established for each instance using a constant velocity or constant acceleration motion model. The state vector includes three-dimensional position, velocity, and acceleration. When the lidar briefly loses the target due to obstruction, the Kalman filter can predict the position at the next moment based on the acceleration of the previous few frames, enabling the tracking to quickly re-establish itself after obstruction and significantly improving continuity. Subsequently, the fixed space is divided into a voxel grid with a side length of 0.2 meters. The obstacle center position updated by the Kalman filter is mapped to the corresponding voxel. Then, based on the instance's attributes such as length 4.8 meters, width 1.9 meters, and height 1.6 meters, the set of all voxels it actually covers is calculated and marked as occupied. For example, a car typically occupies about 200 voxels.

[0083] It should be noted that the areas penetrated by the lidar rays are simultaneously marked as idle voxels. By comparing occupied voxels with idle voxels, a precise occupied voxel grid representation for the current moment is formed, while voxels that are not covered and have no ray penetration are marked as unknown. This method can effectively distinguish narrow gaps between two vehicles traveling side by side, providing centimeter-level reliable occupancy information for path planning, avoiding misjudgments caused by sparse point clouds, and significantly improving the safety and smoothness of dynamic obstacle avoidance in complex urban environments.

[0084] Step S102: Based on the occupancy status representation, obtain the time series data of the obstacle's movement trajectory, calculate the probability density function of the future position using the Monte Carlo method, and determine the collision probability value of each candidate path.

[0085] Specifically, by correlating the occupancy status and movement trajectory data, information on the obstacle's positional changes at different time points is obtained. Time series analysis tools are used to compile continuous trajectory data, resulting in a historical movement path record for the obstacle. Based on this historical movement path record, the predicted demand for future positions is extracted. Combined with probability density distribution calculations, the possible location range of the obstacle in the next time period is generated, determining a potential location distribution set. For this location distribution set, the overlap between candidate paths and the possible obstacle locations is analyzed. If the overlap area exceeds a preset threshold, the path is marked as a high-risk path, identifying path options that need to be avoided. From the high-risk path options, the remaining candidate path data is obtained. Combining the collision probability calculation logic, a Monte Carlo method is used to evaluate the probability of each path, determining the collision risk value for each path. Based on the collision probability evaluation results, relevant information for path selection is extracted. If the collision risk value of a path is lower than a preset safety threshold, it is classified as a feasible path, resulting in a priority path set.

[0086] For example, in the dynamic obstacle avoidance system of autonomous vehicles, the precise positional changes of pedestrians or vehicles at continuous points in time can be accurately obtained by correlating the occupancy status and movement trajectory data.

[0087] Specifically, the system associates the marked occupancy states in the aforementioned VGA voxel grid with the ID of each obstacle instance obtained by Kalman filtering, thereby forming a time-stamped trajectory sequence.

[0088] In this embodiment, the time series analysis tool first sorts the center point positions of the same ID at 0.1-second intervals, and after removing noise points, it obtains smooth continuous trajectory data. For example, the trajectory of a bicycle in front in the past 5 seconds is shown as moving from grid coordinates (120,80) to (160,95) at a speed of 8m / s along a straight line, with a trajectory length of about 40 voxel units.

[0089] It should be noted that when extracting future location prediction requirements based on historical movement path records, the system will analyze the speed and acceleration trends within the last 3 seconds.

[0090] For example, if an acceleration abruptly changes from 0.5 m / s² to -2 m / s², it's determined that the bicycle is about to decelerate and turn. At this point, using a Gaussian probability density distribution calculation, multiple possible position ranges are generated within the next second. Typically centered on the current velocity vector, with a standard deviation of 15% of the velocity, this forms an elliptical set of position distributions covering approximately 150 voxel grids. This prediction method effectively captures position shifts caused by uncertainty, contributing to improved robustness of obstacle avoidance decisions.

[0091] In this embodiment, when analyzing the overlap of candidate paths based on the location distribution set, the planning module generates three candidate paths: one going straight along the current lane, one slightly deviating to the left, and one changing lanes to the right. If the first straight path overlaps with the predicted ellipse of the bicycle by 65%, exceeding the preset 50% threshold, it is immediately marked as a high-risk path and excluded, thus avoiding potential collision areas in advance. This rapid judgment based on area proportion significantly reduces the computational burden while ensuring safety.

[0092] Specifically, when using the Monte Carlo method to assess the collision probability from the remaining candidate paths, the system samples 1000 particles on each path, and each particle randomly shifts to the obstacle position according to the predicted distribution and performs a forward simulation for 0.8 seconds.

[0093] For example, a slightly left-leaning path only experienced 12 voxel-level collisions in 1000 samples, with a collision risk value of only 1.2%, far below the safety threshold of 5%. Therefore, it was classified as a feasible path and assigned the highest priority. Conversely, a right-lane-changing path, due to partial overlap with the distribution of static obstacles on the other side, had a collision risk value of 8.7% and was downgraded. This multi-particle statistical approach makes risk assessment closer to the real-world uncertainty distribution, ensuring that the ultimately selected path maintains an extremely low actual collision probability in dynamic and complex scenarios, thereby significantly improving driving safety and ride comfort.

[0094] Step S103: If the collision probability value exceeds the preset threshold, Bayesian inference is introduced to update the obstacle threat assessment function, determine the path confidence, and obtain the adjusted safety probability optimization result.

[0095] Specifically, if the collision probability exceeds a preset threshold, a Bayesian inference method is used to update the obstacle threat level, resulting in adjusted threat assessment data. Based on the updated threat assessment data, path confidence is analyzed to determine the safety of the current path and obtain a preliminary safety probability value. This preliminary safety probability value is further processed and combined with dynamic environmental information to determine the optimized safety probability result. If the optimized safety probability result is lower than the safety standard, the path confidence is re-verified to obtain adjusted path credibility data. Based on the adjusted path credibility data, the collision probability is recalculated to determine whether the threat assessment still needs updating, providing the final safety decision basis. Using this final safety decision basis, reference data for path planning is generated, outputting optimized results suitable for the current environment.

[0096] For example, in autonomous driving path planning, when the Monte Carlo method calculates that the collision probability of a candidate path has reached 0.32, exceeding the system's preset safety threshold of 0.25, the system will immediately trigger the Bayesian threat update mechanism to upgrade the threat level of the obstacle from the previous medium risk to high risk.

[0097] In this embodiment, the Bayesian update process first uses a prior threat level of 0.4 as a basis, and combines the latest observed sudden acceleration behavior of the obstacle as likelihood evidence. Through simplified posterior calculation, the threat level is quickly adjusted to 0.68. This value reflects the obstacle's stronger aggressive movement intention, thereby providing more accurate prior information for subsequent path evaluation.

[0098] Specifically, the updated threat level of 0.68 is directly mapped to the path confidence correction coefficient. For example, if the initial safety probability of a certain path was originally 0.82, after multiplying by (1 - threat level weighting factor), the optimized safety probability is 0.71. This process effectively incorporates dynamic environmental changes, enabling safety assessments to no longer rely solely on historical trajectories but to respond in real time to abnormal behavior of obstacles.

[0099] It should be noted that if the optimized safety probability of 0.71 is still lower than the system's required standard of 0.75, a second verification stage will be initiated. At this time, the system will retrieve multiple frames of sensor data from the last 5 seconds to reassess the consistency of the obstacle's movement trend. For example, if it finds that the lateral offset rate has increased from 0.8 m / s to 1.6 m / s, it will further confirm the rationality of the increased threat level and ultimately lower the path confidence level to 0.62.

[0100] In this embodiment, based on the confidence level of 0.62 after secondary verification, the system will recalculate the lightweight collision probability. This time, instead of performing full Monte Carlo sampling, the system will use an importance sampling method to focus on high-threat areas and quickly obtain a new collision probability of 0.28. Although the value is slightly lower, it is still higher than the 0.25 threshold, so the system maintains the high-threat judgment and marks the path as needing to be forcibly abandoned.

[0101] Understandably, through the aforementioned Bayesian iterative update and secondary verification mechanism, the path planner can complete the full-link risk correction from initial evaluation to final evaluation within 0.3 seconds, significantly improving the response speed to sudden obstacle behavior, while reducing the misjudgment rate by about 40%, ensuring that in complex urban intersection scenarios, vehicles can promptly select truly low-risk detour routes, thereby greatly improving overall driving safety.

[0102] Step S104: Based on the safety probability optimization results, a spatial clustering algorithm is applied to multiple obstacle scenarios to aggregate obstacles that are close in distance and have similar motion directions into groups, resulting in composite obstacles represented by an envelope boundary model.

[0103] Specifically, by collecting data from multi-obstacle scenarios, the location and trajectory information of obstacles in the environment are obtained to determine the initial distribution. A spatial clustering algorithm is used to group obstacles that are close in location and move in the same direction, resulting in multiple obstacle groups. Based on the grouped obstacle groups, the spatial extent data of each group is extracted to construct the corresponding envelope boundary, determining the boundary range of each group. For the constructed envelope boundary, geometric analysis tools are applied to smooth the boundary data, obtaining an optimized boundary description. If the optimized boundary description deviates from the actual obstacle distribution, the boundary point positions are adjusted through comparative analysis to determine the final boundary model. Using the final boundary model, combined with the movement direction and consistency of the obstacle groups, composite obstacles are generated to determine their dynamic influence range in the scene.

[0104] For example, in dynamic scenarios with multiple obstacles, environmental data is first collected in real time using LiDAR and visual sensors to obtain the precise position coordinates and velocity vectors of all obstacles.

[0105] Specifically, suppose that 15 pedestrian and vehicle obstacles are detected at the current moment, and their positions are distributed in a rectangular area of ​​30 meters × 20 meters in front of the vehicle.

[0106] In this embodiment, the DBSCAN spatial clustering algorithm is used to group these obstacles. This algorithm uses a distance threshold (e.g., within 2 meters) and a velocity direction angle of less than 20 degrees as consistency criteria, quickly grouping obstacles with similar movement trends into one category.

[0107] For example, after actual clustering, three main groups were obtained: the first group contained 6 pedestrians crossing the road side by side, the second group contained 4 electric vehicles traveling slowly in the same direction, the third group contained 3 riders who were scattered but riding in the same direction, and the remaining 2 isolated vehicles did not form a group.

[0108] Understandably, after extracting the spatial occupancy point cloud for each group, the convex hull algorithm is used to construct the initial envelope boundary. Taking the first pedestrian group as an example, its convex hull boundary consists of 8 vertices, and the overall shape is approximately rectangular, with a width of about 5 meters and a length of about 10 meters. However, the edge of the initial convex hull has obvious jagged protrusions, which do not completely match the actual area occupied by the group.

[0109] For example, to improve the accuracy of boundary description, the Alpha-Shape algorithm was then applied to smooth the convex hull concavely. The Alpha radius was set to 1.5 meters, allowing the boundary to shrink inward and enclose the voids within the population, resulting in a more compact and natural smooth envelope boundary. At this point, the number of boundary vertices was reduced to 6, and the area was reduced by approximately 15% compared to the convex hull, more closely resembling the dynamic occupancy range of a real population.

[0110] In this embodiment, the optimized boundary is compared with the latest frame of point cloud for deviation. If it is found that there are still sporadic point clouds outside the right boundary that exceed 0.8 meters, the control points of the corresponding boundary are automatically expanded outward until all point cloud points in the group are included in the boundary with an error of less than 0.3 meters, thereby determining the final accurate boundary model.

[0111] Understandably, by combining the average velocity vector and directional consistency coefficient of each group—for example, the first group has a consistency coefficient as high as 0.92—it indicates a high degree of coordination in group movement. In this case, the entire group can be abstracted as a composite obstacle. Its dynamic influence range can be defined by expanding the boundary model outwards along the direction of movement with a safety margin of 2 meters to form an elliptical predicted occupancy area. This composite obstacle significantly reduces the number of individuals that need to be individually processed in path planning, from 15 to about 5, significantly reducing computational complexity. Simultaneously, by considering the characteristics of coordinated group movement, the predicted dynamic influence range is more accurate, avoiding the risks of over-conservatism or missed judgments caused by traditional individual processing, thus improving path safety and traffic efficiency in densely populated scenarios.

[0112] Step S105: Based on the composite obstacles, obtain the uniformly marked area in the VGA grid, use the probabilistic occupancy grid map construction mechanism to reduce the number of detections during path search, and determine the processing status of the aggregated obstacle area.

[0113] Specifically, data is collected from complex obstacle scenes to obtain raw spatial distribution information in the environment, constructing an initial spatial data structure and obtaining a preliminary environmental representation. Based on this preliminary environmental representation, a probabilistic occupancy method is used to process the spatial data, generating a probabilistic occupancy raster map and determining the occupancy probability value of each raster unit. For the probabilistic occupancy raster map, a grid division operation is performed, dividing the space into multiple VGA grid units to obtain the resulting grid structure data. From the divided grid structure data, regions meeting unified labeling conditions are extracted, and adjacent similar probability units are merged to determine the preliminary labeled area range. For the preliminary labeled area range, a region aggregation operation is performed to integrate the scattered obstacle areas into continuous whole units, obtaining the aggregated obstacle area. By performing state analysis on the aggregated obstacle area, if the occupancy probability of a certain area is higher than a preset threshold, it is marked as a high-risk area, determining the final processing state.

[0114] Specifically, in dynamic scenarios with complex obstacles, real-time environmental data acquisition can obtain raw point cloud data returned by LiDAR or visual sensors. For example, at a given moment, the 3D coordinates and velocity vectors of hundreds of obstacle points can be collected, thus constructing an initial spatial data structure. This structure typically uses an octree or KD-tree format for fast storage and retrieval of spatial distribution information. For example...

[0115] In this embodiment, it is assumed that the autonomous vehicle is traveling in a congested area. The sensors collect point cloud data every 100ms, obtaining approximately 5000 valid obstacle points. These points are mainly concentrated around vehicles, pedestrian groups ahead, and non-motorized vehicle areas on both sides. At this point, the initial data structure directly reflects the discrete state of the original distribution, providing a foundation for subsequent probability processing.

[0116] Specifically, when using the probability occupancy method, each spatial voxel is assigned an occupancy probability value between 0 and 1.

[0117] It should be noted that probability calculations are usually based on ray tracing and Bayesian update rules. When multiple sensor rays detect reflections multiple times within the same voxel, the probability of that voxel will gradually increase from the initial 0.5.

[0118] For example, if a group of pedestrians in front of you are blocking each other, some voxels will only be hit by a portion of the rays, and their probability may remain around 0.7. However, voxels with strong reflection from the metal surface of a vehicle will quickly reach 0.98, thus generating a more accurate probability of occupying the grid map.

[0119] In this embodiment, after completing the probability map, grid division is implemented, dividing the space into VGA grid units with a side length of 0.2m. In the scenario of an urban intersection, the entire perception range of 60m×60m will be divided into approximately 90,000 grid units, among which a large area of ​​20m×10m in front contains grids with a probability exceeding 0.8. These grids correspond to actual pedestrian clusters and parked vehicles.

[0120] Preferably, when extracting regions that meet the unified labeling conditions from the divided grid data, a probability threshold of 0.75 is set as the labeling basis, and connectivity analysis is performed on adjacent grids with similar probabilities.

[0121] For example, if two adjacent grid groups with probabilities of 0.82 and 0.79 are detected, they will be merged into the same labeled region because the difference is less than 0.1 and they are 8-connected, thereby reducing fragmented representation.

[0122] Understandably, when performing region aggregation operations, morphological closing operations and flooding filling algorithms are used to integrate scattered high-probability small regions into a continuous whole.

[0123] For example, several isolated grid points with a probability of 0.85, which were originally caused by sensor noise, are aggregated and connected with the surrounding area to form a complete obstacle area covering an area of ​​about 12 square meters, corresponding to a group of pedestrians walking closely side by side in reality.

[0124] For example, when performing state analysis on aggregated obstacle areas, if the occupancy probability of more than 90% of the grids in a certain area is higher than 0.9, it is marked as a high-risk area. In real-world scenarios, a sudden dense group of electric vehicles ahead, due to their uniform speed and spacing of less than 0.5m, has an average aggregation probability of 0.94, far exceeding the threshold of 0.85, and is therefore marked as high-risk. This allows the planning system to take larger lateral spacing or deceleration measures in advance, effectively improving obstacle avoidance safety and traffic efficiency in multi-dynamic obstacle environments.

[0125] Step S106: If the aggregated obstacle area processing status shows that the group is scattered, the aggregation status is de-aggregated by real-time tracking data to restore the independent avoidance processing of individual obstacles and obtain updated dynamic detection parameters.

[0126] Specifically, based on the processed state of the aggregated obstacle area, the dispersion of the group is determined. If the group is dispersed, real-time tracking data is acquired from the sensor to obtain preliminary dispersed unit identifiers. Based on the preliminary dispersed unit identifiers, the occupancy values ​​of adjacent grids are compared using the real-time tracking data to determine the boundaries of independent obstacle units and obtain a set of separated obstacle units. Using the set of separated obstacle units, the aggregation state is deactivated to restore the avoidance trajectory of a single obstacle unit, resulting in updated avoidance trajectory data. From the updated avoidance trajectory data, speed thresholds and position offsets are extracted to determine dynamic detection parameters.

[0127] For example, when processing the state of aggregated obstacle areas, determining the dispersion of the group is a crucial step. Suppose in an indoor navigation scenario, multiple obstacle areas are initially aggregated into a single entity, but state analysis reveals that some areas are not densely distributed. In this case, real-time data can be acquired through sensors, such as using LiDAR to scan the surrounding environment, to obtain the location information of each small area, thereby identifying the initial dispersed units. For instance, if there are three small units within a certain area, located at coordinates A, B, and C, and the distance between them exceeds a preset threshold of 2 meters, then they can be preliminarily identified as dispersed units.

[0128] For example, based on the initial dispersed unit identification, comparing the occupancy values ​​of adjacent grid cells using real-time tracking data can further determine the boundaries of independent obstacle units. Suppose that in the above scenario, the grid cell near point A has a high occupancy value of 0.8, while the adjacent grid cell value is only 0.2. Then, it can be inferred that point A is the center of an independent unit, with a boundary range of a 1-meter radius centered on A. Using a similar method, the boundaries of points B and C are determined separately, ultimately forming the separated set of obstacle units. This method can clearly distinguish the ranges of different units, laying the foundation for subsequent processing.

[0129] For example, de-aggregating the separated obstacle units to restore the avoidance trajectories of individual obstacle units is a crucial step in ensuring navigation accuracy. In the indoor scenario described above, assuming the original avoidance trajectory of the aggregated area was a curve bypassing the entire area, but after separation, points A, B, and C are found to be independent. Therefore, it is necessary to replan the trajectory for each unit. For instance, for point A, the avoidance trajectory is adjusted to bypass from the left, maintaining a distance of 1.5 meters; for point B, the trajectory is adjusted to bypass from the right, maintaining a distance of 1 meter. In this way, updated avoidance trajectory data is obtained, avoiding path redundancy caused by aggregation and improving navigation efficiency.

[0130] For example, extracting speed thresholds and position offsets from updated obstacle avoidance trajectory data to determine dynamic detection parameters is crucial for adapting to real-time changes in the environment. Taking indoor navigation as an example, suppose the speed threshold for the trajectory near point A is set at 0.5 meters per second, and the position offset is 0.3 meters. This means that when approaching point A, the device needs to decelerate and maintain a certain offset distance to ensure safety. Meanwhile, the speed threshold for point B is 0.8 meters per second, and the offset is 0.2 meters, indicating a lower risk, allowing for a more appropriate increase in speed. Through this parameter adjustment, dynamic detection can more flexibly respond to the characteristics of different obstacle units, ensuring the overall stability of operation. Through the above analysis and examples, it can be seen that the entire process, from group dispersion judgment to dynamic parameter determination, is interconnected, with each step supporting the improvement of the navigation system's adaptability and safety. Especially in complex environments, refining the handling of obstacle units can effectively optimize path planning, reduce unnecessary detours or stops, and improve the user experience.

[0131] Step S107: Obtain the numerical input for path replanning through the updated dynamic detection parameters, integrate the probability optimization results using the threat weight coefficient adjustment mechanism, determine the lowest collision option among multiple feasible paths, and obtain the final path selection strategy.

[0132] Specifically, the dynamic detection module acquires real-time data from the environment, processes it to form structured input parameters, and determines the initial dataset for path replanning. Based on the input parameters, a preset threat weight model is used to calculate the potential risk value of each path, obtaining a threat score for each feasible path. For the threat score, a coefficient adjustment mechanism is applied to correct the score, obtaining the corrected risk assessment result and determining the priority ranking of each path. If multiple paths have similar risk values ​​in the corrected risk assessment result, a probabilistic optimization method is used to perform a secondary analysis on these paths to determine the path option with the lowest collision risk. Based on the path option with the lowest collision risk, combined with the final strategy generation module, a specific path selection scheme is obtained, outputting detailed planning data for the optimal path. Using the specific path selection scheme, a real-time monitoring system is invoked to collect data on path execution, obtaining dynamic feedback information during the execution process. Based on the dynamic feedback information, if a change in threat weight is detected during path execution, the path replanning module is triggered to recalculate and determine the updated optimal path scheme.

[0133] For example, in the field of dynamic path planning for autonomous vehicles, real-time data in the environment is obtained through dynamic detection modules, mainly relying on point cloud and image information fused from LiDAR and cameras. After filtering and timestamp alignment, structured input parameters are formed, such as the distance distribution of obstacles within 50 meters ahead, relative velocity vectors, and road curvature. These parameters together constitute the initial dataset for path replanning.

[0134] In this embodiment, the system inputs these parameters into a preset threat weight model, which is usually a weighted summation model, where the distance weight accounts for 0.4, the speed weight accounts for 0.35, and the type weight accounts for 0.25, thereby quickly calculating the potential risk value of each feasible path and obtaining a numerical threat score. For example, the left lane change path scores 28, the middle following path scores 15, and the right lane change path scores 42.

[0135] Specifically, when applying the coefficient adjustment mechanism to correct the threat level score, dynamic coefficients are introduced based on the current vehicle speed and weather conditions. For example, the overall risk coefficient is increased by 1.2 times in rainy weather, and the distance weight is increased by 10% when driving at high speed, so as to obtain a more realistic corrected risk assessment result. At this time, the middle following path may be adjusted from 15 to 18, while the right lane change path is increased to more than 50, making the priority ranking more reasonable.

[0136] Understandably, if multiple paths still have similar risk values ​​after correction, such as two paths having scores between 20 and 22, a secondary analysis is performed using a probability optimization method. Typically, Monte Carlo sampling is used to simulate the relative motion trajectory within 3 seconds of the next 1000 times, and the actual collision probability is statistically analyzed. The path option with the lowest collision risk can often reduce the probability to below 0.3%, while the other path may have a probability as high as 2.1%.

[0137] In this embodiment, based on the path option with the lowest collision risk, and in conjunction with the final strategy generation module, the system comprehensively considers the vehicle's current acceleration limits and comfort requirements to output detailed planning data for the optimal path. This data includes position points every 0.1 seconds, the desired speed curve, and the steering wheel angle sequence. This data is directly sent to the underlying controller for execution. Through the specific path selection scheme, the real-time monitoring system continuously collects lateral deviation, longitudinal speed error, and the latest threat distance detected by radar as dynamic feedback information. If, during execution, a sudden deceleration of the vehicle in front causes a 30% surge in the speed component of the threat weight, the system immediately triggers the path replanning module to re-enter the calculation loop. Typically, a new optimal path scheme is generated within 150 milliseconds, thus achieving closed-loop adaptive adjustment. This mechanism significantly improves driving safety and traffic efficiency on complex urban roads, especially in congested following scenarios, reducing the frequency of emergency braking by approximately 40% while maintaining a relatively high average vehicle speed.

[0138] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A VGA real-time path replanning method for dynamic obstacles, characterized in that, Includes the following steps: The raw data of dynamic obstacles is collected and converted into discrete raster values ​​to obtain the occupancy status representation of the obstacles in the VGA grid space; Based on the occupancy status, the time series data of the obstacle's movement trajectory are obtained, the Monte Carlo method is used to predict the probability density function of the obstacle's future position, and the collision probability value of each candidate path is calculated. If the collision probability value exceeds a preset threshold, Bayesian inference is introduced to update the obstacle threat assessment function, assess the path confidence, and output the safety probability optimization result. Based on the safety probability optimization results, obstacles are aggregated into groups to construct composite obstacles represented by an envelope boundary model; Based on the composite obstacles, a unified area is marked in the VGA grid, and a probabilistic occupancy grid map construction mechanism is used to simplify the number of detections for path search, and the processing status of the aggregated obstacle area is determined. If the processing status shows that the group is scattered, the aggregation status is deactivated based on real-time tracking data, the independent avoidance processing of individual obstacles is restored, and updated dynamic detection parameters are obtained. Based on the updated dynamic detection parameters, and combined with the threat weight coefficient adjustment mechanism, the security probability optimization results are integrated to generate the final path selection strategy.

2. The method according to claim 1, characterized in that, The obtained representation of the occupancy state of the obstacle in the VGA grid space includes: Raw data of dynamic obstacles is collected using sensors to obtain point cloud sequences; The displacement difference of the point cloud sequence is calculated using a vectorization method to obtain the velocity vector and acceleration vector. The point cloud sequence is clustered based on location information, velocity vector, acceleration vector, volume attribute, and shape attribute to form multiple independent obstacle instances; The state vector of each individual obstacle instance is tracked using a Kalman filter, and its position, velocity, and acceleration information are updated. A fixed spatial region is divided using a voxel grid, and the tracked position information is mapped to the corresponding voxel grid. Calculate the set of voxel grids covered by each independent obstacle instance based on its volume and shape attributes, and mark it as occupied. The occupancy status of the voxel grid space at the current moment is determined by comparing the set of occupied voxel grids with the set of free voxel grids.

3. The method according to claim 1, characterized in that, The calculation of the collision probability value for each candidate path includes: Based on the correlation data between the occupancy status and the movement trajectory, information on the positional changes of obstacles at different time points is obtained, and then processed into continuous trajectory data through time series analysis. Based on historical trajectory data, predict the possible location range of obstacles in the next time period and generate a set of location distributions; For the set of location distributions, the overlap between candidate paths and possible obstacle locations is analyzed. If the overlap area exceeds a preset threshold, the path is marked as a high-risk path. For candidate paths that are not high-risk, the Monte Carlo method is used to perform probability assessment and calculate the collision risk value of each path. Based on the collision risk value, paths below the safety threshold are selected to form a set of feasible paths.

4. The method according to claim 1, characterized in that, The evaluation path confidence and output of the safety probability optimization results include: If the collision probability value exceeds the preset threshold, the obstacle threat assessment function is updated using the Bayesian inference method to obtain the adjusted threat data. Based on the updated threat data, assess the confidence level of the current path to obtain a preliminary security probability value; The initial safety probability value is optimized by combining dynamic environmental information to obtain the optimized safety probability result; If the optimized safety probability result is still lower than the safety standard, the path confidence is checked a second time and the path confidence data is updated. Based on the updated path credibility data, the collision probability is recalculated to determine whether the threat assessment needs to be updated again, forming the basis for the final security decision. Based on the reference data for generating path planning in the final security decision, the system outputs a security probability optimization result applicable to the current environment.

5. The method according to claim 1, characterized in that, The construction of the composite obstacle represented by the envelope boundary model includes: Acquire the position and trajectory information of obstacles in a multi-obstacle scene to determine their distribution; A spatial clustering algorithm is used to group obstacles that are close in location and move in the same direction, resulting in multiple obstacle groups; For each obstacle group, extract its spatial extent data and construct an envelope boundary to determine the boundary extent of the group; Geometric analysis methods are used to smooth the envelope boundary to obtain an optimized boundary description; If the optimized boundary description deviates from the actual distribution, the boundary point positions are adjusted through comparative analysis to determine the final boundary model. Based on the final boundary model and the motion consistency of the group, composite obstacles are generated, and their dynamic influence range in the scene is determined.

6. The method according to claim 1, characterized in that, The determination of the processed state of the aggregated obstacle region includes: Based on the spatial distribution information of complex obstacles, an initial spatial data structure is constructed to form a preliminary environmental representation. Spatial data is processed using a probabilistic occupancy method to generate a probabilistic occupancy raster map and determine the occupancy probability value of each raster cell. The space is divided into multiple VGA grid cells to obtain grid structure data; Regions that meet the unified labeling conditions are extracted from the grid structure data, and adjacent similar probability units are merged to determine the range of the initial labeled region; An aggregation operation is performed on the initially marked area to integrate the scattered obstacle areas into a continuous whole unit, resulting in the aggregated obstacle area; Analyze the state of the obstacle area after aggregation. If the probability of a certain area being occupied is higher than a preset threshold, it is marked as a high-risk area, and the final processing state is determined.

7. The method according to claim 1, characterized in that, The updated dynamic detection parameters include: The dispersion of the group is determined based on the processing status of the obstacle area after aggregation. If dispersion occurs, real-time tracking data is obtained from the sensor to identify the initial dispersion unit. By comparing the occupancy values ​​of adjacent grids with real-time tracking data, the boundaries of independent obstacle units are determined, and the set of separated obstacle units is obtained. The aggregation state is deactivated, and the avoidance trajectory is replanned for each individual obstacle unit to obtain updated avoidance trajectory data. The velocity threshold and position offset information are extracted from the updated avoidance trajectory data to determine the dynamic detection parameters.

8. The method according to claim 1, characterized in that, The final path selection strategy includes: Numerical inputs for path replanning are obtained based on dynamically detected parameters; A threat weighting coefficient adjustment mechanism is adopted, and the weighted risk score of each feasible path is calculated in combination with the security probability optimization results; Prioritize each path based on its weighted risk score; If multiple paths have similar risk scores, a secondary analysis is performed using a probability optimization method to select the path with the lowest collision probability. Generate the final path selection strategy and detailed planning data based on the selected path; During path execution, environmental changes are monitored in real time. If the threat weight changes significantly, a replanning mechanism is triggered to update the optimal path.

Images