A method and system for unmanned vehicle close-range reconnaissance based on dynamic masking path planning

Abstract

This invention relates to a method and system for close-range reconnaissance of unmanned vehicles (UAVs) based on dynamic masking path planning, belonging to the field of autonomous navigation and path planning technology for UAVs. The method includes: identifying candidate masking objects based on real-time environmental data collected by the UAV; filtering effective masking objects according to validity criteria and generating their position and shape information; planning a reconnaissance path on an environmental cost map by combining the current position of the UAV, the target position, and the masking object information; controlling the UAV to travel along the path, and monitoring the path based on real-time updated environmental data; replanning the path and controlling the UAV's movement when replanning conditions are met. This addresses the problems of low accuracy in identifying masking objects in the environment and high exposure risk caused by the inability to plan the most concealed travel path in real time during close-range reconnaissance by existing UAVs.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and path planning technology for unmanned vehicles, and in particular to a method and system for close-range reconnaissance of unmanned vehicles based on dynamic masking path planning. Background Technology

[0002] In close-range reconnaissance missions involving unmanned vehicles, stealth is a key element in ensuring mission success and platform safety. Existing technologies mainly rely on preset routes or simple obstacle avoidance algorithms for path planning, typically treating obstacles as negative factors to be avoided, rather than making full use of naturally occurring cover resources in the environment, such as rocks, bushes, and buildings.

[0003] Existing methods for identifying cover often employ a single sensor modality, limiting their accuracy and making it difficult to accurately acquire the 3D geometry and spatial orientation of cover objects. This results in path planning being unable to effectively assess the actual cover effectiveness. Regarding path optimization objectives, traditional methods focus on the shortest distance or time, failing to quantify exposure risk into the cost function or establish a dynamic correlation model between cover type, distance, azimuth, and cover effectiveness. More importantly, existing systems lack online replanning mechanisms, failing to adjust the path in real time when the environment changes. This causes the autonomous vehicle to be exposed to the reconnaissance target's field of vision during movement, significantly increasing the probability of detection. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide an unmanned vehicle close-range reconnaissance method and system based on dynamic masking path planning, in order to solve the problems of low recognition accuracy of masking objects in the environment and high exposure risk caused by the difficulty in real-time planning of the most concealed travel path during the close-range reconnaissance process of existing unmanned vehicles.

[0005] On one hand, embodiments of the present invention provide a method for close-range reconnaissance of unmanned vehicles based on dynamic masking path planning, including:

[0006] Based on real-time environmental data collected by unmanned vehicles, candidate occlusions in the environment are identified;

[0007] Based on preset validity criteria, select effective shielding objects that meet the shielding requirements from the candidate shielding objects;

[0008] Generate the location information and shape information of the effective shielding object based on the effective shielding object;

[0009] Based on the current location of the unmanned vehicle, the target location, the location information of the cover, and the shape information of the cover, a reconnaissance path is planned on a pre-constructed environmental cost map;

[0010] The unmanned vehicle is controlled to travel along the reconnaissance path, and the reconnaissance path is monitored based on real-time updated environmental data. When the preset replanning conditions are met, the reconnaissance path is replanned, and the unmanned vehicle is controlled to travel along the replanned reconnaissance path.

[0011] Furthermore, based on real-time environmental data collected by the autonomous vehicle, candidate occlusions in the environment are identified, including:

[0012] Acquire image data collected by the visual sensors on the autonomous vehicle and point cloud data collected by the LiDAR;

[0013] The image data is input into a pre-trained deep learning recognition model to extract visual recognition results, which include the target region and its category confidence.

[0014] Cluster analysis is performed on the point cloud data to obtain multiple point cloud clusters, and the geometric feature parameters of each point cloud cluster are calculated.

[0015] The target region is spatially matched with the point cloud cluster, and consistency verification is performed based on the category confidence and the geometric feature parameters. The target region that passes the verification is determined as the candidate occluder.

[0016] Further, consistency verification is performed based on the category confidence level and the geometric feature parameters, and the target region that passes the verification is determined as the candidate occluder, including:

[0017] Calculate the spatial overlap between each target region and each point cloud cluster, and establish a matching relationship between target regions and point cloud clusters with spatial overlap greater than a preset overlap threshold.

[0018] For target regions and point cloud clusters with matching relationships, a multimodal consistency score is calculated based on the category confidence of the target region and the geometric feature parameters of the point cloud clusters.

[0019] When the multimodal consistency score is greater than the preset score threshold, the consistency verification is deemed to be successful, and the target area is identified as a candidate occluder.

[0020] Furthermore, generating the location information of the effective shielding object includes:

[0021] Obtain the point cloud cluster corresponding to each effective shading object;

[0022] Based on the spatial coordinates of all points in the point cloud cluster, determine the coordinate reference value of the point cloud cluster in each spatial dimension;

[0023] The three-dimensional centroid of the point cloud cluster is determined based on coordinate references in each spatial dimension;

[0024] The three-dimensional centroid is used as the location information of the effective shading object.

[0025] Furthermore, generating the shape information of the effective shielding object includes:

[0026] Obtain the point cloud cluster corresponding to each of the aforementioned effective shielding objects;

[0027] Based on the spatial coordinates of each point in the point cloud cluster, calculate the covariance matrix of the point cloud cluster in space.

[0028] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues ​​and eigenvectors corresponding to each eigenvalue. The three eigenvectors are taken as three mutually perpendicular principal directions, including a first principal direction, a second principal direction and a third principal direction.

[0029] Project all points in the point cloud cluster onto the first principal direction, the second principal direction, and the third principal direction respectively, and calculate the projection length in each principal direction to obtain the length, width, and height of the effective shield, wherein the length is the projection length in the first principal direction, the width is the projection length in the second principal direction, and the height is the projection length in the third principal direction.

[0030] The first main direction is defined as the orientation of the effective shield;

[0031] The length, width, height, and orientation of the effective shield are used as the shape information of the effective shield.

[0032] Furthermore, the reconnaissance path is planned on the pre-constructed environmental cost map, including:

[0033] Starting from the current location of the autonomous vehicle and ending at the target location, the path search space is determined on the environmental cost map;

[0034] The path search space is discretized into multiple path nodes;

[0035] Based on the location information and shape information of the obstruction, the exposed area in the path search space is determined;

[0036] The exposure risk cost of each path node is calculated based on the distance from each path node to the nearest exposed area, the distance to the nearest effective shield, and the path curvature between the path node and the previous path node.

[0037] The cover score for each path node is calculated based on the distance from each path node to the effective cover and the relative angle between the effective cover and the target being reconnoitered.

[0038] Based on the exposure risk cost and cover score, a reconnaissance path is planned in the path search space, and a reconnaissance path is searched in the path search space.

[0039] Furthermore, the exposure risk cost of each path node is calculated as shown in the following formula;

[0040] ;

[0041] in, The exposure risk cost of path node n, Let n be the distance from path node n to the nearest exposed area. Let n be the distance from path node n to the nearest effective shading object. This is the path curvature penalty term for path node n. , , These are the weighting coefficients.

[0042] Furthermore, the cover score for each path node is calculated as shown in the following formula;

[0043] ;

[0044] in, The cover score for path node n. Let i be the type weight of the i-th effective occlusion. Let n be the distance decay function from path node n to the i-th effective shield. Let be the relative angular effectiveness function between the i-th effective cover and the direction of the reconnaissance target.

[0045] Furthermore, the preset replanning conditions include:

[0046] A new obstacle is detected on the reconnaissance path or an effective cover on the reconnaissance path disappears.

[0047] On the other hand, embodiments of the present invention provide an unmanned vehicle close-in reconnaissance system based on dynamic masking path planning, comprising:

[0048] The identification module is used to identify candidate occlusions in the environment based on real-time environmental data collected by the unmanned vehicle.

[0049] The filtering module is used to filter out effective shielding objects that meet the shielding requirements from the candidate shielding objects according to preset validity discrimination conditions;

[0050] The generation module is used to generate the location information and shape information of the effective shielding object based on the effective shielding object;

[0051] The planning module is used to plan a reconnaissance path on a pre-built environmental cost map based on the current position of the unmanned vehicle, the target position, the position information of the cover object, and the shape information of the cover object.

[0052] The control module is used to control the unmanned vehicle to travel along the reconnaissance path, monitor the reconnaissance path based on real-time updated environmental data, and replan the reconnaissance path when the preset replanning conditions are met, and control the unmanned vehicle to travel along the replanned reconnaissance path.

[0053] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0054] 1. By fusing visual sensor image data and LiDAR point cloud data, and extracting target region category confidence based on a deep learning recognition model, combined with point cloud clustering analysis and geometric feature parameter calculation, multimodal consistency verification is achieved, improving the accuracy of occlusion recognition. This solves the problem of unstable recognition by a single sensor in complex environments, which can easily generate false alarms or missed detections, ensuring that unmanned vehicles can reliably discover and utilize occlusion resources with practical cover value in the environment.

[0055] 2. By calculating the centroid of three-dimensional space based on the coordinates of point cloud clusters as the location information of the occlusion object, and performing eigenvalue decomposition on the covariance matrix to obtain the principal direction and projected length, the geometric features such as the length, width, height and orientation of the occlusion object are accurately extracted, providing accurate spatial description parameters for path planning. This overcomes the problem of distorted assessment of occlusion effectiveness caused by the lack or coarseness of geometric information of the occlusion object in the existing technology, and reduces the exposure risk of unmanned vehicles.

[0056] 3. By constructing a multi-objective optimized path planning model that integrates exposure risk cost and cover success, the distance from path nodes to the exposed area, the distance to the cover, the path curvature, the weight of cover type, distance decay and angle are evaluated. A real-time monitoring and replanning mechanism is established to dynamically respond to environmental changes such as the appearance of new obstacles or the disappearance of cover, which effectively solves the problems that static planning is difficult to adapt to dynamic environments and path concealment cannot be quantified.

[0057] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0058] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0059] Figure 1 This is a schematic flowchart of an unmanned vehicle close-in reconnaissance method based on dynamic masking path planning, provided as an embodiment of the present invention. Detailed Implementation

[0060] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0061] A specific embodiment of the present invention discloses a method for close-range reconnaissance of unmanned vehicles based on dynamic masking path planning, such as... Figure 1 As shown, it includes:

[0062] S1. Based on real-time environmental data collection by unmanned vehicles, identify candidate occlusions in the environment.

[0063] During its journey, the autonomous vehicle continuously acquires raw data about its surroundings through onboard sensors, including spatial information such as terrain, vegetation, and buildings. The data is then analyzed and processed using pre-defined algorithms to identify objects or terrain structures that may provide occlusion as candidate occlusion objects. These candidate occlusion objects are those that, based on their visual features or geometric structure, could provide cover for the autonomous vehicle, such as buildings, walls, large rocks, and undulating terrain.

[0064] Furthermore, identifying candidate occlusions in the environment includes:

[0065] S11. Acquire image data collected by the visual sensors on the unmanned vehicle and point cloud data collected by the lidar.

[0066] Visual sensors can include multispectral cameras and infrared cameras. Multispectral cameras acquire the spectral characteristics of the environment in different bands, collecting two-dimensional image information of the environment, while infrared cameras can detect thermal radiation information under low-light conditions. LiDAR generates point cloud data containing three-dimensional spatial coordinate information by emitting laser beams and receiving reflected signals.

[0067] S12. Input the image data into the pre-trained deep learning recognition model and extract the visual recognition results, which include the target region and its category confidence.

[0068] Deep learning-based object detection models can employ convolutional neural network-based models, such as the YOLO series or Mask R-CNN. Pre-trained deep learning models are trained on a large number of images labeled with occlusion categories, enabling them to automatically identify potential occluded target regions from input images. The target region is the pixel area in the image where the detected candidate object is located, and the category confidence score is the model's probability estimate that the region belongs to the occlusion category, typically ranging from 0 to 1.

[0069] S13. Perform cluster analysis on the point cloud data to obtain multiple point cloud clusters, and calculate the geometric feature parameters of each point cloud cluster.

[0070] Point cloud clustering can employ Euclidean distance-based clustering algorithms, such as the DBSCAN algorithm, to group points that are close in distance into the same cluster. After clustering, each point cloud cluster corresponds to a potential physical entity. The geometric characteristic parameters of each point cloud cluster are calculated, including height, thickness, and point density. Height is obtained by calculating the vertical range of the point cloud cluster, i.e., the difference between the maximum and minimum Z-coordinate values. Thickness is obtained by calculating the minimum bounding thickness of the point cloud cluster in the horizontal direction, i.e., calculating the shorter side length of the minimum bounding rectangle of the point cloud cluster's projection onto the horizontal plane. Point density is obtained by the ratio of the number of points within the point cloud cluster to their occupied volume; the occupied volume is estimated by multiplying the ranges of the point cloud cluster in the three directions.

[0071] S14. Spatial matching is performed between the target region and the point cloud cluster, and consistency verification is performed based on the category confidence and geometric feature parameters. The target regions that pass the verification are identified as candidate occluders.

[0072] The target region in the image coordinate system is mapped to a three-dimensional spatial coordinate system and compared with the spatial position of point cloud clusters to determine whether they correspond to the same physical entity. Specifically, the pixel coordinates of the target region are converted into a view frustum in three-dimensional space using camera calibration parameters, and the spatial intersection of this view frustum with each point cloud cluster is calculated. Simultaneously, multimodal fusion verification is performed by combining the category confidence of visual recognition with the geometric feature parameters of point cloud analysis. Only target regions that match both visual recognition and point cloud analysis results are confirmed as candidate occluders.

[0073] Furthermore, consistency verification is performed based on category confidence and geometric feature parameters, and the target regions that pass the verification are identified as candidate occluders, including:

[0074] S141. Calculate the spatial overlap between each target region and each point cloud cluster, and establish a matching relationship between target regions and point cloud clusters with spatial overlap greater than a preset overlap threshold.

[0075] Spatial overlap calculation involves projecting the pixel boundaries of the target region into 3D space using camera calibration parameters to form a 3D view frustum. The intersection ratio between this view frustum and the spatial region occupied by the point cloud cluster is then calculated. For example, the proportion of points in the point cloud cluster that fall within the view frustum is calculated as an approximation of the spatial overlap. A preset overlap threshold can be set according to the actual application scenario; for example, setting it to 0.5 means that if the intersection ratio exceeds 50%, the two are considered to correspond.

[0076] S142. For target regions and point cloud clusters with matching relationships, calculate the multimodal consistency score based on the category confidence of the target region and the geometric feature parameters of the point cloud clusters.

[0077] Based on the category of the target region in the visual recognition results, the expected range of geometric feature parameters corresponding to that category is determined. The expected range of geometric feature parameters is obtained from a geometric feature database pre-established for various types of occlusions. This database stores the expected height, thickness, and point density ranges corresponding to each category. Each expected range is set statistically or empirically based on the typical physical characteristics of various types of occlusions.

[0078] For example, if the target area is identified as a wall, its corresponding point cloud cluster is expected to have a relatively large height, typically over 2 meters, a thickness of approximately 0.2 to 0.5 meters, and a high point density, with a smooth surface and continuous, dense point cloud. If the target area is identified as a large rock, its corresponding point cloud cluster is expected to have an irregular shape, a height typically over 1.5 meters, a thickness of approximately 1 to 3 meters, extremely high point density, a solid structure, and no obvious voids. If the target area is identified as shrubs, its corresponding point cloud cluster is expected to have a height of 0.5 to 1.5 meters, a thickness of approximately 0.5 to 2 meters, a relatively sparse point density, and a branch-like structure. This results in laser penetration and point cloud dispersion. If the target area is identified as a building, its corresponding point cloud cluster is expected to have a relatively large height, usually over 3 meters, a thickness of about 2 to 10 meters, extremely high point density, complete structure, and regular surface. If the target area is identified as a tree, its corresponding point cloud cluster is expected to have a height of 2 to 10 meters, a thickness of about 1 to 5 meters, medium point density, denser point clouds in the canopy, and sparser point clouds in the trunk. If the target area is identified as terrain undulation, such as a slope or ridge, its corresponding point cloud cluster is expected to have a height of 0.5 to 3 meters, thickness varying with the terrain, high point density, continuous terrain, and no voids.

[0079] The actual geometric feature parameters of the point cloud cluster, namely height, thickness, and point density, are compared with the expected range. When the actual value is within the expected range, the score for that dimension is 1; when it deviates from the expected range, the score decreases linearly according to the degree of deviation, with a lower score for a larger deviation. The scores of each dimension are weighted and summed to obtain the geometric shape conformity score, which ranges from 0 to 1. The larger the value, the more the geometric shape of the point cloud cluster matches the category recognized by vision.

[0080] Finally, the category confidence score of the target region is multiplied by the geometric shape conformity score to obtain the multimodal consistency score. Only when the category confidence score is high and the point cloud geometry matches the recognized category will the score be close to 1, indicating that the visual recognition and point cloud analysis point to the same entity and have consistent features; if the visual recognition result is unreliable, or the point cloud geometry does not match the recognized category, the score will be low, indicating that there may be a mismatch between the two.

[0081] S143. When the multimodal consistency score is greater than the preset score threshold, the consistency verification is deemed to be passed, and the target area is identified as a candidate occluder.

[0082] The preset scoring threshold can be adjusted according to the system's tolerance for false alarms and missed detections, for example, set to 0.7. Only target areas that pass verification are identified as candidate obstructions.

[0083] S2. Select effective shielding objects that meet the shielding requirements from the candidate shielding objects according to the preset validity judgment conditions.

[0084] Each candidate shield is evaluated to determine its ability to provide effective cover for the autonomous vehicle. The effectiveness criteria include geometric dimensions, structural stability, positional relationships, and traffic impact.

[0085] The geometric dimensions require that the shielding object has sufficient height and thickness to block the line of sight from the reconnaissance target. Specifically, the height of the candidate shielding object is compared with a preset height threshold, and the thickness is compared with a preset thickness threshold. For example, the height threshold can be set to 2 meters (a low height that can block the line of sight of common reconnaissance equipment), and the thickness threshold can be set to 0.5 meters (to ensure that the shielding object itself has sufficient physical thickness to block detection).

[0086] The structural stability condition requires that the occluder be a static entity that is not easily moved. This is determined by analyzing the spatial position changes of the point cloud clusters corresponding to the candidate occluder in multiple consecutive frames of data. If the position change is less than a preset threshold, it is determined to be a static entity. For example, if the centroid displacement of the point cloud cluster is less than 0.1 meters in 10 consecutive frames of data, the occluder can be considered to be a fixed structure such as a building or wall, rather than a temporarily parked vehicle or moving person.

[0087] The positional relationship condition requires that the obstruction be located between the unmanned vehicle's path and the direction of the reconnaissance target. This is determined by calculating the azimuth angle of the obstruction relative to the unmanned vehicle and the reconnaissance target. For example, the condition is considered met when the line connecting the unmanned vehicle, the obstruction, and the reconnaissance target is approximately a straight line and the obstruction is located in the middle.

[0088] The requirement for obstructing passage is that the obstruction should not impede the normal passage of the autonomous vehicle, meaning there must be sufficient passage space around it. This is determined by analyzing the relationship between the shape parameters of the obstruction and the surrounding passable area. For example, for an autonomous vehicle with a width of 1.5 meters, if there are passages wider than 1.8 meters on either side or one side of the obstruction, it is considered not to impede passage. Candidate obstructions that simultaneously meet all the above conditions are determined to be valid obstructions.

[0089] S3. Generate the location information and shape information of the effective shielding objects.

[0090] The location information of the occluder is used to characterize the coordinates of the occluder in three-dimensional space, usually with its geometric center point as the location reference; the shape information of the occluder is used to describe the spatial outline of the occluder, including its length, width, height and main extension direction in space.

[0091] Furthermore, generating the location information of the effective shielding objects includes:

[0092] S31. Obtain the point cloud cluster corresponding to each effective shading object.

[0093] After the effective occlusions are selected, each effective occlusion may be associated with one or more point cloud clusters, and the raw point data of these point cloud clusters are read.

[0094] S32. Based on the spatial coordinates of all points in the point cloud cluster, determine the coordinate reference value of the point cloud cluster in each spatial dimension.

[0095] For the X-axis direction, calculate the arithmetic mean of the X-coordinates of all points as the coordinate reference value to obtain the X-axis reference value. Similarly, perform the same processing for the Y-axis and Z-axis directions to obtain the Y-axis and Z-axis coordinate reference values.

[0096] S33. Determine the three-dimensional spatial centroid of the point cloud cluster based on the coordinate reference values ​​in each spatial dimension, and use the three-dimensional spatial centroid as the location information of the effective occlusion.

[0097] The coordinate reference values ​​in the X, Y, and Z directions are combined into a three-dimensional coordinate point, which is the three-dimensional centroid of the point cloud cluster, representing the center position of the point cloud cluster in space.

[0098] Furthermore, generating the shape information of the occlusion based on the effective occlusion includes:

[0099] S41. Obtain the point cloud cluster corresponding to each effective shading object, and calculate the covariance matrix of the point cloud cluster in space based on the spatial coordinates of each point in the point cloud cluster.

[0100] The covariance matrix describes the spatial distribution of points in a point cloud cluster. Let the point cloud cluster contain n points, and the coordinates of each point be... First, calculate the mean vector of the point cloud cluster;

[0101] ,in The average of the X coordinates of all points. The average of the Y-coordinates of all points. This is the average Z-coordinate of all points.

[0102] The covariance matrix is ​​a 3×3 matrix, represented as:

[0103]

[0104] The formulas for calculating each element are as follows:

[0105] ，

[0106] ，

[0107] ,

[0108] ,

[0109] ,

[0110] ,

[0111] Where n is the number of points in the point cloud cluster. Let be the variance of the point cloud in the X direction; Let be the variance of the point cloud in the Y direction; Let be the variance of the point cloud in the Z direction; Let X be the covariance of the point cloud in the X and Y directions; Let be the covariance of the point cloud in the X and Z directions; Let be the covariance of the point cloud in the Y-direction; the larger the variance value, the more dispersed the point cloud is in its corresponding direction. The covariance reflects the correlation between the point cloud in two directions. As an example, if... A positive value indicates that points with larger X-coordinates tend to have larger Y-coordinates, and the point cloud extends along the region enclosed by the positive X-axis and positive Y-axis in the horizontal plane. A negative X-coordinate indicates that points with larger X-coordinates tend to have smaller Y-coordinates. The point cloud extends across the region bounded by the positive X-axis and negative Y-axis in the horizontal plane. A value close to 0 indicates that the X and Y directions are independent of each other.

[0112] S43. Perform eigenvalue decomposition on the covariance matrix to obtain three eigenvalues ​​and the corresponding eigenvectors of each eigenvalue. Take the three eigenvectors as three mutually perpendicular principal directions, including the first principal direction, the second principal direction, and the third principal direction.

[0113] Eigenvalue decomposition can be achieved using functions in numerical computation libraries such as Eigen or LAPACK. The decomposition yields three eigenvalues ​​and their corresponding eigenvectors. The magnitude of the eigenvalues ​​reflects the dispersion of the point cloud cluster along the corresponding eigenvector direction. The eigenvector corresponding to the largest eigenvalue is set as the primary extension direction of the point cloud cluster, i.e., the first principal direction.

[0114] S44. Project all points in the point cloud cluster onto the first principal direction, the second principal direction, and the third principal direction respectively. Calculate the projection length in each principal direction to obtain the length, width, and height of the effective shading object. The first principal direction is used as the orientation of the effective shading object. The length, width, height, and orientation of the effective shading object are used as the shape information of the effective shading object.

[0115] The projection length is calculated by performing a dot product operation between the coordinates of each point and the unit vector of the principal direction to obtain the projection value of that point in the principal direction. The difference between the maximum and minimum values ​​of all projection values ​​is the projection length in that direction. The projection length in the first principal direction is the length of the effective shielding object, the projection length in the second principal direction is the width of the effective shielding object, and the projection length in the third principal direction is the height of the effective shielding object.

[0116] S4. Based on the current location of the unmanned vehicle, the target location, the location information of the obstruction, and the shape information of the obstruction, plan the reconnaissance path on the pre-built environmental cost map.

[0117] The environmental cost map is a pre-built rasterized map based on prior environmental information. Each raster cell corresponds to a region in the actual environment, and the raster size can be set according to the task's accuracy requirements, for example, 0.2 meters × 0.2 meters. Each raster cell in the environmental cost map stores the basic access cost for that location, such as obstacle access cost and terrain access cost. The obstacle access cost is determined based on the distance between the raster cell and the obstacle. For example, the distance from each raster cell to the nearest obstacle can be calculated using a distance transformation algorithm. Raster cells with a distance less than the autonomous vehicle's safe radius are marked as impassable, and their costs are set to a maximum value; the greater the distance, the lower the cost. The terrain access cost is determined based on the terrain type of the raster cell. For example, the slope can be analyzed using a digital elevation model; areas with a slope greater than 15 degrees are assigned higher costs, swamps and deep water are assigned extremely high costs, and flat roads are assigned lower costs. After the environmental cost map is built, it is stored in the form of a two-dimensional array or matrix for use by the path planning module. During path planning, based on the real-time identified effective cover information, exposure risk costs are dynamically overlaid on the environmental cost map to form a complete path planning cost model.

[0118] Using the current location of the autonomous vehicle as the starting point and the target location specified in the task as the ending point, and utilizing identified effective cover resources, a path search is performed on an environmental cost map. The goal of path planning is to find a path from the starting point to the ending point that minimizes the overall exposure risk and maximizes the level of cover, thereby achieving stealthy movement.

[0119] Furthermore, the reconnaissance path is planned on the pre-constructed environmental cost map, including:

[0120] S41. Starting from the current position of the autonomous vehicle and ending at the target position, determine the path search space on the environmental cost map.

[0121] The path search space typically refers to a map area within a certain range centered on the line connecting the start and end points. The size of this area can be adjusted according to task requirements and computing resources. For example, it can be a rectangular area along the direction of the line connecting the start and end points, expanded to both sides by a certain number of grid cells. For instance, if the length of the line connecting the start and end points is L, the search space can be set as a rectangular area with this line as its axis and a width of L / 3.

[0122] S42. Discretize the path search space into multiple path nodes.

[0123] Path nodes are the basic units in pathfinding algorithms, with each node corresponding to a grid cell in the map. In graph search algorithms such as A*, nodes are connected by edges, and the cost of an edge is the cost of moving from one node to an adjacent node, typically calculated using Euclidean distance or Manhattan distance. For example, when moving from node A to an adjacent node in the up, down, left, or right direction, the cost of the edge is 1 grid cell size; when moving to a diagonal node, the cost of the edge is 1.414 grid cells size.

[0124] S43. Based on the location and shape information of the occlusion, determine the exposed area in the path search space.

[0125] Exposed areas refer to regions within the path search space that are not covered by effective cover objects. The projected coverage area of ​​each effective cover object is calculated based on its location and shape on the map. For each grid cell, if it falls within the projected coverage area of ​​any effective cover object, it is marked as an exposed area; otherwise, it is marked as an exposed area. For example, for wall-type cover objects, the set of grid cells it occupies on the grid map can be calculated based on its length, width, and orientation, and these cells are marked as the cover area; for terrain-relief cover objects, the coverage area can be determined based on the projected range of its outline on the horizontal plane.

[0126] S44. Calculate the exposure risk cost of each path node based on the distance from each path node to the nearest exposed area, the distance to the nearest effective shield, and the path curvature between the path node and the previous path node.

[0127] Exposure risk cost reflects the likelihood that the autonomous vehicle will be detected by the target when passing through a node. For each path node, the distance to the nearest exposed area grid cell is calculated. The larger the distance, the further the node is from the open area, and the lower the exposure risk. The distance to the nearest effective cover is also calculated. The closer the distance, the lower the risk. A path curvature penalty is calculated. The curvature penalty is obtained by calculating the absolute value of the angle between the current node's direction of travel and the direction of travel of its parent node. The smaller the angle, the larger the penalty, in order to suppress the exposure risk caused by straight-line travel.

[0128] The specific formula for calculating the cost of exposure risk is as follows:

[0129] ;

[0130] in, The exposure risk cost of path node n, Let n be the distance from path node n to the nearest exposed area. Let n be the distance from path node n to the nearest effective shading object. This is the path curvature penalty term for path node n. , , These are weighting coefficients used to balance the impact of various factors on the cost of exposure risk. They can be dynamically adjusted according to the mission scenario; for example, they can be increased when the target of reconnaissance poses a high threat. value.

[0131] S45. Calculate the cover score for each path node based on the distance from each path node to the effective cover and the relative angle between the effective cover and the target.

[0132] The cover score reflects the coverage effect that a node can obtain. For each path node, all effective cover is traversed, and the contribution value is calculated based on the distance from the node to the cover and the relative angle between the cover and the direction of the reconnaissance target. The closer the node is to an effective cover, and the more the cover is located between the node and the reconnaissance target, the greater the contribution value. The contribution values ​​of all covers are summed to obtain the cover score.

[0133] The formula for calculating the score for a cover is:

[0134] ;

[0135] Path node The screener scored. For the first The type of effective shelter has a weight, and different types of shelters have different shielding effects. For example, reinforced concrete buildings have a higher weight than natural terrain. The specific value can be preset according to factors such as the material and structural strength of the shelter. Path node To the The distance attenuation function of an effective shield can be expressed in Gaussian or exponential form. The closer the distance, the smaller the attenuation and the greater the contribution to the score.

[0136] Let be the relative angular effectiveness function between the i-th effective cover and the direction of the reconnaissance object. When the cover is located between the node and the reconnaissance object, Larger values ​​result in larger values, while smaller values ​​indicate deviations. The relative angle validity function can be calculated based on the range of occlusion angles; a larger effective occlusion angle results in a higher score. For example, it can be defined as... , The angle between the direction of the path node relative to the effective cover and the direction of the path node relative to the reconnaissance object.

[0137] S46. Based on the exposure risk cost and cover score, plan a reconnaissance path in the path search space and search for a reconnaissance path in the path search space.

[0138] The A* algorithm can be used as the path search algorithm. Nodes are sorted and selected using an evaluation function to obtain the reconnaissance path. The evaluation function is:

[0139] ；

[0140] in, This refers to the path node currently being evaluated. The actual path cost from the origin to node n is given by the nodes in the environmental cost map. The basic passage cost is obtained by summing the basic passage costs of the grid cells. The basic passage cost reflects the cost that the autonomous vehicle needs to pay to pass through the grid cell, including obstacle passage cost and terrain passage cost. For example, if a node If located on a passable, flat road surface, then The accumulated value is small; if the node If located in an area with a steep slope or dense vegetation, then The accumulated value is large; if the node If located in an impassable area where the obstacle is located, then The accumulated value is set to the maximum value, so that the node will not be selected. For the node Heuristic cost estimation to the target point typically uses Euclidean distance or Manhattan distance as the estimate to guide the search direction, prioritizing expansion towards the target point. For example, if the node... If the Euclidean distance between the target point and the target point is 10 meters, then... The value is 10 multiplied by the unit grid cost, which can be preset based on the average accessibility of the environment.

[0141] Starting from the starting point as the current node, calculate the value of each adjacent node. The node with the smallest f(n) value is selected as the next node to move to. The newly selected node is then used as the current node, and the above process is repeated until the target position is reached.

[0142] S5. Control the unmanned vehicle to travel along the reconnaissance path and monitor the reconnaissance path based on real-time updated environmental data. When the preset replanning conditions are met, replan the reconnaissance path and control the unmanned vehicle to travel along the replanned reconnaissance path.

[0143] The autonomous vehicle's motion control system adjusts its steering, speed, and direction in real time according to the planned path instructions, enabling the vehicle to move along the predetermined path. During travel, sensors continuously collect environmental data, and the processing unit constantly compares the real-time data with the environmental information used during path planning to determine if the environment has changed. When new obstacles are detected on the path, existing effective cover disappears, or other situations prevent safe and concealed travel, the system automatically triggers a replanning process, re-executing steps S1 to S4 to generate a new path and controlling the autonomous vehicle to travel along the new path.

[0144] Specifically, the preset replanning conditions include the detection of new obstacles on the reconnaissance path or the disappearance of effective cover on the reconnaissance path.

[0145] During the autonomous vehicle's movement, the environmental data collected in real time by sensors is compared with the environmental information used during path planning. When an obstacle not originally present is detected on the current path, causing a path blockage, replanning is triggered. Similarly, when an effective cover previously used as a concealment resource in the path planning disappears—for example, a temporary cover is removed or a static entity is destroyed—the original path's concealment may no longer meet requirements, also triggering replanning. Furthermore, replanning conditions may also include situations where the current exposure risk value of the path exceeds a preset risk threshold, or the autonomous vehicle's actual position deviates from the preset path beyond the allowable error range.

[0146] The present invention also provides an unmanned vehicle close-in reconnaissance system based on dynamic masking path planning, the system comprising:

[0147] The identification module is used to identify candidate occlusions in the environment based on real-time environmental data collected by the unmanned vehicle.

[0148] The filtering module is used to filter out effective shielding objects that meet the shielding requirements from the candidate shielding objects according to preset validity discrimination conditions;

[0149] The generation module is used to generate the location information and shape information of the effective shielding object based on the effective shielding object;

[0150] The planning module is used to plan a reconnaissance path on a pre-built environmental cost map based on the current position of the unmanned vehicle, the target position, the position information of the cover object, and the shape information of the cover object.

[0151] The control module is used to control the unmanned vehicle to travel along the reconnaissance path, monitor the reconnaissance path based on real-time updated environmental data, and replan the reconnaissance path when the preset replanning conditions are met, and control the unmanned vehicle to travel along the replanned reconnaissance path.

[0152] It is understandable that the modules recorded in this unmanned vehicle close-in reconnaissance system based on dynamic masking path planning are similar to those in the reference system. Figure 1 The steps described correspond to those in the unmanned vehicle close-in reconnaissance method based on dynamic masking path planning. Therefore, the operations, features, and beneficial effects described above for the unmanned vehicle close-in reconnaissance method based on dynamic masking path planning are also applicable to the unmanned vehicle close-in reconnaissance system based on dynamic masking path planning and its included modules, and will not be repeated here.

[0153] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for close-range reconnaissance of unmanned vehicles based on dynamic masking path planning, characterized in that, include: Based on real-time environmental data collected by unmanned vehicles, candidate occlusions in the environment are identified; Based on preset validity criteria, select effective shielding objects that meet the shielding requirements from the candidate shielding objects; Generate the location information and shape information of the effective shielding object based on the effective shielding object; Based on the current position of the unmanned vehicle, the target position, the location information of the cover, and the shape information of the cover, a reconnaissance path is planned on a pre-constructed environmental cost map. This planning includes: determining a path search space on the environmental cost map, starting from the current position of the unmanned vehicle and ending at the target position; discretizing the path search space into multiple path nodes; determining the exposed area in the path search space based on the location and shape information of the cover; calculating the exposure risk cost of each path node based on the distance from each path node to the nearest exposed area, the distance to the nearest effective cover, and the path curvature between the path node and the previous path node; calculating the cover score of each path node based on the distance from each path node to the effective cover and the relative angle between the effective cover and the reconnaissance target; and planning a reconnaissance path in the path search space based on the exposure risk cost and the cover score. The unmanned vehicle is controlled to travel along the reconnaissance path, and the reconnaissance path is monitored based on real-time updated environmental data. When the preset replanning conditions are met, the reconnaissance path is replanned, and the unmanned vehicle is controlled to travel along the replanned reconnaissance path.

2. The method according to claim 1, characterized in that, Based on real-time environmental data collected by autonomous vehicles, candidate occlusions in the environment are identified, including: Acquire image data collected by the visual sensors on the autonomous vehicle and point cloud data collected by the LiDAR; The image data is input into a pre-trained deep learning recognition model to extract visual recognition results, which include the target region and its category confidence. Cluster analysis is performed on the point cloud data to obtain multiple point cloud clusters, and the geometric feature parameters of each point cloud cluster are calculated. The target region is spatially matched with the point cloud cluster, and consistency verification is performed based on the category confidence and the geometric feature parameters. The target region that passes the verification is determined as the candidate occluder.

3. The method according to claim 2, characterized in that, Consistency verification is performed based on the category confidence level and the geometric feature parameters. Target regions that pass the verification are identified as candidate occluders, including: Calculate the spatial overlap between each target region and each point cloud cluster, and establish a matching relationship between target regions and point cloud clusters with spatial overlap greater than a preset overlap threshold. For target regions and point cloud clusters with matching relationships, a multimodal consistency score is calculated based on the category confidence of the target region and the geometric feature parameters of the point cloud clusters. When the multimodal consistency score is greater than the preset score threshold, the consistency verification is deemed to be successful, and the target area is identified as a candidate occluder.

4. The method according to claim 1, characterized in that, The generation of the location information of the effective shielding object includes: Obtain the point cloud cluster corresponding to each effective shading object; Based on the spatial coordinates of all points in the point cloud cluster, determine the coordinate reference value of the point cloud cluster in each spatial dimension; The three-dimensional centroid of the point cloud cluster is determined based on the coordinate reference values ​​in each spatial dimension; The three-dimensional centroid is used as the location information of the effective shading object.

5. The method according to claim 1, characterized in that, The generation of the shape information of the effective shielding object includes: Obtain the point cloud cluster corresponding to each of the aforementioned effective shielding objects; Based on the spatial coordinates of each point in the point cloud cluster, calculate the covariance matrix of the point cloud cluster in space. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues ​​and eigenvectors corresponding to each eigenvalue. The three eigenvectors are taken as three mutually perpendicular principal directions, including a first principal direction, a second principal direction and a third principal direction. Project all points in the point cloud cluster onto the first principal direction, the second principal direction, and the third principal direction respectively, and calculate the projection length in each principal direction to obtain the length, width, and height of the effective shield, wherein the length is the projection length in the first principal direction, the width is the projection length in the second principal direction, and the height is the projection length in the third principal direction. The first main direction is taken as the orientation of the effective shield; The length, width, height, and orientation of the effective shield are used as the shape information of the effective shield.

6. The method according to claim 1, characterized in that, The exposure risk cost for each path node is calculated as shown in the following formula; ； in, The exposure risk cost of path node n, Let n be the distance from path node n to the nearest exposed area. Let n be the distance from path node n to the nearest effective shading object. This is the path curvature penalty term for path node n. , , These are the weighting coefficients.

7. The method according to claim 1, characterized in that, The cover score for each path node is calculated as shown in the following formula; ； in, The cover score for path node n. Let i be the type weight of the i-th effective occlusion. Let n be the distance decay function from path node n to the i-th effective shield. Let be the relative angular effectiveness function between the i-th effective cover and the direction of the reconnaissance target, where , Path node Relative to the first The direction and path nodes of an effective shield The angle between the directions of the target being reconnaissance.

8. The method according to claim 1, characterized in that, The preset replanning conditions include: A new obstacle is detected on the reconnaissance path or an effective cover on the reconnaissance path disappears.

9. An unmanned vehicle close-in reconnaissance system based on dynamic masking path planning, characterized in that, The system includes: The identification module is used to identify candidate occlusions in the environment based on real-time environmental data collected by the unmanned vehicle. The filtering module is used to filter out effective shielding objects that meet the shielding requirements from the candidate shielding objects according to preset validity discrimination conditions; The generation module is used to generate the location information and shape information of the effective shielding object based on the effective shielding object; The planning module is used to plan a reconnaissance path on a pre-built environmental cost map based on the current position of the unmanned vehicle, the target position, the location information of the cover, and the shape information of the cover. Planning the reconnaissance path on the pre-built environmental cost map includes: determining a path search space on the environmental cost map, starting from the current position of the unmanned vehicle and ending at the target position; discretizing the path search space into multiple path nodes; determining the exposed area in the path search space based on the location information and shape information of the cover; calculating the exposure risk cost of each path node based on the distance from each path node to the nearest exposed area, the distance to the nearest effective cover, and the path curvature between the path node and the previous path node; calculating the cover score of each path node based on the distance from each path node to the effective cover and the relative angle between the effective cover and the reconnaissance target; and planning a reconnaissance path in the path search space based on the exposure risk cost and the cover score. The control module is used to control the unmanned vehicle to travel along the reconnaissance path, monitor the reconnaissance path based on real-time updated environmental data, and replan the reconnaissance path when the preset replanning conditions are met, and control the unmanned vehicle to travel along the replanned reconnaissance path.

Images