A multi-resolution autonomous composition method and system for microrobots

Abstract

The application discloses a kind of multi-resolution autonomous composition methods and systems for micro robot, and the distance image is obtained by pre-processing original environment data;Integral image and minimum pooling image are generated based on distance image;According to the logic origin of local map and physical boundary determined by real-time pose of robot, convert local map coordinates into global discrete index, and then linear addressing index in annular buffer zone;Incremental 3D ray casting is performed at low-resolution block level, and potential occupancy endpoint set is screened;According to the resolution of the unit in the set, project the unit to be updated to the distance image, perform adaptive projection verification, and output the updated map structure.Through annular buffer zone mechanism, complete decoupling of memory consumption and exploration range is realized, and extremely low resource overhead is achieved.The dual acceleration strategy of integral image and minimum pooling pyramid is introduced, and the hybrid update mechanism from coarse to fine is used, which effectively avoids redundant calculation in empty area.

Description

Technical Field

[0001] This invention relates to intelligent navigation and environmental perception for robots, and more particularly to a multi-resolution autonomous mapping method and system for microrobots. Background Technology

[0002] Miniature robots (MAVs) play a crucial role in complex tasks such as infrastructure inspection and disaster search and rescue, and efficient and high-precision environmental perception is the foundation for their safe operation and autonomous obstacle avoidance. Occupational grid mapping is currently the mainstream environmental representation method. However, due to the limitations of the size, payload, and battery capacity of miniature robots, edge computing units often struggle to support the huge memory footprint and computational overhead required for high-precision reconstruction.

[0003] In existing research, while UAV-centric sliding map methods (such as FIIMap and ROG-Map) can prevent memory overflow by limiting the local map extent, these methods typically use a fixed resolution. In practical operations, the measurement accuracy of sensors (such as depth cameras or LiDAR) on microrobots varies at different distances, and fixed-resolution maps cannot flexibly adapt to changes in sensor characteristics. For example, at long distances, sensor accuracy decreases, and a fixed high-resolution map will generate a large amount of redundant computation; while at close range, it cannot provide sub-voxel level detail descriptions.

[0004] On the other hand, while tree-based multi-resolution mapping methods (such as OctoMap and WaveMap) can achieve multi-scale representation, their core logic lies in maintaining a global map. As the exploration range expands, the number of map nodes surges, and memory requirements grow linearly or even exponentially, making it difficult to achieve long-term exploration in large-scale environments on microrobots with extremely limited resources.

[0005] Therefore, how to achieve efficient mapping that balances high-precision perception and large-scale exploration while ensuring strictly bounded memory consumption is a pressing technical challenge facing current autonomous navigation systems for microrobots. Summary of the Invention

[0006] Purpose of the invention: In view of the above problems, the purpose of this invention is to provide an efficient and robust multi-resolution autonomous mapping method and system for mapping under limited carrier resources.

[0007] Technical Solution: To solve the above problems, the present invention provides a multi-resolution autonomous mapping method for microrobots, comprising the following steps:

[0008] Step 1: Obtain the raw environmental data, preprocess the raw environmental data to obtain the distance image; and generate a multi-scale integral image and a minimum pooling image pyramid based on the distance image;

[0009] Step 2: Obtain the real-time pose of the microrobot, determine the logical origin and physical boundary of the local map based on the real-time pose of the microrobot, convert the local map coordinates into a global discrete index, and map the global discrete index into a linear addressing index in a circular buffer.

[0010] Step 3: Perform incremental 3D ray casting at the low-resolution block level to determine the block hit by the ray tip as a unit in the potential occupancy endpoint set; dynamically select the resolution based on the distance from the unit to the detection point; based on the selected resolution, project the unit to be updated onto the distance image, perform adaptive projection verification, and output the updated map structure.

[0011] Furthermore, the raw environmental data includes raw images acquired by a depth camera and raw point clouds acquired by a lidar.

[0012] For the raw image acquired by the depth camera, the pixel coordinates of the raw image are projected onto the sensor coordinate system using the pinhole model to obtain the three-dimensional coordinates of the spatial point. The pixel coordinates are used as the row and column indices of the distance to the image, and the distance from the spatial point to the depth camera is used as the distance value to the image.

[0013] The three-dimensional coordinates of spatial points are obtained from the original point cloud acquired by the lidar. A spherical projection model is used to convert the three-dimensional coordinates of spatial points into horizontal and vertical angles. The horizontal and vertical angles are used as row and column indices of the distance image, and the distance from the spatial point to the lidar is used as the distance value of the distance image.

[0014] Furthermore, the multi-scale integral image includes: a distance sum integral image, a distance square sum integral image, and an effective pixel count integral image; each of the multi-scale integral images uses the row and column of pixels as the index, and uses the sum of distance values, the square of distance values, and the number of effective pixels in the projection area within a preset time period as the pixel value;

[0015] The minimum pooling image pyramid includes Layered images, where the distance image For the initial layer, the first layer pixel values By examining the first layer The minimum value within the region is determined.

[0016] Furthermore, the global discrete index is obtained by rounding down the spatial points in the world coordinate system according to the corresponding hierarchical resolution, and the local index is... ,in, It is a global discrete index. The logical origin that is dynamically updated with the robot. This refers to the size of the local map in the spatial dimension.

[0017] The linearized addressing index is ,in, , and For local indexes in three spatial dimensions, , This refers to the size of the local map in two spatial dimensions.

[0018] Furthermore, the resolution is dynamically selected based on the Euclidean distance from the center point of the cell in the potential occupied endpoint set to the detection point. When the distance is less than the near distance threshold, the sub-voxel level is selected; when the distance is greater than the far distance threshold, the block level is selected; and between the near distance threshold and the far distance threshold, the voxel level is selected.

[0019] Furthermore, the adaptive projection verification includes the following steps:

[0020] Calculate the depth range of the unit to be updated in the sensor coordinate system. And projected onto the distance image Determine pixel region ;

[0021] Calculating pixel regions using multi-scale integral images The statistical moments compare the distance value of each effective pixel in the pixel region with the cell depth range. Effective pixels that fall within the cell depth range are considered internal pixels, effective pixels that are below the cell depth range are considered front occluded pixels, and effective pixels that exceed the cell depth range are considered rear free pixels.

[0022] The proportions of internal pixels, occluded pixels in front, and free pixels behind in the pixel region are compared with a threshold. The Bayesian log probability is updated based on the comparison result, and the map state is finally updated based on the Bayesian log probability.

[0023] Furthermore, the pixel region The formula for calculating statistical moments is:

[0024] ;

[0025] in, The mean depth of the region. The standard deviation is denoted as .

[0026] Furthermore, the proportions of internal pixels, front-occluded pixels, and rear-free pixels within the pixel region are compared with a threshold, specifically including:

[0027] like If the cell is in an occupied state, then the Bayesian log probability is updated. ,in This represents the percentage of internal pixels within a pixel region. For adaptive efficiency threshold, To occupy the Bayesian log probability increment of the observation, This represents the upper bound of the Bayesian logarithmic probability;

[0028] like The cell is determined to be in an idle state, and the Bayesian log probability is updated. ,in This represents the percentage of idle pixels in the pixel area. The preset idle threshold is used to determine the idle time. This represents the Bayesian log probability increment for missed observations. This is the lower bound of the Bayesian logarithmic probability;

[0029] If the percentage of effective pixels is lower than the effective threshold or If the occlusion ratio threshold is reached, it is judged as complete occlusion, and the Bayesian log probability remains unchanged.

[0030] The storage unit outputs the Bayesian log probability when the resolution reaches the occupied threshold, and the Bayesian log probability of sub-voxels, voxels, and blocks is passed up through aggregation.

[0031] Furthermore, the formula for calculating the adaptive efficiency threshold is as follows:

[0032] ;

[0033] in, This is the preset minimum efficiency threshold. To normalize the standard deviation, For consistency sensitivity, This is the distance attenuation coefficient. For observation distance.

[0034] The present invention discloses a multi-resolution autonomous mapping system for microrobots, comprising:

[0035] The data preprocessing module is used to acquire raw environmental data, preprocess the raw environmental data to obtain a distance image, and generate a multi-scale integral image and a minimum pooling image pyramid based on the distance image.

[0036] The map management module is used to acquire the real-time pose of the microrobot, determine the logical origin and physical boundary of the local map based on the real-time pose of the microrobot, convert the local map coordinates into a global discrete index, and map the global discrete index into a linear addressing index in a circular buffer.

[0037] The map update module performs incremental 3D ray casting at a low-resolution block level, identifies the block hit by the ray as a unit in the potential occupancy endpoint set, dynamically selects the resolution based on the distance from the unit to the detection point, projects the unit to be updated onto the distance image based on the selected resolution, performs adaptive projection verification, and outputs the updated map structure.

[0038] Beneficial Effects: Compared to existing technologies, the significant advantage of this invention is that it achieves complete decoupling of memory consumption and exploration range through a circular buffer mechanism. Experimental data shows that, under the same high-resolution configuration, this method significantly reduces memory usage from 9321.66MB to 239.12MB compared to mainstream algorithms, achieving extremely low resource overhead.

[0039] By introducing a dual acceleration strategy of integral images and minimum pooling pyramids, coupled with a coarse-to-fine hybrid update mechanism, redundant computation in open areas is effectively avoided. Map update latency is reduced from 99.47ms to 44.85ms, demonstrating excellent computational efficiency and perfectly meeting the requirements of real-time perception.

[0040] The synergy between adaptive resolution selection and dynamic thresholding improves mapping quality in complex environments. In static scene cross-validation, the reconstruction accuracy is improved from 55.41% to 84.32% compared to the traditional OctoMap algorithm, and it can effectively suppress artifact noise caused by dynamic obstacles, demonstrating high reconstruction accuracy and robustness.

[0041] The unified distance image model enables independent and seamless processing of depth camera and LiDAR data, enhancing the system's applicability in different indoor and outdoor operating environments and providing high compatibility. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall method of the present invention.

[0043] Figure 2 This is a schematic diagram of the explicit layered and ring buffer mapping structure in this invention.

[0044] Figure 3 This is a detailed logic flowchart of the hybrid occupancy update and adaptive projection verification in this invention. Detailed Implementation

[0045] like Figure 1 As shown in the figure, this embodiment presents a multi-resolution autonomous mapping method for microrobots. The implementation flowchart of the method is as follows. Figure 1 As shown, it includes the following steps:

[0046] Step (1): Obtain the raw environmental data and construct a unified distance image, integral image, and pooling pyramid, including the following steps:

[0047] (1-1) Multi-sensor projection and alignment, where multiple sensors include a depth camera and a lidar. The raw images acquired by the depth camera and the raw point cloud data acquired by the lidar on the microrobot are obtained and converted into a unified distance image. .

[0048] For the raw images captured by the depth camera, a pinhole model is used to obtain the pixel coordinates of the raw images. Projected onto a three-dimensional point The calculation formula is as follows:

[0049]

[0050] in, and The coordinates of the camera's optical center pixel. and For the camera's focal length parameter, This represents the depth value.

[0051] For LiDAR, three-dimensional spatial points are obtained from the original point cloud data. The coordinates are obtained using a spherical projection model. The formula for calculating a bidirectional mapping is as follows:

[0052]

[0053] in, The distance from the 3D point to the sensor origin (i.e., the center distance) is θ, where θ is the horizontal angle and φ is the pitch angle, calculated from the lateral coordinates (x), longitudinal coordinates (y), and height coordinates (z) of the 3D point. Then, based on the horizontal and vertical field of view and angular resolution of the lidar, θ and φ are converted into pixel row and column indices u and v in the distance image. Conversely, given pixels u and v, the direction can be deduced and the 3D coordinates restored using the above relationship, thus achieving bidirectional mapping.

[0054] For data from depth cameras and LiDAR, a minimum distance preservation strategy is used to handle conflicts projected onto the same pixel. The final distance value of that pixel is defined as: .

[0055] (1-2) Multi-scale information generation. This involves obtaining the distance image. Subsequently, the processor constructs three types of two-dimensional prefix integral images in parallel: a distance sum integral image, a distance square sum integral image, and an effective pixel count integral image. All three are two-dimensional prefix statistical images indexed by pixel rows and columns, used to calculate the distance sum within the projected rectangular region within a preset time period. Sum of squared distances and number of effective pixels At the same time, generate Mini-pooling image pyramid of layers. Based on distance image. For the initial layer, the first Pixel values ​​of the layer pyramid By the layer The minimum value within the region is determined, and the calculation process is as follows:

[0056]

[0057] Step (2): Combining the robot's pose, the map boundary is maintained using a circular buffer mechanism and addressing mapping is implemented, including the following steps:

[0058] (2-1) Construct an explicit hierarchical structure. The explicit hierarchical logic of the environment space is as follows: Figure 2 As shown on the right, the system recursively discretizes the environmental space into blocks, voxels, and sub-voxels based on the observation distance. The hierarchical depth from sub-voxels to voxels is... The depth of the voxel to block hierarchy is Then each voxel contains 2 on one side. The power of sub-voids, a total of 2 of 3 A power of sub-elements, each block containing 2 on one side. Subvoxels. In one implementation of this embodiment, a subvoxel resolution of 100 mm is used. , Each voxel contains 4 sub-voxels on one side (64 in total), and each block contains 2 voxels on one side (8 in total), with corresponding dimensions of 0.05m for sub-voxels, 0.2m for voxels, and 0.4m for blocks.

[0059] (2-2) Ring buffer address mapping. For example... Figure 2 As shown in the mapping process on the left, the current pose information of the microrobot is used to map the global coordinates in the world coordinate system. Convert to global discrete index Through the logical origin that is dynamically updated with the robot. Calculate local indexes in memory The global discrete index is obtained by rounding down the points in the world coordinate system according to the corresponding level resolution; the logical origin is defined as the starting block index of the local map window in the global block index space, which is taken as the current block index of the robot minus half of the number of local map blocks, so that the robot stays near the center area of ​​the local map window.

[0060]

[0061]

[0062]

[0063] in, These represent the size of the local map in three spatial dimensions. Figure 2 Visually demonstrates the use of The modulo operation enables the index to circulate within the circular buffer, ultimately allowing for fast access to the corresponding memory storage unit via the following linear addressing formula: The local map dimension is a cache dimension determined by the map's physical extent and the block side length. Figure 2 The circular buffer in the code does not mean that the geometry is circular, but rather that the global block index is reused cyclically in a fixed-length storage array after modulo mapping.

[0064]

[0065] This achieves decoupling between world coordinates and memory addresses. This linear address is used to locate the corresponding block, voxel, or sub-voxel storage unit during subsequent ray casting, projection verification, and Bayesian probability updates, allowing for the reuse of fixed memory during map sliding.

[0066] The detailed execution flow of the hybrid occupancy update and probability determination logic described in steps (3) to (5) is as follows: Figure 3 As shown, the specific implementation process is as follows:

[0067] Step (3): Perform incremental 3D ray casting at the low-resolution level to coarsely screen potential free space and footprint:

[0068] correspond Figure 3 In the coarse screening stage, at a low-resolution block level, odometry is used to calculate the robot's pose changes between adjacent time steps. 3D ray casting is performed only on newly observed blocks or blocks that have undergone state changes. Blocks through which the rays pass are marked as a set of potential free space blocks. Mark the block hit by the end of the light beam as a potential space-occupying set. .

[0069] Step (4) Perform dynamic resolution selection for potential occupied cells and calculate the statistical moments of the projected region:

[0070] correspond Figure 3 The fine screening and verification stage in the process targets the set The cell to be updated in the data is used to calculate the Euclidean distance from its center point to the sensor. .in accordance with Perform dynamic resolution selection (sub-voxel level is selected when the distance is less than the near distance threshold, block level is selected when the distance is greater than the far distance threshold, and voxel level is selected in between). Project the cell to be updated onto the distance image generated in step (1). The corresponding pixel bounding box region is obtained from the above. Specifically, this involves calculating the depth range of the unit to be updated in the sensor coordinate system. And projected onto the distance image Determine pixel region ;

[0071] The statistical moments within the region are quickly calculated using the integral image, and the resulting standard deviation is used as the normalized standard deviation. This standard deviation is then substituted into the adaptive efficiency threshold formula in step (5-1).

[0072]

[0073] in, The mean depth of the region. The standard deviation is then used. The distance value of each effective pixel within the projection area is then compared with the cell depth range. Pixels falling within this range are counted as pixels within the unit (the proportion is denoted as...). (), distance value less than Count as pixels occluded in front of the unit (the proportion is denoted as ). (Distance value greater than) Counted as the empty pixels behind the unit (the proportion is denoted as ). The effective pixel refers to a pixel that is within the effective ranging range of the sensor and has an effective distance value. The above three ratios are used to determine the occupied, idle, or occluded state of the unit in the next step. .

[0074] Step (5): Combining the adaptive efficiency threshold, update the state and output the map using a Bayesian probabilistic framework:

[0075] (5-1) Calculate the adaptive efficiency threshold. Introduce the adaptive threshold. To handle sensor measurement noise at different detection distances:

[0076]

[0077] in, This is the preset minimum efficiency threshold (e.g., 0.15). The normalized standard deviation is obtained through the statistical moments of the pixel region. and These are the set consistency sensitivity and distance attenuation coefficient, respectively.

[0078] (5-2) Bayesian logarithmic probability state update and map output. (Corresponding to...) Figure 3 The final logical judgment branch uses a Bayesian logarithmic probability framework to update the map state: that is, by using logarithmic probability, it accumulates the occupied and idle evidence of the same storage unit from multiple frames of observation, and avoids probability oversaturation by truncating the upper and lower limits.

[0079] like If so, the cell is determined to be in an occupied state, and its logarithmic probability is updated. This condition refers to the proportion of effective pixels within the depth range of the projected area. Reaching the adaptive efficiency threshold The update amount is the log probability increment corresponding to the occupancy observation probability. In the formula... To occupy the log probability increment corresponding to the observation, Occupying the upper limit of probability, The logarithmic probability of a unit.

[0080] like The cell is determined to be in an idle state, and then updated. This condition refers to the proportion of effective pixels located behind the cells within the projection area. When the idle threshold is reached, the update amount is the log probability increment corresponding to the probability of a missed observation. Where... The preset idle threshold is used to determine the idle time. This represents the logarithmic probability increment corresponding to the missed observation. The lower bound of the probability of occupancy.

[0081] The effective pixel ratio is the ratio of the number of effective pixels to the total number of pixels sampled in the projection area. If the effective pixel ratio is too low, below the effective threshold, or If the occlusion ratio threshold is reached, meaning there is already an observation surface in front of the cell and most effective pixels are located in front of the near end of the cell, then it is determined to be in an occlusion state or an uncertain state, and the original probability value remains unchanged, which is different from the occupancy state determination. In the formula... This is the preset occlusion detection threshold.

[0082] The storage unit state is obtained by comparing the updated log probability with the occupancy threshold: when the log probability of a block, voxel, or sub-voxel reaches the occupancy threshold, it is output as a storage unit at the corresponding resolution; units below the idle threshold or not activated are not output. The log probability of a sub-voxel is passed up to voxels and blocks through aggregation to output a multi-resolution occupancy map.

Claims

1. A multi-resolution autonomous mapping method for microrobots, characterized in that, Includes the following steps: Step 1: Obtain the raw environmental data, preprocess the raw environmental data to obtain the distance image; and generate a multi-scale integral image and a minimum pooling image pyramid based on the distance image; Step 2: Obtain the real-time pose of the microrobot, determine the logical origin and physical boundary of the local map based on the real-time pose of the microrobot, convert the local map coordinates into a global discrete index, and map the global discrete index into a linear addressing index in a circular buffer. Step 3: Perform incremental 3D ray casting at the low-resolution block level to determine the block hit by the ray tip as a cell in the potential occupancy endpoint set; dynamically select the resolution based on the distance from the cell to the detection point; Based on the selected resolution, the cells to be updated are projected onto the distance image, adaptive projection verification is performed, and the updated map structure is output.

2. The multi-resolution autonomous mapping method for microrobots according to claim 1, characterized in that, The raw environmental data includes raw images acquired by a depth camera and raw point clouds acquired by a lidar. For the raw image acquired by the depth camera, the pixel coordinates of the raw image are projected onto the sensor coordinate system using the pinhole model to obtain the three-dimensional coordinates of the spatial point. The pixel coordinates are used as the row and column indices of the distance to the image, and the distance from the spatial point to the depth camera is used as the distance value to the image. The three-dimensional coordinates of spatial points obtained from the original point cloud acquired by the lidar are converted into horizontal and vertical angles using a spherical projection model. The horizontal and vertical angles are used as row and column indices of the distance image, and the distance from the spatial point to the lidar is used as the distance value of the distance image.

3. The multi-resolution autonomous mapping method for microrobots according to claim 2, characterized in that, The multi-scale integral image includes: a distance sum integral image, a distance square sum integral image, and an effective pixel count integral image; each multi-scale integral image is indexed by the row and column of pixels, and the pixel value is the sum of distance values, the square of distance values, and the number of effective pixels in the projection area within a preset time period. The minimum pooling image pyramid includes Layered images, where the distance image For the initial layer, the first layer pixel values By examining the first layer The minimum value within the region is determined.

4. The multi-resolution autonomous mapping method for microrobots according to claim 1, characterized in that, The global discrete index is obtained by rounding down the spatial points in the world coordinate system according to the corresponding hierarchical resolution. The local index is... ,in, It is a global discrete index. As the logical origin that is dynamically updated with the robot, This refers to the size of the local map in the spatial dimension. The linearized addressing index is ,in, , and For local indexes in three spatial dimensions, , This refers to the size of the local map in two spatial dimensions.

5. The multi-resolution autonomous mapping method for microrobots according to claim 4, characterized in that, The resolution is dynamically selected based on the Euclidean distance from the center point of the cell in the potential occupied endpoint set to the detection point. When the distance is less than the near distance threshold, the sub-voxel level is selected; when the distance is greater than the far distance threshold, the block level is selected; and the voxel level is selected between the near distance threshold and the far distance threshold.

6. The multi-resolution autonomous mapping method for microrobots according to claim 3, characterized in that, The adaptive projection verification includes the following steps: Calculate the depth range of the unit to be updated in the sensor coordinate system. And projected onto the distance image Determine pixel region ; Calculating pixel regions using multi-scale integral images The statistical moments compare the distance value of each effective pixel in the pixel region with the cell depth range. Effective pixels that fall within the cell depth range are considered internal pixels, effective pixels that are below the cell depth range are considered front occluded pixels, and effective pixels that exceed the cell depth range are considered rear free pixels. The proportions of internal pixels, occluded pixels in front, and free pixels behind in the pixel region are compared with a threshold. The Bayesian log probability is updated based on the comparison result, and the map state is finally updated based on the Bayesian log probability.

7. The multi-resolution autonomous mapping method for microrobots according to claim 6, characterized in that, The pixel region The formula for calculating statistical moments is: ； in, The mean depth of the region. The standard deviation is denoted as .

8. The multi-resolution autonomous mapping method for microrobots according to claim 7, characterized in that, The proportions of internal pixels, front-occluded pixels, and rear-free pixels within the pixel region are compared with a threshold, specifically including: like If the cell is in an occupied state, then the Bayesian log probability is updated. ,in This represents the percentage of internal pixels within a pixel region. For adaptive efficiency threshold, To occupy the Bayesian log probability increment of the observation, This represents the upper bound of the Bayesian logarithmic probability; like The cell is determined to be in an idle state, and the Bayesian log probability is updated. ,in This represents the percentage of idle pixels in the pixel area. The preset idle threshold is used to determine the idle time. This represents the Bayesian log probability increment for missed observations. This is the lower bound of the Bayesian logarithmic probability; If the percentage of effective pixels is lower than the effective threshold or If the occlusion ratio threshold is reached, it is judged as complete occlusion, and the Bayesian log probability remains unchanged. The storage unit outputs the Bayesian log probability when the resolution reaches the occupied threshold, and the Bayesian log probability of sub-voxels, voxels, and blocks is passed up through aggregation.

9. The multi-resolution autonomous mapping method for microrobots according to claim 8, characterized in that, The formula for calculating the adaptive efficiency threshold is: ； in, This is the preset minimum efficiency threshold. To normalize the standard deviation, For consistency sensitivity, This is the distance attenuation coefficient. For observation distance.

10. A multi-resolution autonomous mapping system for microrobots, characterized in that, include: The data preprocessing module is used to acquire raw environmental data, preprocess the raw environmental data to obtain a distance image, and generate a multi-scale integral image and a minimum pooling image pyramid based on the distance image. The map management module is used to acquire the real-time pose of the microrobot, determine the logical origin and physical boundary of the local map based on the real-time pose of the microrobot, convert the local map coordinates into a global discrete index, and map the global discrete index into a linear addressing index in a circular buffer. The map update module performs incremental 3D ray casting at a low-resolution block level, identifies the block hit by the ray as a cell in the potential occupancy endpoint set, and dynamically selects the resolution based on the distance from the cell to the detection point. Based on the selected resolution, the cells to be updated are projected onto the distance image, adaptive projection verification is performed, and the updated map structure is output.

Images