AGV free space generation method based on three-mode obstacle segmentation and conditional diffusion inflation

Abstract

The application discloses an AGV free space generation method based on three-mode obstacle segmentation and conditional diffusion inflation, and relates to the technical field of warehouse automation and mobile robot navigation. Based on a design blueprint and a scanning point cloud, a common control point is used to unify a terrace coordinate system, and a logical grid is generated along a building main shaft. Unlike traditional single-scale projection, a vertical cutting interest area is carried out on the height of a shelf bottom beam and the highest working surface of a vehicle, and density adaptive voxelization is carried out in the interest area, so that a multi-resolution hierarchical structure is realized, high-density obstacle areas are subdivided, sparse channel areas are aggregated, additional semantic labels and local geometric attributes are added, and vertical details and overall data volume control are considered.

Description

Technical Field

[0001] This invention relates to the fields of warehouse automation and mobile robot navigation technology, specifically to a method that integrates one-time 3D point cloud and design blueprint multi-resolution rasterization, three-channel feature extraction, conditional diffusion safety buffer generation, and free space vectorization. This method can quickly construct accurate and interactive access topology base maps for AGVs / AMRs in high-shelf dense environments and supports online safety obstacle avoidance and path planning. Background Technology

[0002] In the warehousing and logistics field, navigation maps are typically generated from a two-dimensional occupancy grid based on vehicle-mounted laser SLAM. The process includes: stitching together point clouds using laser radar, projecting the dense point cloud onto the ground, distinguishing obstacles from vacant areas based on reflection intensity, and then establishing an occupancy table with a fixed resolution. To ensure navigation safety, buffer zones are often constructed using morphological dilation or Euclidean distance transformation, suitable for real-time obstacle avoidance in low- to medium-density environments.

[0003] To address the vertical occlusion problem in high-shelf scenarios, some methods introduce octree voxelization, compressing 3D point clouds into multi-layer models and retaining only voxels within a specific height range for navigation. Some studies further use voxel FCN for semantic segmentation to distinguish structures such as shelves and walls, but these still primarily rely on single-scale planar images, making it difficult to balance accuracy and data efficiency.

[0004] Path planning often employs sampling trees or wavefront search, supplemented by skeleton extraction and cyclotron radius pruning to optimize narrow passage passages. Typically, the extracted skeleton path is converted into a simplified vector form (such as a polyline or curve connecting a series of key point coordinates) and then directly stored in the database. Subsequently, to support map version management and iterative updates, the system compares these vectorized skeletons with the new version of the map to determine which areas have changed and require updates or replanning of paths.

[0005] However, existing methods generally suffer from the following problems: loss of 3D structure projection, difficulty in balancing detail and efficiency with fixed grid scales, coarse buffer construction, and lack of flexibility in skeleton extraction. In complex environments, they are prone to misjudging narrow alleyways or wasting buffer zones, and are difficult to adjust in real time according to vehicle size or safety strategies. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems by providing a free space generation method for AGVs based on trimodal obstacle segmentation and conditional diffusion expansion. In warehouse environments with dense, multi-scale structures and high-bay racking, this method automatically constructs a continuous, spike-free safety buffer zone based on a single acquisition of a 3D point cloud and design blueprint. This buffer zone precisely conforms to complex obstacle geometry and can be dynamically adjusted according to the vehicle's outline and safety margin. Simultaneously, it generates a sparse and topologically complete free space skeleton. Furthermore, by expanding the buffer zone to form multi-level nested buffers, it achieves graded speed and path constraints for areas with different risk levels, supporting real-time planning and dynamic online updates for the mobile robot.

[0007] The technical solution of the present invention is as follows: A method for generating free space for AGVs based on trimodal obstacle segmentation and conditional diffusion dilation includes the following steps: Multi-resolution rasterization preprocessing: The architectural design blueprints and the point cloud data obtained by one-time scanning are aligned with the coordinates of the preset landmarks and unified to the ground reference system. The point cloud is then vertically cropped to form adaptive voxels and the aisles / shelves are marked. Three-channel feature extraction and geometric alignment: The point data is copied to the point channel, slice channel and top view channel respectively. Feature information is extracted under a unified coordinate reference. The three-channel features are projected onto the world grid and dynamically fused according to spatial continuity to form a voxel dataset with obstacle labels and local curvature information. Voxel patching and obstacle planarization: The voxel dataset is flattened into a two-dimensional top view in the height direction, and obstacle occupancy map, numbering map and curvature map are generated by weighted image processing methods; Conditional diffusion safety buffer generation: The outer dimensions and safety margin of the transport vehicle are mapped to a multi-channel distance map, which is then aligned with the obstacle occupancy map, curvature map, and numbering map and combined into a grid array with channel semantics. Initial noise is injected into the diffusion deformation network and the reverse sampling process is started. In each denoising loop, the channel attention module is called and the features are gated and scaled according to the radial-tangential tension calculated by the boundary estimator. The final step is threshold extraction to obtain the binary map of the safety buffer. Free space simplification and travel backbone extraction: The candidate free space is obtained by performing a difference operation on the binary map of the safety buffer and the world grid. The continuous free grid is scanned and recorded grid by grid to generate a travel line index table. Based on the minimum available passage length threshold of the vehicle, the minor travel lines are pruned, and the distance between the obstacles on both sides is verified according to the minimum turning radius. The free space backbone grid that has passed the verification is retained. Free space vectorization and traffic attributes: The backbone mesh is input into an eight-neighbor Markov boundary tracker to obtain the sequence of free connected block boundary vertices. The number of vertices is compressed using a piecewise variable Douglas thinning algorithm. The obstacle redundancy width on both sides of each segment is calculated and mapped to the achievable maximum speed and recommended steering angle to generate a polygon-attribute table. Furthermore, a hierarchical safety redundancy region with vehicle dynamic state and environmental conditions as variables is introduced, and nested buffers are generated through multi-level elliptical Minkowski expansion. Front-end interaction and result backtracking: Load the binary image of the safety buffer and the polygon-attribute table structure into the front-end rendering layer, provide the functions of adjusting the safety factor and dragging the boundary control points, listen for interaction events, write back the grid array with channel semantics and trigger fast backsampling to refresh the safety buffer, and update the result serialization file for the navigation module to call after user confirmation.

[0008] Furthermore, the generation of the conditional diffusion safety buffer includes the following steps: Length of the automated guided vehicle Outer width and surrounding safety margin Pixel-by-pixel mapping to a four-channel distance map ; curvature diagram Perform pixel-level alignment and overlay to form the fifth channel; then number the images. As the sixth channel, it is sequentially combined to obtain a grid array of multi-channel semantics. ; Top-down obstacle occupancy map Mapping to Consistent resolution, with position encoding added to each pixel. Then, zero-mean Gaussian noise is added to obtain the initial state array. ;Will Input the diffusion deformation network and initiate the backsampling process; In each denoising loop, the channel attention module is called to... Dynamically inject backbone features to enable the network to synchronously perceive the grooves and the actual size of the vehicle, and constrain the expansion range to not exceed the preset seat belt. Number each obstacle Calculate the minimum inscribed circle radius Then based on the linear mapping function Convert to backsampling steps: If If the channel is small and the passage is narrow, allocate more steps to refine the solution; if If the area is large and open, reduce the number of steps to solve it quickly; Map the end features of the denoising backbone Copy into two paths and and in A tension boundary estimator is added at the tail end to assess the boundary forces. For tension coefficients greater than the threshold Pixel-level local smoothing is performed, target dilation sampling is added to low-density corners, wall corners and column base shadows are automatically backfilled, and safety buffer zones are completed. In the final step of backsampling, a binary image of the safety buffer is extracted according to a fixed threshold. And perform a contour back scan along the boundary, so that It should fit snugly against the outer edges of shelves, walls, and column bases, without any gaps or sharp edges; Safety buffer Tension field and obstacle number Package and output.

[0009] Furthermore, the construction of the boundary estimator specifically includes: For candidate boundary pixels Along the normal element vector Constructing a three-cell radial neighborhood ,right Find the maximum gradient to obtain the radial tension; Along the tangential element vector The tangential tension is obtained by taking a symmetrical point and calculating the first-order difference. Overall curvature The two types of tension are weighted to form a tension coefficient. ;

[0010] in, and For experience weights; For projection onto pixels The curvature; Indicates along The three-grid neighborhood; in accordance with Generate gated scaling factor And on Scaling is applied to the corresponding location features:

[0011] in Tension threshold; Used to suppress acute angles; Used to magnify the indentation; Tension diagram after scale adjustment Back to The boundary orientation and buffer width of subsequent backsampling are updated synchronously to ensure that the generated safety buffer is close to the shelf and wall and free of punctures.

[0012] Furthermore, the free space simplification and route backbone extraction specifically include: Binary graph of the safety buffer With full-frame world grid Performing a difference operation inverts the occupancy state at the raster level, directly generating candidate free space. ; right Using a directional, frame-by-frame scanning method, along , , , The system scans sequentially in four fixed directions. During the scan, the start-point coordinates, end-point coordinates, and direction labels of each consecutive free grid segment are recorded to generate a travel line index table. ; Calculate the diagonal threshold of the automated guided vehicle (AGV) based on its overall length and width. ,in and These are the outer length and outer width of the automated guided vehicle, respectively. The minimum available aisle length for the automated guided vehicle (AGV); this threshold serves as the minimum aisle length through which the vehicle can pass; traverse all travel line index tables. If a certain line segment Length less than the diagonal threshold If so, mark it as a twig and delete it; Preserve travel route Calculate the distance between the two obstacles Map it to the turning safety factor ;like Less than the vehicle's minimum turning radius The travel line will be temporarily marked as pending deletion, and a final decision on whether to delete it will be made after global alternative pathway verification. The travel line checked by rotation is written back to free space to generate a lightweight occupied-free mesh. .

[0013] Furthermore, the free space vectorization and access attribute annotation specifically include: Will occupy - free grid The Markov boundary tracking module is input, and for each freely connected region, the topmost pixel is selected as the initial starting point from top to bottom, initiating an eight-neighbor traversal. During the traversal, the forward movement direction vector is recorded using a first-order directional Markov chain model. For candidate pixels within the eight neighbors of the current pixel, the consistency of its direction with the forward vector, the edge gradient magnitude, and the redundancy safety margin outside the current pixel are calculated respectively. Based on these three indicators, the transition probability is calculated comprehensively, and the candidate pixel with the highest transition probability is selected as the next boundary point. When the current boundary point coincides with the initial starting point, the boundary of the connected region is considered to have completed a closed loop. In areas with dense shelving or sharp corners, a progressive smoothing strategy is introduced to automatically remove minor jitter in the boundary path, outputting a boundary vertex sequence with good connectivity and no obvious jagged edges. ; Boundary vertex sequence Input the piecewise variable Douglas thinning algorithm: traverse the entire ring and calculate the curvature magnitude at each vertex. The polygon is divided into several line segments based on the peak curvature. Then, the average curvature of each segment is calculated. And according to the following formula, the adaptive threshold is mapped. For segments with low curvature, a large threshold is used for significant compression, while for segments with high curvature, a small threshold is used to preserve details. Then, Douglas iteration is performed independently on each segment, removing points within the threshold and retaining only segment endpoints and high-curvature inflection points. This ultimately generates a simplified polygon set with significantly fewer points and a smoother trajectory. Satisfying the vehicle's minimum turning constraint:

[0014] in, and These are the preset maximum and minimum thinning thresholds, respectively; and For set Maximum and minimum values; For polygon sets Calculate passage attributes line by line: Map the redundant width of obstacles on both sides to the maximum achievable speed. And estimate the recommended steering angle based on the local obstacle density. Finally, the geometric vertex sequence and attribute parameters are written into the polygon-attribute table. This is used for the rendering of the visual interface and for the scheduling system to call; Multi-level obstacle avoidance zone layer generation: Based on the calculated obstacle redundancy width on both sides of each segment and mapped to the achievable maximum speed and recommended steering angle, a layered safety redundancy zone is further introduced so that navigation planning can adopt differentiated constraints under different risk levels.

[0015] Furthermore, the multi-level obstacle avoidance area layer generation specifically includes: The layering is based on the vehicle's dynamic state and environmental conditions, forming a nested multi-level buffer set, and maintaining the corresponding numbering consistency with the polygon-attribute table.

[0016] To characterize the immediate impact of speed, braking, and system latency on safe distance, the vehicle's current dynamic parameters are set as follows:

[0017] in To perceive the total delay of control, The equivalent friction coefficient of the road surface. The acceleration is due to gravity; the corresponding reaction distance and braking distance are:

[0018]

[0019] Introduce basic buffers related to vehicle body dimensions and dynamic margin :

[0020]

[0021] in For the vehicle's external dimensions, The coefficients are determined empirically; this gives the stratification threshold:

[0022]

[0023]

[0024] And form a hierarchical definition:

[0025]

[0026]

[0027] in For point Distance to the boundary of the obstacle, It can be adjusted in real time by a safety factor or actual working strategy; anisotropic elliptical structural elements are used to perform Minkowski expansion on the polygonal boundaries; the lateral and longitudinal semi-axes of the ellipse are:

[0028] in This is the horizontal magnification factor; Calculate the distance transformation of the obstacle polygon and then... Threshold truncation directly yields nested sets. This set is a three-level binary buffer graph. ; Apply a hierarchical strategy to multi-level buffers: Firstly, Considered as a strongly constrained unreachable region.

[0029] Secondly, in Apply stronger constraints and speed suppression; Thirdly, in Apply mild constraints and gradually relax the speed limit; the speed limit is approximately:

[0030] in Nominal cruising speed; Further considering lateral comfort and curvature constraints, an upper limit constraint on curvature is added. Take the minimum value.

[0031] Furthermore, the voxel repair and obstacle planarization specifically include: Three-dimensional voxel set according to The axis is decomposed into columnar units, with barrier labels on the top and bottom layers of the same column. The connectivity status is detected; if connectivity between the upper and lower layers is detected but there is no connectivity in the middle section... If a fracture occurs, all elements within that height range are continuously filled in to reconstruct a complete column. Select voxel sets whose curvature coefficient exceeds a preset threshold. The voxels are categorized into specific sets. Subsequently, each voxel in the set undergoes an expansion aggregation operation within its three-dimensional neighborhood and propagates along the wall normal direction. Repaired voxel set along Axial compression leveling floor grid Accumulate the number of voxels for each grid cell and apply a logarithmic density weight. Weighted scaling is applied to enhance the details of the shelf base and column base, resulting in a high-fidelity planar grid. ; For planar grid Perform hole filling and boundary closure to generate a gap-free top-view obstacle occupancy map. Furthermore, each obstacle was assigned a number in sequence. And project the curvature to generate a curvature map. ,Will Output to the safety buffer expansion phase.

[0032] Furthermore, the aforementioned front-end interaction and result feedback specifically include: Polygon-attribute table V, safety buffer binary map and three-level binary buffer map Loaded into the Qt front-end rendering layer, colors are set for the three types of elements (obstacles, buffers, and free elements) in the binary image B of the safety buffer. Colors are also applied to the three-level binary buffer image. The border is displayed with different colored boundary lines; a slider is provided on the interface side for dynamically adjusting the safety factor. It allows users to drag and fine-tune the editable control points on the boundaries; after the event listener captures the above interaction, it writes the changes back to the grid array with channel semantics. Synchronously triggers the fast backsampling chain to refresh the buffer, and the visualization results are displayed in real time while maintaining a stable frame rate; When the user clicks "confirm," the backend service serializes the updated polygon set, attribute fields, and security parameters into XML and pushes it to the navigation module for use.

[0033] Furthermore, the multi-resolution rasterization preprocessing specifically includes: Design blueprint Contour point cloud obtained from a single survey Align with known landmarks to a unified ground coordinate system And construct a mesh with side length of based on the principal axis direction. equal-amplitude logic grid ; For the original point cloud Perform planar partitioning based on the grid boundaries to obtain a set of spatial blocks. ; height of the shelf bottom beam With the highest point of the automated guided vehicle body Write to the vertical safety plane , For each space block The inner point cloud is vertically clipped according to its height coordinates, retaining only those located at the top. and The effective point set between them forms the cross-sectional block related to passage. ; For each cross-section block Statistical point density Adaptive setting of voxel size based on density increment strategy :when When the value is above a threshold, refine the voxels to preserve the details of the shelf uprights. When the data volume is below a certain threshold, voxels are coarsened to compress the amount of data in empty channels; this is combined with blueprint channel masks. Determine the center position of the blocks and assign the function label "Channel Block" to the blocks located in the passageway. The remaining blocks are assigned to the shelf blocks. The final output is a point set with multi-resolution, multi-functional labels. and will and Write block-level metadata.

[0034] Furthermore, the three-channel feature extraction and geometric alignment specifically include: Output multi-resolution pixel blocks Based on function labels Choose the kernel scale: when the channel block Set the core side length When the shelf block Time setting ; Synchronous copy Point Channel Slice Channel With overhead passage Based on uniform voxel size Establish a three-channel coordinate reference system; Point-to-point channel The two-stage processing, feature self-attention and multi-scale window scanning unit, is performed sequentially. For slice channels Divide the origin block into sections along the direction of vehicle travel. The system employs sequential convolution with embedded gated recurrent units to process each sector, preserving the long-distance connectivity of the wall and railing and outputting sequential state features. ; View the channel points from above Input a fan-shaped sliding window encoding / decoding network to complete a full planar mapping in polar coordinates; for Cylindrical coordinates Rearrange and divide into sectors according to a fixed angle. Then, each sector is expanded into radial strips of uniform length using a radial straightening operator; Subsequently, multi-scale sliding windows were constructed along the radial direction of the strip, with a window length of... Increment is defined as:

[0035] in, Based on the radial sampling step size, For the first The radial dimensions of the fan-shaped blocks are used to balance close-range details and distant sparse points; Then extract the height distribution for each sliding window block. Reflectance density With radial curvature The local feature vectors are generated and fed into the encoder; the encoder aggregates these features through three layers of residual convolution. To highlight the long, narrow shape of the shelves against the smooth wall surface; The feature map is copied by one cell at the corner boundary and shares the convolution weights to ensure the continuity of the cross-sector structure; the decoding end adopts an upsampling structure symmetrical to the encoding end, and restores the spatial resolution through cross-layer jumpers; An angular attention gate is inserted after each level of decoding. The gate coefficient is adaptively calculated based on the correlation of edge features of adjacent sectors. The weight is relaxed for low-correlation areas to smooth the broken edges around the pillar. Finally, the reconstructed multi-scale features are merged into a top-down raster feature map. The diagram fully preserves the outline of the shelf strips, the concave column bases, and the landmark lines, providing a consistent global view for the geometric alignment process. In the geometric alignment gating layer , and Projected onto a unified world grid And based on the continuous temperature coefficient Dynamically weight cross-channel features; write barrier labels to each voxel. and local spatial curvature Generate the input library required for the expansion phase. This completes the fusion and alignment of the three-channel features and geometric information.

[0036] Compared with existing technologies, the advantages of this invention are: 1. Traditional methods typically project a 3D point cloud into a single-scale 2D element and then use a fixed dilation operation to generate a safety buffer. In contrast, this application utilizes a more intelligent multi-level voxel structure to automatically adjust the resolution based on the point cloud density and focuses only on the area between the bottom beam of the rack and the highest point of the vehicle. This controls the amount of data while preserving the details of the rack uprights and beams, making navigation more accurate in judging narrow passages and sharp turns. 2. Most current technologies use fixed distances or uniform expansion kernels to expand the obstacle area, which can lead to blind spots in sharp grooves and excessive contraction in open areas. This application introduces a conditional diffusion network and tension field to dynamically adjust the buffer zone width according to the curvature of the obstacle, the shape of the vehicle, and the preset safety margin, making the boundary more continuous and accurate, and can be continuously optimized through learning. 3. Regarding free space abstraction, existing methods often use heuristic skeleton pruning to directly write to the database, and the entire process needs to be rerun when the parameters are modified. This application first calculates the difference set of the safe zone, and then combines the run index and the flexible Douglas compression algorithm to quickly generate the backbone vector. It also quickly backsamples the original data, so that when the front-end interaction is modified, only the local data is refreshed, and there is no need to redo the point cloud stitching and feature extraction, which greatly improves efficiency and flexibility. 4. This application aligns 3D point clouds with design blueprints, then proceeds to multi-resolution voxelization, generates a safety buffer using conditional diffusion networks, and finally achieves backbone vectorization and real-time interactive updates. It can quickly construct accurate, coherent, and dynamically adjustable access maps in warehouses with dense shelving. This significantly shortens mapping and maintenance time for warehouse automation, reduces manual debugging and recalculation, and improves input-output efficiency. 5. This application can flexibly adapt to different vehicle sizes and safety requirements, making it convenient for hardware manufacturers to develop various vehicle models based on a unified map format; combined with cloud-based digital twins and scheduling systems, the real-time updated vector backbone can also support simulation and optimization, creating more commercial value; 6. This application has a wide range of applications, not only suitable for warehouses, but also applicable to complex environments such as manufacturing workshops, airport baggage sorting, hospital logistics, and parking lots. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the method described in this application.

[0038] Figure 2 This is a schematic diagram of the voxel repair and obstacle planarization process of the method in this application.

[0039] Figure 3 This is a schematic diagram illustrating the process of generating the conditional diffusion safety buffer for the method described in this application.

[0040] Figure 4 This is a schematic diagram of the free space simplification and process backbone extraction process of the method in this application.

[0041] Figure 5 This is a schematic diagram of the free space vectorization and access attribute process of the method in this application.

[0042] Figure 6 This is a schematic diagram of the front-end interaction and result backtracking process of the method in this application. Detailed Implementation

[0043] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0044] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0045] Please see Figure 1-6 A method for generating free space for AGVs based on trimodal obstacle segmentation and conditional diffusion dilatation, such as... Figure 1 As shown, it includes the following steps: Step 1: Multi-resolution rasterization preprocessing Step 1.1: Design blueprint Contour point cloud obtained from a single survey Align with known landmarks to a unified ground coordinate system And construct a mesh with side length of based on the principal axis direction. equal-amplitude logic grid ; For the original point cloud Perform planar partitioning based on the grid boundaries to obtain a set of spatial blocks. .

[0046] Step 1.2: Adjust the height of the bottom beam of the shelving. With the highest point of the automated guided vehicle body Write to the vertical safety plane , For each space block The inner point cloud is vertically clipped according to its height coordinates, retaining only those located at the top. and The effective point set between them forms the cross-sectional block related to passage. .

[0047] Step 1.3: For each cross-section block Statistical point density Adaptive setting of voxel size based on density increment strategy :when When the value is above a threshold, refine the voxels to preserve the details of the shelf uprights. When the data volume is below a certain threshold, voxels are coarsened to compress the amount of data in empty channels; this is combined with blueprint channel masks. Determine the center position of the blocks and assign the function label "Channel Block" to the blocks located in the passageway. The remaining blocks are assigned to the shelf blocks. The final output is a point set with multi-resolution, multi-functional labels. and will and Write block-level metadata.

[0048] Step 2: Three-channel feature extraction and geometric alignment Step 2.1: Convert the multi-resolution pixel blocks output from the previous stage into... Based on function labels Choose the kernel scale: when the channel block Set the core side length When the shelf block Time setting ; Synchronous copy Point Channel Slice Channel With overhead passage Based on uniform voxel size Establish a three-channel coordinate reference system.

[0049] Step 2.2: Point-to-point channel The two-stage processing of feature self-attention and multi-scale window scanning unit is performed sequentially.

[0050] The feature self-attention is as follows: inputting point blocks into a point feature attention network, and focusing on the center point... and its spherical neighborhood Calculate attention based on reflection intensity and orientation information And complete the aggregation of neighborhood features.

[0051] The multi-scale window scanning unit is constructed by inserting a multi-scale window scanning unit at the end of the network, and the unit is constructed according to the following steps: Step 2.2.1: Split the attention network output array into radial sub-channels in polar coordinates. With tangential sub-channel This ensures that the direction of subsequent displacement can be clearly mapped to spatial orientation; Step 2.2.2: Slide the three layers of windows sequentially on each sub-channel. ,in Voxel dimensions. For the center point of the window. Translate along the corresponding axis Sampling yields feature vectors ; Step 2.2.3: Combine the three-layer feature vectors cascade:

[0052] in Abrupt changes in the same radial or tangential reflection intensity can produce a high-amplitude response at edge corners; Step 2.2.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] After being concatenated with atomic channel features, the data is input into a convolutional layer and fused to output a multi-channel array. The array carries both reflection gradient and local normal information, providing high-resolution corner feature support for subsequent multimodal alignment.

[0053] Step 2.3: Slice the channel Divide the origin block into sections along the direction of vehicle travel. The system employs sequential convolution with embedded gated recurrent units to process each sector, preserving the long-distance connectivity of the wall and railing and outputting sequential state features. .

[0054] Step 2.4: Place the top-view channel blocks Input a fan-shaped sliding window encoder-decoder network to achieve a complete planar mapping in polar coordinates. Cylindrical coordinates Rearrange and divide into sectors according to a fixed angle. Then, each sector is expanded into radial strips of uniform length using a radial straightening operator.

[0055] Subsequently, multi-scale sliding windows were constructed along the radial direction of the strip, with a window length of... Increment is defined as

[0056] in, Based on the radial sampling step size, For the first The radial dimensions of the fan-shaped blocks are used to balance close-range details with distant sparse points.

[0057] Then extract the height distribution for each sliding window block. Reflectance density With radial curvature This forms local feature vectors, which are then fed into the encoding end. The encoding end aggregates these vectors through three layers of residual convolution. To highlight the long, narrow shape of the shelves against the smooth wall surface; The feature map is copied by one cell at the corner boundary and shares the convolution weights to ensure structural continuity across sectors. The decoder uses an upsampling structure symmetrical to the encoder and restores spatial resolution through cross-layer jumpers. An angular attention gate is inserted after each level of decoding. The gate coefficient is adaptively calculated based on the correlation of edge features of adjacent sectors. The weight is relaxed for low-correlation areas to smooth the broken edges around the pillar. Finally, the reconstructed multi-scale features are merged into a top-down raster feature map. The diagram fully preserves the outline of the shelf strips, the concave column bases, and the landmark lines, providing a consistent global view for the geometric alignment process.

[0058] Step 2.5: In the geometric alignment gating layer, , and Projected onto a unified world grid And based on the continuous temperature coefficient Dynamically weight cross-channel features; write barrier labels to each voxel. and local spatial curvature Generate the input library required for the expansion phase. This completes the fusion and alignment of the three-channel features and geometric information.

[0059] Step 3: Voxel Repair and Obstacle Planarization Step 3.1: Assemble the three-dimensional voxel set according to The axis is decomposed into columnar units, with barrier labels on the top and bottom layers of the same column. The connectivity status is detected; if connectivity between the upper and lower layers is detected but there is no connectivity in the middle section... If a fracture occurs, all elements within that height range are continuously replenished to reconstruct a complete column.

[0060] Step 3.2: Select voxel sets whose curvature coefficient exceeds a preset threshold. The voxels are categorized into specific sets. Subsequently, each voxel in the set undergoes an expansion aggregation operation within its three-dimensional neighborhood and advances along the wall normal direction to quickly fill sparse voids caused by materials such as soft door curtains, thereby ensuring the continuity between the wall and the side of the shelf.

[0061] Step 3.3: Assemble the repaired voxels along Axial compression leveling floor grid Accumulate the number of voxels for each grid cell and apply a logarithmic density weight. Weighted scaling is applied to enhance the details of the shelf base and column base, resulting in a high-fidelity planar grid. .

[0062] Step 3.4: For the planar grid Perform hole filling and boundary closure to generate a gap-free top-view obstacle occupancy map. Furthermore, each obstacle was assigned a number in sequence. And project the curvature to generate a curvature map. ,Will Output to the safety buffer expansion phase.

[0063] Step 4: Conditional Diffusion Safety Buffer Generation Step 4.1: Measure the overall length of the automated guided vehicle. Outer width and surrounding safety margin Pixel-by-pixel mapping to a four-channel distance map ; The curvature diagram given in the previous stage Perform pixel-level alignment and overlay to form the fifth channel; then add an obstacle number map. As the sixth channel, it is sequentially combined to obtain a grid array of multi-channel semantics. This allows for the simultaneous depiction of vehicle volume constraints and local boundary irregularities during the subsequent diffusion stage.

[0064] Step 4.2: Map the obstacle occupancy from above. Mapping to Consistent resolution, with position encoding added to each pixel. Then, zero-mean Gaussian noise is added to obtain the initial state array. ;Will Input the diffusion deformation network and start the inverse sampling process.

[0065] Step 4.3: In each denoising loop, call the channel attention module to... Dynamically injecting backbone features enables the network to synchronously perceive the gaps and grooves and the actual size of the vehicle, thus constraining the expansion range to not exceed the preset seat belt.

[0066] Step 4.4: Number each obstacle Calculate the minimum inscribed circle radius Then based on the linear mapping function Convert to backsampling steps: If If the channel is small and the passage is narrow, allocate more steps to refine the solution; if If the area is large and open, reduce the number of steps to solve it quickly.

[0067] Step 4.5: Map the terminal features of the denoising backbone Copy into two paths and and in A tension boundary estimator is added at the tail end to evaluate the boundary forces. This boundary estimator is a first-time invention in this patent, and its specific construction process is as follows: ① For candidate boundary pixels Along the normal element vector Constructing a three-cell radial neighborhood ,right Find the maximum gradient to obtain the radial tension; ② Along the tangential element vector The tangential tension is obtained by taking a symmetrical point and calculating the first-order difference. ③Comprehensive curvature The two types of tension are weighted to form a tension coefficient. .

[0068]

[0069] In the formula, and For experience weights; For step 3.4, project onto pixels The curvature; Indicates along The three-grid neighborhood.

[0070] ④ Basis Generate gated scaling factor And on Scaling is applied to the corresponding location features:

[0071] in Tension threshold; Used to suppress acute angles; Used to magnify indentations.

[0072] ⑤ Tension diagram after scale adjustment Back to The boundary orientation and buffer width of subsequent backsampling are updated synchronously to ensure that the generated safety buffer is close to the shelf and wall and free of punctures.

[0073] Step 4.6: For tension coefficients greater than the threshold The pixel performs local smoothing, adds target dilatation sampling to low-density corners, automatically backfills dead corners and column base shadows, and completes the safety buffer zone.

[0074] Step 4.7: Extract the binary image of the safety buffer at the final step of backsampling using a fixed threshold. And perform a contour back scan along the boundary, so that It should fit snugly against the outer edges of shelves, walls, and column bases, without any gaps or sharp edges.

[0075] Step 4.8: Set the safety buffer Tension field and obstacle number Package and output.

[0076] Step 5: Simplifying Free Space and Extracting the Backbone of the Journey Step 5.1: Convert the binary image of the safety buffer. With full-frame overhead grid Performing a difference operation inverts the occupancy state at the raster level, directly generating candidate free space. The obstacle number association information from step 3.4 is retained in the difference set so that subsequent sparsification will not destroy the verified security boundary.

[0077] Step 5.2: For Using a directional, frame-by-frame scanning method, along , , , The system scans sequentially in four fixed directions. During the scan, the start-point coordinates, end-point coordinates, and direction labels of each consecutive free grid segment are recorded to generate a travel line index table. .

[0078] Step 5.3: First, based on the vehicle's overall length and width defined in Step 4.1, calculate the vehicle's diagonal threshold. ,in and These are the vehicle's outer length and outer width as defined in step 4.1. This is the minimum available lane length for vehicles. This threshold serves as the minimum lane length that a vehicle can pass through. Then, all travel line index tables are traversed. If a certain line segment Length less than the diagonal threshold If a branch is found to be a short branch, it is marked as a minor branch and deleted. This operation preserves only the backbone in the center of the free space, maintaining the topological connectivity of the pathways while eliminating invalid short branches, thus simplifying the structure of the free space.

[0079] Step 5.4: Preserve the travel lines Calculate the distance between the two obstacles Map it to the turning safety factor ;like Less than the vehicle's minimum turning radius The travel line is temporarily marked as pending deletion. After global alternative pathways are verified, the travel line is finally determined whether to delete it, in order to avoid excessive pruning that could lead to a broken link.

[0080] Step 5.5: Write the travel line that has passed the rotation check back to free space to generate a lightweight occupied-free mesh. The grid significantly reduces the number of lattice cells while maintaining the original topology, providing a sparse and clear base map for subsequent vectorization and access attribute annotation.

[0081] Step 6: Free Space Vectorization and Passage Attribute Labeling Step 6.1: Occupied-Free Grid The Markov boundary tracking module is input, and for each freely connected region, the topmost pixel is selected as the initial starting point from top to bottom, initiating an eight-neighbor traversal. During the traversal, a first-order directional Markov chain model is introduced to record the forward movement direction vector. For candidate pixels within the eight neighbors of the current pixel, the consistency of its direction with the forward vector, the edge gradient magnitude, and the redundancy safety margin outside the current pixel are calculated respectively. Based on the above three indicators, the transition probability is calculated comprehensively, and the candidate pixel with the highest transition probability is selected as the next boundary point. When the current boundary point coincides with the initial starting point, the boundary of the connected region is considered to have completed a loop closure. To improve the continuity and stability of the boundary extraction results, a progressive smoothing strategy is introduced in areas with dense shelving or sharp corners to automatically remove minor jitter in the boundary path, outputting a boundary vertex sequence with good connectivity and no obvious jagged edges. .

[0082] Step 6.2: The sequence Input the piecewise variable Douglas thinning algorithm: traverse the entire ring and calculate the curvature magnitude at each vertex. The polygon is divided into several line segments based on the peak curvature. Then, the average curvature of each segment is calculated. And according to the following formula, the adaptive threshold is mapped. For segments with low curvature, a large threshold is used for significant compression, while for segments with high curvature, a small threshold is used to preserve details. Then, Douglas iteration is performed independently on each segment, removing points within the threshold and retaining only segment endpoints and high-curvature inflection points. This ultimately generates a simplified polygon set with significantly fewer points and a smoother trajectory. Satisfying the vehicle's minimum turning constraint:

[0083] In the formula, and These are the preset maximum and minimum thinning thresholds, respectively; and For set The maximum and minimum values.

[0084] Step 6.3: For the set of polygons Calculate passage attributes line by line: Map the redundant width of obstacles on both sides to the maximum achievable speed. And estimate the recommended steering angle based on the local obstacle density. Finally, the geometric vertex sequence and attribute parameters are written into the polygon-attribute table. It is used for visual interface rendering and scheduling system calls.

[0085] Step 6.4: Generation of multi-level obstacle avoidance zones Based on the calculation of the obstacle redundancy width on both sides of each segment in step 6.3 and its mapping to the achievable maximum speed and recommended steering angle, a layered safety redundancy area is further introduced so that navigation planning can adopt differentiated constraints under different risk levels.

[0086] The layering is based on the vehicle's dynamic state and environmental conditions, forming a nested multi-level buffer set, and maintaining the corresponding numbering consistency with the polygon-attribute table.

[0087] To characterize the immediate impact of speed, braking, and system delay on safe distance, the vehicle's current dynamic parameters are defined as follows:

[0088] in To perceive the total delay of control, The equivalent friction coefficient of the road surface. This is the acceleration due to gravity. The corresponding reaction distance and braking distance are:

[0089]

[0090] Introduce basic buffers related to vehicle body dimensions and dynamic margin :

[0091]

[0092] in The vehicle's external dimensions are from step 4.1. The coefficients are determined empirically. This leads to the stratification threshold:

[0093]

[0094]

[0095] And form a hierarchical definition:

[0096]

[0097]

[0098] in For point Distance to the boundary of the obstacle, It can be adjusted in real time by a safety factor or actual working strategy. To balance the direction of travel and lateral margin, anisotropic elliptical structural elements are used to perform Minkowski dilation on the polygonal boundaries. During calculation, boundary points can be selectively selected to reduce computational load while preserving key geometric features. The lateral and longitudinal semi-axes of the ellipse are:

[0099] in This is the horizontal magnification factor.

[0100] Calculate the distance transformation for the obstacle polygon in step 6.3 and then... Threshold truncation directly yields nested sets. This set is a three-level binary buffer graph. .

[0101] Apply a hierarchical strategy to multi-level buffers: Firstly, Considered as a strongly constrained unreachable region.

[0102] Secondly, in Apply stronger constraints and speed suppression; Thirdly, in Apply mild constraints and gradually relax the speed limit. The speed limit is approximately:

[0103] in This is the nominal cruising speed.

[0104] Further considering lateral comfort and curvature constraints, an upper limit constraint on curvature is added. Take the minimum value.

[0105] Step 7: Combine the polygon-attribute table V and the binary image of the safety buffer. and three-level binary buffer map Loaded into the Qt front-end rendering layer, colors are set for the three types of elements (obstacles, buffers, and free elements) in the binary image B of the safety buffer. Colors are also applied to the three-level binary buffer image. The border is displayed with different colored boundary lines; a slider is provided on the interface side for dynamically adjusting the safety factor. It allows users to drag and fine-tune the editable control points on the boundaries; after the event listener captures the above interaction, it writes the changes back to the grid array with channel semantics. It synchronously triggers a fast backsampling chain to refresh the buffer, and the visualization results are displayed in real time while maintaining a stable frame rate.

[0106] When the user clicks "confirm," the backend service serializes the updated polygon set, attribute fields, and security parameters into XML and pushes it to the navigation module for use.

[0107] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for generating free space for AGVs based on three-modal obstacle segmentation and conditional diffusion dilation, characterized in that, Includes the following steps: Multi-resolution rasterization preprocessing: The architectural design blueprints and the point cloud data obtained by one-time scanning are aligned with the coordinates of the preset landmarks and unified to the ground reference system. The point cloud is then vertically cropped to form adaptive voxels and the aisles / shelves are marked. Three-channel feature extraction and geometric alignment: The point data is copied to the point channel, slice channel and top view channel respectively. Feature information is extracted under a unified coordinate reference. The three-channel features are projected onto the world grid and dynamically fused according to spatial continuity to form a voxel dataset with obstacle labels and local curvature information. Voxel patching and obstacle planarization: The voxel dataset is flattened into a two-dimensional top view in the height direction, and obstacle occupancy map, numbering map and curvature map are generated by weighted image processing methods; Conditional diffusion safety buffer generation: The outer dimensions and safety margin of the transport vehicle are mapped to a multi-channel distance map, which is then aligned with the obstacle occupancy map, curvature map, and numbering map and combined into a grid array with channel semantics. Initial noise is injected into the diffusion deformation network and the reverse sampling process is started. In each denoising loop, the channel attention module is called and the features are gated and scaled according to the radial-tangential tension calculated by the boundary estimator. The final step is threshold extraction to obtain the binary map of the safety buffer. Free space simplification and travel backbone extraction: The candidate free space is obtained by performing a difference operation on the binary map of the safety buffer and the world grid. The continuous free grid is scanned and recorded grid by grid to generate a travel line index table. Based on the minimum available passage length threshold of the vehicle, the minor travel lines are pruned, and the distance between the obstacles on both sides is verified according to the minimum turning radius. The free space backbone grid that has passed the verification is retained. Free space vectorization and passage attributes: Input the backbone mesh into the eight-neighbor Markov boundary tracker to obtain the sequence of free connected block boundary vertices, use the piecewise variable Douglas thinning algorithm to compress the number of vertices, calculate the obstacle redundancy width on both sides of each segment and map it to the achievable maximum speed and recommended steering angle, and generate a polygon-attribute table. Furthermore, a hierarchical safety redundancy region is introduced with vehicle dynamic state and environmental conditions as variables, and nested buffers are generated through multi-level elliptical Minkowski expansion. Front-end interaction and result backtracking: Load the binary image of the safety buffer and the polygon-attribute table structure into the front-end rendering layer, provide the functions of adjusting the safety factor and dragging the boundary control points, listen for interaction events, write back the grid array with channel semantics and trigger fast backsampling to refresh the safety buffer, and update the result serialization file for the navigation module to call after user confirmation.

2. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 1, characterized in that, The generation of the conditional diffusion safety buffer includes the following steps: Automated Guide Cart Outer Length Outer width and surrounding safety margin Pixel-by-pixel mapping to a four-channel distance map ; curvature diagram Pixel-level alignment is performed and the images are superimposed to form the fifth channel; then, the numbered images are used as the sixth channel, and the images are combined sequentially to obtain a multi-channel semantic grid array. ; Top-down obstacle occupancy map Mapping to Consistent resolution, with position encoding added to each pixel. Then, zero-mean Gaussian noise is added to obtain the initial state array. ;Will Input the diffusion deformation network and initiate the backsampling process; In each denoising loop, the channel attention module is called to... Dynamically inject backbone features to enable the network to synchronously perceive the grooves and the actual size of the vehicle, and constrain the expansion range to not exceed the preset seat belt. Number each obstacle Calculate the minimum inscribed circle radius Then based on the linear mapping function Convert to backsampling steps: If If the channel is small and the passage is narrow, allocate more steps to refine the solution; if If the area is large and open, reduce the number of steps to solve it quickly; Map the end features of the denoising backbone Copy into two paths and and in A tension boundary estimator is added at the tail end to assess the boundary forces. For tension coefficients greater than the threshold Pixel-level local smoothing is performed, target dilation sampling is added to low-density corners, wall corners and column base shadows are automatically backfilled, and safety buffer zones are completed. In the final step of backsampling, a binary image of the safety buffer is extracted according to a fixed threshold. And perform a contour back scan along the boundary, so that It should fit snugly against the outer edges of shelves, walls, and column bases, without any gaps or sharp edges; Binary graph of the safety buffer Tension field and obstacle numbering Package and output.

3. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 2, characterized in that, The construction of the boundary estimator specifically includes: For candidate boundary pixels Along the normal element vector Constructing a three-cell radial neighborhood ,right Find the maximum gradient to obtain the radial tension; Along the tangential element vector The tangential tension is obtained by taking a symmetrical point and calculating the first-order difference. Overall curvature The two types of tension are weighted to form a tension coefficient. ; in, and For experience weights; For overall curvature; Represents a three-cell radial neighborhood; in accordance with Generate gated scaling factor And on Scaling is applied to the corresponding location features: in Tension threshold; Used to suppress acute angles; Used to magnify the indentation; Tension diagram after scale adjustment Back to The boundary orientation and buffer width of subsequent backsampling are updated synchronously to ensure that the generated safety buffer is close to the shelf and wall and free of punctures.

4. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 3, characterized in that, The free space simplification and route backbone extraction specifically include: Binary graph of the safety buffer With full-frame world grid Performing a difference operation inverts the occupancy state at the raster level, directly generating candidate free space. ; right Using a directional, frame-by-frame scanning method, along , , , The system scans sequentially in four fixed directions. During the scan, the start-point coordinates, end-point coordinates, and direction labels of each consecutive free grid segment are recorded to generate a travel line index table. ; Calculate the diagonal threshold of the automated guided vehicle (AGV) based on its overall length and width. ,in and These are the outer length and outer width of the automated guided vehicle, respectively. The minimum available aisle length for the automated guided vehicle (AGV); this threshold serves as the minimum aisle length through which the vehicle can pass; traverse all travel line index tables. In the line segment, if a certain travel line Length less than the diagonal threshold If so, mark it as a twig and delete it; Preserve travel route Calculate the distance between the two obstacles Map it to the turning safety factor ;like Less than the vehicle's minimum turning radius The travel line will be temporarily marked as pending deletion, and a final decision on whether to delete it will be made after global alternative pathway verification. The travel line checked by rotation is written back to free space to generate a lightweight occupied-free mesh. .

5. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 4, characterized in that, The free space vectorization and access attribute annotation specifically include: Will occupy - free grid The Markov boundary tracking module is input, and for each freely connected region, the topmost pixel is selected as the initial starting point from top to bottom, initiating an eight-neighbor traversal. During the traversal, the forward movement direction vector is recorded using a first-order directional Markov chain model. For candidate pixels within the eight neighbors of the current pixel, the consistency of its direction with the forward vector, the edge gradient magnitude, and the redundancy safety margin outside the current pixel are calculated respectively. Based on these three indicators, the transition probability is calculated comprehensively, and the candidate pixel with the highest transition probability is selected as the next boundary point. When the current boundary point coincides with the initial starting point, it is determined that the boundary of each freely connected region has completed a closed loop. In areas with dense shelving or sharp corners, a progressive smoothing strategy is introduced to automatically remove minor jitter in the boundary path, outputting a boundary vertex sequence with good connectivity and no obvious jagged edges. ; Boundary vertex sequence Input the piecewise variable Douglas thinning algorithm: traverse the entire ring and calculate the curvature magnitude at each vertex. The polygon is divided into several line segments based on the peak curvature. Then, the average curvature of each segment is calculated. And according to the following formula, the adaptive threshold is mapped. For segments with low curvature, a large threshold is used for significant compression, while for segments with high curvature, a small threshold is used to preserve details. Then, Douglas iteration is performed independently on each segment, removing points within the threshold and retaining only segment endpoints and high-curvature inflection points. This ultimately generates a simplified polygon set with significantly fewer points and a smoother trajectory. Satisfying the vehicle's minimum turning constraint: in, and These are the preset maximum and minimum thinning thresholds, respectively; and For set Maximum and minimum values; For simplified polygon sets Calculate passage attributes line by line: Map the redundant width of obstacles on both sides to the maximum achievable speed. And estimate the recommended steering angle based on the local obstacle density. Finally, the geometric vertex sequence and attribute parameters are written into the polygon-attribute table. This is used for the rendering of the visual interface and for the scheduling system to call; Multi-level obstacle avoidance zone layer generation: Based on the calculated obstacle redundancy width on both sides of each segment and mapped to the achievable maximum speed and recommended steering angle, a layered safety redundancy zone is further introduced so that navigation planning can adopt differentiated constraints under different risk levels.

6. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 5, characterized in that, The multi-level obstacle avoidance area layering generation specifically includes: The layering is based on the vehicle's dynamic state and environmental conditions, forming a nested multi-level buffer set, and maintaining the corresponding numbering consistency with the polygon-attribute table. To characterize the immediate impact of speed, braking, and system latency on safe distance, the vehicle's current dynamic parameters are set as follows: in To perceive the total delay of control, The equivalent friction coefficient of the road surface. The acceleration is due to gravity; the corresponding reaction distance and braking distance are: Introduce basic buffers related to vehicle body dimensions and dynamic margin : in To automatically transport the outer length and width, The coefficients are determined empirically; this gives the stratification threshold: And form a hierarchical definition: in For point Distance to the boundary of the obstacle, It can be adjusted in real time by a safety factor or actual working strategy; anisotropic elliptical structural elements are used to perform Minkowski expansion on the polygonal boundaries; the lateral and longitudinal semi-axes of the ellipse are: in This is the horizontal magnification factor; Calculate the distance transformation of the obstacle polygon and then... Threshold truncation directly yields nested sets. This set is a three-level binary buffer graph. ; Apply a hierarchical strategy to multi-level buffers: Firstly, Treating it as an unreachable region with strong constraints triggers emergency braking / forced replanning; Secondly, in Apply stronger constraints and speed suppression; Thirdly, in Apply mild constraints and gradually relax the speed limit; the speed limit is approximately: in Nominal cruising speed; Further considering lateral comfort and curvature constraints, an upper limit constraint on curvature is added. Take the minimum value.

7. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilatation according to claim 1, characterized in that, The voxel repair and obstacle planarization specifically include: Three-dimensional voxel set according to The axis is decomposed into columnar units, with barrier labels on the top and bottom layers of the same column. The connectivity status is detected; if connectivity between the upper and lower layers is detected but there is no connectivity in the middle section... If the columnar element is fractured, then all elements are continuously filled within the fracture height range of the columnar unit to reconstruct a complete columnar shape. Select voxel sets whose curvature coefficient exceeds a preset threshold. The voxels are grouped into sets. Subsequently, each voxel in the set undergoes an expansion aggregation operation within its three-dimensional neighborhood and propagates along the wall normal direction. Repaired voxel set along Axial compression leveling floor grid Accumulate the number of voxels for each grid cell and apply a logarithmic density weight. Weighted scaling is applied to enhance the details of the shelf base and column base, resulting in a high-fidelity planar grid. ; For planar grid Perform hole filling and boundary closure to generate a gap-free top-view obstacle occupancy map. Furthermore, each obstacle was assigned a number in sequence. And project the curvature to generate a curvature map. ,Will Output to the safety buffer expansion phase.

8. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 1, characterized in that, The aforementioned front-end interaction and result feedback specifically include: Polygon-attribute table V, safety buffer binary map and three-level binary buffer map Loaded into the Qt front-end rendering layer, the binary image in the safe buffer is... Colors are assigned to three categories of elements: obstacles, buffers, and freedom, for a three-level binary buffer map. The border is displayed with different colored boundary lines; a slider is provided on the interface side for dynamically adjusting the safety factor. It allows users to drag and fine-tune the editable control points on the boundaries; after the event listener captures the above interaction, it writes the changes back to the grid array with channel semantics. Synchronously triggers the fast backsampling chain to refresh the buffer, and the visualization results are displayed in real time while maintaining a stable frame rate; When the user clicks "confirm," the backend service serializes the updated polygon set, attribute fields, and security parameters into XML and pushes it to the navigation module for use.

9. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 1, characterized in that, The multi-resolution rasterization preprocessing specifically includes: Design blueprint Contour point cloud obtained from a single survey Align with known landmarks to a unified ground coordinate system And construct a mesh with side length of based on the principal axis direction. equal-amplitude logic grid ; For the original point cloud Perform planar partitioning based on the grid boundaries to obtain a set of spatial blocks. ; height of the shelf bottom beam With the highest point of the automated guided vehicle body Write to the vertical safety plane , For each space block The inner point cloud is vertically clipped according to its height coordinates, retaining only those located at the top. and The effective point set between them forms a multi-resolution point patch related to passage. ; For each multi-resolution pixel block Statistical point density Adaptive setting of voxel size based on density increment strategy :when When the value is above a threshold, refine the voxels to preserve the details of the shelf uprights. When the data volume is below a certain threshold, voxels are coarsened to compress the amount of data in empty channels; this is combined with blueprint channel masks. Determine the center position of the blocks and assign function labels to the blocks located in the passageway. Channel blocks, the remaining blocks are assigned function labels. Shelf blocks; the final output is a set of points with multi-resolution, multi-functional labels. and will and Write block-level metadata.

10. The AGV free space generation method based on three-modal obstacle segmentation and conditional diffusion dilation according to claim 9, characterized in that, The three-channel feature extraction and geometric alignment specifically include: Output multi-resolution pixel blocks Based on function labels Select the kernel scale: when the function label When setting the core side length of the channel block When the function label Shelf block time setting ; Synchronous copy Point Channel Slice Channel With overhead passage Based on uniform voxel size Establish a three-channel coordinate reference system; Point-to-point channel The two-stage processing, feature self-attention and multi-scale window scanning unit, is performed sequentially. For slice channels The origin block is divided into fixed-angle sectors along the vehicle's direction of travel. Sequential convolution with embedded gated recurrent units is used to process each sector, preserving the long-distance connectivity of the wall and guardrail, and outputting the sequence state features. ; View the channel points from above Input a fan-shaped sliding window encoding / decoding network to complete a full planar mapping in polar coordinates; for Cylindrical coordinates Rearrange and divide into sectors according to a fixed angle. Then, each sector is expanded into radial strips of uniform length using a radial straightening operator; Subsequently, multi-scale sliding windows were constructed along the radial direction of the strip, with a window length of... Increment is defined as: in, Based on the radial sampling step size, For the first The radial dimensions of the fan-shaped blocks are used to balance close-range details and distant sparse points; Then extract the height distribution for each sliding window block. Reflectance density With radial curvature The local feature vectors are generated and fed into the encoder; the encoder aggregates these features through three layers of residual convolution. To highlight the long, narrow shape of the shelves against the smooth wall surface; The feature map is copied by one cell at the corner boundary and shares the convolution weights to ensure the continuity of the cross-sector structure; the decoding end adopts an upsampling structure symmetrical to the encoding end, and restores the spatial resolution through cross-layer jumpers; An angular attention gate is inserted after each level of decoding. The gate coefficient is adaptively calculated based on the correlation of edge features of adjacent sectors. The weight is relaxed for low-correlation areas to smooth the broken edges around the pillar. Finally, the reconstructed multi-scale features are merged into a top-down raster feature map. The diagram fully preserves the outline of the shelf strips, the concave column bases, and the landmark lines, providing a consistent global view for the geometric alignment process. In the geometric alignment gating layer , and Projected onto a unified world grid And based on the continuous temperature coefficient Dynamically weight cross-channel features; write barrier labels to each voxel. and local spatial curvature Generate the input library required for the expansion phase. This completes the fusion and alignment of the three-channel features and geometric information.

Images