A roof landing evaluation method and system considering multiple aircraft types

Abstract

This invention discloses a roof landing assessment method and system that considers multiple aircraft types. Specifically, it utilizes multi-source data, including lidar point clouds and imagery / raster data, to extract candidate landing areas, estimate physically interpretable parameters, and conduct conditional safety assessments for multiple aircraft types. This invention enables interpretable and scalable multi-aircraft landing safety assessments under large-scale scenarios and complex roof conditions.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional spatial information processing and low-altitude flight safety assurance, and particularly relates to a roof landing assessment method that takes into account multiple aircraft types. Background Technology

[0002] With the rapid development of the low-altitude economy, the safety assessment of "candidate landing areas" has become a critical foundational capability when low-altitude aircraft / UAVs perform take-off and landing / forced landings on rooftops of unknown or semi-unknown buildings, or when rapidly selecting landing sites at the beginning of a mission. Rooftop landing safety is influenced not only by geometric factors (such as slope, undulation, flatness, obstacles, and available space dimensions), but also by surface materials and microstructures. In large-scale scene information acquisition, LiDAR and image sensors can provide three-dimensional geometric information from point clouds and color and texture information from images, providing a rich data foundation for rooftop site interpretation and risk identification. Point clouds are beneficial for extracting geometric features such as flatness and slope, while images are beneficial for extracting texture and material clues; the two complement each other and can be used to establish more reliable landing condition assessment models. However, in practical applications, it is often difficult to directly obtain the key physical quantities and engineering constraint parameters of candidate landing sites, such as the available landing space dimensions (e.g., the maximum inscribed circle / maximum circumscribed circle), load capacity / load margin, tilt angle and tipping threshold, and slip risk. This results in a lack of supervised physical quantity sources for "physically assisted deep learning modeling," while purely data-driven methods are prone to problems such as physical inconsistencies, failure of cross-domain information extraction and fusion.

[0003] Low-altitude flight mission systems typically require parallel comprehensive evaluation of multiple aircraft specifications at the same site. Different aircraft models vary significantly in weight, size, support contact area, center of gravity height, and support radius. Without interpretable physical feature learning and constraints, even if deep learning networks achieve high accuracy on offline data, they may not truly learn the correct "landability physical characteristics." This could lead to safety risks during engineering deployments, such as insufficient load-bearing capacity, inadequate landing radius, or excessive slope causing tipping. Therefore, given only candidate region point clouds and texture images, and lacking explicit physical quantities such as load-bearing capacity, space size, and slope, how to inject physical priors into deep learning networks and use aircraft specifications and site conditions as queries to achieve conditional evaluation of multiple aircraft models, thereby enabling landing point densification and alternative landing point selection, becomes a crucial fundamental issue for ensuring the economical and safe operation of low-altitude flight. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problems existing in the prior art, the present invention provides a roof landing assessment method that takes into account multiple aircraft models.

[0005] Technical solution: This invention provides a method for dynamic energy allocation in remote sensing satellite mission planning, the method comprising the following steps:

[0006] S1, Extraction of candidate landing area features and landing physical indicators from lidar point cloud and imagery.

[0007] S2, multi-category aircraft constraints;

[0008] S3 evaluates the site landing conditions comprehensively based on site conditions and aircraft.

[0009] Extraction of candidate landing area features and landing physics indicators from lidar point clouds and imagery:

[0010] First, local point cloud and texture images of the candidate landing areas on the roof are obtained based on point cloud and image segmentation and target extraction methods. To address the difficulty in outputting a fixed-shape tensor due to the "asymmetry, inconsistent size, and irregular shape" of the candidate regions, point cloud height values ​​are mapped to image pixels and aligned with color information to form a fused image input. Feature maps are then extracted using a pyramid backbone network. Next, a bounding rectangle is selected for each candidate region on the feature map, and masked region of interest (ROI) alignment is used to bilinearly sample the region features to a fixed grid size. The mask alignment results are used to weight the features within the ROI to suppress background noise, resulting in a candidate region fusion tensor for batch training and unified inference. Simultaneously, a global token representing the candidate region is obtained through masked weighted aggregation for subsequent inter-site communication and aircraft type lookup.

[0011] Simultaneously, the original 3D point cloud of the candidate region is voxelized and input into a 3D sparse convolutional backbone network to extract and aggregate geometric tokens. To compensate for the adverse effects of point cloud holes and missing features on the extraction of geometric features such as steps and undulations, the sparse voxel features are densified into a regular 3D raster feature map, and geometric completion is achieved by outputting the occupancy probability / density through a 3D convolutional head. Based on the completed point cloud, local plane fitting is performed to obtain geometric quantities such as normal vectors. Furthermore, a physical index extraction module is constructed: a spatial weight map is learned based on the regional features obtained by Mask ROIAlign and weighted converged to predict the load-bearing related proxy physical parameters / load capacity of the candidate region; feasible planar regions are obtained by clustering based on normal vector consistency and then projected to 2D to construct a binary raster mask of the feasible region and obtain the maximum inscribed circle radius as a landing size index through Euclidean distance transformation; the tilt degree / tilt angle of the candidate region is calculated based on the normal vector to form the tilt-related physical quantities. The above-mentioned physical quantities and multimodal features together constitute the candidate site condition representation, providing input for interpretable evaluation.

[0012] To address the need for multiple aircraft to operate in parallel in low-altitude environments, an aircraft category and structural parameter library is established. Basic attributes such as mass and size of different aircraft models are input as a scalar set, and can be expanded to include landing gear / support parameters related to ground contact stability (such as total contact area of ​​the footpads, support radius, and center of gravity height). These aircraft parameters are encoded into aircraft constraint query vectors (Query) of the same dimension as the site token through embedding / linear mapping or lightweight MLP. This is used for conditional retrieval of site information, enabling the model to automatically focus on key areas and factors such as "insufficient size, insufficient load capacity, and tipping risk" for specific aircraft models, forming a consistent evaluation entry point for multiple aircraft models.

[0013] The site token obtained in step S1 (containing point cloud features, image features, and physical indicators such as load capacity / maximum inscribed circle radius / tilt angle, or their proxy parameters) is used as the Key / Value pair. The aircraft type constraint query vector obtained in step S2 is used as the Query. These are input into a cross-attention Transformer for site-aircraft interaction, outputting a fused representation for each aircraft type. Based on this fused representation and the physical margins (e.g., load capacity margin, dimensional margin, tilt margin) output by the rule physics module, the decision head is input together to obtain a comprehensive evaluation result of "aircraft type-site suitability score / feasibility probability / risk level," which can be used for candidate region ranking, landing point encryption, and alternative landing point selection. To improve reliability and stability, physical consistency constraints and ranking supervision can be introduced during training to ensure that the score for high-quality sites is significantly higher than that for low-quality sites. The magnitude of the residual correction term is also limited to prevent the model from deviating excessively from explicit physical rules. This enables interpretable and scalable multi-aircraft landing safety assessment under large-scale scenarios and complex roof conditions.

[0014] Beneficial Effects: This invention, while maintaining information integrity, encodes point cloud and image features of irregular regions into tensors of a fixed number and dimension to support efficient multi-batch training and unified model inference. It also suppresses background noise introduced by the bounding rectangle through methods such as masking and region of interest alignment. This ensures that the fused features not only represent the typical geometric and texture attributes of candidate regions but also serve subsequent physical quantity mining and risk assessment. On the point cloud side, due to the sparsity of the original 3D point cloud and the potential for holes and missing data due to reflectivity differences, directly relying on incomplete point clouds and images is insufficient to fully characterize key geometric features such as steps and local undulations. Therefore, it is necessary to utilize methods such as sparse convolution to efficiently learn and aggregate features from the 3D voxel representation, and, when necessary, densify and complete the sparse voxel features to support spatial dimension calculations such as normal vector estimation, local plane fitting, feasible region extraction, and the calculation of spatial dimensions such as the maximum inscribed circle. Building upon this foundation, key physical quantities for landing safety assessment (such as load capacity, landing radius / size, tilt angle, and tipping constraints) should be modeled in an interpretable manner. On one hand, a physical head predicts "proxy physical parameters" from fused features that can be used for engineering judgment, and calculates safety margins for tipping, load capacity, and slippage accordingly. On the other hand, to compensate for complex environmental factors that rule-based models cannot fully cover, residual correction and uncertainty modeling can be introduced, while applying physical consistency regularization constraints to improve overall stability and reliability. Furthermore, since parallel assessment of multiple aircraft types is a core requirement, structural parameters such as aircraft weight and size (and extending to parameters related to ground contact stability such as landing gear contact area, support radius, and center of gravity height) should be encoded as query vectors. These vectors interact with site tokens through mechanisms such as cross-attention, enabling the model to automatically focus on key risk areas for "specific aircraft types" and output conditional feasibility scores / risk levels or multi-dimensional margin indicators. This allows for reliable screening, ranking, and alternative point recommendation of candidate landing areas in unknown scenarios, providing interpretable and scalable technical support for downstream applications such as real-time decision-making, progressive transmission, and rendering in large-scale scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention;

[0016] Figure 2 A schematic diagram for extracting site features through point cloud and image fusion;

[0017] Figure 3 This is a schematic diagram for extracting site features from point clouds. Detailed Implementation

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] The flowchart of this invention is as follows Figure 1 , 2 As shown in Figure 3, specifically:

[0020] Step 1: Physical Enhancement of Candidate Site Features

[0021] The aim is to jointly utilize lidar point cloud and image information to construct a unified feature representation that can characterize "landability" for each candidate landing area, and to explicitly enhance the characterization of physical factors closely related to landing safety, such as load, size and tilt, at the feature level, so as to provide a robust and interpretable input basis for subsequent multi-aircraft conditional assessment.

[0022] Step 11 completes multimodal alignment and unified encoding of candidate regions;

[0023] First, the point cloud of the candidate region is spatially aligned with the image. Nearest neighbor mapping is used to map the height values ​​of the point cloud to the image pixel grid, thus constructing a fused image input containing "color texture + height channel". Subsequently, this fused image is input into a pyramid backbone network to extract multi-scale feature maps, enabling the network to simultaneously capture texture differences and geometric variations, thereby improving its ability to perceive small obstacles, undulating boundaries, and areas of abrupt material changes on the roof surface. Since candidate landing areas often exhibit asymmetrical polygonal shapes with large scale differences, directly using rectangular ROI convergence would introduce significant background noise and weaken the focus on the internal structure of the candidate region. Therefore, this invention further employs Mask ROIAlign to achieve unified tensor quantization of irregular candidate regions: a bounding rectangle is taken for the candidate polygon on the feature map and sampled onto a fixed-size grid (e.g., ...). This process obtains a uniformly sized region feature tensor. Simultaneously, candidate region masks are aligned to the same grid and used as soft weights to weight the region features, suppressing background information from non-candidate regions within the circumscribed rectangle. Through this mechanism, even with varying candidate region shapes, batch training and inference can still be performed in the network using fixed-dimensional tensors, preserving the spatial distribution structure within the candidate regions and providing stable input for subsequent weak region mining and size-feasible region extraction.

[0024] Based on the unified regional coding, a regional token is further generated for site interaction and device model query. Specifically, in one embodiment, in addition to outputting the regional grid feature tensor, the masked weighted regional features are aggregated to obtain a global token, which is used for subsequent "inter-site communication" and "multi-device model cross-attention query". To ensure that the token can represent the overall conditions of the candidate region and continuously focus on the effective range, this invention uses a masked normalized weighted average method to generate the token, which takes the following form (Formula 1):

[0025]

[0026] in, For the first The grid features of candidate regions after masking suppression For the aligned mask weights, To prevent minute amounts from being divided by zero, the polymerization method yields... It comprehensively characterizes the material and geometric conditions of the candidate region as a whole, while avoiding interference from the background region on the token, thereby providing a stable global representation for subsequent cross-region feature interaction and multi-device query.

[0027] Step 12: After completing the unified encoding of the image side and the region side, this invention introduces point cloud 3D geometric feature extraction and optional completion to enhance the perception of geometric risk factors. Since roof landing safety is highly dependent on geometric characteristics (e.g., steps, local undulations, slope changes, sharp edge transitions, and obstacles), and point cloud data is sparse and irregular, to improve efficiency and robustness, this invention employs a sparse voxel convolution method in one embodiment: the candidate region point cloud is voxelized and input into a sparse convolution backbone network to extract multi-scale 3D geometric features, which are then aggregated into a geometric token, forming a joint representation of the candidate site together with the image-side token from step 11. Furthermore, when there are point cloud gaps due to occlusion, reflectivity differences, or scanning angle, local holes reduce the reliability of plane fitting, slope estimation, and step recognition. Therefore, this invention provides an optional completion branch: after scattering sparse voxel features onto a fixed 3D grid, a lightweight 3D CNN is used to predict voxel occupancy / density to complete geometric continuity, and geometric quantities such as normal vectors and curvature are estimated based on neighborhood fitting. This branch can be enabled based on real-time requirements: offline high-precision evaluation can be enabled to improve robustness, while online fast evaluation can be simplified to reduce computational overhead.

[0028] Step 2: Physical Proxy Parameter Mining and Physical Margin Construction

[0029] The purpose of step 2 is to extract physical factors that are strongly related to landing safety from the multimodal features obtained in step 1 in the absence of strong physical annotations such as explicit load capacity and size thresholds, and organize them into an interpretable "physical margin" output, thereby providing an auditable basis for engineering decisions.

[0030] Step 21: Construct a load-bearing related proxy parameter extraction module. Considering that load-bearing capacity is often determined by local weak areas, directly averaging the features of candidate areas would dilute the risk of weak points and lead to an overly optimistic assessment. In one embodiment, this invention adopts a load-bearing head of "learning weight graph + weighted aggregation": the network first learns spatial weights on the grid features of candidate areas, focusing attention on potential weak areas (such as material changes, crack textures, abnormal point cloud undulations, etc.), and then weighted aggregation of features to obtain a load-bearing related representation, outputting load-bearing proxy parameters (which can be load-bearing capacity per unit area, load-bearing level, or load-bearing risk score); these proxy parameters are then combined with the aircraft weight, contact area, etc., to obtain the load margin, thereby logically realizing the interpretable determination that "the weakest point in the local area dominates the overall risk" and providing structured input for subsequent residual correction.

[0031] In step 22, the present invention performs size-related available space estimation. Since UAV landing can usually be approximated as a "circular footprint constraint," it is necessary to estimate the radius and center position of the "maximum safe circle" that can be accommodated within the candidate region. In one embodiment, the present invention first filters feasible planar regions based on point cloud normal vector consistency and slope threshold, and projects them onto a two-dimensional grid to form a binary feasible mask; then, a distance transformation is performed on the mask to obtain the shortest distance from each grid point to the boundary / obstacle, and the position with the maximum distance field is taken as the center, and the maximum distance is taken as the radius of the maximum inscribed circle. The available radius obtained is then compared with the minimum safe radius required by the aircraft model to construct the size margin. This process can adapt to irregular candidate regions without complex derivation, and naturally supports "landing point encryption": the vicinity of the center of the maximum inscribed circle is usually a better landing point candidate, and multiple alternative points can be obtained by further encrypting the sampling around it, improving mission redundancy and safety.

[0032] In step 23, the present invention constructs a tilt-related margin to constrain slope risk. Tilting risk is mainly related to the local tilt of the roof and the stability of the aircraft model. Therefore, in one embodiment, the present invention calculates the tilt of candidate areas based on normal vectors or local plane fitting, and obtains allowable tilt thresholds for different aircraft models based on their stability parameters such as support radius, center of gravity height, and weight distribution. The tilt margin is output by comparing the site tilt with the aircraft model threshold, and is used as a hard constraint in the screening process. When the tilt margin is negative, it is determined that the aircraft model is not suitable for landing in the candidate area, even if the dimensions and load-bearing capacity are met, it should be excluded to avoid the risk of tilting and instability when the slope is too steep.

[0033] Step 3: Multi-model conditional assessment and scoring output.

[0034] Step 3 aims to perform conditional interactive modeling of candidate site conditions and multi-aircraft constraints, outputting a feasibility score and risk explanation for each aircraft type on each candidate site for ranking and recommendation. In Step 31, this invention encodes aircraft parameters as queries: for multiple types of aircraft, aircraft parameter inputs (such as length, width, height, weight, and can be expanded to include contact area, support radius, center of gravity height, etc.) are established and encoded into query vectors of the same dimension as the site token through an embedding layer or lightweight MLP. This allows the query to carry clear aircraft type requirements, enabling the model to output differentiated judgments for different aircraft types on the same site. In Step 32, this invention uses cross-attention to achieve aircraft type-site interaction: using the aircraft type query as the query end and the site token as the key value end, the cross-attention module reads site information according to aircraft type requirements, enabling the model to automatically focus on the most critical regional features for the current aircraft type. For example, large aircraft focus more on size boundaries and available space shape, heavy aircraft focus more on areas with weak load-bearing capacity, and high center of gravity aircraft focus more on slope and undulation changes, thereby achieving interpretable differentiated decision-making basis.

[0035] In step 33, the present invention outputs a comprehensive score and performs ranked supervised training. The comprehensive score is jointly determined by "physical margin (load capacity / size / tipping / slippage, etc.) + machine type-site interaction characteristics", and its form is as follows (Formula 2):

[0036]

[0037] in, For model At the venue On the compatibility rating; , , These are safety margins for tipping, bearing capacity, and sliding, respectively. The model-site interaction representation obtained through cross-attention; The scoring function can be implemented using a lightweight MLP. To address the lack of absolute scoring labels, this invention employs relative ranking supervision: if the field is under the same model... Superior to the venue Then it should satisfy (Formula 3):

[0038]

[0039] Based on this, a pairwise ranking loss is constructed to differentiate between superior and inferior venues, in the form of (Formula 4):

[0040]

[0041] in The interval parameter is used to improve ranking stability. During training, in addition to the ranking loss, auxiliary constraints can be set on the consistency of physically related outputs and the magnitude of residual correction to enhance interpretability and physical consistency; the total loss is expressed as a weighted sum (Formula 5):

[0042]

[0043] in Used to ensure that physics-related outputs are consistent with weak annotations / rules. Used to limit residual correction and network output fluctuations These are the weighting coefficients.

[0044] Step 4: Encrypting Landing Points and Recommending Alternative Landing Points

[0045] This invention generates multiple candidate landing points within the candidate roof area (e.g., uniform sampling within a feasible planar area, or locally encrypted sampling around the center of the largest inscribed circle). A small-area token is formed for each candidate landing point, and the multi-aircraft evaluation in step 3 is repeated to obtain a landing point-level score heatmap. The system combines the comprehensive score with hard physical thresholds to output Top-K candidate landing points and provides different strategies based on the mission scenario: in emergency landing mode, it prioritizes meeting hard constraints such as tipping / size / load capacity and selects landing points with higher scores after meeting these constraints; in regular mission mode, it prioritizes the landing point with the highest comprehensive score while meeting the hard constraints, and provides several alternative points for trajectory planning and redundancy assurance.

[0046] Step 5: Output and Interpretability Presentation

[0047] The output of this invention not only includes "feasible / infeasible" or "scoring and ranking", but also simultaneously outputs interpretable information for auditing and downstream systems, including the maximum available size (maximum inscribed circle radius and center position), site inclination and tilt margin given by aircraft type, load proxy parameters and load margin given by aircraft type, and outputs skid margin and risk warning when skid risk modeling is enabled; at the same time, it outputs the final Top-K alternative landing points and their main risk sources (such as insufficient size, insufficient load, excessive slope, etc.) so that flight control decision-making, mission planning and safety review can quickly call and trace them.

[0048] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A roof landing assessment method considering multiple aircraft types, characterized in that, Specifically, this involves using lidar point cloud and multi-source image / raster data to extract candidate landing areas, estimate physically interpretable parameters, and conduct conditional safety assessments for multiple aircraft types.

2. The roof landing assessment method for multiple aircraft types according to claim 1, characterized in that, Specifically: Step 1: Physically enhance the features of the candidate site; Step 2: Constraints on multiple types of aircraft; Step 3: Evaluate the site landing conditions comprehensively based on site conditions and aircraft.

3. The roof landing assessment method for multiple aircraft types according to claim 2, characterized in that, Specifically: Step 1.1: Multimodal alignment and unified encoding of candidate regions to obtain a global token; Step 1.2: Introduce point cloud 3D geometric feature extraction and optional completion to obtain geometric tokens. Combine the geometric tokens and global tokens to form a joint representation of the candidate site.

4. The roof landing assessment method for multiple aircraft types according to claim 2, characterized in that, Step 2 specifically involves: Step 2.1: Calculate the bearer proxy parameters based on the features extracted in Step 1; Step 2.2: Estimate the dimensional margin that can be accommodated within the candidate region; Step 2.3: Establish slope risk with relevant tilting margin constraints.

5. The roof landing assessment method for multiple aircraft types according to claim 2, characterized in that, Step 3 specifically involves: Step 3.1: Collecting parameters for any model. Step 3.2: Retrieve key information about the location for the specific model of the aircraft according to its requirements; Step 3.3 Calculate the evaluation results based on the extracted information.

6. A roof landing assessment system that takes into account multiple aircraft models, characterized in that, The system utilizes multi-source data, including lidar point clouds and imagery / grid data, to extract candidate landing areas, estimate physically interpretable parameters, and conduct conditional safety assessments for multiple aircraft types.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a roof landing assessment method for multiple aircraft types as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a roof landing assessment method for multiple aircraft types as described in any one of claims 1 to 5.

Images