Low-altitude aircraft flight obstacle avoidance method in dense building area

Abstract

The application relates to a low-altitude aircraft dense building area flight obstacle avoidance method, which first cooperatively collects point cloud and image data through an airborne laser radar and a vision sensor, generates a high-precision fusion observation set through multi-step space-time alignment and space registration, and extracts a space-stable geometric-semantic entity in combination with multi-modal semantic coding. Subsequently, a semantic decision tree and a path anchor point sequence are constructed by using the spatial relationship and dynamic attributes of the entity, path matching and conflict resolution are carried out through dimensions such as semantic consistency, topological rationality and time sequence continuity. Finally, the system outputs an optimal obstacle avoidance path with environmental adaptability according to the judgment result, and real-time feedback optimization decision tree and confidence model after path execution, so that closed-loop adaptive updating is realized. The application improves the path planning robustness and accurate obstacle avoidance capability of the aircraft in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent obstacle avoidance and path planning technology for low-altitude aircraft, and in particular to a method for obstacle avoidance when flying in densely built-up areas with low-altitude aircraft. Background Technology

[0002] Currently, when low-altitude aircraft perform obstacle avoidance and path planning in densely built-up areas, mainstream technologies generally adopt a sequential "perception-map building-path search-global verification" paradigm. Typical approaches include using airborne LiDAR, visual cameras, and other sensors to collect environmental point cloud and image data in real time, constructing local and global maps through rasterization, topological mapping, or fuzzy logic-based potential field methods, and then using traditional algorithms such as A* and DWA to complete path search and obstacle avoidance. Meanwhile, to further enhance the ability to cope with complex environmental changes, some solutions in recent years have introduced deep learning semantic segmentation, path planning based on behavioral intent prediction, and multi-sensor fusion technologies to achieve functions such as building surface type recognition and dynamic obstacle prediction. With the rapid development of urban air traffic, aircraft obstacle avoidance path decision-making is trending towards multi-source information fusion, real-time dynamic adjustment, and global target collaborative guidance.

[0003] While existing intelligent obstacle avoidance and path planning technologies for low-altitude aircraft can achieve basic obstacle avoidance functions, they have significant shortcomings in local path decision optimization and multi-source path conflict resolution. Specifically, these shortcomings are reflected in the following aspects:

[0004] Due to limitations in perception field of view, aircraft can typically only generate multiple feasible path segments based on a local observation area. When the number of candidate paths increases and environmental obstacles change frequently, the lack of an efficient and semantically relevant path conflict decision-making mechanism leads to the path selection process relying on redundant parameters or non-intelligent weights, easily resulting in selection ambiguity and reduced mission robustness. Current technologies mostly rely on global map verification, grid reconstruction, or manually set priorities for branch decisions, which cannot dynamically adapt to changes in environmental semantics and are easily affected by temporary obstacles, causing them to fail.

[0005] There is a lack of a mission-goal-oriented local-global semantic coordination mechanism. Existing methods are disconnected from local path obstacle avoidance and global mission achievement. Local perception results are not effectively transformed into a global intent-driven path selection skeleton. Especially when there is no clear main navigation topology, the path selection process lacks the ability to dynamically align mission objectives with key semantic anchors, which affects flight efficiency and safety.

[0006] There is insufficient credibility assessment of key semantic anchor points under multi-source perception. When facing highly dynamic scenarios such as changes in lighting, reflections from building materials, and temporary construction structures, aircraft lack historical memory caching and multi-factor weighting optimization strategies. Interference with anchor points often leads to path selection confusion, increasing flight risks. Summary of the Invention

[0007] This application provides a method for obstacle avoidance when a low-altitude aircraft flies through densely built-up areas, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0008] This application provides a method for obstacle avoidance in densely built-up areas for low-altitude aircraft, specifically including:

[0009] S1: Acquire raw point cloud data and image frame data collected by airborne sensors of low-altitude aircraft, and perform spatiotemporal synchronization calibration on the raw point cloud data and image frame data to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference.

[0010] S2: Based on the multimodal local observation dataset, perform online parsing to identify and abstract spatially stable geometric-semantic hybrid entities, and assign a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity to generate a real-time semantic anchor sequence with attribute annotations.

[0011] S3: Based on the current task endpoint coordinates and the real-time semantic anchor sequence, plan and generate multiple candidate local path segments that meet the motion feasibility, and generate a path anchor sequence based on the candidate local path segments.

[0012] S4: Using the task endpoint coordinates as the root node, and based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor point sequence, dynamically construct a semantic decision tree that pulls backward from the global goal to generate a semantic guidance skeleton structure for evaluating path compliance.

[0013] S5: Input the path anchor sequence into the semantic guidance skeleton structure, calculate the matching score of each candidate local path segment in three dimensions: anchor type consistency, spatial topological rationality, and temporal evolution continuity, so as to generate a semantic matching score vector that represents the degree of alignment between each candidate local path segment and the global task intent.

[0014] S6: Based on the semantic matching score vector, identify conflict candidate paths with functional redundancy and similar scores, call the stability records of the same type of geometric semantic mixed entities in the historical memory cache under similar environmental conditions, and perform a credibility re-evaluation process on the key semantic anchors that cause ambiguity to generate a conflict resolution decision result after dynamic weight correction.

[0015] S7: Based on the conflict resolution decision, select the target local path with the highest score and that has passed the verification, convert it into an aircraft control command and output it to drive the low-altitude aircraft to complete the obstacle avoidance maneuver according to the target local path.

[0016] S8: Based on the execution feedback information of the target local path, update the state of the terminal node of the semantic decision tree and the stability record in the historical memory cache to generate an optimized semantic guidance skeleton structure and a confidence prior library.

[0017] The obstacle avoidance method for low-altitude aircraft flying over densely built-up areas provided in this application has the following beneficial effects:

[0018] (1) To address the issues of poor real-time path planning and delayed obstacle avoidance response in densely built-up areas for traditional low-altitude aircraft, this solution significantly improves the efficiency and robustness of path generation in dynamic environments by constructing a closed-loop collaborative decision-making mechanism mediated by semantic anchors. Existing methods generally rely on a sequential process of "perception → mapping → search → verification," which leads to risks of replanning delays or even loss of control in complex urban canyon scenarios due to map update delays or sudden changes in local obstacles. This solution abandons global grid map construction and traditional algorithms (such as A* and DWA), and instead extracts lightweight multimodal semantic anchor points (such as rooftop edge segments and glass curtain wall reflection areas) with navigation significance online from the original point cloud and image frames. Only key geometric semantic hybrid entities and their relative pose relationships are retained, which significantly reduces the data processing dimension and memory usage. On this basis, the path planning module directly generates candidate local path segments that meet dynamic constraints based on the current observation, and realizes structured guidance from the task objective through semantic decision trees. This avoids the dependence on modeling the complete environment, enabling the system to complete the closed-loop response of perception and decision within hundreds of milliseconds. It effectively copes with challenges such as weak GPS signals and frequent temporary obstacles, significantly advances the timing of obstacle avoidance intervention, and improves flight safety and trajectory smoothness.

[0019] (2) To address the decision ambiguity caused by insufficient semantic stability under multi-source interference, this solution innovatively introduces a semantic matching degree fusion evaluation and conflict resolution engine, achieving continuous and reliable intent alignment capability in uncertain environments. Traditional methods are prone to misjudging the same physical structure as different objects when faced with varying materials (such as specular reflection), light fluctuations, or partial occlusion, resulting in path oscillation or cyclical hesitation. This solution jointly calculates the matching degree between candidate paths and semantic decision trees through three dimensions: anchor point type consistency, spatial topology rationality, and temporal evolution continuity, ensuring that the selected path not only conforms to the semantic logic of the task (such as prioritizing the selection of anchor point sequences that can land on platforms), but also follows the spatial causal laws of the real world (such as having to bypass obstructions to approach the target); when multiple paths have similar scores, the system automatically triggers a conflict resolution mechanism, identifies key semantic anchor points that cause ambiguity (such as glass curtain walls affected by strong light), and calls the stability records of similar anchor points under similar conditions in the historical memory cache, integrates the current multi-sensor confidence for dynamic weight correction, thereby eliminating severely interfered branch paths. This mechanism can achieve adaptive anti-interference decision-making without introducing fuzzy logic controllers or potential field functions, significantly enhancing the robustness and interpretability of the system in high dynamic and low signal-to-noise ratio environments.

[0020] (3) This solution constructs an adaptive path planning system with cross-scenario generalization capability and continuous learning potential through a lightweight decision-making architecture driven by semantic anchors. Unlike traditional systems that rely on fixed maps or preset rules, all decisions in this solution are based on the semantic structure stimulated by the aircraft's own observations. The semantic decision tree only maintains the relative reachability and time decay weights between anchors and does not store specific coordinate information, thus naturally supporting incremental updates and knowledge transfer during long-term operation. With the accumulation of flight experience, the system can gradually optimize the recognition accuracy and stability model of various semantic anchors, forming behavioral priors for typical urban structures (such as construction cranes and ventilation openings), further improving the prediction accuracy and response sensitivity in complex scenarios. The entire framework does not require complex parameter tuning and has good engineering deployment feasibility. It is particularly suitable for mission scenarios that require frequent crossing of low-altitude dense areas, such as campuses, parks, and emergency rescue. It can support the aircraft to navigate autonomously between resources inside and outside the campus efficiently, in a variety of ways, and with high accuracy, providing a scalable technical paradigm for the intelligent evolution of future urban air traffic systems.

[0021] In summary, this solution overcomes multiple bottlenecks in real-time performance, robustness, and generalization ability of traditional path planning methods by designing semantic anchor abstraction, intent alignment-driven approach, and closed-loop collaborative decision-making mechanism. Without relying on global mapping and classical algorithms, it achieves high-precision, low-latency, and highly adaptable path planning and obstacle avoidance decision-making for aircraft in low-altitude densely built-up areas, demonstrating significant technological advancement and practical application value. Attached Figure Description

[0022] Figure 1 This is the main flowchart of a method for obstacle avoidance when flying low-altitude aircraft in densely built-up areas.

[0023] Figure 2 This is a sub-flowchart of a method for obstacle avoidance when low-altitude aircraft fly over densely built-up areas.

[0024] Figure 3 This is another sub-flowchart of a method for obstacle avoidance in densely built-up areas for low-altitude aircraft.

[0025] Figure 4 This is an application environment diagram of a method for obstacle avoidance in densely built-up areas for low-altitude aircraft, as described in one embodiment.

[0026] Figure 5 An internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0028] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0029] like Figure 1 As shown, this application provides a method for obstacle avoidance when a low-altitude aircraft flies through densely built-up areas, specifically including:

[0030] S1: Acquire raw point cloud data and image frame data collected by airborne sensors of low-altitude aircraft, and perform spatiotemporal synchronization calibration on the raw point cloud data and image frame data to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference.

[0031] S2: Based on the multimodal local observation dataset, perform online parsing to identify and abstract spatially stable geometric-semantic hybrid entities, and assign a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity to generate a real-time semantic anchor sequence with attribute annotations.

[0032] S3: Based on the current task endpoint coordinates and the real-time semantic anchor sequence, plan and generate multiple candidate local path segments that meet the motion feasibility, and generate a path anchor sequence based on the candidate local path segments.

[0033] S4: Using the task endpoint coordinates as the root node, and based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor point sequence, dynamically construct a semantic decision tree that pulls backward from the global goal to generate a semantic guidance skeleton structure for evaluating path compliance.

[0034] S5: Input the path anchor sequence into the semantic guidance skeleton structure, calculate the matching score of each candidate local path segment in three dimensions: anchor type consistency, spatial topological rationality, and temporal evolution continuity, so as to generate a semantic matching score vector that represents the degree of alignment between each candidate local path segment and the global task intent.

[0035] S6: Based on the semantic matching score vector, identify conflict candidate paths with functional redundancy and similar scores, call the stability records of the same type of geometric semantic mixed entities in the historical memory cache under similar environmental conditions, and perform a credibility re-evaluation process on the key semantic anchors that cause ambiguity to generate a conflict resolution decision result after dynamic weight correction.

[0036] S7: Based on the conflict resolution decision, select the target local path with the highest score and that has passed the verification, convert it into an aircraft control command and output it to drive the low-altitude aircraft to complete the obstacle avoidance maneuver according to the target local path.

[0037] S8: Based on the execution feedback information of the target local path, update the state of the terminal node of the semantic decision tree and the stability record in the historical memory cache to generate an optimized semantic guidance skeleton structure and a confidence prior library.

[0038] Step S1: Acquire raw point cloud data and image frame data collected by the airborne sensors of the low-altitude aircraft, and perform spatiotemporal synchronization calibration processing on the raw point cloud data and image frame data to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference. Specifically, this includes:

[0039] S1.1: Acquire the raw point cloud data stream collected by the airborne lidar sensor of the low-altitude aircraft and the raw image frame data stream collected by the vision camera sensor. Perform noise reduction filtering on the raw point cloud data stream and extract the point cloud timestamp sequence. At the same time, perform distortion correction processing on the raw image frame data and extract the image frame timestamp sequence to generate a time-stamped preprocessed point cloud set and a time-stamped preprocessed image frame set containing geometric and texture information.

[0040] An independent data channel is used to read the raw point cloud data stream generated by the airborne lidar sensor of the low-altitude aircraft, ensuring that the data frame order and sampling frequency are consistent during the acquisition process, and the raw point cloud data is cached in a high-speed data cache module for subsequent processing.

[0041] Statistical filtering is applied to the original point cloud data stream. The number of neighboring points and the distance threshold are set as filtering control parameters. Isolated points and outliers are removed based on the neighborhood density of each point, thereby reducing the interference of noise on subsequent geometric feature extraction.

[0042] The filtered point cloud data stream is parsed to extract the point cloud frame header information containing the sensor's internal timestamps, and the original timestamp sequence is extracted and stored in the order of acquisition to form a point cloud time reference set.

[0043] The raw image frame data stream acquired by the airborne vision camera sensor of the low-altitude aircraft is read through the channel to ensure that the image frame data is not lost or misaligned due to transmission delay, and is cached in the image processing pre-buffer.

[0044] The distortion correction algorithm based on the pinhole camera model is called on the original image frame data. The pre-calibrated radial distortion coefficient and tangential distortion coefficient are set, and the deformation caused by the lens structure is eliminated by pixel position remapping to obtain a geometrically accurate corrected image frame.

[0045] Extract the frame header timestamp information from the distortion-corrected image frames, construct an image time reference set and keep it consistent with the acquisition sequence for subsequent spatiotemporal synchronization.

[0046] Through the above filtering and distortion correction methods, the original lidar point cloud data and visual camera image data are transformed into a time-stamped preprocessed point cloud set and a time-stamped preprocessed image frame set containing geometric and texture information, thereby achieving preliminary optimization of local observation data quality and preservation of time-stamp integrity.

[0047] For example, for a quadcopter low-altitude aircraft equipped with a Velodyne VLP-16 LiDAR and a global shutter industrial camera, under the conditions of a LiDAR sampling frequency of 10Hz and a camera frame rate of 20Hz, statistical filtering with a neighborhood point count of 10 and a distance threshold of 0.5 meters was used for point cloud filtering. After filtering out noise points, the average number of points per frame decreased from 28,200 to 27,350, and the spatial distribution became more uniform. During distortion correction, radial distortion coefficients k1=−0.23, k2=0.12 and tangential distortion coefficients p1=−0.001, p2=0.0007 were used. The straight line structure error of the image edge was significantly reduced after correction, and the average pixel deviation was reduced to 0.85 pixels. In the frame timestamp extraction stage, both the point cloud time reference set and the image time reference set were timed using the sensor's internal clock to ensure no cross-frame misalignment during subsequent linear interpolation alignment. After this sub-step, the output time-stamped preprocessed point cloud set retains complete spatial geometric features, and the time-stamped preprocessed image frame set retains accurate texture information. Both have a unified time reference, laying the foundation for data accuracy for time reference alignment in step S1.2.

[0048] S1.2: Based on the point cloud timestamp sequence in the time-stamped preprocessed point cloud set and the image frame timestamp sequence in the time-stamped preprocessed image frame set, a time reference alignment operation is performed using linear interpolation. The relative time offset between each data frame is calculated and resampling compensation is performed to generate time-synchronized point cloud data blocks and time-synchronized image frame data blocks with a unified global time reference.

[0049] S1.3: Based on the pre-calibrated extrinsic parameter matrix of the airborne sensor, perform rigid body transformation processing on the time-synchronized point cloud data block to map the point cloud coordinate system to the camera coordinate system, and perform pixel-level projection association operation on the time-synchronized image frame data block to establish the correspondence between three-dimensional spatial points and two-dimensional image pixels, so as to generate spatially registered point cloud image pairs with spatial position consistency.

[0050] For time-synchronized point cloud data blocks and time-synchronized image frame data blocks under a unified global time reference, based on the pre-calibrated extrinsic parameter matrix of the airborne sensor, the three-dimensional rigid body transformation method is used to map the three-dimensional coordinate system of the point cloud data blocks to the camera coordinate system according to the calibration matrix, thereby achieving the consistency of the spatial reference.

[0051] For the point cloud coordinate data after rigid body transformation, the two-dimensional pixel coordinates of each spatial point on the camera imaging plane are calculated using a perspective projection model. The core calculation formula of the perspective projection model is:

[0052]

[0053] Where p is a two-dimensional pixel coordinate vector, K is the camera intrinsic parameter matrix, R is the rotation matrix, t is the translation vector, and P is the homogeneous coordinate vector of a point in three-dimensional space.

[0054] Pixel-level correlation matching is performed on all point cloud projection results. By searching for the consistency of color gradient and depth gradient within adjacent pixel intervals, a correspondence matrix between 3D point cloud points and 2D image pixels is established.

[0055] For the established correspondence matrix, the occlusion culling operator is applied to remove spatial points that cannot be fully presented in the camera's field of view, ensuring the visibility and integrity of the registration data.

[0056] The above processing method transforms the time-synchronized multimodal data from the previous step into spatially registered point cloud image pairs with spatial location consistency, enabling high-precision input for subsequent feature matching and spatial residual evaluation.

[0057] S1.4: Based on the overlapping field of view in the spatially registered point cloud image pair, spatial residual evaluation is performed using feature matching to calculate the spatial deviation value between the point cloud projection edge and the image feature edge, and the external parameter matrix is ​​fine-tuned and optimized according to the spatial deviation value to generate high-precision spatially aligned point cloud image fusion data after spatial error correction.

[0058] Based on the overlapping field of view region in the spatially registered point cloud image pair, the input conditions include the pixel correspondence data of the 3D point cloud and 2D image after time synchronization and spatial coordinate system mapping from the airborne LiDAR and visual camera, as well as the initial values ​​of the pre-calibrated sensor extrinsic parameter matrix. Two-dimensional feature points with high repeatability and spatial positioning significance are extracted from the image data using feature detection operators. A projection edge dataset is generated using the corresponding 3D point cloud projection, and candidate matching pairs of 2D feature points and 3D projection points are established in the overlapping field of view region. Robust feature matching is used to filter out a set of point pairs with high matching confidence from the candidate pairs, and the spatial residual vector between each matching point pair is calculated. The residual vector consists of the difference between the 3D projection point and the corresponding 2D feature point obtained by backprojecting the pixel coordinates of the 3D point. The sum of the squares of the residual vectors is then averaged to calculate the spatial residual evaluation value. Based on this spatial residual evaluation value, an error gradient vector is constructed. The gradient vector is used as the derivative of the objective function for extrinsic parameter optimization, and iterative optimization (Levenberg-Marquardt) is called to adjust the extrinsic parameter matrix component by component until the error evaluation value converges to a preset tolerance range. During each iteration of the optimization process, the mapping relationship between 3D points and 2D projections is updated using the latest extrinsic parameter matrix, and the spatial residuals are recalculated to ensure that the optimization direction is consistent with the actual spatial geometry. After the iteration is completed, the optimized extrinsic parameter matrix is ​​integrated to perform fusion mapping on the time-synchronized point cloud and time-synchronized image data, generating high-precision spatially aligned point cloud image fusion data with improved matching accuracy after spatial error correction. Through the above processing method, the results of the previous step are transformed into multimodal fusion data with low residuals and high consistency, thereby improving the spatial accuracy of the subsequent semantic parsing stage.

[0059] For example, in an urban low-altitude vehicle obstacle avoidance scenario, the airborne LiDAR point cloud resolution is set to 0.05 meters, the visual camera sampling frequency is 20 frames per second, and the overlapping field of view area is approximately 85%. 500 SIFT feature points are extracted from the 2D image, and 500 corresponding projection point candidates are found at the edge of the 3D point cloud projection. After RANSAC matching to remove outliers, 450 valid matching point pairs are retained. The residual evaluation value is calculated, and the average of the squared differences in the matching point coordinates is 0.0125 meters. 2 The corresponding spatial residual evaluation value was 0.112 meters. Iterative optimization was used to adjust the step size of the rotation component of the extrinsic parameter matrix to 0.01 degrees and the step size of the translation component to 0.2 millimeters. After 15 iterations, the residual evaluation value decreased to 0.028 meters, significantly improving the spatial alignment accuracy. In the high-precision spatially aligned point cloud image, the fusion result showed a significant improvement in the overlap between image edges and point cloud projection edges, meeting the requirements of spatial consistency and positioning accuracy for subsequent semantic anchor point extraction.

[0060] S1.5: Encapsulate the high-precision spatially aligned point cloud image fusion data according to a unified data structure, bind the current aircraft's pose status information with the unified global time reference, and perform data serialization and packaging operations to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference.

[0061] Step S2: Based on the multimodal local observation dataset, perform online parsing to identify and abstract spatially stable geometric-semantic hybrid entities. Assign a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity to generate a real-time semantic anchor point sequence with attribute annotations. Specifically, this includes:

[0062] S2.1: Based on the multimodal local observation dataset, a cross-modal attention mechanism is used to perform feature-level fusion processing on the original point cloud data fragments and image frame data fragments to generate a multimodal fusion feature map containing depth and texture information.

[0063] Based on the time-synchronized point cloud data blocks and time-synchronized image frame data blocks in the multimodal local observation dataset, the weight parameters of the point cloud feature encoder and image feature encoder are initialized by calling the cross-modal attention mechanism, and a unified feature embedding dimension is set as the matching benchmark for subsequent fusion. Geometric feature extraction operators are performed on the point cloud data blocks to generate a high-dimensional geometric feature matrix containing surface normals, curvature, and local voxel distribution. Simultaneously, convolutional feature extraction operators are performed on the image frame data blocks to generate a high-dimensional visual feature matrix containing color gradients, texture patterns, and edge directions. Frame-level mapping is performed between the point cloud feature matrix and the image feature matrix according to the time-synchronized sequence, and cross-modal similarity attention weights are calculated. A self-attention mechanism based on scaled dot products is used to normalize the cross-modal weight vector, enabling priority to be given to semantically relevant local regions during feature matching. The normalized cross-modal attention weights are applied to the point cloud feature matrix and the image feature matrix to generate a weighted feature vector of the fused feature subspace. A depth modulation factor is introduced during the fusion process to explicitly retain the depth quantization information provided by the LiDAR. The output value of each dimension of the fused feature is calculated using the following formula:

[0064]

[0065] Where F is the fused feature vector, A is the cross-modal attention weight matrix, p is the depth modulation factor exponent, and T is the normalized point cloud-image joint feature matrix. Feature domain smoothing filtering is performed on the fused feature vector to remove isolated noise components caused by cross-modal alignment errors, and the output is a unified multimodal fused feature map containing depth and texture information. Through feature-level fusion processing, the high-precision spatially aligned point cloud image fusion data from the previous step is transformed into a multimodal fused feature map that can be used for subsequent instance segmentation, achieving semantic alignment of cross-modal geometric-visual information. For example, in an environment where a low-altitude aircraft is equipped with a 32-line LiDAR and a global shutter camera, the input to this sub-step is a block of time-synchronized image frames with a resolution of 1920×1080 and approximately 1.5×10⁸ pixels per frame. 5 Point cloud data blocks were used, with the fusion embedding dimension set to 256 and the depth modulation factor p set to 2.1. In point cloud feature extraction, the voxel size was chosen to be 0.05m, outputting a total of 64 dimensions for normal vectors and curvature features. In image feature extraction, the convolution kernel size was 3×3, extracting a total of 128 dimensions for color gradient, texture pattern, and edge direction features. Cross-modal attention weights were calculated using a scaled dot product mechanism with a scaling factor of 1 / √256. After normalization, the weight matrix was multiplied by the feature matrix, and the fused features were calculated according to the formula. The depth information of the fused feature map remained continuous in each frame, and the texture details were significantly improved. Validation results showed that the anchor point localization deviation of the fused feature map in subsequent instance segmentation was significantly reduced, and the robustness was significantly improved.

[0066] S2.2: Based on the multimodal fusion feature map, perform instance segmentation to extract candidate object regions with geometric continuity and semantic interpretability to generate a preliminary geometric-semantic hybrid entity set.

[0067] Based on the multimodal fusion feature map, a 3D instance segmentation model is invoked to perform convolution processing on the fusion features, mapping depth information and texture information into a feature response matrix with geometrically significant boundaries.

[0068] A multi-scale anchor box generation mechanism is used in the feature response matrix to deploy candidate detection windows for spatial regions of different scales, and low-consistency candidate windows are screened out by cross-modal feature consistency scoring.

[0069] Region feature extraction is performed on the selected candidate windows. The point cloud density distribution within the window and the image texture gradient are used to construct a joint feature vector. A mask-based instance segmentation network is then called to generate semantic category predictions and corresponding spatial boundary masks.

[0070] In the generated spatial boundary mask and semantic category prediction, geometric continuity constraint detection is used to perform connectivity analysis on the boundary mask and remove candidate instances with breaks or isolated points to ensure that the extracted objects have complete geometric contours.

[0071] For instances that pass the geometric continuity test, the semantic interpretability verification operator is invoked. By comparing the semantic prediction results with the semantic definition sets of the preset building category and obstacle category, candidate instances that do not conform to the preset semantic category are eliminated.

[0072] The above processing method transforms the multimodal fusion feature map from the previous step into a preliminary geometric-semantic hybrid entity set containing geometric continuity and semantic interpretability, thereby achieving accurate extraction of candidate object regions and providing reliable input for subsequent temporal consistency screening.

[0073] For example, in the multimodal fusion feature map of an urban canyon flight mission, a 3D convolutional kernel size of (3,3,3) and stride of 1 is used to extract the feature response matrix. This generates multi-scale anchor boxes with scales of 32, 64, and 128 in the 640×480 resolution image field of view. For each scale anchor box, a cross-modal consistency score is calculated in the fusion features, with a consistency threshold of 0.75. Only candidate windows with scores higher than this value are retained. Within the candidate window, a point cloud density threshold of 500 points per cubic meter and a texture gradient threshold of 0.2 are used to construct a joint feature vector. A mask segmentation network is then called to output the semantic category and boundary mask. The mask connectivity threshold is set to 85% boundary point connectivity to ensure geometric continuity. In the semantic interpretability verification stage, the predicted category is compared with pre-defined navigation building categories such as "rooftop," "suspended beam," "ventilation duct opening," and "tower crane boom end," and instances that cannot be matched are removed. Through this step, 12 preliminary geometric-semantic hybrid entities were obtained in the current field of view, including 4 rooftop edge line segment entities with strong navigation significance and 2 temporary construction boom endpoint entities. These entities showed significantly improved retention rate and spatial consistency in the subsequent temporal stability screening, supporting the intelligent goal orientation of path planning.

[0074] S2.3: Based on the preliminary geometric semantic hybrid entity set, the entity position distribution in the current frame and historical frames is compared using temporal consistency verification to filter out target geometric semantic hybrid entities that maintain spatial stability within a continuous time window, so as to generate a list of highly stable geometric semantic hybrid entities.

[0075] Based on the preliminary geometric-semantic hybrid entity set, the frame index management module is invoked to load the three-dimensional position and attitude data of all geometric-semantic hybrid entities in the current observation frame and the historical observation frames within the set time window, forming an entity position distribution matrix for time series consistency analysis.

[0076] The current observation frame is mapped to the entity location distribution matrix in each historical observation frame. The nearest neighbor matching rule and unique type label constraint are used to achieve one-to-one pairing. The quantized values ​​of the Euclidean distance and attitude difference angle between the paired entities are calculated.

[0077] For each paired entity, the distance difference and attitude difference are quantified, and a multi-threshold stability determination is performed. The distance threshold is set according to the space resolution accuracy of the aircraft, and the attitude angle threshold is set according to the range of sensor viewpoint variation. The determination result is recorded as a binary stability flag sequence.

[0078] For each entity's stability flag sequence, perform continuous stability integration within a time window and calculate the stability score using the following formula:

[0079]

[0080] Where S is the stability score, s i Let T be the stability flag value for the i-th frame, and T be the total number of frames within the time window.

[0081] Entities with stability scores below a preset stability threshold are removed, while entities with stability scores above or equal to the threshold are retained as high-stability targets.

[0082] A list of highly stable geometric-semantic hybrid entities is established based on the unique type label and three-dimensional location index of the highly stable target entities, which serves as the input for subsequent generation of unique spatiotemporal identifiers and confidence calculation.

[0083] By using a temporal consistency verification process, the preliminary geometric-semantic hybrid entity set from the previous step is transformed into a highly stable geometric-semantic hybrid entity list, achieving the expected technical effects of spatial stability purification and dynamic interference removal of entity recognition results.

[0084] S2.4: Based on the highly stable list of geometric-semantic hybrid entities, a globally unique encoding generation strategy is executed to assign a unique spatiotemporal identifier to each target geometric-semantic hybrid entity, and the observation confidence label of each entity is calculated in combination with the sensor noise model to generate attributed geometric-semantic hybrid entities with spatiotemporal identifiers and confidence labels.

[0085] Based on the input of the highly stable geometric-semantic hybrid entity list, a globally unique encoding generation strategy is invoked. By combining hash functions and index mapping, feature concatenation encoding is performed on the spatial location parameters, semantic category labels, and timestamp data of each target geometric-semantic hybrid entity to form an original unique encoding candidate set.

[0086] A collision detection operation is performed on the original unique coding candidate set. The codes with hash collisions are re-encoded by adding pose perturbation components and category feature check bits to eliminate coding duplication caused by similar spatial locations or similar labels, thus forming a collision-free unique coding sequence.

[0087] The conflict-free unique encoded sequence is used as a spatiotemporal identifier allocation table and bound one-to-one with a list of highly stable geometric-semantic hybrid entities, so that each target geometric-semantic hybrid entity has a globally unique spatiotemporal identifier that can be tracked across frames.

[0088] For each target geometric-semantic hybrid entity with an assigned spatiotemporal identifier, a pre-built sensor noise model is invoked to calculate a confidence label based on the combined characteristics of its multimodal observation redundancy, signal strength attenuation curve, and observation distance.

[0089] The confidence labels and corresponding spatiotemporal identifiers are combined to form attributed annotation units, which are then attached to the structural description of highly stable geometric-semantic hybrid entities, forming an attributed geometric-semantic hybrid entity set with spatiotemporal identifiers and confidence labels.

[0090] By using unique encoding generation and noise model calculation processing, the highly stable geometric and semantic hybrid entity list from the previous step is transformed into attributed entity data with cross-time period tracking capabilities and reliability quantification indicators, enabling precise control of anchor point association and weight allocation in subsequent path planning.

[0091] S2.5: Based on the attributed geometric semantic hybrid entity with spatiotemporal identifier and confidence label, perform serialization and reorganization operation in the order of timestamps to generate a real-time semantic anchor sequence with complete attribute annotations.

[0092] like Figure 2 As shown, step S3 involves: based on the current task endpoint coordinates and the real-time semantic anchor point sequence, planning and generating multiple candidate local path segments that satisfy motion feasibility, and generating a path anchor point sequence based on the candidate local path segments. Specifically, this includes:

[0093] S3.1: Obtain the coordinate data of the current task endpoint and the relative pose information of each geometric-semantic hybrid entity in the real-time semantic anchor sequence, apply a gravitational potential field to the task endpoint and a repulsive potential field to each geometric-semantic hybrid entity to generate a composite potential field gradient map that represents the distribution of global target traction force and local obstacle repulsion force.

[0094] Based on the relative pose information of geometric-semantic hybrid entities in the real-time semantic anchor sequence and the current task endpoint coordinate data, an input dataset containing the global pose matrix and local anchor 3D coordinates is selected as the processing object.

[0095] The mission endpoint coordinates are converted into the spacecraft's inertial coordinate system position vector under a unified spatial reference. The spatial potential field is used to construct the initial position and intensity of the gravitational potential field source point, and the gravitational potential field intensity parameters are set according to the mission urgency.

[0096] For each geometric-semantic hybrid entity in the real-time semantic anchor sequence, its unique spatiotemporal identifier and spatial envelope model are read, and the repulsive potential field intensity coefficient is assigned according to the differences in obstacle size, shape and material.

[0097] During the potential field construction process, the range of gravitational and repulsive forces is limited by the distance decay function. The radius of the gravitational potential field is set to the globally reachable area, and the radius of the repulsive potential field is set to a linear combination of the obstacle radius and the safety margin.

[0098] The total potential value at each spatial point is calculated using the superposition method of composite potential fields.

[0099] Numerical differentiation of the composite potential field gradient yields a gradient map matrix, which indicates the potential field direction and magnitude at each spatial location. The output is a composite potential field gradient map representing the distribution of global target traction force and local obstacle repulsion force, thus realizing the force field constraints for local path planning.

[0100] Through the above processing method, the pose and spatial position information of the previous step are transformed into a composite potential gradient map that can be used for trajectory search, so as to realize the path feasibility assessment under the joint constraints of gravity and repulsion.

[0101] S3.2: Based on the composite potential field gradient map, perform a multi-branch trajectory search operation in the unobstructed area. At the same time, introduce the maximum turning radius, maximum climb rate and maximum acceleration of the aircraft as dynamic constraint boundary conditions for pruning to generate a set of multiple discrete candidate local path segments that meet the kinematic feasibility requirements.

[0102] Based on the gravitational and repulsive potential field distribution parameters in the composite potential field gradient map, the random sampling extended tree trajectory generation module is invoked to perform multi-branch trajectory search operations within the marked unobstructed spatial region. The composite potential field gradient map is used as the spatial cost reference plane for the random sampling extended tree (RRT). By defining a sampling probability function on this plane, the generation probability of sampling nodes is inversely proportional to the local potential field gradient intensity, ensuring a reduction in sampling density in high-drag regions. For each candidate node generated by sampling, its spatial Euclidean distance to the previous node is calculated, and the curvature of the line connecting the node to the previous node is determined based on the constraint of the aircraft's maximum turning radius. The spatial height difference between candidate nodes is calculated, and combined with the vertical velocity threshold set by the aircraft's maximum climb rate, it is determined whether the height change satisfies the climb constraint. The acceleration change from the candidate node to the previous node is calculated, and a horizontal acceleration threshold set by the aircraft's maximum acceleration is used for discrimination. Nodes exceeding any constraint threshold are deleted through the pruning module, retaining nodes that satisfy all dynamic constraints and are located in the unobstructed region as valid extensions. The set of effective extended nodes is connected by multiple branches to form a set of feasible discrete candidate local path segments, which are then output to the subsequent spatial projection detection module. By combining random sampling extended trees with dynamic constraint pruning, the composite potential field gradient map generated in the previous step is transformed into multiple discrete path segments that meet the physical characteristics of the aircraft and environmental safety requirements, thereby optimizing the feasibility of path search and improving the accuracy of local obstacle avoidance.

[0103] S3.3: For each discrete candidate local path segment in the set of discrete candidate local path segments, perform a spatial projection overlap detection operation, compare the path trajectory point cloud with the spatial envelope of each geometric semantic hybrid entity in the real-time semantic anchor sequence, so as to identify the list of target geometric semantic hybrid entities that each discrete candidate local path segment traverses or is adjacent to.

[0104] For each path segment in the discrete candidate local path segment set, the spatial envelope data of the geometric-semantic hybrid entities in the path trajectory point cloud and the real-time semantic anchor point sequence are loaded as matching input objects.

[0105] The spatial envelope construction method is adopted to describe the shape boundary of the geometric semantic hybrid entity as a closed envelope composed of polygonal patches, and the envelope structure is represented by the vertex set in the world coordinate system.

[0106] The three-dimensional spatial projection operator is invoked to map the path trajectory point cloud to a coordinate system consistent with the spatial envelope of the geometric semantic hybrid entity, ensuring that the path trajectory and the obstacle surface are accurately compared in the same spatial reference system.

[0107] Perform envelope overlap detection calculations and determine whether the trajectory point is located within the envelope or the shortest distance to the envelope boundary is lower than a preset safety threshold by calculating the directed distance between the point cloud and the envelope patch.

[0108] If the trajectory point overlaps with or is adjacent to the envelope patch, the geometric-semantic hybrid entity corresponding to the envelope is recorded in the target entity list associated with the path and the association type information is marked.

[0109] By using spatial envelope comparison and overlap detection, the discrete candidate local path segments generated in the previous step are transformed into a list of target geometric semantic hybrid entities containing information on traversed or adjacent semantic anchor points, thereby achieving accurate extraction of path spatial association information.

[0110] S3.4: Based on the target geometric-semantic hybrid entity list, extract the unique spatiotemporal identifier and semantic category label corresponding to each target geometric-semantic hybrid entity, and perform a serialization and arrangement operation according to the time sequence of the path travel direction to generate the original path anchor point sequence representing the path traversal order.

[0111] Based on the target geometric-semantic hybrid entity list, the unique identifier extraction module is invoked to parse the globally unique spatiotemporal identifier and corresponding semantic category label of each target entity, forming anchor attribute data pairs. The anchor attribute data pairs are then ordered according to the time series index of the path travel direction to ensure that the path anchor point sequence is consistent with the actual movement sequence of the aircraft. Duplicate detection and deduplication are performed on the ordered anchor attribute data pairs to eliminate multiple records of the same entity caused by overlapping sensor observations, thereby reducing the interference of redundant anchor points on the path sequence. A sequence encoder is applied to the deduplicated anchor attribute data pairs to generate the original path anchor point sequence, which contains three core fields: timestamp, unique spatiotemporal identifier, and semantic category label. Through the above processing, the target geometric-semantic hybrid entity list from the previous step is transformed into an original path anchor point sequence with path sequence characteristics, achieving precise binding between path segments and semantic anchors.

[0112] S3.5: Based on the original path anchor point sequence, perform path smoothing optimization processing and bind the corresponding discrete candidate local path segment spatial coordinate data to convert discrete path points into continuous executable trajectory descriptions, so as to generate a final standardized path anchor point sequence containing complete semantic anchor point type annotations and sequence information.

[0113] like Figure 3 As shown, step S4 involves using the task endpoint coordinates as the root node and dynamically constructing a semantic decision tree that pulls backward from the global goal based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor sequence, to generate a semantic guidance skeleton structure for evaluating path compliance. Specifically, this includes:

[0114] S4.1: Obtain the coordinate data of the current task endpoint and the relative pose information of key anchor points in the real-time semantic anchor point sequence. Use the spatial Euclidean distance calculation method and the line-of-sight occlusion detection operator to perform accessibility quantification evaluation processing, calculate the unobstructed passage probability value and spatial connectivity status label of each key anchor point relative to the task endpoint, and generate a key anchor point accessibility feature set containing accessibility confidence parameters.

[0115] Using the mission endpoint coordinate data and the relative pose information of key anchor points in the real-time semantic anchor point sequence as input, and combining the spatial registration point cloud and image fusion data output by the spacecraft local perception module, a three-dimensional relative position vector between the mission endpoint and the key anchor points is established.

[0116] For the three-dimensional relative position vector of each key anchor point, the spatial Euclidean distance calculation method is called to calculate its straight-line distance to the task endpoint, thus completing the distance quantization.

[0117] Using a line-of-sight occlusion detection operator, a ray is projected onto the straight-line distance vector along the path from the task endpoint to the critical anchor point. The system detects whether a spatial envelope intersects with the ray. If an intersection point exists and the distance from the intersection point to the task endpoint is less than the distance from the critical anchor point to the task endpoint, the system is determined to be in an unreachable state.

[0118] Based on the above direct or inaccessible status, construct spatial connectivity status markers for key anchor points, and simultaneously calculate the accessibility probability P by combining the distance value with an accessibility probability model.

[0119] For all key anchor points, generate a set of reachability features containing the probability value P of unobstructed passage and the spatial connectivity status marker, and assign corresponding reachability confidence parameters as the input basis for subsequent time decay weight calculation.

[0120] By using the above processing method, the spatial relationship between the task endpoint and each key anchor point is transformed into quantifiable reachability parameters and connectivity status, thereby strengthening the reachability constraints of key nodes in the semantic decision tree construction process.

[0121] S4.2: Based on the spatial connectivity status markers in the set of key anchor point reachability features, perform time-sensitivity weighting on the timestamp data of each key anchor point, and calculate the time decay weight coefficient that decreases over time by combining the sensor observation interval, so as to generate a dynamic time decay weight vector that characterizes the freshness and reliability of environmental information.

[0122] Based on the spatial connectivity status markers in the set of key anchor point reachability features, time-sensitivity weighting is performed on the timestamp data of each key anchor point to form an initial time decay factor vector for different observation times.

[0123] By combining sensor observation interval data, a normalized reference quantity for the time decay coefficient is calculated to ensure that different types of sensors maintain a consistent scale standard during the time weighting process.

[0124] Based on the mathematical expression of the exponential decay function

[0125]

[0126] Where Δt is the difference between the current timestamp and the reference timestamp, and τ is the time constant estimated from the observation interval. A multiplicative weighted operation is performed on the reachability confidence parameter of each key anchor point to generate a timeliness correction value.

[0127] By linearly synthesizing the aforementioned timeliness correction value with the original accessibility confidence parameter, a comprehensive weight matrix that integrates spatial connectivity and time decay effects is obtained.

[0128] A vectorization and recombination operation is performed on the comprehensive weight matrix, and the order of arrangement is based on the anchor point type classification and spatial distribution location index to form a dynamic time decay weight vector that characterizes the freshness and reliability of environmental information.

[0129] By using exponential decay function weighting and observation interval correction, the key anchor point reachability feature set from the previous step is transformed into a dynamic time decay weight vector with spatiotemporal consistency constraints, thereby enabling real-time updates of node weights during the semantic decision tree construction process.

[0130] For example, in a low-altitude flight scenario within an city, the key anchor point reachability feature set contains 50 anchor points, each with a spatial connectivity status marker and an reachability confidence parameter ranging from 0 to 1. The sensor observation interval is 0.5 seconds, and the time constant τ is set to 2 seconds. For each anchor point, Δt is calculated for different cases such as 0.5 seconds, 1.0 seconds, and 1.5 seconds, and then applied to the formula to obtain, for example, attenuation coefficients of ≈0.7788 for Δt=0.5 seconds, ≈0.6065 for Δt=1.0 seconds, and ≈0.4724 for Δt=1.5 seconds. The attenuation coefficients are multiplied by the initial reachability confidence parameters (e.g., 0.85, 0.92, 0.76) to obtain a timeliness correction value, which is then linearly synthesized with the original confidence parameters at a ratio of 0.6:0.4 to form a comprehensive weight value. All anchor point weights are sorted by type (e.g., "platform", "occlusion", "reflection") and spatial location index to form a dynamic time-decaying weight vector. In the subsequent semantic decision tree construction, this weight vector guides the node connection priority, ensuring that anchor points with recent observations and high stability are preferentially included in path generation, effectively improving the spatiotemporal consistency of the decision tree and the reliability of obstacle avoidance paths.

[0131] S4.3: Establish a root node instance based on the task endpoint coordinate data, use the unobstructed passage probability value in the key anchor point reachability feature set as the branch connection criterion, and use reverse topology growth to expand the node connection relationship layer by layer from the root node to the local observation area, and map key anchor points with high reachability confidence parameters as intermediate branch nodes to generate a preliminary global target reverse traction topology connection map.

[0132] Using the task endpoint coordinate data as initial input, the coordinate information is defined as a root node instance in the topology construction module, and the starting point identifier of the global target in the semantic connection structure is established.

[0133] For each anchor point in the set of key anchor point reachability features, the unobstructed passage probability value is used as the criterion for branch connection between nodes. Based on the combination of this probability value and the spatial connectivity status label, anchor points that meet the minimum connectivity requirements are selected as candidate expansion objects.

[0134] During the reverse topology growth process, the root node instance is used as the starting point for topology generation. The node connection relationship is extended layer by layer to the spatial range of the local observation area through recursive expansion. Each layer of extension maps the key anchor point that satisfies the high reachability confidence parameter to the intermediate branch node and records its spatial connection attribute with the parent node.

[0135] For each newly generated intermediate branch node, the topology growth rule is applied to determine whether there is a further reachable downstream anchor point. If there is, the anchor point is included in the next round of node expansion calculation, thereby forming a preliminary multi-level, multi-branch topology connection structure.

[0136] By hierarchically expanding the topology generation, key highly accessible anchor points are accurately located in the spatial semantic connectivity graph to construct a preliminary global target topology connectivity graph that pulls the target from the destination to the local observation area, thus realizing the basic semantic skeleton required for path evaluation.

[0137] By using the reverse topology growth method, the reachability feature set of the previous step is transformed into a preliminary global target reverse traction topology connection graph containing node hierarchical relationships and spatial connection attributes, thereby realizing the construction of a semantic guidance path skeleton from key anchor points to the task endpoint.

[0138] For example, in an urban canyon flight scenario, the mission endpoint coordinates are defined as (125.4, 68.2, 15.0) meters. The key semantic anchor point reachability feature set includes 8 anchor points with unobstructed passage probabilities of 0.92, 0.85, 0.61, 0.40, 0.75, 0.89, 0.95, and 0.58, respectively. The branch connection criterion is set to a probability value ≥ 0.8, and anchor points 1, 2, 6, and 7 are selected for topology expansion. The Euclidean distance between the root node and anchor point 7 is calculated using the formula:

[0139]

[0140] Where x, y, and z are the coordinate components of the task endpoint, and x', y', and z' are the coordinate components of anchor point 7. The calculated distance is 21.6 meters, which is lower than the set threshold of 30 meters. Therefore, a connection edge is directly established from the root node to anchor point 7. Following the reverse topology growth, the downstream anchor points of anchor point 7 are further expanded. The connectivity between anchor point 6 and anchor point 7 is marked as True, and the reachability probability is 0.89, meeting the expansion conditions. A connection edge is then established from anchor point 7 to anchor point 6, ultimately generating a preliminary topology connection graph containing the root node, two layers of intermediate nodes (anchor point 7 and anchor point 6), and their spatial connectivity attributes. Verification results show that this topology structure can significantly improve the accuracy of path evaluation in this scenario and provides a precise skeletal foundation for subsequent edge weight assignment and decision tree optimization.

[0141] S4.4: For each node connection edge in the preliminary global target reverse traction topology connection graph, perform edge weight assignment operation by integrating the corresponding weight coefficient in the dynamic time decay weight vector, eliminate weak connection branches below the preset connectivity threshold and optimize the parent-child hierarchical relationship between nodes, so as to generate a semantic decision tree structure with spatiotemporal consistency constraints.

[0142] Based on the established node connection relationships and corresponding reachability confidence parameters in the preliminary global target reverse traction topology connection graph, the weight coefficients in the dynamic time decay weight vector are called as timeliness constraint inputs. Edge weight assignment calculations are performed on the connection edges of each node, and the reachability probability value and the time decay weight coefficient are multiplied to generate a comprehensive connectivity strength index.

[0143] The comprehensive connectivity strength index is compared with the preset connectivity threshold. Weak connection branches with insufficient continuity or low timeliness are eliminated. After the deletion operation, the parent-child hierarchy of the remaining connection edges is rearranged to ensure that the high confidence anchor points in the topology are at a higher level and that the reachability and timeliness of the path between the terminal leaf node and the root node meet the preset standards.

[0144] A topology consistency optimization operator is used to verify the connectivity of the adjusted parent-child relationship, correct the isolated nodes caused by branch removal, and ensure the integrity of the reverse traction path from the global target to the local key anchor point by supplementing connections or merging nodes.

[0145] Hierarchical recursive optimization is performed using the edge weight data of the nodes to ensure that the cumulative edge weight of each path is in the optimal state in terms of both spatial connectivity and time decay. At the same time, the parent node references and child node lists of the nodes are updated to form a semantic decision tree structure with spatiotemporal consistency constraints.

[0146] Through the above processing method, the preliminary topology connection graph is transformed into a data structure that meets the decision-making requirements in terms of both spatial reachability and temporal freshness, realizing an efficient semantic guidance skeleton that can be directly called by the path evaluation module.

[0147] S4.5: Based on the semantic decision tree structure, extract the complete path skeleton information from the leaf node to the root node, encapsulate the node type definition, relative reachability parameter and time decay factor attribute, perform data structure serialization and packaging operation to generate a semantic guidance skeleton structure for evaluating path compliance and output it to the subsequent scoring module.

[0148] Step S5: Input the path anchor sequence into the semantic guidance skeleton structure, calculate the matching score of each candidate local path segment in three dimensions: anchor type consistency, spatial topological rationality, and temporal evolution continuity, to generate a semantic matching score vector representing the degree of alignment between each candidate local path segment and the global task intent. Specifically, this includes:

[0149] S5.1: Based on the geometric semantic mixed entity type labels in the path anchor point sequence and the node type definitions in the semantic guidance skeleton structure, perform type feature comparison processing to generate initial evaluation parameters of anchor point type consistency that characterize the consistency between the candidate local path segment and the global target expectation in terms of semantic category.

[0150] S5.2: Using the initial evaluation parameters of anchor point type consistency as basic constraints, and combining the relative pose coordinate data of each geometric-semantic hybrid entity in the real-time semantic anchor point sequence, a spatial topological rationality evaluation index is generated to characterize whether the candidate local path segment conforms to the implicit spatial avoidance logic.

[0151] Using the initial evaluation parameters of anchor point type consistency as constraints, the set of three-dimensional relative pose coordinates of each geometric-semantic hybrid entity in the real-time semantic anchor point sequence is imported into the spatial topology analysis module.

[0152] The spatial connectivity construction operator is invoked to generate a connectivity graph of local path candidate segments in three-dimensional space based on the Euclidean distance and direction vector between each pair of anchor points in the sequence.

[0153] An implicit space avoidance logic rule set is introduced, and the path segments in the connectivity graph are compared with the rule set one by one to filter out the node edge connection relationships that violate the preset avoidance order or detour strategy.

[0154] By utilizing line-of-sight occlusion detection and combining it with the anchor point envelope model, collision prediction is performed on the spatial interaction between path segments and potential occlusions, and path segments with obstructed lines of sight or that cannot be safely passed are marked.

[0155] Based on the remaining valid connections, the overall spatial topological rationality index of the path is calculated, and the score of the path segment in terms of compliance with avoidance logic is quantified by weight accumulation.

[0156] Through the above chain processing method, the anchor point consistency result of the previous step is transformed into a quantitative spatial topological rationality evaluation index, so as to determine whether the candidate path conforms to the implicit avoidance logic.

[0157] S5.3: Based on the key geometric semantic hybrid entities locked by the spatial topology rationality evaluation index, call the trajectory records of the corresponding unique spatiotemporal identifiers in the historical frame data, and perform temporal smoothness difference operation to generate a time evolution continuity metric value that characterizes the stability of the motion trend of candidate local path segments within a continuous observation period.

[0158] Based on the key geometric semantic hybrid entities identified by the spatial topology rationality evaluation index, the trajectory records with corresponding unique spatiotemporal identifiers in the historical frame data are called, and the trajectory point sequences are sorted and indexed under a unified global time reference to form a stable trajectory dataset within a continuous observation period.

[0159] The stable trajectory dataset is subjected to position coordinate interpolation completion processing. Missing points are filled in using a cubic spline-based interpolation model to ensure that the trajectory sequence is uninterrupted in the time domain and to provide a uniform sampling interval constraint for subsequent difference operations.

[0160] The changes in position coordinates within each adjacent time window are calculated on the completed trajectory sequence, and the trajectory displacement sequence is extracted using the Euclidean distance formula.

[0161] Perform temporal smoothness difference operations on the displacement sequence to construct a continuity metric calculation model.

[0162] By normalizing the smoothness difference values ​​to a fixed interval, a time evolution continuity measure that characterizes the stability of the motion trend of candidate local path segments within a continuous observation period is obtained.

[0163] By using the above differential and normalization processing methods, the spatial topological rationality assessment results of the previous step are transformed into a time evolution continuity measurement index, thereby achieving the technical effects of path trend stability quantification and multi-dimensional matching degree calculation.

[0164] S5.4: Based on the initial evaluation parameters of anchor point type consistency, the evaluation index of spatial topology rationality, and the measure of temporal evolution continuity, a multi-dimensional weighted fusion model is constructed and normalized aggregation calculation is performed to generate a semantic matching score vector that ultimately represents the degree of alignment between each candidate local path segment and the global task intent.

[0165] Based on the initial evaluation parameters of anchor point type consistency, the evaluation index of spatial topology rationality, and the measure of temporal evolution continuity, before constructing the multidimensional weighted fusion model, the three types of indicators are sequentially input into the weighted coefficient allocation module. The weight calibration function is called to set the initial weight matrix according to the contribution of each indicator to the task success rate in historical path selection. Normalization is performed on the initial weight matrix, and each weight value is adjusted proportionally to the total weight sum, ensuring that the weighted sum of the three types of indicators remains constant. The multidimensional fusion operation unit is called to perform product operations on the normalized weights and indicator values ​​within the same calculation cycle, and the product results are input into the accumulator to complete vector summation. Normalized aggregation calculation is performed on the fusion results output by the accumulator. By setting the denominator to the sum of squares of the weight coefficients of the three types of indicators plus a bias term, numerical offsets under different indicator dimensions are eliminated. The following weighted fusion formula is used to calculate the matching degree score:

[0166]

[0167] Where D is the semantic matching score, T is the initial evaluation parameter for anchor type consistency, S is the spatial topological rationality evaluation index, C is the temporal evolution continuity measure, and w t w s w c These are the normalized weights for the three categories of indicators. The normalized aggregation result is encapsulated as a single element of the semantic matching score vector, and the entire score vector is filled sequentially according to the candidate local path segment index order. Through multi-dimensional weighted fusion and normalized aggregation processing, the multi-dimensional evaluation results of the previous step are transformed into matching score data under a unified scale, realizing the quantitative integration of path scoring and accurate alignment determination.

[0168] For example, in an urban low-altitude logistics flight scenario, let the initial evaluation parameter T for anchor point type consistency be 0.82, the spatial topology rationality evaluation index S be 0.75, the temporal evolution continuity metric C be 0.68, and the weight be set to w. t =0.4, w s =0.35, w c=0.25. Normalize each weight to keep the sum at 1, and substitute it into the fusion formula:

[0169]

[0170] The calculated fusion result was 0.767, the normalized baseline value was 0.574, and the final matching score D was 1.337. This score was filled into the matching vector of the target path, and compared with the scores of other candidate paths. It was found that this path significantly improved the output score in the current scenario, verifying the superiority of multi-dimensional weighted fusion and normalized aggregation, which greatly improved the path selection accuracy of the aircraft under limited perception conditions.

[0171] Step S6: Based on the semantic matching score vector, identify conflict candidate paths with functional redundancy and similar scores. Call the stability records of similar geometric semantic hybrid entities in the historical memory cache under similar environmental conditions, and perform a credibility reassessment process on the key semantic anchors causing ambiguity to generate a conflict resolution decision result after dynamic weight correction. Specifically, this includes:

[0172] S6.1: Based on the semantic matching score vector, perform threshold comparison and variance analysis to identify a set of conflict candidate paths whose score difference is less than a preset tolerance range and whose spatial topological overlap is higher than a set threshold, and extract a list of key semantic anchors that cause ambiguity in path selection as input objects for subsequent evaluation.

[0173] Based on the semantic matching score vector, a dual judgment condition matrix for screening conflict candidate paths is constructed by loading preset tolerance threshold parameters and spatial topological overlap threshold parameters.

[0174] The semantic matching score vector is subjected to element-wise difference calculation. The score difference of each pair of candidate paths is used as input, and the threshold comparison operator is used to check whether it is lower than the preset tolerance range to generate a score difference judgment mark matrix.

[0175] Spatial overlap calculation is performed on the set of anchor point trajectories of the candidate paths in three-dimensional space. A path overlap matrix is ​​generated by ratio operation based on spatial envelope intersection. A threshold comparison operator is used to check whether the spatial overlap of each pair of paths is higher than a set threshold to generate a spatial overlap judgment mark matrix.

[0176] Perform a logical AND operation between the score difference determination mark matrix and the spatial overlap determination mark matrix to generate a set of path pairs that simultaneously satisfy the conditions of low tolerance for score difference and high threshold for spatial overlap.

[0177] For the set of path pairs, the sequence of path anchor points involved in each path is extracted and common anchor point identification processing is performed. Geometric and semantic hybrid entities that appear simultaneously in the path pairs are selected as candidate sets of key semantic anchor points, and the final list of key semantic anchor points is generated through frequency statistics and type uniqueness processing.

[0178] By using threshold comparison and variance analysis, the results of the previous step are transformed into a set of conflict candidate paths and a list of key semantic anchors, thus preparing the input data for the subsequent credibility reassessment process.

[0179] For example, considering a low-altitude logistics aircraft performing a mission in an urban canyon environment, the input semantic matching score vector is [0.865, 0.872, 0.879, 0.920], with a preset tolerance threshold of 0.010 and a spatial topological overlap threshold of 0.75. The score difference between path 1 and path 2 is calculated, yielding a difference of 0.007, which is below the tolerance threshold and is marked as meeting the condition. The spatial envelope intersection volume ratio between path 1 and path 2 is 0.80, which is above the set threshold and is also marked as meeting the condition. A logical AND operation is used to identify path 1 and path 2 as conflict candidate path pairs. The path anchor point sequences of path 1 and path 2 are extracted. Common anchor points include "rooftop edge segment" (ID=102) and "ventilation duct opening" (ID=215), each appearing once and uniquely typed, generating a key semantic anchor point list [102, 215]. In practical applications, this list was used for historical stability record retrieval in S6.2. Verification results showed that anchor point 102 had high stability under similar lighting conditions, while anchor point 215 exhibited significant stability fluctuations. This output allows the subsequent weight correction stage to reduce false optimal path scores caused by the instability of anchor point 215, thereby significantly improving the accuracy of conflict resolution.

[0180] S6.2: For each geometric semantic hybrid entity in the key semantic anchor list, use its unique spatiotemporal identifier as the retrieval key to query and extract the historical stability record sequence of the same type of geometric semantic hybrid entity under similar lighting conditions, observation angles and weather environments in the historical memory cache, so as to construct a priori confidence benchmark library for current state comparison.

[0181] The historical memory cache is a dedicated database used in this invention to persistently store the historical stability records of geometric-semantic hybrid entities (i.e., key semantic anchors) under different environmental conditions. This cache uses a unique spatiotemporal identifier as an index and records, in time series, the positional fluctuations, feature intensity changes, confidence labels, and corresponding environmental condition parameters (light intensity, observation angle, weather conditions, etc.) of each entity during past observation periods.

[0182] The history cache uses a hybrid storage structure combining key-value pairs and relational tables, and includes the following core data tables:

[0183] Entity Index Table: Using a unique spatiotemporal identifier as the primary key, it stores the type label, first discovery time, last update time, and current active status of entities with mixed geometric and semantic characteristics.

[0184] Stability record table: Each record corresponds to one historical observation event and contains the following fields:

[0185] Unique spatiotemporal identifier (foreign key);

[0186] Observation timestamp (precision in milliseconds);

[0187] Position fluctuation value (unit: meters, representing the average offset of the geometric center of an entity across consecutive frames).

[0188] Feature intensity change rate (unit: percentage per minute, e.g., rate of change of reflection intensity or texture gradient).

[0189] Confidence level label (dimensionless, ranging from 0 to 1, indicating the reliability of the observation);

[0190] Environmental condition parameters: light intensity (lux), observation angle (angle with the normal to the entity, degrees), weather conditions (discrete values ​​such as sunny / cloudy / light fog / rain).

[0191] Statistical characteristic table: Grouped by entity identifier and environmental conditions, pre-stores historical mean, variance, sample count and most recent update time to accelerate prior confidence calculation.

[0192] Construction method:

[0193] Offline initialization: During the system deployment phase, stability records for each geometric-semantic hybrid entity are extracted using existing historical flight datasets (such as simulated flight logs and actual road test records). After cleaning and deduplication, these records are filled into the cache. For missing environmental condition fields, they are filled in using sensor data interpolation.

[0194] Online incremental construction: During each flight mission, after step S2 generates a real-time semantic anchor sequence with confidence labels, the system writes the entities in the sequence and their observation data into the cache; if the entity already exists, the new observation record is appended to the stability record table, and the mean and variance in the statistical feature table are updated.

[0195] Environmental condition index: A multidimensional hash index is built for each entity according to the combination of environmental conditions (e.g., light range, angle range, weather type) to ensure that historical records under similar conditions can be returned in sub-milliseconds during the retrieval in step S6.

[0196] Lifecycle management: Archives or deletes entity records that exceed the preset storage duration (e.g., 30 days) and have a sample count below a threshold to prevent cache bloat. Manual cleanup of inactive entities is also supported.

[0197] Dynamic updates and maintenance: After each task, step S8 incrementally updates the cache based on actual execution feedback, gradually converging the historical data to a true stable distribution. Online hot updates are supported; cached data resides in memory, and a read-write lock mechanism ensures data consistency during concurrent multi-threaded access. An asynchronous persistence strategy can be configured to periodically write in-memory data to non-volatile storage to prevent data loss.

[0198] Through the above design, the historical memory cache provides a cumulative and traceable experience knowledge base for low-altitude aircraft to avoid obstacles in densely built-up areas, which significantly improves the robustness of conflict resolution and path decision-making and the ability to adapt to the environment.

[0199] For each geometric-semantic hybrid entity in the key semantic anchor list, a unique spatiotemporal identifier is used as a high-precision retrieval key. Type matching and condition filtering are performed in the historical memory cache data index structure to lock the corresponding historical record index range. Relying on the global cache index table, similar environmental conditions are filtered by combining semantic category labels. The combination parameters of lighting conditions, observation angle, and weather environment are limited to a preset tolerance window to eliminate the interference of non-matching scene data on subsequent comparison results. For the filtering results, the historical stability data reading module is called to extract the continuous observation state sequence of the geometric-semantic hybrid entity under the target conditions in batches. This sequence includes the evolution record of position fluctuation value, feature intensity change rate, and confidence label over time. The extracted stability record sequence is subjected to intra-time window statistical calculation. A prior confidence description vector is constructed using multi-dimensional statistical features such as mean, standard deviation, and fluctuation range. This vector completely describes the state stability benchmark of the target anchor under a specific environmental combination. The aforementioned prior confidence description vectors are output to the prior library construction module, archived by unique spatiotemporal identifiers, and their mapping relationship with environmental conditions is maintained to form a prior confidence benchmark library for comparison with the current state, thereby achieving effective connection between historical states and current observations.

[0200] By employing precise retrieval, conditional filtering, batch extraction, and statistical feature construction based on unique spatiotemporal identifiers, the list of key semantic anchors identified in the previous step is transformed into a priori confidence benchmark library containing multidimensional stability indicators, thus providing a basis for comparing historical and current states in subsequent dynamic reassessments.

[0201] For example, in a low-altitude aircraft flight mission, a key semantic anchor point is the "glass curtain wall reflection area," with its unique spatiotemporal identifier being GMA-2024-015. In this embodiment, the historical memory cache contains observation records from the past two years under different conditions. The illumination condition filter range is set to a brightness value between 550 and 650 lux, the observation angle tolerance is ±5 degrees, and the weather conditions are limited to clear skies and wind speeds below 4 m / s. Matching historical record sequences are obtained through index retrieval. Each record contains three types of parameters supporting the calculation: position fluctuation value (in meters), reflection intensity change rate (in percentages per minute), and historical confidence label (dimensionless). In a record sequence with a time window of 300 seconds, the calculated mean position fluctuation is 0.08 meters with a standard deviation of 0.02 meters, the mean reflection intensity change rate is 3.5 with a fluctuation range of 1.2 to 5.0, and the mean historical confidence label is 0.92. Using these statistical features, a prior confidence description vector of [0.08, 0.02, 3.5, 1.2, 5.0, 0.92] is constructed, stored in a priori library, and mapped to the identifier GMA-2024-015 and the aforementioned combination of environmental conditions. This priori library is invoked in subsequent step S6.3, and multi-factor weighted fusion calculation is performed with the real-time observed confidence labels. This significantly improves the accuracy of the stability reassessment of the glass curtain wall reflective zone under the current clear sky and low wind speed conditions, thereby effectively supporting path conflict resolution decisions.

[0202] S6.3: Based on the prior confidence benchmark library and the current real-time sensor confidence labels of the key semantic anchors, perform multi-factor weighted fusion calculation processing, and generate a dynamic reassessment confidence value that represents the reliability of each key semantic anchor in the current environment by comparing the consistency between historical fluctuation characteristics and current instantaneous characteristics.

[0203] Based on the prior confidence benchmark library and the current real-time sensor confidence labels of the key semantic anchors, an input vector for the confidence fusion calculation model is constructed, where each element consists of the historical stability mean, historical stability variance, and real-time confidence label.

[0204] The historical stability fluctuation characteristics are converted into frequency domain energy distribution indicators, and the energy peak positions and relative amplitudes of each key semantic anchor point in the historical sequence are calculated by fast Fourier transform.

[0205] Real-time sensor confidence labels are mapped to short-term stability indices for time series, and the sliding window method is used to extract the confidence fluctuation amplitude within the window as a reference for instantaneous stability.

[0206] By using multi-factor weighted fusion, weight coefficients are assigned to historical mean, historical fluctuation range and real-time stability respectively. The weight coefficients are determined according to the environmental similarity calculation formula. The environmental similarity is comprehensively evaluated by cosine similarity and matching degree of illumination and observation angle.

[0207] The calculated environmental similarity is used as a weight adjustment factor in the multi-factor weighted fusion formula:

[0208]

[0209] Where C is the dynamic reassessment confidence score, H is the historical mean stability, V is the historical fluctuation stability, R is the real-time stability, and α, β, and γ are the weighting coefficients adjusted for environmental similarity.

[0210] The above calculations yield a dynamic reassessment confidence value that reflects the consistency between historical fluctuation characteristics and current instantaneous characteristics. This value will then be used as the basis for reliability in subsequent path weight adjustments.

[0211] By using a multi-factor weighted fusion calculation method, the prior confidence benchmark library formed in the previous step and the current real-time confidence labels are transformed into dynamic re-evaluation confidence indicators, thereby realizing a quantitative assessment of the reliability of key semantic anchors in the current environment.

[0212] S6.4: Based on the dynamic reassessment confidence value, the weights of the path anchor sequence contained in each path in the conflict candidate path set are dynamically adjusted to reduce the scoring weight of path branches composed of high-volatility key semantic anchors and increase the scoring weight of path branches composed of low-volatility key semantic anchors, so as to generate a weight-optimized path competition ranking result.

[0213] The input objects are the dynamic reassessment confidence vector generated in step S6.3 and the sequence of path anchors in the conflict candidate path set.

[0214] For each path anchor sequence in the conflict candidate path set, the key semantic anchor identifiers contained therein are parsed and indexed and matched with the dynamic re-evaluation confidence value vector, and the confidence value of each anchor is mapped to the path score weight correction factor matrix.

[0215] The elements corresponding to the high volatility anchors in the correction factor matrix are subjected to weight decay calculation. The high volatility anchor score weights are adjusted by decreasing the product of their confidence value and the preset decay coefficient using an exponential decay model. The decay coefficient is set based on the historical stability fluctuation amplitude.

[0216] Weight enhancement operations are performed on the elements of low volatility anchors. A linear enhancement model is used to incrementally adjust the low volatility anchor score weights by multiplying their confidence values ​​by a preset enhancement coefficient, where the enhancement coefficient is set based on the historical consistency level.

[0217] Perform a normalization operation on the weight factor matrix after attenuation and enhancement processing.

[0218] The normalized weight vector is element-wise multiplied with the original path score vector to calculate the corrected path score.

[0219] Through the above dynamic weight correction and fusion operation, the dynamic reassessment confidence value generated in the previous step is transformed into an optimized path competition ranking result, thereby enhancing the reliability of conflict path scoring and improving selection robustness.

[0220] For example, in a low-altitude flight environment in an urban canyon, the conflict candidate path set contains 3 paths, each with an anchor point sequence length of 5. The dynamic reassessment confidence values ​​are set to [0.35, 0.78, 0.60, 0.40, 0.85], where 0.35 and 0.40 correspond to high volatility anchor points, and 0.78, 0.60, and 0.85 correspond to low volatility anchor points. The attenuation coefficient for high volatility anchor points is set to 0.5, and the enhancement coefficient for low volatility anchor points is set to 1.2. The calculated attenuated weight vector is [0.175, 0.936, 0.72, 0.20, 1.02], and the normalized weight vector is [0.067, 0.358, 0.276, 0.077, 0.392]. The original path score vector is [2.0, 3.5, 3.2, 1.8, 4.0]. After element-wise multiplication, the corrected path score components are [0.134, 1.253, 0.883, 0.138, 1.568], with a total of 3.976. After performing the same process on each path, the corrected score shows that the second path has the highest score, and its key anchor point confidence is stable. The output is the path competition ranking result after weight optimization, ensuring that the aircraft performs obstacle avoidance actions along paths with high stability in this scenario, significantly improving mission completion rate and safety.

[0221] S6.5: Based on the path competition ranking result after weight optimization, perform maximum value filtering and logical verification processing to eliminate false high-scoring paths caused by the instability of key semantic anchors, and finally output the conflict resolution judgment result with the highest credibility and verified by conflict resolution, as the sole decision basis for determining the target local path.

[0222] Based on the weighted and optimized path competition ranking results, the path priority filtering module is invoked to perform maximum value extraction operations on the score vectors of all candidate paths, locking the path index corresponding to the peak score and obtaining its anchor point sequence attribute set. Combining the path matching score and the dynamic confidence value of key semantic anchor points, the logical consistency verification operator is invoked to verify the stability and coherence of each anchor point in the peak path, and the path stability verification value is calculated.

[0223] The path stability check value is compared with a preset stability threshold. If it is lower than the threshold, the path is discarded, and the second-highest-scoring path is located in the ranking results. This stability check process is repeated until a path that meets the stability threshold is found. For paths that pass the stability check, the spatial topology legality verification module is invoked to perform a global goal consistency verification on the anchor point order and pose relationship, verifying that the path does not violate the spatial constraints of the global semantic guidance skeleton. Based on the verification results, a conflict resolution decision with the highest confidence level that passes the conflict resolution verification is output.

[0224] By combining maximum value extraction, stability verification, and spatial topology review, the path competition ranking result after weight optimization is transformed into a unique target local path decision index, achieving the expected technical effects of conflict path elimination and high-confidence path selection.

[0225] Step S7: Based on the conflict resolution decision result, the target local path with the highest score and that has passed verification is selected, converted into aircraft control commands, and output to drive the low-altitude aircraft to complete obstacle avoidance maneuvers according to the target local path. Specifically, this includes:

[0226] S7.1: Based on the path anchor point sequence contained in the target local path and its corresponding spatial coordinate reference, the discrete path anchor point sequence is smoothed using cubic spline interpolation to generate a smooth reference trajectory curve with continuous position, velocity and acceleration characteristics.

[0227] S7.2: Based on the curvature characteristics of the smooth reference trajectory curve at the current moment and the dynamic constraints of the aircraft, the theoretical attitude angle command and theoretical thrust command required to maintain trajectory tracking are calculated using the model predictive control algorithm to generate a basic control vector containing roll angle, pitch angle and total thrust components.

[0228] Based on the curvature characteristics of the smooth reference trajectory curve at the current moment and the aircraft dynamic constraints, a trajectory tracking error state vector is constructed and input into the model predictive control solution module. The curvature characteristics are discretized, and the rate of change of the trajectory tangent vector and normal vector is calculated as a direct input for attitude adjustment requirements. The aircraft dynamic parameter library is called to extract the maximum roll rate, maximum pitch rate, and thrust change response time constant, forming a constraint parameter set. Based on the constraint parameter set and the error state vector, the state evolution model within the predicted horizon length is solved, generating a pose prediction set for multiple future sampling moments. A trajectory error cost function is evaluated for each predicted pose, with the cost function based on the weighted sum of attitude deviation, position deviation, and thrust deviation. Based on the optimization result of minimizing the cost function, the control solutions for the first control cycle of the predicted horizon are solved, including roll angle commands, pitch angle commands, and total thrust components. This solution is then checked for compliance with preset dynamic constraints, and out-of-limit values ​​are eliminated and the solution is re-solved to ensure execution safety. The model predictive control algorithm transforms the trajectory smoothing result from the previous step into theoretical attitude angle and thrust commands that conform to the dynamic constraints of the aircraft, thereby achieving optimal suppression of trajectory tracking error within the prediction range.

[0229] For example, in a low-altitude logistics vehicle trajectory tracking task in an urban canyon scenario, the maximum curvature of the smooth reference trajectory curve is set to 0.08 m⁻¹, and the vehicle dynamics constraints include a maximum roll rate of 150 degrees / s, a maximum pitch rate of 120 degrees / s, and a thrust change response time constant of 0.15 s. A state evolution model with a predicted horizon length of 20 steps is constructed, assigning weights of 0.6 and 0.3 to the trajectory position deviation and attitude deviation, respectively, and a weight of 0.1 to the thrust deviation. The current error state vector is input, the predicted pose set is calculated, and a cost function is evaluated for each pose according to the above formula, resulting in the lowest cost for the predicted pose in step 4. The control solution for this step is extracted, outputting a roll angle command of 4.5 degrees, a pitch angle command of 2.8 degrees, and a total thrust component of 38.2 N. Compliance verification showed that all values ​​were within the constraints. When the combination of commands was applied to drive the aircraft to perform obstacle avoidance maneuvers, the verification results showed that the trajectory deviation range was significantly reduced under gust interference conditions, the path tracking accuracy was significantly improved, and the dynamic response process was stable.

[0230] S7.3: To address the execution deviations that may occur in the basic control vector due to gusts of wind in the building area, an adaptive gain scheduling mechanism is introduced to dynamically compensate and correct the basic control vector, so as to generate a final aircraft control command set with anti-disturbance capability.

[0231] For the attitude angle and thrust components in the basic control vector, a gust interference prediction model is constructed, using the instantaneous wind speed vector and turbulence intensity index output from the local wind speed sensor and airflow disturbance monitoring module in the building area as the interference calculation benchmark. Based on the interference prediction model, the deviation amplitude and phase delay index of each control component under the action of the wind speed vector are calculated using a dynamic response function, and the corresponding deviation vector is generated. For the deviation vector, adaptive gain scheduling is used to adjust the gain coefficient in real time. The gain scheduling benchmark depends on the aircraft attitude error vector, thrust error vector, and the rate of change of the interference, forming a control adjustment coefficient matrix. Matrix weighted operations are used to fuse the control adjustment coefficient matrix and the basic control vector element-wise, correcting the amplitude and direction of each control component to offset the instantaneous execution deviation caused by gust interference in the building area. The stability criteria are used to verify the disturbance resistance capability of the modified control command set. Adjustment schemes that do not meet the preset stability threshold are eliminated and command combinations with excellent disturbance resistance performance are retained. Through an adaptive gain scheduling mechanism, the results of the previous step are transformed into the final aircraft control command set with the characteristics of resisting instantaneous gust interference, so as to achieve stable obstacle avoidance performance in low-altitude dense building environment.

[0232] S7.4: The final aircraft control command set is transmitted to the motor drive unit in the dynamic control execution module through the airborne flight control bus to drive each rotor motor of the low-altitude aircraft to rotate differentially according to the command speed and complete obstacle avoidance maneuvers.

[0233] Step S8: Based on the execution feedback information of the target local path, update the state of the terminal node of the semantic decision tree and the stability record in the historical memory cache to generate an optimized semantic guidance skeleton structure and a confidence prior library. Specifically, this includes:

[0234] S8.1: Acquire multi-source sensor telemetry data and external environment observation sequence during the flight of the low-altitude aircraft along the local path of the target, and use spatiotemporal registration to map the multi-source sensor telemetry data and the external environment observation sequence to a unified world coordinate system to generate an execution feedback dataset containing actual trajectory deviation and environmental feature changes.

[0235] S8.2: Based on the execution feedback dataset, extract the measured pose information and confidence fluctuation curves of key geometric-semantic hybrid entities, and use the Bayesian filtering estimation method to smooth the measured pose information and confidence fluctuation curves to generate anchor point state correction vectors that characterize the true stability of each semantic anchor point under the current environmental conditions.

[0236] S8.3: Identify the terminal nodes in the semantic decision tree affected by execution deviations based on the anchor point state correction vector, and adaptively adjust the connection weights and time decay factors of the terminal nodes using the graph topology reconstruction operator to generate an optimized semantic guidance skeleton structure that reflects the latest environmental reachability relationship.

[0237] S8.4: Call the historical stability record in the historical memory cache that matches the key geometric semantic hybrid entity type, and use the incremental learning strategy to fuse the anchor point state correction vector into the historical stability record to generate a confidence prior library containing cross-time environmental adaptation characteristics.

[0238] Retrieve historical stability records from the historical memory cache that match the current key geometric semantic hybrid entity type. Each record contains the entity's position fluctuation value, feature intensity change rate, confidence label, and corresponding environmental condition parameters (light intensity, observation angle, weather conditions, etc.) at the historical observation time.

[0239] The anchor point state correction vector output from step S8.2 is used as the new sample for the current observation period. This vector represents the actual stability performance of the geometric-semantic hybrid entity under the current environmental conditions. An exponentially weighted moving average method is employed to fuse this new sample with the corresponding statistics in the historical stability record, updating the historical mean and historical variance. This results in recent observation data receiving a higher fusion weight, while older data gradually loses influence through exponential decay. Specifically, the new mean is obtained by weighting the old mean and the current sample value using a preset forgetting factor, and the new variance is updated based on the old variance, the deviation between the current sample and the new mean, and the same forgetting factor.

[0240] Once a sufficient number of fused samples of the same type of entity have been accumulated under various environmental conditions, the updated statistics will be grouped and stored according to the combination of environmental conditions (lighting range, observation angle range, weather type), forming a mapping table from multidimensional environmental states to stable priors. Simultaneously, each group of prior data will be appended with a sample count and a timestamp for subsequent confidence smoothing and aging-out.

[0241] Finally, the updated group statistics are merged with the original prior library, and invalid entries with insufficient sample size or outdated data that have exceeded their lifespan are removed, resulting in a confidence prior library that includes cross-time period and multi-environment adaptability characteristics.

[0242] S8.5: Based on the optimized semantic guidance skeleton structure and the confidence prior library, perform a global consistency check operation to eliminate node state conflicts, output the final version of the semantic guidance skeleton structure and the highly available confidence prior library that adapts to the next moment's perception data parsing requirements, and complete the closed-loop update process.

[0243] This application provides a method for obstacle avoidance in densely built-up areas for low-altitude aircraft, which can be applied to, for example... Figure 4 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0244] In one embodiment, a low-altitude aircraft obstacle avoidance device for densely built-up areas is provided, comprising: a data acquisition module, a semantic parsing module, a path planning module, a decision tree module, a matching and scoring module, a correction module, an obstacle avoidance execution module, and a closed-loop update module, wherein:

[0245] Data acquisition module: acquires raw point cloud data and image frame data collected by airborne sensors of low-altitude aircraft, and performs spatiotemporal synchronization calibration on the raw point cloud data and image frame data to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference.

[0246] Semantic parsing module: Based on the multimodal local observation dataset, it performs online parsing operations to identify and abstract spatially stable geometric-semantic hybrid entities, and assigns a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity to generate a real-time semantic anchor sequence with attribute annotations.

[0247] Path planning module: Based on the current task endpoint coordinates and real-time semantic anchor point sequence, it plans and generates multiple candidate local path segments that meet the motion feasibility, and generates a path anchor point sequence based on the candidate local path segments.

[0248] Decision Tree Module: Using the task endpoint coordinates as the root node, and based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor point sequence, a semantic decision tree is dynamically constructed from the global goal in reverse order to generate a semantic guidance skeleton structure for evaluating path compliance.

[0249] Matching and scoring module: The path anchor sequence is input into the semantic guidance skeleton structure, and the matching score of each candidate local path segment is calculated in three dimensions: anchor type consistency, spatial topological rationality and temporal evolution continuity, so as to generate a semantic matching degree score vector that represents the degree of alignment between each candidate local path segment and the global task intent.

[0250] Correction module: Based on the semantic matching score vector, identify conflict candidate paths with functional redundancy and similar scores, call the stability records of the same type of geometric semantic mixed entities in the historical memory cache under similar environmental conditions, and perform a credibility re-evaluation process on the key semantic anchors that cause ambiguity to generate a conflict resolution decision result after dynamic weight correction.

[0251] Obstacle avoidance execution module: Based on the conflict resolution decision results, the target local path with the highest score and that has passed the verification is selected, converted into aircraft control commands and output, so as to drive the low-altitude aircraft to complete the obstacle avoidance maneuver according to the target local path.

[0252] Closed-loop update module: Based on the execution feedback information of the target local path, update the state of the terminal node of the semantic decision tree and the stability record in the historical memory cache to generate an optimized semantic guidance skeleton structure and a confidence prior library.

[0253] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for obstacle avoidance in densely built-up areas for low-altitude aircraft.

[0254] In one embodiment, a computer device is provided, the computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods in the above method embodiments.

[0255] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the methods described in the above method embodiments.

[0256] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0257] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0258] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0259] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for obstacle avoidance when low-altitude aircraft fly over densely built-up areas, specifically including: S1: Acquire raw point cloud data and image frame data collected by airborne sensors of low-altitude aircraft to generate a multimodal local observation dataset; S2: Based on a multimodal local observation dataset, identify and abstract geometric-semantic hybrid entities, assign a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity, and generate a real-time semantic anchor sequence; S3: Based on the current task endpoint coordinates and the real-time semantic anchor point sequence, plan and generate multiple candidate local path segments that meet the motion feasibility, and generate a path anchor point sequence based on the candidate local path segments; S4: Using the task endpoint coordinates as the root node, construct a semantic decision tree based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor point sequence, and generate a semantic guidance skeleton structure. S5: Input the path anchor sequence into the semantic guidance skeleton structure, calculate the matching score of each candidate local path segment in three dimensions: anchor type consistency, spatial topological rationality and temporal evolution continuity, and generate a semantic matching degree score vector. S6: Based on the semantic matching degree score vector, identify conflict candidate paths with functional redundancy and similar scores, call the stability records of the same type of geometric semantic mixed entities in the historical memory cache under similar environmental conditions, perform a credibility re-evaluation process on the key semantic anchors that cause ambiguity, and generate conflict resolution judgment results. S7: Based on the conflict resolution decision, select the target local path with the highest score and that has passed verification, convert it into aircraft control commands, and output it.

2. The obstacle avoidance method for low-altitude aircraft flying over densely built-up areas according to claim 1, characterized in that, The historical memory cache includes an entity index table, a stability record table, and a statistical feature table.

3. The obstacle avoidance method for low-altitude aircraft flying over densely built-up areas according to claim 1, characterized in that, Following S7, the following also includes: S8: Based on the execution feedback information of the target local path, update the state of the terminal node of the semantic decision tree and the stability record in the historical memory cache to generate an optimized semantic guidance skeleton structure and a confidence prior library.

4. The obstacle avoidance method for low-altitude aircraft flying over densely built-up areas according to claim 1, characterized in that, S3 specifically includes: The system acquires the coordinate data of the current task endpoint and the relative pose information of each geometric-semantic hybrid entity in the real-time semantic anchor sequence. It applies a gravitational potential field to the task endpoint and a repulsive potential field to each geometric-semantic hybrid entity to generate a composite potential field gradient map that characterizes the distribution of global target traction force and local obstacle repulsion force. Based on the composite potential field gradient map, a multi-branch trajectory search operation is performed in the unobstructed area. At the same time, the maximum turning radius, maximum climb rate and maximum acceleration of the aircraft are introduced as dynamic constraint boundary conditions for pruning to generate a set of discrete candidate local path segments that meet the kinematic feasibility requirements. For each discrete candidate local path segment in the discrete candidate local path segment set, a spatial projection overlap detection operation is performed to compare the path trajectory point cloud with the spatial envelope of each geometric semantic hybrid entity in the real-time semantic anchor point sequence in order to identify the list of target geometric semantic hybrid entities that each discrete candidate local path segment traverses or is adjacent to. Based on the target geometric-semantic hybrid entity list, extract the unique spatiotemporal identifier and semantic category label corresponding to each target geometric-semantic hybrid entity, and perform a serialization and arrangement operation according to the time sequence of the path travel direction to generate the original path anchor point sequence representing the path traversal order; Based on the original path anchor point sequence, path smoothing optimization is performed and the corresponding discrete candidate local path segment spatial coordinate data is bound to it. This converts the discrete path points into continuous executable trajectory descriptions, thereby generating a final standardized path anchor point sequence containing complete semantic anchor point type annotations and sequence information.

5. A method for obstacle avoidance in densely built-up areas for low-altitude aircraft according to claim 1, characterized in that, S4 specifically includes: The system acquires the coordinates of the current task endpoint and the relative pose information of key anchor points in the real-time semantic anchor point sequence. It then uses the spatial Euclidean distance calculation method and the line-of-sight occlusion detection operator to perform accessibility quantification evaluation. The system calculates the unobstructed passage probability value and spatial connectivity status label of each key anchor point relative to the task endpoint to generate a key anchor point accessibility feature set containing accessibility confidence parameters. Based on the spatial connectivity status markers in the set of key anchor point reachability features, the timestamp data of each key anchor point is subjected to time-sensitive weighting calculation, and the time decay weight coefficient that decreases over time is calculated in combination with the sensor observation interval, so as to generate a dynamic time decay weight vector that characterizes the freshness and reliability of environmental information. Based on the coordinate data of the task endpoint, a root node instance is established. The unobstructed passage probability value in the set of key anchor point reachability features is used as the branch connection criterion. The node connection relationship is expanded layer by layer from the root node to the local observation area. Key anchor points with high reachability confidence parameters are mapped as intermediate branch nodes to generate a preliminary global target reverse traction topology connection map. For each node connection edge in the preliminary global target reverse traction topology connection graph, the corresponding weight coefficient in the dynamic time decay weight vector is fused to perform edge weight assignment operation, weak connection branches below the preset connectivity threshold are eliminated and the parent-child hierarchical relationship between nodes is optimized to generate a semantic decision tree structure with spatiotemporal consistency constraints. Based on the semantic decision tree structure, the complete path skeleton information from the leaf node to the root node is extracted, the node type definition, relative reachability parameter and time decay factor attribute are encapsulated, and the data structure serialization and packaging operation is performed to generate a semantic guidance skeleton structure for evaluating path compliance and output it to the subsequent scoring module.

6. The obstacle avoidance method for low-altitude aircraft flying over densely built-up areas according to claim 1, characterized in that, S5 specifically includes: Based on the geometric semantic hybrid entity type labels in the path anchor point sequence and the node type definitions in the semantic guidance skeleton structure, type feature comparison processing is performed to generate initial evaluation parameters for anchor point type consistency. Using the anchor point type consistency initial evaluation parameters as basic constraints, and combining the relative pose coordinate data of each geometric-semantic hybrid entity in the real-time semantic anchor point sequence, a spatial topological rationality evaluation index is generated to characterize whether the candidate local path segment conforms to the implicit spatial avoidance logic. Based on the key geometric-semantic hybrid entities locked by the spatial topology rationality assessment index, the trajectory records with corresponding unique spatiotemporal identifiers in the historical frame data are called, and the temporal smoothness difference operation is performed to generate a temporal evolution continuity metric that characterizes the stability of the motion trend of candidate local path segments within a continuous observation period. Based on the initial evaluation parameters of anchor point type consistency, the spatial topology rationality evaluation index, and the temporal evolution continuity metric, a multi-dimensional weighted fusion model is constructed and normalized aggregation calculation is performed to generate a semantic matching score vector that ultimately represents the degree of alignment between each candidate local path segment and the global task intent.

7. A method for obstacle avoidance in densely built-up areas for low-altitude aircraft according to claim 6, characterized in that, The initial evaluation parameter for anchor point type consistency characterizes the consistency between the candidate local path segment and the global target expectation in terms of semantic category.

8. A flight obstacle avoidance device for low-altitude aircraft in densely built-up areas, characterized in that, The device includes: Data acquisition module: acquires raw point cloud data and image frame data collected by airborne sensors of low-altitude aircraft, and performs spatiotemporal synchronization calibration on the raw point cloud data and image frame data to generate a multimodal local observation dataset with a unified timestamp and spatial coordinate reference. Semantic parsing module: Based on the multimodal local observation dataset, it performs online parsing operations to identify and abstract spatially stable geometric-semantic hybrid entities, and assigns a unique spatiotemporal identifier and confidence label to each geometric-semantic hybrid entity to generate a real-time semantic anchor sequence with attribute annotations; Path planning module: Based on the current task endpoint coordinates and real-time semantic anchor point sequence, it plans and generates multiple candidate local path segments that meet the motion feasibility, and generates a path anchor point sequence based on the candidate local path segments; Decision tree module: Using the task endpoint coordinates as the root node, and based on the relative reachability and time decay weight of key anchor points in the real-time semantic anchor point sequence, dynamically construct a semantic decision tree that pulls backward from the global goal to generate a semantic guidance skeleton structure for evaluating path compliance. Matching and scoring module: Input the path anchor sequence into the semantic guidance skeleton structure, calculate the matching score of each candidate local path segment in three dimensions: anchor type consistency, spatial topological rationality and temporal evolution continuity, so as to generate a semantic matching score vector that represents the degree of alignment between each candidate local path segment and the global task intent. Correction module: Based on the semantic matching score vector, identify conflict candidate paths with functional redundancy and similar scores, call the stability records of the same type of geometric semantic mixed entities in the historical memory cache under similar environmental conditions, and perform a credibility re-evaluation process on the key semantic anchors that cause ambiguity to generate conflict resolution decision results after dynamic weight correction. Obstacle avoidance execution module: Based on the conflict resolution decision results, the target local path with the highest score and that has passed the verification is selected, converted into aircraft control commands and output, so as to drive the low-altitude aircraft to complete the obstacle avoidance maneuver according to the target local path.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that... When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images