A bamboo forest low-altitude unmanned aerial vehicle flyable passage discrimination and navigation method and system

Abstract

The application belongs to the technical field of low-altitude unmanned aerial vehicle vision navigation, and relates to a bamboo forest low-altitude unmanned aerial vehicle flyable passage discrimination and navigation method and system. In view of the problems that the flyable passage marking is insufficient when flying at low altitude in the bamboo forest, the narrow gap between bamboo stems, the dark area blocked by branches and leaves and the light and shadow patch are easily regarded as the flyable passage, the application first regularizes the candidate area, and then establishes a flyable passage configuration record table around the passage gap configuration, the surrounding obstacle arrangement and the forest environment. With the passage mask and the configuration record as constraints, the passage samples under different widths, directions, blockages and illumination conditions are supplemented; meanwhile, the misjudgment area of the current discrimination model is reviewed in the limited passage scene to form a misguide passage configuration record table. The actual passage samples, the configuration supplemented samples and the same-form different-line samples are used for passage discrimination training to obtain a flight usability map, a flyable passage and local navigation control instructions.

Description

Technical Field

[0001] This application belongs to the field of visual navigation technology for low-altitude unmanned aerial vehicles (UAVs), and relates to a method and system for identifying and navigating flight paths for low-altitude UAVs in bamboo forests. Background Technology

[0002] Low-altitude flight within bamboo forests relies on airborne vision for continuous assessment of the space ahead, and can be used for forestry inspections, understory resource surveys, disaster monitoring, mountain patrols, and emergency search and rescue. When a drone flies through a bamboo forest, its forward-facing image, binocular image, or depth image is simultaneously affected by the camera's perspective, flight attitude, bamboo density, foliage obstruction, understory lighting, and terrain undulations. The slender and repetitive nature of bamboo stalks often obstructs the field of vision, and the location and scale of flight paths change with the viewing angle. Adjacent bamboo stalks, foliage obstacles, and gaps collectively determine whether passage is possible in a given area, and their effective width, direction of extension, and continuous length also change with the flight direction and observation distance.

[0003] Identifying flyable paths during UAV visual navigation within bamboo forests is not easy. Uneven lighting, foliage obstruction, motion blur, unclear boundaries, and local structural breaks can all render narrow paths incomplete. Narrow gaps between bamboo stalks, dark areas created by branches and leaves, canopy projections, bright spots in the sky, ground reflections, abrupt changes in terrain, and image ghosting can also appear as flyable paths. Especially when multiple gaps are close together, or when the path ahead is partially obscured by branches and leaves, the system is prone to connecting separate areas into one, leading to incorrect navigation decisions.

[0004] Existing visual navigation methods for unmanned aerial vehicles (UAVs) in bamboo forests mostly employ geometric rules, traditional feature classification, semantic segmentation, or end-to-end local planning models. Rule-based methods rely on texture, edge, depth, or morphological parameters, and their stability is limited under conditions of dense bamboo forests, abrupt changes in lighting, and multiple types of obstacles. Learning-based methods can improve pixel-level judgment capabilities, but they are heavily reliant on labeled samples. When there are insufficient samples of flyable paths, and the path width, occlusion level, and lighting conditions are not adequately covered, the model is prone to missing true paths; conversely, when misleading paths closely resemble true paths, the model may mistake treacherous areas for flyable ones. Summary of the Invention

[0005] To address the issues of insufficient path samples, misleading paths being easily identified as flyable, and unstable path boundaries in complex bamboo forests, this invention provides a method and system for identifying and navigating flyable paths for low-altitude UAVs in bamboo forests.

[0006] A method for identifying and navigating flight paths of low-altitude unmanned aerial vehicles (UAVs) in bamboo forests includes the following steps: S1. Obtain the bamboo forest navigation image to be processed, and preprocess the bamboo forest navigation image to obtain a standardized bamboo forest navigation image; S2. Perform candidate region regularization on the flyable pathway area in the standardized bamboo forest navigation image to obtain flyable pathway blocks; extract pathway gap configuration features, surrounding obstacle arrangement features and forest environment features from the flyable pathway blocks to establish a flyable pathway configuration record table. Using the current flight path discrimination model on restricted access bamboo forest scenario data, inference is performed to obtain misleading path candidate areas that are misjudged as flight paths; the same candidate area regularization as the flight path area is performed on the misleading path candidate areas to obtain misleading path blocks; the path gap configuration features, surrounding obstacle arrangement features and forest environment features corresponding to the flight path blocks are extracted from the misleading path blocks to establish a misleading path configuration record table; S3. Based on the flight path configuration record table, the configuration constraint supplementary model is trained using the flight path mask as a spatial constraint and the path configuration description as a supplementary sample input to generate supplementary samples with matching configurations, i.e., configuration supplementary samples; the configuration supplementary samples are used to increase the morphological coverage of flight paths under different bamboo forest conditions. S4. In the path configuration comparison library composed of the flightable path configuration record table and the misleading path configuration record table, samples with similar shapes but different configurations are selected according to the path configuration comparison score. The samples with similar shapes but different configurations refer to restricted passage area samples whose appearance, surrounding obstacle arrangement, or path gap configuration is similar to a flightable path, but whose effective width, forward continuous distance, minimum safe clearance, or turning margin does not meet the flight requirements. The actual flightable path samples, the configuration supplementary samples, and the samples with similar shapes but different configurations are used to form a training batch to train the flightable path discrimination model for path identification. S5. Use the trained flightable path discrimination model to discriminate flightable paths in the input bamboo forest navigation image, and perform spatial weighted fusion on the block flight availability prediction results or multi-scale prediction results in the inference stage to output the flight availability map and flightable paths; then generate local waypoints or navigation control commands from the flightable paths based on the safe distance, minimum channel width, target direction and UAV kinematic constraints.

[0007] Preferably, after obtaining standardized bamboo forest navigation images Then, based on the target markings or the initial set of flyable path masks... Candidate regions of flyable pathways are regularized to obtain a set of structurally stable flyable pathway blocks. : ; in, Indicates the first One flyable path block Indicates the number of flyable areas; Candidate region normalization is denoted as a function For the first A flight path mask The candidate region normalization process is represented as: ; in, Indicates a mask for a flightable path. Apply the mask after candidate region normalization. This represents the smallest bounding rectangle. This indicates that the boundary of the outer rectangle is extended. Indicates navigation from standardized bamboo forest images Cut and trim flyable path blocks in the middle; For mask Perform connected component decomposition to obtain a set of candidate connected regions: ; in, Indicates the first The first of the flight-accessible path mask A connected region, Indicates the number of connected components; for each connected component It is divided into several grid cells within its bounding rectangle. ,in Indicates the number of grid cells, the first The foreground pixel density of each grid cell is: ; in, Indicates the number of pixels in a grid cell. This indicates the foreground pixel density within the grid cell; if If so, the grid cell is considered a local fragment or edge burr and is discarded; in, The density threshold is used to select the flight path mask. After density screening and morphological screening based on area, aspect ratio, and compactness, a regularized flight path mask is obtained. .

[0008] Preferred, flight path configuration record table establishment: Obtaining the channel gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector Then, construct the first Configuration description vectors for each flightable path: ; This represents the one-hot encoding function. For scene semantic tags; Generate configuration description text Text input for configuration constraint sampling models: ; in, This represents a configuration description text encoding or template generation function used to convert the pathway configuration description vector... Convert to natural language description or text conditions; Write the flight path number, flight path mask, flight path block, configuration description vector, configuration description text, and sample reference information into the flight path configuration record table: ; in, Indicates the area number of the flightable access route. This indicates the original flight path mask. Indicates a flyable path block. This represents a vector describing the configuration of a flightable pathway. This represents the configuration description text. This indicates referenced information; Flightable Path Configuration Record Table This serves as input for subsequent configuration supplementation and screening of homomorphic but heteromorphic samples.

[0009] Preferably, in order to construct misleading pathway-side discrimination boundary correction samples, in the restricted access bamboo forest scenario dataset... The above utilizes the current flight path discrimination model Reasoning is performed to obtain candidate regions for misleading pathways; For the Image of a bamboo forest scene with restricted access Standardized images are obtained by using the same preprocessing method as the training images. The current flight availability map is output by the flight availability discrimination model: ; in, Indicates pixel position The probability of being predicted as a flightable route; Based on the misjudgment threshold Obtain the misleading region mask: ; in, Indicates the first Misleading path candidate masks in images of bamboo forest scenes with limited access. This indicates the threshold for misleading pathway candidates.

[0010] Preferably, the misleading path block is regular: for the misleading region mask Perform the same candidate region regularization as for flyable paths to obtain a set of misleading path blocks: ; in, Indicates the first A misleading path block, Indicates the number of misleading path blocks; right By performing connected component decomposition, mesh density filtering, and morphological constraint filtering, a structurally stable misleading path mask is obtained. From the corresponding standardized restricted access bamboo forest scene images Misleading path blocks in the middle: ; For each misleading path block Extract the path gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector And construct the configuration description vector of the misleading pathway block. : ; Write the misleading path block number, misleading path block, misleading path mask, configuration description vector, and sample reference information into the misleading path configuration record table: ; in, This indicates a misleading path block number. This indicates a misleading pathway mask. This indicates a misleading path block. This represents a vector describing the misleading pathway configuration. Indicates image source, frame number, or sample path reference information, misleading the pathway configuration record table. This serves as input for subsequent screening of homomorphic but heteromorphic samples.

[0011] Preferably, configuration-constrained supplementary sampling model training and supplementary sample generation: Index based on flightable pathway configuration description Construct a configuration-constrained imputation model with a flyable path mask as the structural constraint and a flyable path configuration description vector or configuration description text as the imputation input; when the configuration-constrained imputation model adopts a conditional diffusion model, its training objective can be expressed as: ; in, This indicates the conditioned diffusion training objective. Indicates the diffusion time step. Indicates Gaussian noise. Indicates the first The sample at the th Noise samples at each diffusion time step This represents the output of the noise prediction network. It is an energy function; Mask consistency loss and configuration conformity loss can be introduced: ; in, Represents the overall training objective of the generative model. This indicates a loss of mask consistency. Indicates configurational consistency loss. and Indicates the weighting coefficient; Supplementary sample set: ; in, For configuration description text, This is a vector describing the configuration of a flightable pathway. Indicates the number of supplementary samples, and the supplementary sample set. Together with real flightable pathway samples, they are used for subsequent pathway identification training to complete the coverage of flightable pathway morphology.

[0012] Preferred, homomorphic heterogeneous sample screening: for flyable pathway areas and misleading pathway blocks Define the pathway configuration comparison score: ; in, Indicates a flyable area With misleading pathway blocks The pathway configuration comparison score between them Indicates the similarity of the forest environment. Indicates the similarity of surrounding arrangements. Indicates the similarity of the channel gap configuration. The weighting coefficients are non-negative and satisfy the following conditions: ; The similarity of forest environments can be expressed as: ; The similarity of surrounding arrangements can be expressed as: ; The similarity of the channel gap configuration can be expressed as: ; in, and These are scale parameters representing the similarity of the surrounding arrangement and the similarity of the channel gap configuration, respectively. For the A flightable pathway region, from the misleading pathway configuration record table Select the top scorers Each misleading pathway block constitutes a set of homomorphic but heterogeneous samples corresponding to the flyable pathway region: ; in, Indicates the first A set of homomorphic but heteromorphic samples corresponding to each flyable pathway region. This indicates the number of homomorphic and heteromorphic samples selected. The homomorphic and heteromorphic sample sets corresponding to all flyable pathway regions are merged to obtain the homomorphic and heteromorphic sample set: ; Homomorphic heterogeneous sample set Together with the actual flightable pathway sample set and the configuration supplementary sample set, they constitute the pathway discrimination training data.

[0013] The preferred set of real flyable pathway samples is denoted as: ; in, Indicates the first A standardized bamboo forest navigation training image, This indicates the corresponding pixel-level flight path label. Indicates the number of actual training samples collected; Configurational supplementary sample set is The homomorphic heterogeneous sample set is Pathway identification training batch Obtained by proportional sampling of three types of samples: ; in, ; The training objectives for pathway discrimination include pathway determination loss, isomorphic-heterogeneous separation loss, and pathway configuration preservation loss: ; in, This indicates the overall goal of pathway identification training. Indicates path determination loss, This represents the loss for separating homomorphic and heteromorphic lines. This indicates a loss in pathway configuration retention. and Indicates the weighting coefficient; Path determination loss can be expressed as: ; in, This represents pixel-level cross-entropy loss. This indicates Dice's loss. Indicates the weighting coefficient; The homomorphic-heterogeneous separation loss is used to separate the flyable path samples from the homomorphic-heterogeneous samples in the feature embedding space. Let the feature embedding of the flyable path sample be... The positive sample feature embeddings that match the flightable pathway configuration are Isomorphic heterogeneous sample feature embedding ,but: ; in, Represents the similarity function. Indicates the temperature coefficient; Pathway configuration retention loss Used to constrain the relative position, orientation, or length relationship between the obstacle boundary and the passage gap, so that the model maintains a consistent passage gap configuration when learning the flyable passage region; ; in, This indicates the model's prediction results. Indicates a true label or configuration constraint label. This represents a function that extracts pathway gap configuration descriptions from prediction results or labels.

[0014] Preferably, during the inference phase, for the large bamboo forest navigation image to be navigated... First, perform the same preprocessing as in the training phase to obtain standardized test images. Subsequently Perform segmented flight availability reasoning or multi-scale reasoning to obtain Individual segmented flight availability prediction results: ; in, Indicates the first Each block at pixel position Flight availability prediction results at the location Indicates the number of blocks participating in the fusion; For each block, construct a spatial weight map where the central region has a high weight and the peripheral region has a low weight. Let the internal coordinates of the block be... The spatial weights are then expressed as: ; in, and They represent along direction and To assign higher weights to the central region and lower weights to the peripheral region of the window function, the following form can be used: ; ; in, Indicates the center coordinates of the block. and They represent the blocks in direction and The normalized scale for the direction is set to zero for positions that are outside the effective range of the block; By fusing all segmented flight availability prediction results according to spatial weights, a flight availability map of the entire bamboo forest navigation image is obtained: ; in, This represents the merged flight availability map. To prevent constants with a denominator of zero; Finally, based on the output threshold Obtain the binary mask for the flight path: ; in, This indicates the final flight path determination result. Indicates the output threshold; For indicator functions; Based on the binary mask of the flightable path, combined with the current position of the UAV, target direction, minimum safe gap, path gap extension length and UAV kinematic constraints, the candidate path is continuously screened to generate local waypoints or navigation control commands.

[0015] A low-altitude unmanned aerial vehicle (UAV) flight path identification and navigation system for bamboo forests includes: a preprocessing module for acquiring and preprocessing bamboo forest navigation images to obtain standardized bamboo forest navigation images; a path configuration recording module for regularizing candidate areas of flightable path regions and extracting path gap configuration features, surrounding obstacle arrangement features, and forest environment features to establish a flightable path configuration record table; a configuration constraint supplementation module for generating configuration supplementation samples based on the flightable path configuration record table, using a flightable path mask as a spatial constraint and a path configuration description as a supplementation input; a misleading area review and regularization module for obtaining misleading path candidate areas on restricted-access bamboo forest scene data using the current flightable path discrimination model, and regularizing the misleading path candidate areas to obtain misleading path blocks; and a misleading path configuration recording module for recording the misleading path configuration of the selected areas. The system extracts the path gap configuration features, surrounding obstacle arrangement features, and forest environment features corresponding to the flightable path blocks to establish a misleading path configuration record table. A configuration comparison module is used to filter homomorphic samples from a path configuration comparison library composed of the flightable path configuration record table and the misleading path configuration record table, based on the path configuration comparison score. A path discrimination training module is used to form a training batch with the collected flightable path samples, configuration supplementary samples, and homomorphic samples, and to train the flightable path discrimination model. A fusion navigation output module is used to use the trained flightable path discrimination model to discriminate flightable paths in the input bamboo forest navigation image, and to perform spatial weighted fusion of the segmented flight availability prediction results or multi-scale prediction results, outputting a flight availability map, flightable paths, local waypoints, or navigation control commands.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: First, the present invention simultaneously records the configuration of the passageway gaps, the arrangement of the surrounding bamboo barriers, and the forest environment, so that flyable passageways and areas prone to misjudgment can be compared under the same set of actual flight conditions, rather than just comparing pixel appearances.

[0017] Second, the present invention supplements the samples according to the description of the pathway mask and configuration, which can cover the flight paths under different bamboo forest densities, light, gap widths and shading conditions, and reduce the omissions caused by the single sample morphology.

[0018] Third, this invention reviews misleading pathways from the misjudged areas of the current model and selects homomorphic samples, enabling the model to identify typical cases of "appearing to fly but actually not flying", reducing misjudgments caused by narrow slits, dark areas and light and shadow patches.

[0019] Fourth, this invention uses real-world flight path samples, configuration supplementary samples, and homomorphic heteromorphic samples for path discrimination training. Through three types of constraints—path determination, homomorphic heteromorphic separation, and path configuration preservation—it improves the ability to distinguish between real paths and restricted passage areas.

[0020] Fifth, this invention reduces edge inconsistencies in block reasoning or multi-scale reasoning through spatial weighted fusion, improves the continuity of the flight availability map, and generates local waypoints or navigation control commands that meet safety constraints.

[0021] Therefore, the present invention can reduce misjudgments caused by misleading pathways in complex bamboo forest backgrounds, improve the reliability of flight path boundaries, pathway gap configurations and local track outputs, and enhance the navigation adaptability of UAVs under different bamboo forest densities, lighting conditions and flight altitudes. Attached Figure Description

[0022] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or conventional adjustments made by those skilled in the art to the steps, parameters, or model structures without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0024] This embodiment provides a method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests. The method takes a bamboo forest navigation image as input and sequentially completes image normalization, path configuration recording, configuration supplementation, misleading region review, configuration comparison, path identification training, and navigation output. The key is not simply adding samples, but comparing real flight paths with misleading paths that the model easily misinterprets under the same configuration scale: the former is used to supplement the real path morphology, and the latter is used to mark areas that appear similar but are actually unsafe to traverse.

[0025] For ease of explanation, the following variable notation is used uniformly in this embodiment.

[0026] Let the original bamboo forest navigation image be The preprocessed standardized bamboo forest navigation image is Let the set of flyable pathway regions be . ,in Indicates the first A flightable access area This represents the number of flyable access areas. Let the number of areas be... The mask for each flightable passage area is Flightable path blocks are The pathway configuration description vector is The pathway configuration description vector is composed of the pathway gap configuration description vector. Peripheral obstacle arrangement description vector Forest environment description vector Composition, namely: ;

[0027] Assume the flight path configuration record table is as follows: The misleading pathway configuration record table is as follows Let the current flight path discrimination model be... The flight path discrimination model trained on the path discrimination is as follows: Let the configuration-constrained sampling model be... .in, This represents the parameters of the current flightable path discrimination model. This represents the parameters of the flight path discrimination model after path discrimination training.

[0028] I. Data and Applicable Objects: This invention is applicable to forward-looking monocular images, binocular images, depth images, or multi-frame temporal images from UAVs. Input bamboo forest navigation image. It can be an RGB image, or an airborne visual image that includes depth, optical flow, semantic priors, or multi-frame fusion features. The objects to be identified include, but are not limited to, flight paths, tree trunks, branches, canopy occlusion areas, shrub and grass obstacles, and areas of abrupt terrain change.

[0029] For training images with target annotations, the set of flyable path masks is denoted as: ; in, Indicates pixel position Does it belong to the first There are one flyable path. For bamboo forest images that do not have flyable paths or have limited access, a restricted access bamboo forest scene dataset is constructed, denoted as: ; in, Indicates the first Images of a bamboo forest scene with restricted access. This indicates the number of images in the bamboo forest scene with restricted access.

[0030] II. Preprocessing and Standardization Implementation: Navigation images of the original bamboo forest Preprocessing is performed to obtain a standardized bamboo forest navigation image. Preprocessing includes one or more of the following: camera distortion correction, pose compensation and viewpoint alignment, dehazing, motion blur suppression, dynamic range compression, and contrast normalization.

[0031] In one implementation, camera distortion correction and geometric resampling are performed on the airborne wide-angle image to reduce wide-angle distortion and scale inconsistency; attitude compensation, viewpoint alignment, and illumination normalization are performed on the forward-looking UAV airborne image to obtain an image representation in uniform coordinates. The geometrically corrected pixel position can be represented as: ; in, This represents the pixel coordinates in the original bamboo forest navigation image. This represents the plane coordinates or measured coordinates after geometric correction of the viewpoint. This represents the camera distortion correction or attitude compensation mapping function.

[0032] Dynamic range compression was performed on the bamboo forest navigation image after viewpoint geometric correction to obtain: ; in, This represents the bamboo forest navigation image after perspective geometric correction. Indicates pixel position, Indicates position The result of dynamic range compression at that location.

[0033] Subsequently, After performing standardization processing, a standardized bamboo forest navigation image is obtained: ; in and They represent The mean and standard deviation, To prevent constants with a denominator of zero. Standardized bamboo forest navigation image. It serves as a unified input for subsequent pathway configuration description, configuration supplementation, misleading pathway back-checking, and flight-ready pathway identification.

[0034] III. Regularization and Configuration Record of Flightable Path Candidate Regions: Obtaining standardized bamboo forest navigation images Then, first, based on the target label or the initial set of flyable path masks... Candidate regions of flyable pathways are regularized to obtain a set of structurally stable flyable pathway blocks: ; in, Indicates the first A flightable path block.

[0035] Candidate region normalization is denoted as a function For the first A flight path mask The candidate region normalization process is represented as: ;

[0036] in, Indicates a mask for a flightable path. Apply the mask after candidate region normalization. This represents the smallest bounding rectangle. This indicates that the boundary of the outer rectangle is extended. Indicates navigation from standardized bamboo forest images Cut the flyable path block in the middle.

[0037] Candidate region regularization includes connected component decomposition, mesh density filtering, and shape constraint filtering. First, the mask... Perform connected component decomposition to obtain a set of candidate connected regions: ; in, Indicates the first The first of the flight-accessible path mask A connected region, This indicates the number of connected components. For each connected component... It is divided into several grid cells within its bounding rectangle. ,in Indicates the number of grid cells. (The first...) The foreground pixel density of each grid cell is: ; in, Indicates the number of pixels in a grid cell. This indicates the foreground pixel density within that grid cell. If If the condition is not met, the mesh cell is considered a local fragment or edge burr and is discarded; where, This represents the density threshold. After density screening and morphological screening based on area, aspect ratio, and compactness, a well-formed flight path mask is obtained. Subsequent pathway configuration descriptions are all based on and the corresponding flyable path block conduct.

[0038] IV. Extraction of Pathway Gap Configuration Description: For the One flyable path block and its enhanced flight path mask Extract the path gap configuration description vector .

[0039] First, in the flyable path block Extract the salient sub-region of obstacles, denoted as Obstacle salient subregions can be obtained through local adaptive threshold segmentation, deep mutation detection, or obstacle response thresholding. ; in, Represents pixels Does it belong to the salient sub-region of the obstacle? Indicates the first Obstacle response threshold in each flyable path block Indicates an indicator function.

[0040] Secondly, accessible gap sub-regions are extracted within the adjacent regions of obstacles or the neighborhood of the main direction of the passage, denoted as... A passable gap sub-region can be represented as: ; in, Represents pixels Does it belong to a passable gap sub-region? This indicates the passable gap response threshold. Represents a significant sub-region with obstacles Starting from and along the direction of travel The candidate passage gap neighborhood obtained by the search Indicates the first The unit vector representing the direction of travel for each flightable path.

[0041] Area of ​​passage gap Represented as: ;

[0042] The perimeter of the flight path outline is denoted as This can be calculated from the number of pixels at the mask boundary or the length of the boundary curve. Flight path compactness. Represented as: ; in, To prevent constants with a denominator of zero.

[0043] Flightable main road Based on the enhanced flight path mask Calculation of the central moment. Let... for of If the order central moment is given, then: .

[0044] Barrier salient subregion centroid and the centroid of the passable void subregion They are represented as follows: ; .

[0045] The displacement vector of the centroid of the obstacle boundary and the passage gap is: .

[0046] The displacement distance between the barrier boundary and the centroid of the passage gap is: .

[0047] The unit vector of the passage direction can be obtained from the displacement of the centroid of the obstacle boundary and the passage gap: .

[0048] Passage gap extension length Represented as: ; in, This represents the vector dot product.

[0049] Therefore, the first... A vector describing the gap configuration of each flyable passageway region: .

[0050] Based on the actual navigation task requirements It can also include information such as the dimensions of the circumscribed rectangle of the passage, aspect ratio, gap width, passage direction angle, and minimum safety gap.

[0051] V. Extraction of surrounding obstacle arrangement features: In obtaining an enhanced mask set for all flightable pathways Then, calculate the obstacle arrangement description vector for each flightable path. First, calculate the... The center of mass of a flightable path: ; by Build a local window around the center Count the number of bamboo plants in this area within the window: .

[0052] Local density of bamboo plants Represented as: ;

[0053] in, Represents a local window The area.

[0054] Nearest neighbor distance Represented as: .

[0055] If there is only one flight path in the current image, then It can be set to a preset maximum distance or a null value marker, and then processed uniformly in the subsequent normalization process.

[0056] Local orientation consistency of arrangement direction Represented as: ; in, Represents a local window Inner The main axis direction of a flightable path Display window The average or main direction of the main direction of the internal flight path.

[0057] Therefore, the first... A vector describing the arrangement of obstacles around a flyable area: ; in, Used to describe the local density of bamboo plants. Used to describe nearest neighbor relationships Used to describe the consistency of orientation in the arrangement direction. Used to describe the degree of local aggregation.

[0058] VI. Extraction of Forest Environment Description: Forest environmental features are used to describe the bamboo forest scene type or navigation semantic category where the flight path is located. For the first One flyable path block Scene semantic labels can be obtained based on rule classifiers, lightweight classifiers, or manual annotation. Its value range is: ; in, This indicates the number of scene semantic categories. Scene semantic categories can include bamboo forest scene types such as dense bamboo forest, sparse bamboo forest, understory shrub and grass area, sloping bamboo forest, forest edge and road boundary area, and area with sudden change in light, as well as navigation semantic categories such as flight path, tree trunk, branches, canopy shading area, shrub and grass obstacle, and terrain change area.

[0059] Scene semantic tags Convert to a semantic configuration description vector: ; in, This represents the one-hot encoding function. Indicates the first A vector describing the forest environment of each flyable pathway. If learnable semantic embedding is used, it can also be... Input embedding layer .

[0060] VII. Establishment of Flightable Path Configuration Record Table: Obtaining the channel gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector Then, construct the first Configuration description vectors for each flightable path: .

[0061] Further generate configuration description text Text input for configuration constraint sampling models: ; in, This represents a configuration description text encoding or template generation function used to convert configuration description vectors... Converted into natural language descriptions or text conditions. For example, configuration description text can describe the type of bamboo forest scene where the flight path is located, the morphology of the salient area of ​​obstacles, the extension length of the passage gap, the local density of bamboo plants, and the adjacency relationship.

[0062] Write the flight path number, flight path mask, flight path block, configuration description vector, configuration description text, and sample reference information into the flight path configuration record table: ; in, Indicates the area number of the flightable access route. This indicates the original flight path mask. Indicates a flyable path block. This represents a pathway configuration description vector. This represents the configuration description text. This indicates reference information such as image source, frame number, or sample path. Flightable Path Configuration Record Table This serves as input for subsequent configuration supplementation and screening of homomorphic but heteromorphic samples.

[0063] VIII. Configuration-Constrained Supplementation Model Training and Supplementation Sample Generation:

[0064] Index based on flightable pathway configuration description A configuration-constrained imputation model is constructed, using a flyable path mask as the structural constraint and flyable path configuration description vectors or configuration description text as imputation inputs. For flyable paths... Masking the flight path Flightable pathway configuration description vector and configuration description text Input the conditional encoder to obtain the conditional code. :

[0065] ; in, Indicates the conditional encoder for the supplemental sampling. Indicates the first The generation condition code corresponding to each flightable path.

[0066] In one implementation, a configuration-constrained complementation model With random noise Conditional coding As input, generate supplementary configuration samples that match the configuration: ; in, Indicates the first Additional samples of each configuration This represents random noise or latent variables.

[0067] Equivalently, the generation process can also be represented as: ; Among them, flight path mask Used to constrain the location, shape, and scale of pathways in generated samples, flyable pathway configuration description vector and configuration description text Used to constrain the bamboo forest texture, lighting conditions, passageway morphology, and bamboo plant adjacency relationships in the generated samples.

[0068] When the configuration-constrained sampling model adopts the conditional diffusion model, its training objective can be expressed as: ; in, This indicates the conditioned diffusion training objective. Indicates the diffusion time step. Indicates Gaussian noise. Indicates the first The sample at the th Noise samples at each diffusion time step This represents the output of the noise prediction network.

[0069] To ensure that the generated samples are consistent with the flight path mask and flight path configuration description, mask consistency loss and configuration consistency loss can be introduced: ; in, Represents the overall training objective of the generative model. This indicates a loss of mask consistency. Indicates configurational consistency loss. and This represents the weighting coefficient.

[0070] After generation, a supplementary sample set is obtained: ; in, Indicates the number of supplementary samples. Supplementary sample set Together with real flightable pathway samples, they are used for subsequent pathway identification training to complete the coverage of flightable pathway morphology.

[0071] IX. Misleading Channel Review To construct misleading path discrimination boundary correction samples, a restricted-access bamboo forest scenario dataset was used. The above utilizes the current flight path discrimination model Reasoning is performed to obtain candidate regions for misleading pathways.

[0072] For the Image of a bamboo forest scene with restricted access First, standardized images are obtained by using the same preprocessing method as the training images. The current flight availability map is output by the flight availability discrimination model. ; in, Indicates pixel position The probability of being predicted as a flightable route.

[0073] Based on the misjudgment threshold Obtain the misleading region mask: ; in, Indicates the first Misleading path candidate masks in images of bamboo forest scenes with limited access. This indicates the threshold for candidate misleading pathways. Candidate areas for misleading pathways typically originate from visually misleading pathway regions such as narrow gaps between bamboo trunks, dark areas formed by shading from branches and leaves, canopy projections, bright spots in the sky, ground reflections, motion blur, or abrupt changes in local topography.

[0074] 10. Misleading Path Block Regularization and Configuration Recording:

[0075] Misleading area mask Perform the same candidate region regularization as for flyable paths to obtain a set of misleading path blocks: ; in, Indicates the first A misleading path block, Indicates the number of misleading path blocks.

[0076] Specifically, for By performing connected component decomposition, mesh density filtering, and morphological constraint filtering, a structurally stable misleading path mask is obtained. Subsequently, images of the corresponding standardized, restricted-access bamboo forest scene were used. Misleading path blocks in the middle: .

[0077] For each misleading path block The same configuration description extraction method as that used for flightable pathways was employed to extract the configuration description vectors of the pathway gaps. Peripheral obstacle arrangement description vector Forest environment description vector And construct the configuration description vector of the misleading pathway block: ; Among them, the superscript " "Used to represent the configuration description of the misleading pathway side. Write the misleading pathway block number, misleading pathway block, misleading pathway mask, configuration description vector, and sample reference information into the misleading pathway configuration record table:" ; in, This indicates a misleading path block number. This indicates a misleading pathway mask. This indicates a misleading path block. This represents a vector describing the misleading pathway configuration. This indicates reference information such as image source, frame number, or sample path. Misleading Pathway Configuration Record Table This serves as input for subsequent screening of homomorphic but heteromorphic samples.

[0078] XI. Screening of Isomorphic but Different-Line Samples: Based on flight path configuration record table and misleading pathway configuration record table In the pathway configuration comparison library, samples with similar appearance and structure to flightable pathways but which cannot actually meet flight conditions are screened.

[0079] For areas with flightable access and misleading pathway blocks Define the pathway configuration comparison score: ; in, Indicates a flyable area With misleading pathway blocks The pathway configuration comparison score between them Indicates the similarity of the forest environment. Indicates the similarity of surrounding arrangements. Indicates the similarity of the channel gap configuration. The weighting coefficients are non-negative and satisfy the following conditions: .

[0080] The similarity of forest environments can be expressed as: ; The similarity of surrounding arrangements can be expressed as: ; The similarity of the channel gap configuration can be expressed as: ; in, and These are scale parameters representing the similarity of the surrounding arrangement and the similarity of the channel gap configuration, respectively.

[0081] For the A flightable pathway region, from the misleading pathway configuration record table Select the top scorers Each misleading pathway block constitutes a set of homomorphic but heterogeneous samples corresponding to the flyable pathway region: ; in, Indicates the first A set of homomorphic but heteromorphic samples corresponding to each flyable pathway region. This indicates the number of homomorphic and heteromorphic samples selected. The homomorphic and heteromorphic sample set is obtained by merging the sets of homomorphic and heteromorphic samples corresponding to all flyable pathway regions. .

[0082] Homomorphic heterogeneous sample set Together with the actual flightable pathway sample set and the configuration supplementary sample set, they constitute the pathway discrimination training data.

[0083] 12. Pathway Identification Training: The sample set of real flyable pathways is denoted as: ; in, Indicates the first A standardized bamboo forest navigation training image, This indicates the corresponding pixel-level flight path label. This indicates the number of actual training samples collected.

[0084] Configurational supplementary sample set is The homomorphic heterogeneous sample set is Pathway identification training batch Obtained by proportional sampling of three types of samples: ; in, .

[0085] Flightable Path Discrimination Model Semantic segmentation networks, segmentation networks, or other pixel-level flight availability prediction networks can be used. Pathway discrimination training objectives include path determination loss, isomorphic / heterogeneous path separation loss, and path configuration preservation loss. ; in, This indicates the overall goal of pathway identification training. Indicates path determination loss, This represents the loss for separating homomorphic and heteromorphic lines. This indicates a loss in pathway configuration retention. and This represents the weighting coefficient.

[0086] Path determination loss can be expressed as: ; in, This represents pixel-level cross-entropy loss. This indicates Dice's loss. This represents the weighting coefficient.

[0087] The homomorphic-heterogeneous separation loss is used to separate flyable path samples from homomorphic-heterogeneous samples in the feature embedding space. Let the feature embedding of the flyable path sample be... The positive sample feature embeddings that match the flightable pathway configuration are Isomorphic heterogeneous sample feature embedding ,but: ; in, Represents the similarity function. This represents the temperature coefficient.

[0088] Pathway configuration retention loss This constraint is used to constrain the relative position, orientation, or length relationship between obstacle boundaries and passage gaps, ensuring the model maintains consistent passage gap configuration when learning flyable passage regions. This constraint allows for the recalculation of the corresponding configuration description based on the predicted obstacle boundaries and passage gaps in the passage region, and comparison with the actual or supplementary passage configuration description. Its general form can be expressed as: ; in, This indicates the model's prediction results. Indicates a true label or configuration constraint label. This represents a function that extracts pathway gap configuration descriptions from prediction results or labels.

[0089] Through the above-mentioned path discrimination training, the actual flightable path samples provide basic supervision, the configuration supplementary samples increase the path morphology under different conditions, and the homomorphic heteromorphic samples improve the model's ability to distinguish easily confused restricted passage areas.

[0090] Thirteen, Reasoning, Spatial Weighted Fusion and Navigation Output:

[0091] During the inference phase, for the large bamboo forest navigation image to be navigated... First, perform the same preprocessing as in the training phase to obtain standardized test images. Subsequently, on Perform segmented flight availability reasoning or multi-scale reasoning to obtain Individual segmented flight availability prediction results:

[0092] ; in, Indicates the first Each block at pixel position Flight availability prediction results at the location This indicates the number of blocks participating in the fusion.

[0093] To mitigate inconsistent responses at block boundaries, a spatial weight map is constructed for each block, with higher weights in the central region and lower weights in the peripheral region. Let the internal coordinates of the block be... The spatial weights are then expressed as: ; in, and They represent along direction and Window function weights based on direction. To ensure high weights in the central region and low weights in the peripheral region, the following form can be used: ; ; in, Indicates the center coordinates of the block. and They represent the blocks in direction and Normalization scale for direction. For positions outside the effective range of the block, the weight is set to zero.

[0094] By fusing all segmented flight availability prediction results according to spatial weights, a flight availability map of the entire bamboo forest navigation image is obtained: ; in, This represents the merged flight availability map. To prevent constants with a denominator of zero.

[0095] Finally, based on the output threshold Obtain the binary mask for the flight path: ; in, This indicates the final flight path determination result. This represents the output threshold. Based on the binary mask of the flyable path, combined with the UAV's current position, target direction, minimum safe clearance, path gap extension length, and UAV kinematic constraints, candidate paths are continuously filtered to generate local waypoints or navigation control commands.

[0096] XIV. System Implementation Method: This embodiment also provides a flight path identification and navigation system for low-altitude UAVs in bamboo forests. The system includes a preprocessing module, a path configuration recording module, a configuration constraint supplementation module, a misleading area review and normalization module, a misleading path configuration recording module, a configuration comparison module, a path identification training module, and a fusion navigation output module.

[0097] The preprocessing module is used to receive the raw bamboo forest navigation image. It outputs standardized bamboo forest navigation images. .

[0098] The pathway configuration recording module uses standardized bamboo forest navigation images. and flight path mask set As input, candidate region normalization is performed on the flyable pathway region, and a set of flyable pathway blocks is output. Pathway configuration description vector set Flightable Path Configuration Record Table .

[0099] Configuration-constrained sampling module with flight path configuration record table As input, generate a configuration supplementary sample set. .

[0100] Misleading region lookup and regularization module for restricted access bamboo forest scene dataset And the current flightable path discrimination model As input, output a set of misleading path blocks. .

[0101] The misleading pathway configuration recording module uses a set of misleading pathway blocks. As input, extract the configuration description vector of the misleading pathway block and output the misleading pathway configuration record table. .

[0102] Configuration comparison module records flight path configurations. and misleading pathway configuration record table Using this as input, homomorphic heteromorphic samples are selected from the pathway configuration reference library, and the output is a set of homomorphic heteromorphic samples. .

[0103] The pathway identification training module uses a real-world sample set of flightable pathways. Configuration supplementary sample set and homomorphic sample sets As input, train the flight path discrimination model to identify flight paths, and output the trained discrimination model. .

[0104] The fusion navigation output module is used in conjunction with the trained flight path discrimination model. The image of the bamboo forest to be navigated is taken as input, and the fused flight availability map is output. Binary mask for flight paths And local waypoints or navigation control commands.

[0105] By sequentially connecting the above modules, this invention realizes a complete process from image preprocessing, pathway configuration recording, configuration supplementation, misleading region back-checking, isomorphic and heteromorphic sample screening, pathway discrimination training to navigation output.

[0106] XV. System Deployment and Hardware Implementation

[0107] The bamboo forest low-altitude UAV flight path identification and navigation system provided by this invention is as follows: Figure 1 As shown, the system includes a preprocessing module, a pathway configuration recording module, a configuration constraint supplementation module, a misleading region backtracking and normalization module, a misleading pathway configuration recording module, a configuration comparison module, a pathway discrimination training module, and a fusion navigation output module. The modules are connected in the following order: "standardized image—pathway configuration recording—configuration supplementation—misleading region backtracking—configuration comparison—pathway discrimination training—navigation output".

[0108] The preprocessing module is used to receive the original bamboo forest navigation image. The image is then processed using one or more of the following methods: camera distortion correction, pose compensation and viewpoint alignment, dehazing, motion blur suppression, dynamic range compression, and contrast normalization, to output a standardized bamboo forest navigation image. Standardized bamboo forest navigation image It serves as a unified input for subsequent pathway configuration descriptions, misleading pathway lookups, and online inference.

[0109] The pathway configuration recording module is used to receive standardized bamboo forest navigation images. and flight path mask assembly Candidate region regularization is performed on flyable pathways to obtain a set of flyable pathway blocks. Subsequently, the module extracts the passage gap configuration description vector from the flyable passage block. Peripheral obstacle arrangement description vector Forest environment description vector Construct a flight path configuration description vector And form a flight path configuration record table. The flight path configuration record table It also serves as input for the configuration constraint sampling module and the configuration comparison module.

[0110] The configuration constraint sampling module is used to receive the flight path configuration record table. And based on the flight path mask within it. Flightable pathway configuration description vector and configuration description text Generate configurational supplementary samples that match the configuration. This forms a supplementary sample set. It is used to complete the coverage of traverseable paths and serves as the input for the path discrimination training module.

[0111] The Misleading Region Retrospective and Regularization module is used to receive datasets of restricted-access bamboo forest scenes. And the current flightable path discrimination model This module first utilizes the current flight path discrimination model. Inference is performed on images of restricted-access bamboo forest scenes to obtain candidate masks for misleading pathways. Subsequently, connected component decomposition, mesh density filtering, and morphological constraint filtering are performed on the candidate masks for misleading pathways to obtain a set of structurally stable misleading pathway blocks. .

[0112] The misleading path configuration recording module is used to receive a set of misleading path blocks. And using the same configuration description extraction method as the flightable path, for each misleading path block Extracting the channel gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector Constructing a configuration description vector for the candidate region of the misleading pathway And form a misleading pathway configuration record table. The misleading pathway configuration record table As input to the configuration comparison module.

[0113] The configuration comparison module is used to receive the flight path configuration record table. and misleading pathway configuration record table In the pathway configuration comparison library composed of the two, the samples were compared in segments according to the similarity of the forest environment, the similarity of the surrounding arrangement, and the similarity of the pathway gap configuration, to obtain a set of isomorphic but heteromorphic samples that are similar to the configuration of aviable pathways but are actually impassable. The homomorphic heterogeneous sample set As input to the pathway identification training module.

[0114] The pathway discrimination training module is used to receive a sample set of real flightable pathways. Supplementary sample set and homomorphic sample sets The three types of samples were used to construct training batches for pathway discrimination. This module trains the flyable path discrimination model using path determination loss, homomorphic heteromorphic separation loss, and path configuration preservation loss, resulting in a trained recognition model. Among them, real flyable path samples are used to provide pixel-level supervision, configuration supplementary samples are used to complete the morphological coverage of flyable paths, and homomorphic heteromorphic samples are used to enhance the model's ability to distinguish easily confused restricted passage areas.

[0115] The fusion navigation output module is used to receive the trained flight path discrimination model. The module first performs preprocessing on the bamboo forest image to be navigated, consistent with the training phase, to obtain a standardized test image. Subsequently, segmented flight availability inference or multi-scale inference is performed on it to obtain multiple segmented flight availability prediction results. The fusion navigation output module further constructs a spatial weight map for each block. The flight availability prediction results of each segment are weighted and fused to obtain the flight availability map of the entire bamboo forest navigation image. Binary mask for flight paths Then output local waypoints or navigation control commands.

[0116] In one hardware deployment method, an airborne monocular camera, binocular camera, depth camera, or multi-sensor vision module outputs raw airborne image frames via an airborne data interface, Ethernet interface, flight control serial port, or UAV internal data bus. The timestamp and flight status information are sent to the computing unit. The computing unit includes one or more of a central processing unit (CPU), graphics processing unit (GPU), dedicated inference accelerator, field-programmable gate array (FPGA), or system-on-a-chip (SoC). The preprocessing module can be deployed in the CPU or FPGA and can be used for flight path discrimination models. It can be deployed in GPUs, dedicated inference accelerators, or SoCs, and the fusion navigation output module can be deployed in CPUs, GPUs, or embedded processing units.

[0117] During the training or offline update phase, the system is based on a real flightable path sample set. Dataset of restricted access bamboo forest scenes and flight path mask set A flight path configuration record table is generated sequentially. Supplementary sample set Misleading Pathway Configuration Record Table and homomorphic sample sets And complete the flight path discrimination model. The training process involves path discrimination. The model parameters, flight path configuration record table, misleading path configuration record table, preprocessing parameters, and fusion parameters obtained from the training are stored in non-volatile memory and can be updated via a host computer, network link, or offline storage medium.

[0118] During the online inference phase, the system receives real-time airborne image frames from the airborne vision sensors. Standardized test images are obtained through the preprocessing module. Then, the trained flight path discrimination model is used. The system outputs segmented flight availability prediction results or multi-scale prediction results. The fusion navigation output module performs spatial weighted fusion of the prediction results to obtain a flight availability map. And generate a binary mask for a flight path based on the threshold. Based on this, the system combines the current flight status, target direction, minimum safe clearance, and UAV kinematic constraints to output one or more navigation information, including local waypoints, heading angles, speeds, flight altitudes, or obstacle avoidance constraints.

[0119] In one application, the output results can be visualized via a display terminal or transmitted back to a host computer, task control terminal, or UAV flight controller via serial port, network port, data bus, or wireless link. When the system is linked with bamboo forest inspection UAVs, forestry operation UAVs, mountain patrol UAVs, or ground robots, the location of flyable paths, path outlines, local waypoints, or navigation control commands can be used as data input for subsequent local path planning, risk verification, obstacle avoidance control, or task recording. Thus, the system achieves a complete deployment process from bamboo forest image acquisition, flyable path identification, result fusion, to navigation control terminal output.

[0120] XVI. Method Feasibility Verification:

[0121] To further illustrate the feasibility and effectiveness of the method of this invention in visual navigation tasks of UAVs in bamboo forests, this embodiment compares and verifies the basic flyable path discrimination model, some module combination schemes, and the complete method of this invention under the same training set, validation set, and test set partitioning conditions. Evaluation indicators include average intersection-over-union ratio (IoU), F1 score, false positive rate, average number of misclassified misleading paths per image, local track executability rate, and inference time. Among these, the IoU and F1 scores reflect the accuracy of flyable path discrimination; the false positive rate and average number of misclassified misleading paths per image reflect the ability to control misjudgments in complex bamboo forest backgrounds; the local track executability rate reflects the proportion of navigation outputs that meet safety constraints; and the inference time reflects online processing efficiency. Table 1 below provides exemplary experimental data to illustrate the effectiveness verification method of this invention and is not intended to limit the scope of protection of this invention.

[0122] Table 1. Comprehensive performance verification results of the method of the present invention (exemplary experimental data) .

[0123] As shown in Table 1, under the same data partitioning and basic flyable path discrimination model, the complete method of this invention outperforms the basic model and the combination schemes of each module in both mIoU and F1 scores, indicating that this invention can improve the accuracy of flyable path identification and the comprehensive navigation perception capability. Meanwhile, the complete method of this invention achieves lower FPR and FP / image scores, indicating that it can effectively reduce misjudgments caused by visually misleading paths in complex bamboo forest backgrounds, such as narrow gaps between bamboo stalks, dark areas obscured by branches and leaves, canopy projection, bright spots in the sky, and ground reflections.

[0124] The results from different module combinations show that supplementing samples based on flyable path masks and configuration descriptions can improve mIoU and F1, indicating that the configuration supplementation samples cover the insufficient path width, void structure, and bamboo forest texture variations in the actual collected data. By reviewing the misleading path candidate regions of the current model and filtering homomorphic but heteromorphic samples by configuration comparison, FPR and FP / image can be significantly reduced, indicating that the model's ability to distinguish easily confused restricted passage areas is improved. Path configuration preservation constraints help maintain the relationship between obstacle boundaries and path voids. The spatial weighted fusion module can improve the continuity and stability of the block flight availability inference results.

[0125] Furthermore, the inference time of the complete method of this invention increases only slightly compared to the basic model, indicating that the recording of flyable path configurations, recording of misleading path configurations, configuration supplementation, and screening of homomorphic and heterogeneous samples are mainly performed during the training or offline update phase; the online inference phase mainly performs preprocessing, flyable path discrimination, spatial weighted fusion, and local track output. Therefore, this invention improves the accuracy of path judgment and suppresses misleading paths while still possessing good engineering deployment feasibility.

Claims

1. A method for identifying and navigating flight paths of low-altitude unmanned aerial vehicles (UAVs) in bamboo forests, characterized in that, Includes the following steps: S1. Obtain the bamboo forest navigation image to be processed, and preprocess the bamboo forest navigation image to obtain a standardized bamboo forest navigation image; S2. Perform candidate region regularization on the flyable pathway area in the standardized bamboo forest navigation image to obtain flyable pathway blocks; extract pathway gap configuration features, surrounding obstacle arrangement features and forest environment features from the flyable pathway blocks to establish a flyable pathway configuration record table. Using the current flight path discrimination model on restricted access bamboo forest scenario data, inference is performed to obtain misleading path candidate areas that are misjudged as flight paths; the same candidate area regularization as the flight path area is performed on the misleading path candidate areas to obtain misleading path blocks; the path gap configuration features, surrounding obstacle arrangement features and forest environment features corresponding to the flight path blocks are extracted from the misleading path blocks to establish a misleading path configuration record table; S3. Based on the flight path configuration record table, the configuration constraint supplementary model is trained using the flight path mask as a spatial constraint and the path configuration description as a supplementary sample input to generate supplementary samples with matching configurations, i.e., configuration supplementary samples; the configuration supplementary samples are used to increase the morphological coverage of flight paths under different bamboo forest conditions. S4. In the path configuration comparison library composed of the flightable path configuration record table and the misleading path configuration record table, samples with similar shapes but different configurations are selected according to the path configuration comparison score. The samples with similar shapes but different configurations refer to restricted passage area samples whose appearance, surrounding obstacle arrangement, or path gap configuration is similar to a flightable path, but whose effective width, forward continuous distance, minimum safe clearance, or turning margin does not meet the flight requirements. The actual flightable path samples, configuration supplementary samples, and the samples with similar shapes but different configurations are combined to form a training batch to train the flightable path discrimination model for path identification. S5. Use the trained flightable path discrimination model to discriminate flightable paths in the input bamboo forest navigation image, and perform spatial weighted fusion on the block flight availability prediction results or multi-scale prediction results in the inference stage to output the flight availability map and flightable paths; then generate local waypoints or navigation control commands from the flightable paths based on the safe distance, minimum channel width, target direction and UAV kinematic constraints.

2. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, Obtaining standardized bamboo forest navigation images Then, based on the target label or the initial set of flyable path masks... Candidate regions of flyable pathways are regularized to obtain a set of structurally stable flyable pathway blocks. : ； in, Indicates the first One flyable path block Indicates the number of flyable areas; Candidate region normalization is denoted as a function For the first A flight path mask The candidate region normalization process is represented as: ； in, Indicates a mask for a flightable path. Apply the mask after candidate region normalization. This represents the smallest bounding rectangle. This indicates that the boundary of the outer rectangle is extended. Indicates navigation from standardized bamboo forest images Cut and trim flyable path blocks in the middle; For mask Perform connected component decomposition to obtain a set of candidate connected regions: ； in, Indicates the first The first of the flight-accessible path mask A connected region, Indicates the number of connected components; for each connected component It is divided into several grid cells within its bounding rectangle. ,in Indicates the number of grid cells, the first The foreground pixel density of each grid cell is: ； in, Indicates the number of pixels in a grid cell. This indicates the foreground pixel density within the grid cell; if If so, the grid cell is considered a local fragment or edge burr and is discarded; The density threshold is used to select the flight path mask. After density screening and morphological screening based on area, aspect ratio, and compactness, a regularized flight path mask is obtained. .

3. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, Establishment of a flight path configuration record table: Obtaining the channel gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector Then, construct the first Configuration description vectors for each flightable path: ； This represents the one-hot encoding function. For scene semantic tags; Generate configuration description text Text input for configuration constraint sampling models: ； in, This represents a configuration description text encoding or template generation function used to convert the pathway configuration description vector... Convert to natural language description or text conditions; Write the flight path number, flight path mask, flight path block, configuration description vector, configuration description text, and sample reference information into the flight path configuration record table: ； in, Indicates the area number of the flightable access route. This indicates the original flight path mask. Indicates a flyable path block. This represents a vector describing the configuration of a flightable pathway. This represents the configuration description text. This indicates referenced information; Flightable Path Configuration Record Table This serves as input for subsequent configuration supplementation and screening of homomorphic but heteromorphic samples.

4. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, To construct misleading pathway-side discrimination boundary correction samples, a restricted-access bamboo forest scenario dataset was used. The above utilizes the current flight path discrimination model Reasoning is performed to obtain candidate regions for misleading pathways; For the Image of a bamboo forest scene with restricted access Standardized images are obtained by using the same preprocessing method as the training images. The current flight availability map is output by the flight availability discrimination model: ； in, Indicates pixel position The probability of being predicted as a flightable route; Based on the misjudgment threshold Obtain the misleading region mask: ； in, Indicates the first Misleading path candidate masks in images of bamboo forest scenes with limited access. This indicates the threshold for misleading pathway candidates.

5. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, Misleading path block regularization: masking of misleading regions Perform the same candidate region regularization as for flyable paths to obtain a set of misleading path blocks: ； in, Indicates the first A misleading path block, Indicates the number of misleading path blocks; right By performing connected component decomposition, mesh density filtering, and morphological constraint filtering, a structurally stable misleading path mask is obtained. From the corresponding standardized restricted access bamboo forest scene images Misleading path blocks in the middle: ； For each misleading path block Extract the path gap configuration description vector Peripheral obstacle arrangement description vector Forest environment description vector And construct the configuration description vector of the misleading pathway block. : ； Write the misleading path block number, misleading path block, misleading path mask, configuration description vector, and sample reference information into the misleading path configuration record table: ； in, This indicates a misleading path block number. This indicates a misleading pathway mask. This indicates a misleading path block. This represents a vector describing the misleading pathway configuration. Indicates image source, frame number, or sample path reference information, misleading the pathway configuration record table. This serves as input for subsequent screening of homomorphic but heteromorphic samples.

6. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, Configuration-constrained supplementation model training and supplementation sample generation: Index based on flightable pathway configuration description Construct a configuration-constrained imputation model with a flyable path mask as the structural constraint and a flyable path configuration description vector or configuration description text as the imputation input; when the configuration-constrained imputation model adopts a conditional diffusion model, its training objective can be expressed as: ； in, This indicates the conditioned diffusion training objective. Indicates the diffusion time step. Indicates Gaussian noise. Indicates the first The sample at the th Noise samples at each diffusion time step This represents the output of the noise prediction network. It is an energy function; Mask consistency loss and configuration conformity loss can be introduced: ； in, Represents the overall training objective of the generative model. This indicates a loss of mask consistency. Indicates configurational consistency loss. and Indicates the weighting coefficient; Supplementary sample set: ； in, For configuration description text, This is a vector describing the configuration of a flightable pathway. Indicates the number of supplementary samples, and the supplementary sample set. Together with real flightable pathway samples, they are used for subsequent pathway identification training to complete the coverage of flightable pathway morphology.

7. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, Screening of homomorphic and heteromorphic samples: for flightable pathway areas and misleading pathway blocks Define the pathway configuration comparison score: ； in, Indicates a flyable area With misleading pathway blocks The pathway configuration comparison score between them Indicates the similarity of the forest environment. Indicates the similarity of surrounding arrangements. Indicates the similarity of the channel gap configuration. The weighting coefficients are non-negative and satisfy the following conditions: ； The similarity of forest environments can be expressed as: ； The similarity of surrounding arrangements can be expressed as: ； The similarity of the channel gap configuration can be expressed as: ； in, and These are scale parameters representing the similarity of the surrounding arrangement and the similarity of the channel gap configuration, respectively. For the A flightable pathway region, from the misleading pathway configuration record table Select the top scorers Each misleading pathway block constitutes a set of homomorphic but heterogeneous samples corresponding to the flyable pathway region: ； in, Indicates the first A set of homomorphic but heteromorphic samples corresponding to each flyable pathway region. This indicates the number of homomorphic and heteromorphic samples selected. The homomorphic and heteromorphic sample sets corresponding to all flyable pathway regions are merged to obtain the homomorphic and heteromorphic sample set: ； Homomorphic heterogeneous sample set Together with the actual flightable pathway sample set and the configuration supplementary sample set, they constitute the pathway discrimination training data.

8. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, The sample set of real flyable pathways is denoted as: ； in, Indicates the first A standardized bamboo forest navigation training image, This indicates the corresponding pixel-level flight path label. Indicates the number of actual training samples collected; Configurational supplementary sample set is The homomorphic heterogeneous sample set is Pathway identification training batch It was obtained by proportional sampling of three types of samples: ； in, ; The training objectives for pathway discrimination include pathway determination loss, isomorphic-heterogeneous separation loss, and pathway configuration preservation loss: ； in, This indicates the overall goal of pathway identification training. Indicates path determination loss, This represents the loss for separating homomorphic and heteromorphic lines. This indicates a loss in pathway configuration retention. and Indicates the weighting coefficient; The path determination loss can be expressed as: ； in, This represents pixel-level cross-entropy loss. This indicates Dice's loss. Indicates the weighting coefficient; The homomorphic-heterogeneous separation loss is used to separate the flyable path samples from the homomorphic-heterogeneous samples in the feature embedding space. Let the feature embedding of the flyable path sample be... The positive sample feature embeddings that match the flightable pathway configuration are Isomorphic heterogeneous sample feature embedding ,but: ； in, Represents the similarity function. Indicates the temperature coefficient; Pathway configuration retention loss Used to constrain the relative position, orientation, or length relationship between the obstacle boundary and the passage gap, so that the model maintains a consistent passage gap configuration when learning the flyable passage region; ； in, This indicates the model's prediction results. Indicates a true label or configuration constraint label. This represents a function that extracts pathway gap configuration descriptions from prediction results or labels.

9. The method for identifying and navigating flight paths of low-altitude UAVs in bamboo forests according to claim 1, characterized in that, During the inference phase, for the large bamboo forest navigation image to be navigated... First, perform the same preprocessing as in the training phase to obtain standardized test images. Subsequently Perform segmented flight availability reasoning or multi-scale reasoning to obtain Individual segmented flight availability prediction results: ； in, Indicates the first Each block at pixel position Flight availability prediction results at the location Indicates the number of blocks participating in the fusion; For each block, construct a spatial weight map where the central region has a high weight and the peripheral region has a low weight. Let the internal coordinates of the block be... The spatial weights are then expressed as: ； in, and They represent along direction and To assign higher weights to the central region and lower weights to the edge region of the window function, the following form can be used: ； ； in, Indicates the center coordinates of the block. and They represent the blocks in direction and The normalized scale for the direction is set to zero for positions that are outside the effective range of the block; By fusing all segmented flight availability prediction results according to spatial weights, a flight availability map of the entire bamboo forest navigation image is obtained: ； in, This represents the merged flight availability map. To prevent constants with a denominator of zero; Finally, based on the output threshold Obtain the binary mask for the flight path: ； in, This indicates the final flight path determination result. Indicates the output threshold; For indicator functions; Based on the binary mask of the flightable path, combined with the current position of the UAV, target direction, minimum safe gap, path gap extension length and UAV kinematic constraints, the candidate path is continuously screened to generate local waypoints or navigation control commands.

10. A flight path identification and navigation system for low-altitude unmanned aerial vehicles in bamboo forests, characterized in that, include: The preprocessing module acquires and preprocesses bamboo forest navigation images to obtain standardized bamboo forest navigation images. The pathway configuration recording module regularizes candidate areas for aviable pathways and extracts pathway gap configuration features, surrounding obstacle arrangement features, and forest environment features to establish an aviable pathway configuration record table. The configuration constraint supplementation module generates configuration supplementation samples based on the aviable pathway configuration record table, using aviable pathway masks as spatial constraints and pathway configuration descriptions as supplementation inputs. The misleading area review and regularization module obtains misleading pathway candidate areas on restricted-access bamboo forest scene data using the current aviable pathway discrimination model, and regularizes the misleading areas... The candidate regions for navigation pathways are regularized to obtain misleading pathway blocks. A misleading pathway configuration recording module is used to extract pathway gap configuration features, surrounding obstacle arrangement features, and forest environment features corresponding to the aviable pathway blocks from the misleading pathway blocks, establishing a misleading pathway configuration recording table. A configuration comparison module is used to filter homomorphic samples based on their configuration comparison scores from a pathway configuration comparison library composed of the aviable pathway configuration recording table and the misleading pathway configuration recording table. A pathway discrimination training module is used to form a training batch from the collected aviable pathway samples, configuration supplementary samples, and homomorphic samples, and to train the aviable pathway discrimination model for pathway discrimination. The fusion navigation output module is used to use the trained flight path discrimination model to determine the flight path of the input bamboo forest navigation image, and to perform spatial weighted fusion of the segmented flight availability prediction results or multi-scale prediction results to output flight availability map, flight path, local waypoints or navigation control commands.

Images