A Method and System for Blind Spot Inspection Route Planning in Forest Areas Based on UAV Collaboration

Abstract

This application discloses a method and system for path planning in forest blind spot inspection based on UAV collaboration, belonging to the field of path planning technology. First, it preprocesses the multi-source sensor data of the UAV swarm using photometric methods, extracts visual features, and establishes a statistical list of illumination levels. Then, it uses feature back-projection to map the visual information into three-dimensional space, quantifies and evaluates the visual confidence level of the area, and constructs an illumination confidence level model characterizing perception reliability. Based on this, it identifies visual degradation areas prone to positioning failure, establishes anchor point mappings by combining visibility search, and generates a navigation cost field that integrates illumination risks. Finally, under the constraints of the cost field, it solves the multi-UAV collaborative path, generating a chain-like trajectory that avoids blind spots while maintaining perception connectivity. In this way, the perception risks caused by drastic changes in illumination can be transformed into hard constraints for path planning, effectively solving the visual positioning drift problem under complex illumination conditions in forest areas and achieving highly reliable collaborative inspection by UAV swarms.

Description

Technical Field

[0001] This application relates to the field of path planning technology, and more specifically, to a method and system for forest blind spot inspection path planning based on UAV collaboration. Background Technology

[0002] Forest resources, as the main body of terrestrial ecosystems, are crucial for maintaining global ecological balance and biodiversity. To ensure the safety of forest resources, routine inspections for tasks such as fire early warning, pest and disease monitoring, and illegal logging are indispensable. Traditional forest inspections mainly rely on manual foot patrols or vehicle patrols. This method is not only limited by complex and rugged terrain, but also inefficient and poses many safety hazards, making it difficult to achieve full coverage monitoring of large forest areas. In recent years, with the rapid development of drone technology, multi-rotor drones, with their advantages of maneuverability, vertical take-off and landing, and low-altitude operation, have gradually replaced traditional manual operations and become the core equipment for modern forestry management. In particular, the swarm operation mode based on multi-drone collaboration can significantly improve the inspection efficiency and coverage of large forest areas through task distribution and cooperation, and has become an important development direction for the construction of smart forestry.

[0003] However, most existing UAV forest patrol path planning schemes focus on obstacle avoidance and shortest path solving based on geometric information, assuming that environmental perception information is continuous and reliable. But in actual forest flight scenarios, the lighting conditions within the forest area are extremely complex due to differences in canopy shading and variations in solar altitude angle, exhibiting a high dynamic range of alternating light and dark. This intense change in lighting can cause overexposure or underexposure in the visual sensors on the UAV, severely undermining the photometric consistency assumption upon which the Visual Simultaneous Localization and Mapping (SLAM) algorithm relies. When navigating through dappled light gaps in the forest or moving from bright light areas into deep shade, the visual front end often struggles to extract stable feature points, leading to feature tracking failures and positioning drift, ultimately causing the planned path to fail due to lost positioning. Current path planning strategies lack the ability to predict this visual degradation risk, failing to identify and avoid geometrically passable but visually unreliable blind spots, making it difficult to effectively cover blind spots in the forest area while ensuring perception stability, thus limiting the in-depth application of UAV swarms in complex forestry scenarios.

[0004] Therefore, there is an urgent need for an optimized method and system for planning blind spot inspection routes in forest areas based on drone collaboration. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a method for planning blind spot inspection routes in forest areas based on UAV collaboration is provided, which includes: S1: Perform photometric preprocessing on the acquired raw sensor data stream of the drone fleet to obtain a visual feature set and a photometric statistics list. The raw sensor data stream of the drone fleet includes camera image streams, inertial navigation data and illumination sensor readings of each drone. S2: Based on the visual feature set and photometric statistics list, perform feature back projection and visual confidence score calculation on the preset three-dimensional voxel grid to obtain the illumination confidence raster map. S3: Use a preset threshold to perform connected component segmentation and visibility search on the illumination confidence raster to obtain a set of visually degraded regions and an anchor point mapping table corresponding to the set of visually degraded regions; S4: Generate a navigation cost field based on the illumination confidence raster map and the set of visually degraded regions; S5: Solve the navigation cost field and anchor point mapping table for multi-aircraft system cooperative escort path to obtain a visual chain trajectory set; S6: Controls the coordinated flight of a swarm of drones based on a visual chain trajectory set.

[0007] According to another aspect of this application, a forest blind spot inspection path planning system based on UAV collaboration is provided, which includes: The photometric preprocessing module is used to perform photometric preprocessing on the acquired raw sensor data stream of the drone fleet to obtain a visual feature set and a photometric statistics list. The raw sensor data stream of the drone fleet includes camera image streams, inertial navigation data and illumination sensor readings of each drone. The confidence mapping module is used to perform feature back projection and visual confidence score calculation on a preset three-dimensional voxel grid based on a visual feature set and a photometric statistics list to obtain an illumination confidence raster map. The illumination visibility analysis module is used to perform connected component segmentation and visibility search on the illumination confidence raster map using preset thresholds to obtain a set of visually degraded regions and an anchor point mapping table corresponding to the set of visually degraded regions. The navigation cost field construction module is used to generate a navigation cost field based on the illumination confidence raster map and the set of visually degraded regions; The cooperative escort path solving module is used to solve the multi-aircraft system cooperative escort path from the navigation cost field and anchor point mapping table to obtain a visual chain trajectory set; The swarm cooperative flight control module is used to control the cooperative flight of a swarm of drones based on a visual chain trajectory set.

[0008] Compared with existing technologies, this application provides a method and system for forest blind spot inspection path planning based on UAV collaboration. First, it preprocesses the multi-source sensor data of the UAV swarm to extract visual features and establish a lighting statistics list. Then, it uses feature back-projection to map the visual information to three-dimensional space, quantifies and evaluates the visual confidence of the area, and constructs a lighting confidence model characterizing perception reliability. Based on this, it identifies visual degradation areas prone to positioning failure, establishes anchor point mappings by combining visibility search, and generates a navigation cost field that integrates lighting risks. Finally, under the constraints of the cost field, it solves the multi-UAV collaborative path, generating a chain-like trajectory that avoids blind spots while maintaining perception connectivity. In this way, the perception risks caused by drastic changes in lighting can be transformed into hard constraints for path planning, effectively solving the visual positioning drift problem under complex lighting conditions in forest areas and achieving highly reliable collaborative inspection by UAV swarms. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram of a forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of sub-step S1 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of sub-step S2 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of sub-step S3 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of sub-step S4 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0016] Figure 7This is a flowchart of sub-step S5 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0017] Figure 8 This is a flowchart of sub-step S6 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application.

[0018] Figure 9 This is a block diagram of a forest blind spot inspection path planning system based on UAV collaboration, according to an embodiment of this application. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] To address the problems mentioned above, this application proposes a method for planning blind spot inspection routes in forest areas based on UAV collaboration. Figure 1 This is a flowchart of a forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 2 This is a data flow diagram of a forest blind spot inspection path planning method based on UAV collaboration, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the method for planning blind spot inspection paths in forest areas based on UAV collaboration includes the following steps: S1: Perform photometric preprocessing on the acquired raw sensor data stream of the UAV swarm to obtain a visual feature set and a photometric statistics list. The raw sensor data stream of the UAV swarm includes camera image streams, inertial navigation data, and illumination sensor readings of each UAV; S2: Based on the visual feature set and photometric statistics list, perform feature back projection and visual confidence score calculation on a preset three-dimensional voxel grid to obtain an illumination confidence raster map; S3: Use a preset threshold to perform connected component segmentation and visibility search on the illumination confidence raster map to obtain a set of visual degradation regions and an anchor point mapping table corresponding to the set of visual degradation regions; S4: Generate a navigation cost field based on the illumination confidence raster map and the set of visual degradation regions; S5: Solve the multi-UAV system collaborative flight path on the navigation cost field and the anchor point mapping table to obtain a visual chain trajectory set; S6: Control the UAV swarm to fly collaboratively according to the visual chain trajectory set.

[0021] In the aforementioned method for planning blind spot inspection paths in forest areas based on UAV collaboration, step S1 involves photometric preprocessing of the acquired raw sensor data stream from the UAV swarm to obtain a visual feature set and a photometric statistics list. The raw sensor data stream includes camera image streams, inertial navigation data, and illumination sensor readings from each UAV. It should be understood that due to the complex environment of forest areas, the raw sensor data stream acquired by the UAV swarm collaborative operation suffers from problems such as multi-source heterogeneity, high noise interference, and inconsistent benchmarks. Direct use of this data stream can lead to map construction misalignment and decreased perception accuracy. Therefore, this application performs photometric preprocessing on the raw sensor data stream from the UAV swarm to remove noise, synchronize spatiotemporal benchmarks, and extract visual features and illumination statistics with environmental characterization value. First, the sensors of each drone in the fleet are uniformly calibrated and synchronized to ensure consistent data acquisition. Then, each drone activates its sensors according to a preset inspection frequency. Cameras continuously capture images of the forest scene at a fixed frame rate to generate a camera image stream. Inertial navigation sensors collect real-time inertial navigation data from the drones, including angular velocity, acceleration, and attitude angle data. Illumination sensors simultaneously record the current ambient light intensity data, thus obtaining illumination sensor readings. This process eliminates heterogeneous data differences and environmental interference, laying a standardized data foundation for the subsequent construction of a high-precision illumination confidence grid map and ensuring stable environmental perception and consistent data interaction within the multi-drone system in complex forest areas.

[0022] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S1 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 3 As shown, step S1 includes: S11, performing data stream parsing and adaptive histogram equalization on the original sensor data stream of the cluster to obtain an enhanced grayscale image stream; S12, performing gradient-based primary feature point extraction on the enhanced grayscale image stream to obtain a candidate feature coordinate set; S13, determining a local neighborhood window in the enhanced grayscale image stream based on the candidate feature coordinate set, and performing local variance statistics and descriptor extraction on the pixel data within the local neighborhood window to obtain a visual feature set and a photometric statistics list.

[0023] Specifically, step S11 involves performing data stream parsing and adaptive histogram equalization on the raw sensor data stream of the aircraft fleet to obtain an enhanced grayscale image stream. It should be understood that due to the extremely large dynamic range of illumination caused by the canopy shading in forest areas, images often exhibit local overexposure or underexposure, resulting in the loss of texture in shadow areas and severely hindering feature extraction. Therefore, this application further performs data stream parsing on the raw sensor data stream of the aircraft fleet and applies adaptive histogram equalization to dynamically adjust the local contrast of the image, restoring texture details in both dark and bright areas while suppressing noise. This effectively solves the imaging quality problem caused by uneven illumination, restores environmental features obscured by light and shadow, provides the algorithm with high-quality images with uniform grayscale distribution, and ensures that effective landmarks can still be captured even in environments with drastic changes in brightness.

[0024] Specifically, in one possible embodiment, the processor converts the parsed color video stream into a single-channel grayscale image to reduce processing dimensionality. Next, a contrast-limited adaptive histogram equalization algorithm is used to divide the image into several non-overlapping rectangular sub-blocks. A grayscale histogram is calculated independently for each sub-block, and a threshold is set to crop high-frequency noise components, uniformly distributing the cropped portion across the entire histogram. Finally, bilinear interpolation is used to smoothly blend the boundaries of each sub-block, eliminating blockiness and synthesizing an enhanced grayscale image stream with balanced overall brightness and sharp, clear local texture details.

[0025] Specifically, step S12 involves performing gradient-based primary feature point extraction on the enhanced grayscale image stream to obtain a candidate feature coordinate set. It should be understood that since visual localization relies on prominent environmental landmarks, and flat, textureless areas cannot provide effective geometric constraints, full-pixel computation would also result in computational redundancy. Therefore, this application further performs primary feature point extraction on the enhanced grayscale image stream based on gradient information to filter out corner points or edge points with drastic grayscale changes and obvious geometric structures, accurately locating tree trunks and branches in forest areas. This filters out invalid backgrounds from massive amounts of data, significantly reducing the computational load and ensuring that the selected features possess good geometric invariance, providing highly reliable candidate coordinates for constructing robust visual odometry.

[0026] Specifically, in one possible embodiment, the algorithm iterates through each frame of the enhanced grayscale image, calculates the gradient values ​​of pixels in the horizontal and vertical directions using a first-order differential operator, constructs a local structure tensor matrix based on the gradient, and solves for its eigenvalues ​​to quantify the corner response intensity. Then, a response threshold is set (e.g., only retaining points with a corner response function value R > 0.01, or retaining the top 500 points with the strongest response in each frame), and pixels with acceptable response values ​​are selected as potential feature points. Subsequently, non-maximum suppression is performed in the local neighborhood (e.g., within a 7×7 pixel sliding window) to remove spatially dense redundant points, retaining only the points with the strongest local response, ultimately generating a set of candidate feature coordinates that are spatially evenly distributed and can represent the key geometric structure of the forest area.

[0027] Specifically, in step S13, a local neighborhood window is determined in the enhanced grayscale image stream based on the candidate feature coordinate set, and local variance statistics and descriptor extraction are performed on the pixel data within the local neighborhood window to obtain a visual feature set and a photometric statistics list. It should be understood that, due to the presence of swaying branches and flickering light spots in forest areas, it is difficult to distinguish stable entities from false light and shadow based solely on geometric location; misusing light spots as landmarks will cause positioning drift. Therefore, this application further determines the local neighborhood window and calculates pixel variance and descriptors to quantify the illumination stability of feature points and generate unique identity fingerprints, achieving deep identification of feature attributes. In this way, it is possible to intuitively determine whether a feature point is in an area of ​​drastic illumination change, and to use descriptors to achieve accurate matching between consecutive frames, thereby constructing a feature set containing illumination attributes, enabling the system to eliminate unstable light and shadow interference and lock onto real landmarks.

[0028] Specifically, in one possible embodiment, the processing unit extracts a rectangular neighborhood window centered on the candidate feature coordinates, calculates the mean and variance of pixel grayscale within the window, and quantifies the degree of illumination fluctuation in the region as a stability index. Simultaneously, using a binary descriptor algorithm, it compares point-to-point grayscale values ​​within the window according to a specific pattern, generating binary feature codes that are rotation- and illumination-invariant. Finally, the 3D coordinates of the feature points, the binary descriptors, and the calculated local photometric variance are associated and bound, encapsulated into a structured visual feature set and a photometric statistics list, and output to the confidence mapping module.

[0029] In the aforementioned method for planning blind spot inspection paths in forest areas based on UAV collaboration, step S2 involves performing feature back-projection and calculating visual confidence scores on a preset three-dimensional voxel grid based on a visual feature set and a photometric statistics list to obtain a light confidence raster map. It should be understood that since the visual feature points acquired by the front-end perception exhibit a discrete and sparse spatial distribution, and a simple set of feature points is insufficient to intuitively represent the continuous variation patterns of light intensity and the distribution of positioning safety within a large-scale forest area, this application further implements a mapping from discrete features to continuous space. It utilizes feature back-projection technology to fill the three-dimensional voxel grid with visual feature points carrying light attributes and performs quantitative calculations of visual confidence scores for each grid cell. This allows the construction of a rasterized environmental map with light risk perception capabilities, transforming abstract feature quality into navigation passability probabilities in spatial location. This enables subsequent path planning algorithms to identify and avoid areas of drastic light changes, much like recognizing physical obstacles, ensuring the positioning stability of the UAV in complex forest shadow environments.

[0030] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S2 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 4 As shown, step S2 includes: S21, based on the boundary constraints of the prior geometric map and preset resolution parameters, discretizing and indexing the three-dimensional coordinates of the visual feature set to obtain a voxel index mapping table; S22, traversing the voxel index mapping table, performing voxel-level photometric statistical feature aggregation on the feature points inside non-empty voxels to obtain a voxel aggregated data package containing the number of stable feature points and the average photometric variance data; S23, performing multi-dimensional visual confidence calculation on the voxel aggregated data package to obtain an illumination confidence raster map.

[0031] Specifically, in step S21, based on the boundary constraints of the prior geometric map and preset resolution parameters, the three-dimensional coordinates of the visual feature set are discretized and projected, and indexed and encoded to obtain a voxel index mapping table. It should be understood that since the coordinates of visual feature points collected by the UAV are continuous floating-point values, directly performing large-scale data retrieval and state updates in continuous space would consume enormous computing resources and be inefficient. Therefore, this application further utilizes the boundary range determined by the prior map and the preset resolution (e.g., setting the voxel side length to a cubic unit of 0.2 meters to 0.5 meters) to perform spatial discretization projection on the massive number of feature point three-dimensional coordinates, and uses hash encoding technology to transform the three-dimensional spatial position into a unique one-dimensional memory index. This establishes an efficient voxel index mapping relationship, enabling rapid conversion from massive unstructured point cloud data to structured grid data, significantly improving the retrieval speed and memory management efficiency of spatial data, and providing a fast data access entry point for subsequent statistical analysis of illumination attributes in specific spatial areas.

[0032] Specifically, in one possible embodiment, the algorithm first reads the extreme values ​​of the bounding box coordinates of the prior geometric map and, combined with a preset voxel side length resolution, constructs a three-dimensional grid coordinate system covering the entire domain. Then, it iterates through the world coordinates of each feature point in the visual feature set, determining its integer coordinates in the grid coordinate system through rounding down and coordinate offset calculations. Next, a spatial hash function is used to map these three-dimensional integer coordinates to unique linear index values, which serve as the keys of the hash table. Simultaneously, a dynamic list is maintained as the hash table's values, writing the identifiers of all feature points mapped to the same index value into this list. Finally, a voxel index mapping table is generated, recording the correspondence between active voxels and their internal feature point sets, completing the structured organization of the spatial data.

[0033] Specifically, in step S22, the voxel index mapping table is traversed, and voxel-level photometric statistical feature aggregation is performed on the feature points within non-empty voxels to obtain a voxel aggregated data package containing the number of stable feature points and the average photometric variance. It should be understood that the photometric data of a single feature point may be subject to random errors due to sensor noise or momentary occlusion, and cannot independently represent the overall illumination level of the spatial area. Therefore, this application further traverses the index table and performs voxel-level data aggregation on non-empty voxels, statistically analyzing the total number of feature points within each voxel and the number that meets the stability threshold, and calculating the statistical average of the photometric variance. This allows for the elimination of random noise interference from individual data through statistical methods, extracting spatially representative regional illumination features, quantifying the feature richness and overall illumination volatility within the spatial unit, providing robust statistical basis for subsequent confidence scoring, and ensuring that map attributes reflect the essence of the environment rather than transient noise.

[0034] Specifically, in one possible embodiment, the data processing unit accesses each non-empty voxel containing feature points one by one according to the voxel index mapping table. For the current voxel, the detailed attributes of all internal feature points are retrieved from the photometric statistics list according to the index list. Subsequently, the algorithm counts the total number of feature points falling into the voxel and compares it with a preset variance threshold to filter out the effective feature points with stable illumination for counting. At the same time, the local photometric variances of all feature points within the voxel are accumulated and the arithmetic mean is calculated to characterize the average illumination intensity of the spatial unit. Finally, the total number of feature points, the number of stable points, and the average photometric variance are packaged to generate an aggregated data package specific to the voxel, which serves as the basic input data for subsequent scoring calculations.

[0035] Specifically, in step S23, multi-dimensional visual confidence calculation is performed on the voxel aggregated data packet to obtain an illumination confidence raster map. It should be understood that feature richness and illumination stability are two key factors affecting the accuracy of visual SLAM positioning in different dimensions, and a single index cannot comprehensively assess the positioning reliability of the environment. Therefore, this application further performs multi-dimensional fusion calculation on the voxel aggregated data packet, comprehensively considering the effective feature density and illumination fluctuation within the region through a weighted model, transforming statistical data into normalized confidence probability values. This generates an illumination confidence raster map that intuitively reflects the visual positioning friendliness of each region, allowing regions with high feature density and stable illumination to receive high scores, while regions with drastic illumination changes or poor feature quality receive low scores. This transforms complex multi-source perception data into a single scalar field that can be directly utilized by path planning algorithms, achieving accurate quantification of environmental perception risks.

[0036] Specifically, in one possible embodiment, the calculation module reads the voxel aggregation data packet and applies a preset confidence scoring model for calculation. This model includes a feature richness term and an illumination stability term. First, it calculates the proportion of stable feature points to the total number of features to evaluate the effectiveness of the features; simultaneously, it uses an exponential decay function to process the average photometric variance, converting the illumination fluctuation value into a stability coefficient. Then, based on preset weighting factors (such as setting a feature richness weighting coefficient), the calculation is performed... =0.6, Light stability weighting coefficient =0.4) The two indicators mentioned above are linearly weighted and summed to obtain the final visual confidence score of the voxel. Then, the score is written into the data structure corresponding to the raster map, and this process is repeated for all voxels. Finally, an illumination confidence raster map containing global spatial positioning reliability distribution information is output for subsequent modules to identify blind spots.

[0037] In the aforementioned method for planning forest blind spot inspection paths based on UAV collaboration, step S3 involves using a preset threshold to perform connected component segmentation and visibility search on the illumination confidence raster map to obtain a set of visually degraded regions and an anchor point mapping table corresponding to the set of visually degraded regions. It should be understood that although the illumination confidence raster map provides a global illumination quality score, discrete voxel scores cannot be directly used as spatial constraints for collaborative tasks, and low-scoring regions without external observation point support will leave the UAV isolated and helpless. Therefore, this application further implements connected component segmentation and visibility search to cluster scattered low-confidence voxels into entity regions with geometric boundaries, and matches the optimal external collaborative observation position for each region. This transforms the abstract illumination risk into a concrete collaborative escort task object, ensuring that when the main UAV enters the visual blind spot, the wingman is already positioned at the pre-calculated optimal anchor point to provide a stable observation link, thereby maintaining the continuity of perception and the robustness of the system.

[0038] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S3 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 5 As shown, step S3 includes: S31, performing connected component segmentation on the illumination confidence raster map based on a preset risk threshold to obtain a set of visually degraded regions; S32, sampling high-confidence candidate points around each region in the set of visually degraded regions to obtain a cluster of candidate anchor points associated with potential observation locations; S33, based on the illumination confidence raster map, performing visibility detection and collaborative observation utility evaluation on the candidate points in the cluster of candidate anchor points based on ray casting technology, and recording the coordinates of the candidate point with the highest utility value in the anchor point mapping table.

[0039] Specifically, in step S31, the illumination confidence raster map is segmented into connected components based on a preset risk threshold to obtain a set of visually degraded regions. It should be understood that due to the dynamic complexity of the forest environment, the illumination confidence map inevitably contains discrete low-volume elements caused by sensor noise or momentary occlusion. If these noise points are not processed, they will interfere with the decision-making logic of the path planner. Therefore, this application further performs threshold-based connected component segmentation to filter high-frequency spatial noise and extract low-illuminance regions with a certain volume that will have a continuous impact on visual positioning as independent connected components. In this way, the complex scalar field can be simplified into several specific risk geometries, enabling the system to clearly identify which are key degraded regions that require collaborative blind spot filling, thereby providing a precise spatial object basis for subsequent resource allocation and path constraints.

[0040] Specifically, in one possible embodiment, the algorithm module first traverses each voxel unit in the illumination confidence raster, comparing the voxel's confidence score with a preset risk threshold (e.g., a normalized value where the confidence score is below 0.4). If the score is below the threshold, it is marked as a risk voxel; otherwise, it is marked as a safe voxel. Then, a 3D connected component labeling algorithm is applied to examine the connectivity of risk voxels in their 3D spatial neighborhood, grouping adjacent risk voxels into the same connected component cluster. Next, the volume of each connected component is calculated, and tiny patches with volumes smaller than a specific threshold (e.g., less than 5 consecutive voxels or less than 0.05 cubic meters) are removed. Finally, the geometric centroid and bounding box of the remaining connected components are calculated, encapsulating them to generate a set of visually degraded regions containing spatial location and geometric morphological information.

[0041] Specifically, in step S32, a surrounding high-confidence candidate point sampling is performed on each region within the visual degradation area concentration to obtain a cluster of candidate anchor points associated with potential observation locations. It should be understood that since coordinated escort requires the wingman to remain in a location with good lighting conditions and a suitable distance, arbitrarily selecting observation points may lead to the wingman itself experiencing positioning instability, or observation failure due to excessive distance. Therefore, this application further performs surrounding high-confidence sampling around the degradation area to screen potential loitering spaces that are both adjacent to the risk area and have high lighting scores. This constructs a safe and effective set of candidate locations, ensuring that the wingman is always in a highly reliable visual positioning environment when performing auxiliary surveillance tasks, thus guaranteeing the cascade stability of the coordinated system from the source.

[0042] Specifically, in one possible embodiment, the planner uses the geometric centroid of each visual degradation region as a reference to define a spherical search space containing a minimum safe distance (e.g., 5 meters) and a maximum observation radius (e.g., 30 meters). Within this space, the algorithm samples discrete points at a preset step size (e.g., 1.0 meter) and reverse-checks the illumination confidence raster to obtain the confidence score for each sampling point. Only sampling points with scores higher than the safe anchoring threshold are retained, while locations with poor illumination conditions are filtered out. Subsequently, a clustering algorithm is used to sparsify the dense qualified sampling points, extracting representative spatial location nodes, and finally generating a cluster of candidate anchor points for the degradation region that is associated with potential safe observation locations.

[0043] Specifically, in step S33, based on the illumination confidence grid map, the candidate points in the regional candidate anchor point cluster are subjected to visibility detection and collaborative observation utility evaluation based on ray casting technology, and the coordinates of the candidate point with the highest utility value are recorded in the anchor point mapping table. It should be understood that since tree trunks and branches commonly obstruct the view in forest areas, anchor points selected solely based on distance and illumination conditions may not be able to see the interior of degraded areas through a straight line of sight, leading to collaborative observation failure. Therefore, this application further implements visibility detection and utility evaluation based on ray casting to accurately calculate the actual observation coverage and signal quality of each candidate point in the target blind zone. This allows for a quantitative evaluation of the tactical value of each candidate location, selecting the best collaborative anchor point that is both "stable" and "clearly visible," ensuring that when the main UAV penetrates deep into the blind zone, the wingman can establish an unobstructed, high-quality data link for auxiliary positioning.

[0044] Specifically, in one possible embodiment, the computation engine traverses each candidate point in the cluster of candidate anchor points in the region, emitting virtual rays from that point to each voxel within the corresponding visual degradation region. Then, using occupancy information from the illumination confidence raster, it detects whether there are obstacles obstructing the ray path and calculates the proportion of unobstructed visible voxels. Simultaneously, a specific collaborative observation utility scoring function is constructed, defined as the weighted sum of the anchor point's own illumination confidence and the proportion of visible voxels, multiplied by a distance-based utility coefficient. Specifically, a Gaussian kernel function with the optimal observation radius as the mean is used to normalize the observation distance score, ensuring that the closer the distance is to the optimal observation radius, the higher the score; if the distance is too far or too close, the score decays rapidly. Finally, the system sorts the utility values ​​of all candidate points, selects the point with the highest utility as the unique designated collaborative anchor point for that region, and writes its 3D coordinates into a mapping table, completing the logical binding from the risk area to the optimal support location.

[0045] Here, when evaluating the utility of collaborative observation, it is necessary to consider the unstructured, semi-transparent occlusion (such as sparse branches and leaves) environment of forest areas. The computational model needs to be robust in complex environments. Therefore, information theory and probabilistic perception models can be introduced to evaluate the utility of collaborative observation through anisotropic information gain calculation based on probabilistic line-of-sight transmission. That is, for the regional candidate anchor point cluster (potential observation station locations) and the visual degradation region set (lesion targets to be observed), if 0 or 1 is used to represent line of sight, since the line of sight is often in a "semi-occluded" state when passing through the edge of the tree canopy in forest areas, binarization processing will lead to the loss of a large number of "partially visible" high-value observation points, or over-evaluation of "pinhole" visible points, making the line of sight easily blocked by wind-blown leaves during actual flight, thus demonstrating the vulnerability of binarized line of sight. Furthermore, if we assume that all voxels in the degraded region have equal observational value, in reality, even in the degraded region (dark area), those voxels with relatively high luminosity variance (i.e. slightly richer texture) contribute far more to visual localization than solid-color (pure black / pure white) voxels. Therefore, anchor points should avoid homogenizing feature quality and prioritize seeing those voxels that have the greatest potential to provide features.

[0046] Therefore, the geometric counting needs to be improved to an integral estimation of the expected visual information content. A probabilistic transmission model is used, leveraging the Beer-Lambert theorem to model visibility as the integral of the "occupancy probability density" along the line-of-sight path. This effectively distinguishes between "complete occlusion," "sparse occlusion (leaves)," and "complete visibility." Simultaneously, the luminosity variance of feature points is introduced as a nonlinear mapping for feature entropy weighting. That is, if a voxel is black but has complex texture (large variance), it should be given a higher observation weight because once the wingman clearly sees it, it can provide stronger positioning constraints. Specifically, in a preferred embodiment, step S33 includes: performing ray step sampling on the line connecting the candidate anchor point and the target voxel in the visual degradation region set; calculating the integral of the occupancy probability density of the voxels passing through the ray path to obtain the cumulative optical thickness; and processing the cumulative optical thickness using an exponential decay function to obtain the probability pass-through factor; obtaining the local photometric variance of the target voxel; and mapping the local photometric variance using a nonlinear activation function to obtain the feature information weight; and calculating the probability visual information gain score based on the anchor point's own confidence, the feature information weight, and the probability pass-through factor, and using the probability visual information gain score as the collaborative observation utility value.

[0047] Specifically, firstly, ray step sampling is performed on the lines connecting candidate anchor points and target voxels in the visual degradation region concentration. The integral of the occupancy probability density of voxels passing through the ray path is calculated to obtain the cumulative optical thickness, and the cumulative optical thickness is processed using an exponential decay function to obtain the probabilistic visibility factor. Then, the target voxels are acquired. Local photometric variance The local luminosity variance is mapped and calculated using the Sigmoid nonlinear activation function to obtain feature information weights. Finally, based on the anchor point's self-confidence, feature information weights, and probabilistic visibility factor, the probabilistic visual information gain score is calculated and used as the collaborative observation utility value, expressed as: in, The probabilistic visual information gain score provides a more precise quantification of what would happen if the wingman landed at the candidate anchor point. It can provide a solution for degraded areas How much effective bit information does the reconstruction and positioning provide? It is the anchor point's own confidence level, used to ensure the observer's own safety. It is a feature information activation function, such as the Sigmoid activation function, which maps the luminance variance of voxels to information weights. The significance is that even if the line of sight is unobstructed, the value of observing a "solid color flat area" will be low, while the value of observing a "textured area" will be high. It represents the local luminance variance of a voxel, reflecting the richness of the voxel texture (the larger the variance, the richer the texture). The term is a probability pass-through factor. It is the line of sight. Upper position The voxel occupancy density at a given location (obtained from an octree or probabilistic raster plot, ranging from [0,1]) is used to calculate the cumulative "optical thickness" along the path via integration. For example, if the path passes through several sparse leaves ( If the exponent term does not become 0, it will become a smaller value (such as 0.6), indicating that "even with occlusion, there is still a probability of successful feature matching". It is a squared distance attenuation, used to preserve the inverse distance characteristic of physical optics. This is a regularization term used to prevent numerical explosion at extremely close distances. It is an optional observation angle gain, in which As the main characteristic direction, To determine the direction of the observation line of sight, dot product calculations can be used to prioritize the observation angle directly facing the feature surface, avoiding feature distortion caused by large-angle tilted observations.

[0048] This avoids abandoning support due to the inability to find a perfectly visible anchor point, or choosing a point that is actually severely obscured by branches. By tolerating soft occlusion, it seeks the optimal solution that "although there is foliage obstruction, the distance is close enough and the texture is good enough, resulting in the highest overall success rate," thus improving the survival rate in complex forest occlusion conditions. Furthermore, through... With the introduction of the function, the system no longer blindly pursues "seeing more volume" but instead pursues "seeing more texture," enabling the wingman to accurately lock onto the feature points that are most effective for the visual SLAM algorithm. This significantly improves the host's positioning accuracy under the same computing power, maximizing the actual effect of collaborative positioning.

[0049] In the aforementioned method for planning blind spot inspection paths in forest areas based on UAV collaboration, step S4 generates a navigation cost field based on the illumination confidence grid map and the set of visually degraded regions. It should be understood that forest environments not only contain explicit physical obstacles such as tree branches and trunks, but also implicit perceptual obstacles that cause visual positioning failure due to drastic changes in illumination. Traditional maps based solely on geometric occupancy information cannot represent such perceptual risks, easily leading to UAVs entering blind spots and becoming lost. Therefore, this application further combines illumination confidence and visually degraded region information to generate a navigation cost field, thereby constructing a comprehensive potential field model that integrates physical obstacle avoidance constraints and visual positioning reliability constraints. This transforms abstract visual positioning risks into quantifiable travel costs for the path planning algorithm, enabling UAVs to automatically avoid poorly lit areas during the planning stage, just like avoiding physical obstacles, thus fundamentally improving the safety and robustness of the inspection path.

[0050] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S4 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 6 As shown, step S4 includes: S41, calculating the nearest obstacle distance for the occupied voxels in the illumination confidence raster and performing geometric repulsion cost calculation in combination with a preset safety radius to obtain a geometric distance field; S42, using the visual degradation region set to determine the spatial attributes of voxels in the illumination confidence raster to determine the region adaptive penalty weight, and performing inverse proportional risk mapping on the illumination confidence scores of the voxels to obtain a visual risk field; S43, performing multi-source cost weighted fusion of the geometric distance field and the visual risk field to obtain a navigation cost field.

[0051] Specifically, in step S41, the nearest obstacle distance is calculated for the occupied voxels in the illumination confidence raster, and a geometric repulsion cost is calculated in conjunction with a preset safety radius to obtain a geometric distance field. It should be understood that since UAVs need to maintain a safe distance from physical obstacles such as tree trunks and branches when navigating through forests, and the binarized occupied raster cannot provide gradient information to guide smooth obstacle avoidance, this application further performs nearest obstacle distance calculation and geometric repulsion cost calculation on the occupied voxels to construct an Euclidean symbolic distance field and transform the discrete obstacle boundaries into a continuously changing repulsion potential energy field. This ensures the formation of a high-cost buffer zone around obstacles, forcing the planned path to remain smooth while moving away from obstacles, avoiding physical collisions caused by flying close to obstacles, and providing basic flight safety for UAVs.

[0052] Specifically, in one possible embodiment, the algorithm first extracts all voxels marked as occupied in the illumination confidence raster as a seed point set. Next, a linear-time Euclidean distance transform algorithm is applied to calculate the Euclidean distance from each idle voxel in the map to the nearest seed point. Then, a repulsion threshold is set based on a preset physical safety radius of the drone (e.g., 1.5 times the drone's maximum wingspan, approximately 0.8 meters). For voxels with a distance less than this threshold, their geometric repulsion cost is calculated using an exponential or inverse proportional function, resulting in a higher cost for closer distances. For voxels with a distance greater than the threshold, their geometric cost is set to zero, ultimately generating a geometric distance field reflecting the distribution of physical collision risk.

[0053] Specifically, in step S42, the spatial attributes of voxels in the illumination confidence raster map are determined using the visual degradation region set to determine the adaptive penalty weights for the regions, and an inverse proportional risk mapping is performed on the illumination confidence scores of the voxels to obtain the visual risk field. It should be understood that since the illumination conditions of different regions have varying degrees of influence on visual localization, and the identified visual degradation regions have an extremely high risk of localization loss, they need to be highlighted and strongly avoided. Therefore, this application further utilizes the visual degradation region set to determine spatial attributes and performs an inverse proportional risk mapping to apply additional adaptive penalty weights to high-risk regions and converts the illumination confidence scores into specific navigation resistance values. This creates a high potential energy barrier for illumination blind spots in the cost field, making the passage cost higher in areas with poorer illumination conditions, thereby guiding the path planning algorithm to actively choose areas with good illumination conditions for passage.

[0054] Specifically, in one possible embodiment, the processing unit traverses each voxel in the illumination confidence raster, querying whether the voxel's coordinates are contained within the spatial range of the visual degradation region set. Then, a truncated inverse barrier function is used to construct the visual potential field: if the voxel is located inside a degradation region, it is assigned a high-gain region penalty weight coefficient; if it is located outside, it is assigned a base weight coefficient. Subsequently, the visual risk cost of the voxel is calculated using the formula: weight coefficient multiplied by (1 / (current illumination confidence + ϵ)), where ϵ is a small relaxation factor to prevent the denominator from being zero. This formula converts the normalized confidence value into a nonlinear high-potential-energy drag value; that is, the lower the confidence, the more rapidly the calculated risk cost increases hyperbolically. Finally, the calculation results for all voxels are mapped to three-dimensional space to form a visual risk field characterizing the distribution of visual positioning failure risk.

[0055] Specifically, step S43 involves multi-source cost weighted fusion of the geometric distance field and the visual risk field to obtain the navigation cost field. It should be understood that since physical obstacle avoidance and visual hazard avoidance are constraints in two dimensions, and an optimal balance needs to be found between them in actual flight, a single field data cannot meet the requirements for solving the globally optimal path. Therefore, this application further implements multi-source cost weighted fusion of the geometric distance field and the visual risk field to unify the heterogeneous physical repulsion force and visual repulsion force under the same mathematical dimension, and balances the weights of safety and perception stability by adjusting coefficients. This generates a scalar field that comprehensively reflects the complexity of the environment, enabling subsequent algorithms to prioritize the planning of flight trajectories with rich visual features and stable lighting, while ensuring no collisions occur.

[0056] Specifically, in one possible embodiment, the fusion algorithm initializes the final navigation cost field and visits the corresponding voxel positions of the geometric distance field and the visual risk field point by point. Then, a visual factor adjustment coefficient (e.g., setting the adjustment weight to 2.5) is introduced as a hyperparameter to weight the visual risk cost, adjusting the algorithm's sensitivity to lighting conditions. Subsequently, the geometric exclusion cost, the weighted visual risk cost, and a preset path length regularization term (e.g., setting the distance cost weight to 0.5) are linearly superimposed. This regularization term guides the path towards the shortest path in unobstructed and well-lit areas. Finally, the fused navigation cost field is output as input data for the multi-machine collaborative path solving module.

[0057] In the aforementioned method for forest blind spot inspection path planning based on UAV collaboration, step S5 involves solving the navigation cost field and anchor point mapping table for multi-UAV collaborative escort path to obtain a visual chain trajectory set. It should be understood that forest inspection tasks involve not only the physical obstacle avoidance of a single UAV, but more importantly, ensuring the stability and non-divergence of its positioning system when the main UAV has to traverse visual blind spots with extremely poor lighting conditions. Simple single-UAV path planning cannot mobilize collaborative resources, while collaboration without spatiotemporal constraints can lead to the breakage of the observation link. Therefore, this application further implements multi-UAV collaborative escort path solving based on the navigation cost field and anchor point mapping table to construct a dual-UAV tactical maneuver scheme that is precisely coupled in time and space, logically binding the main UAV's traversal actions with the wingman's support actions. This ensures that the system can automatically generate a visual chain trajectory in which the main and wingmen rely on each other when facing complex lighting environments. This enables the drone swarm to maintain the continuity and robustness of its perception capabilities when performing high-risk inspection tasks, and achieves a leap from single-machine intelligence to multi-machine swarm intelligence.

[0058] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S5 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 7 As shown, step S5 includes: S51, performing a main task path search based on the navigation cost field to obtain the segmented marked original path; S52, traversing the segmented marked original path, matching corresponding cooperative anchor points for the identified blind zone crossing segments according to the anchor point mapping table, and constructing a main wingman task queue containing relative temporal constraints to obtain a cooperative waypoint pair sequence; S53, performing multi-aircraft spatiotemporal trajectory joint optimization on the cooperative waypoint pair sequence to obtain a visual chain trajectory set.

[0059] Specifically, step S51 involves performing a main task path search based on the navigation cost field to obtain a segmented and marked original path. It should be understood that while the navigation cost field has quantified the repulsive forces of physical obstacles and the visual risk resistance of illumination in the environment, the initial path generated by the planner is merely a series of discrete geometric coordinate points, lacking semantic descriptions of path segment attributes. Without distinguishing between safe and dangerous segments on the path, subsequent coordinated scheduling will be ineffective. Therefore, this application further performs a main task path search based on the cost field and segments the path, thereby accurately identifying which segments on the path are "inevitable" visual blind spots while obtaining the optimal passage path. This deconstructs continuous flight missions into two distinct state sequences: "safe cruise" and "blind spot penetration," providing clear spatiotemporal triggering conditions for accurately triggering the wingman escort support mechanism, avoiding resource waste during safe periods and lack of support during high-risk periods.

[0060] Specifically, in one possible embodiment, the path planning engine first performs a global path solution using a 3D A search algorithm in the 3D navigation cost field. Specifically, the system maintains a minimum priority queue based on the total cost function. The voxel nodes are expanded and sorted. From the starting point to the current node The actual cumulative movement cost is obtained by summing the path integrals of all voxels traversed along the path in the navigation cost field (scalar values ​​that combine geometric repulsion and visual risk weights). For the current node The estimated heuristic cost to the target point is calculated by multiplying the Euclidean distance by a preset heuristic weight coefficient. The algorithm expands nodes outwards using a 26-neighborhood connectivity approach, prioritizing the selection of nodes from all directions. The system iterates through the smallest node to ensure that, while favoring well-lit areas with lower costs, it can calculate a globally optimal path with the lowest overall travel cost (including physical and visual safety) when traversing shadow areas to reach the target point. Subsequently, the system performs a secondary scan along the generated path, checking the overall cost of each voxel traversed. When the cost of a continuous path segment exceeds a preset high-risk threshold, the system marks the path segment as a "blind zone traversal segment" and records its entry and exit point indices; otherwise, it marks it as a "regular cruise segment." Finally, it outputs a segmented original path with segment attribute labels, serving as the basic input for subsequent collaborative logic.

[0061] Specifically, in step S52, the original path segmented and marked is traversed, and corresponding cooperative anchor points are matched for the identified blind zone crossing segments according to the anchor point mapping table. A primary and wingman task queue containing relative temporal constraints is constructed to obtain a cooperative waypoint pair sequence. It should be understood that simple path segmentation only indicates when the primary UAV faces risk, without providing specific instructions on how the wingman should act. Furthermore, the wingman's movement takes time and must be pre-emptively maneuvered to achieve effective coverage. Therefore, this application further traverses the segmented path and matches cooperative anchor points to establish a logical mapping between the primary UAV's risk period and the wingman's safe observation position, and constructs a task queue containing strict relative temporal constraints. This transforms the abstract "support requirement" into a concrete "tactical instruction," forcing the wingman to reach the designated anchor point and hover before the primary UAV enters the blind zone, thereby eliminating the lag in cooperative actions in the time dimension and ensuring that the wingman is in position and has established a stable observation link at the primary UAV's most vulnerable moment.

[0062] Specifically, in one possible embodiment, the task scheduler parses the segmented original path segment by segment. When a path marked as a "blind spot crossing segment" is parsed, the visual degradation region ID corresponding to that segment is extracted, and the corresponding optimal observation anchor point coordinates are looked up in the anchor point mapping table. Then, a wingman's flight task is generated, setting the wingman's target as the anchor point coordinates. Simultaneously, strict timing constraints are established: the wingman's arrival time at the anchor point must be less than or equal to the time the main UAV arrives at the blind spot crossing segment entrance, and the wingman's departure time from the anchor point must be greater than or equal to the time the main UAV leaves the blind spot crossing segment exit. Then, the main UAV's crossing waypoints are paired with the wingman's hovering waypoints, with the aforementioned inequality timing constraints applied, and pushed sequentially into the task queue, ultimately generating a sequence of cooperative waypoint pairs containing dual-aircraft cooperative action commands.

[0063] Specifically, step S53 involves performing multi-aircraft spatiotemporal trajectory joint optimization on the cooperative waypoint pair sequence to obtain a visual chain trajectory set. It should be understood that since the waypoint pairs generated in the preceding steps are only discrete spatial positions with rigid time constraints, direct execution would lead to a broken UAV flight trajectory, abrupt maneuvers, and difficulty in accurately meeting microsecond-level time synchronization requirements, thus affecting the stability of the video stream. Therefore, this application further implements multi-aircraft spatiotemporal trajectory joint optimization on the cooperative waypoint pair sequence to fit discrete waypoints into continuous, smooth high-order polynomial curves, minimizing energy loss and time errors while satisfying dynamic constraints. This generates executable trajectories that conform to the physical characteristics of the UAV, ensuring smooth coordination between the main and wingmen even at high speeds, achieving both precise spatial positioning and seamless temporal synchronization, thereby outputting a high-quality visual chain trajectory set.

[0064] Specifically, in one possible embodiment, the trajectory optimizer receives a sequence of cooperative waypoint pairs and constructs a safe flight corridor consisting of a series of convex polyhedra along the waypoint sequence based on the navigation cost field, transforming it into control point constraints and spatial geometric boundary constraints for a polynomial trajectory. Then, a joint optimization objective function incorporating the states of the master and wingman aircraft is constructed. This function aims to minimize the trajectory jerk to ensure flight smoothness, while introducing a penalty term to minimize synchronization errors at critical time points. The objective function is iteratively solved using a nonlinear programming solver, adjusting the time allocation parameters and polynomial coefficients of each trajectory segment until all position constraints, velocity constraints, timing constraints, and safe flight corridor boundary constraints are satisfied. Finally, the algorithm outputs a set of parameterized time-position curves, i.e., a visual chain trajectory set. This trajectory set defines in detail the precise position, velocity, and acceleration commands of each UAV at any given time, directly used for underlying flight control execution.

[0065] In the aforementioned method for planning blind spot inspection paths in forest areas based on UAV collaboration, step S6 involves controlling the coordinated flight of a UAV swarm based on a visual chain trajectory set. It should be understood that due to the complexity of the forest environment, the swarm must maintain strict spatiotemporal synchronization when traversing blind spots; simple open-loop control cannot cope with trajectory deviations caused by airflow disturbances. Therefore, this application further implements closed-loop coordinated flight control based on a visual chain trajectory set to transform parameterized spatiotemporal curves into precise physical maneuver commands and maintain the relative geometric observation relationship between the UAVs. This ensures that the UAV swarm strictly adheres to the preset tactical formation in a dynamic environment, guaranteeing the continuity of the visual relay link and the accuracy of the inspection task.

[0066] In particular, in one specific embodiment, Figure 8 This is a flowchart of sub-step S6 of the forest blind spot inspection path planning method based on UAV collaboration according to an embodiment of this application. Figure 8 As shown, step S6 includes: S61, performing low-level instruction parsing and tracking on the visual chain trajectory set, and performing extended Kalman filtering collaborative correction based on inter-machine relative observation data when the state estimation covariance exceeds the limit to obtain a real-time observation data stream; S62, performing instantaneous illumination confidence reassessment on the real-time observation data stream to obtain an instantaneous confidence measurement packet reflecting the true illumination conditions at the current location; S63, based on the instantaneous confidence measurement packet, using a recursive Bayesian update strategy and observation uncertainty coefficient to perform online numerical correction on the illumination confidence raster map to obtain an updated local confidence map.

[0067] Specifically, in step S61, the visual chain trajectory set undergoes low-level command parsing and tracking, and when the state estimation covariance exceeds the limit, extended Kalman filtering collaborative correction is performed based on inter-machine relative observation data to obtain a real-time observation data stream. It should be understood that due to drastic changes in illumination under the forest canopy, visual odometry can experience cumulative drift in areas lacking texture, making it highly susceptible to positioning divergence and collisions if relying solely on single-machine state estimation. Therefore, this application further monitors the state covariance in real time during command tracking and triggers extended Kalman filtering collaborative correction based on inter-machine observations when uncertainty exceeds the limit, thereby introducing strong external constraints. This allows the stable relative position information provided by the wingman to suppress the host's positioning error, ensuring high-precision state estimation is maintained even under extreme conditions where visual features fail.

[0068] Specifically, in one possible embodiment, the airborne computer employs a nonlinear model predictive control algorithm as the underlying tracking controller. The controller uses the UAV's position, velocity, and attitude as state variables, and total thrust and body angular velocity as control variables, constructing predictive equations based on a six-degree-of-freedom dynamic model. Within each control cycle, the system uses the reference state in the visual chain trajectory set as the tracking target, constructs a quadratic programming cost function including state tracking error terms and control input smoothing terms, and solves for the optimal control sequence under physical constraints such as maximum thrust and maximum tilt angle, converting the first term of the sequence into a motor speed command. Simultaneously, the system calculates the trace of the state estimation covariance matrix of the visual odometry in real time. When the trace value exceeds a preset safety threshold and is within the cooperative window period, the system automatically activates the inter-aircraft relative positioning link to acquire distance and angle data relative to the anchor point wingman. Subsequently, the extended Kalman filter algorithm is used to weightedly fuse this relative observation value with the internal state prediction value, outputting a corrected real-time observation data stream.

[0069] Specifically, step S62 involves reassessing the instantaneous illumination confidence of the real-time observation data stream to obtain an instantaneous confidence measurement package reflecting the true illumination conditions at the current location. It should be understood that the illumination conditions during prior map construction may differ significantly from the actual environment during inspection, and dynamic factors such as cloud cover can alter local illumination quality in real time. Therefore, this application further reassesses the instantaneous illumination confidence of the images in the real-time observation data stream to obtain the true illumination quality score at the current location at the current moment, i.e., the instantaneous measurement value. This captures the dynamic changes in ambient illumination, providing an objective basis based on measured data for subsequent static map correction and preventing path planning from relying on outdated environmental information.

[0070] Specifically, in one possible embodiment, the current frame image is first extracted from the real-time observation data stream, and the local photometric variance and the number of effective feature points within the field of view are calculated using the same photometric analysis algorithm as in the preprocessing stage. Then, these statistical indicators are converted into normalized instantaneous measurement confidence values ​​according to a preset scoring model. Simultaneously, spatial back-projection is performed in conjunction with the UAV's current corrected pose to determine the raster voxel index corresponding to the current observation. Finally, the voxel index and measurement confidence are packaged to generate an instantaneous confidence measurement package containing accurate spatial labels and illumination quality scores.

[0071] Specifically, in step S63, based on the instantaneous confidence measurement package, the illumination confidence raster map is numerically corrected online using a recursive Bayesian update strategy and observation uncertainty coefficients to obtain an updated local confidence map. It should be understood that since a single instantaneous measurement is easily affected by sporadic factors such as sensor noise or insect occlusion, directly overwriting the original map data can lead to numerical oscillations and instability in the environmental model. Therefore, this application further utilizes a recursive Bayesian update strategy and observation uncertainty coefficients to smooth the map, thereby integrating historical prior information with current observational evidence. This enables the illumination confidence raster map to possess online learning and adaptive evolution capabilities, gradually approximating the true illumination distribution of the environment while suppressing noise.

[0072] Specifically, in one possible embodiment, the map maintenance module reads the historical confidence values ​​of the corresponding voxels in the illumination confidence raster map and calculates the observation uncertainty coefficient based on the current flight attitude stability. Then, it applies a recursive Bayesian update formula or a weighted exponential moving average algorithm to adjust the learning rate with the uncertainty coefficient, fusing new data from the instantaneous confidence measurement package into the historical data. Finally, the calculated posterior confidence value is written into the corresponding storage unit of the raster map to generate an updated local confidence map. Simultaneously, it detects in real time whether the currently unexecuted trajectory segment crosses an area where the updated confidence is below the safety threshold. If the detection result is yes, a global replanning instruction is immediately triggered and the process jumps back to step S5 to regenerate a safe trajectory that avoids newly discovered illumination blind spots; if the detection result is no, the current flight strategy is maintained, thereby achieving dynamic closed-loop control of perception update and path execution.

[0073] In summary, the UAV-based collaborative forest blind spot inspection path planning method based on the embodiments of this application is explained. First, photometric preprocessing is performed on the multi-source sensor data of the UAV swarm to extract visual features and establish a lighting statistics list. Then, feature back-projection is used to map the visual information to three-dimensional space, quantifying and evaluating the visual confidence of the area, and constructing a lighting confidence model characterizing perception reliability. Based on this, visual degradation areas prone to positioning failure are identified, and anchor point mapping is established by combining visibility search to generate a navigation cost field that integrates lighting risks. Finally, multi-UAV collaborative path solving is performed under the constraints of the cost field to generate a chain-like trajectory that avoids blind spots while maintaining perception connectivity. In this way, the perception risks caused by drastic changes in lighting can be transformed into hard constraints for path planning, effectively solving the visual positioning drift problem under complex lighting conditions in forest areas and achieving highly reliable collaborative inspection by UAV swarms.

[0074] Figure 9 This is a block diagram of a forest blind spot inspection path planning system based on UAV collaboration, according to an embodiment of this application. Figure 9 As shown, the forest blind spot inspection path planning system 100 based on UAV collaboration according to an embodiment of this application includes: a photometric preprocessing module 110, used to perform photometric preprocessing on the acquired raw sensor data stream of the UAV group to obtain a visual feature set and a photometric statistics list, wherein the raw sensor data stream of the UAV group includes camera image streams, inertial navigation data, and illumination sensor readings of each UAV; a confidence mapping module 120, used to perform feature back projection and visual confidence score calculation on a preset three-dimensional voxel grid based on the visual feature set and photometric statistics list to obtain an illumination confidence raster map; and an illumination visibility module 120. The system comprises: a performance analysis module 130, used to perform connected component segmentation and visibility search on the illumination confidence raster map using a preset threshold to obtain a set of visual degradation regions and an anchor point mapping table corresponding to the set of visual degradation regions; a navigation cost field construction module 140, used to generate a navigation cost field based on the illumination confidence raster map and the set of visual degradation regions; a cooperative escort path solving module 150, used to solve the multi-aircraft system cooperative escort path on the navigation cost field and the anchor point mapping table to obtain a set of visual chain trajectories; and an aircraft swarm cooperative flight control module 160, used to control the cooperative flight of the UAV swarm according to the set of visual chain trajectories.

[0075] As described above, the UAV-based forest blind spot inspection path planning system 100 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with UAV-based forest blind spot inspection path planning algorithms. In one possible implementation, the UAV-based forest blind spot inspection path planning system 100 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the UAV-based forest blind spot inspection path planning system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the UAV-based forest blind spot inspection path planning system 100 can also be one of many hardware modules of the wireless terminal.

[0076] Alternatively, in another example, the drone-based forest blind spot inspection route planning system 100 and the wireless terminal can also be separate devices, and the drone-based forest blind spot inspection route planning system 100 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0077] Here, those skilled in the art will understand that the specific operations of each step in the above-mentioned UAV-based forest blind spot inspection path planning system have been referenced above. Figures 1 to 8 The method for planning blind spot inspection routes in forest areas based on UAV collaboration has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for planning a forest blind spot inspection path based on cooperation of unmanned aerial vehicles, characterized in that, include: S1: Perform photometric preprocessing on the acquired raw sensor data stream of the drone fleet to obtain a visual feature set and a photometric statistics list. The raw sensor data stream of the drone fleet includes camera image streams, inertial navigation data and illumination sensor readings of each drone. S2: Based on the visual feature set and photometric statistics list, feature back projection and visual confidence score calculation are performed on the preset three-dimensional voxel grid to obtain the illumination confidence raster map. The illumination confidence raster map is a three-dimensional rasterized map with three-dimensional voxels as units, and each voxel contains a normalized visual confidence score calculated based on the weighted fusion of feature richness and illumination stability. S3: Use a preset threshold to perform connected component segmentation and visibility search on the illumination confidence raster to obtain a set of visually degraded regions and an anchor point mapping table corresponding to the set of visually degraded regions. The anchor point mapping table records the correspondence between each visually degraded region and the coordinates of the best collaborative observation position determined after visibility detection and collaborative observation utility evaluation. S4: Generate a navigation cost field based on the illumination confidence raster map and the set of visually degraded regions; S5: Solve the multi-aircraft system collaborative escort path for the navigation cost field and anchor point mapping table to obtain the visual chain trajectory set. The visual chain trajectory set is a set of multi-aircraft parameterized spatiotemporal flight trajectories that satisfy the temporal constraints between the master and wingmen and maintain the visual observation link connection. S6: Control the coordinated flight of a swarm of drones based on a visual chain trajectory set; Step S3 includes: The illumination confidence raster is segmented into connected components based on a preset risk threshold to obtain a set of visually degraded regions. A surrounding high-confidence candidate point sampling method is used to obtain a cluster of candidate anchor points that are associated with potential observation locations for each region where the visual degradation region is concentrated. Based on the illumination confidence raster map, visibility detection and collaborative observation utility evaluation are performed on candidate points in the candidate anchor point cluster using ray casting technology. The coordinates of the candidate point with the highest utility value are recorded in the anchor point mapping table. This includes: performing ray step sampling on the line connecting the candidate anchor point and the target voxel in the visual degradation region, calculating the integral of the occupancy probability density of the voxel passing through the ray path to obtain the cumulative optical thickness, and processing the cumulative optical thickness using an exponential decay function to obtain the probability visibility factor; obtaining the local photometric variance of the target voxel, and mapping the local photometric variance using a nonlinear activation function to obtain the feature information weight; calculating the probability visual information gain score based on the anchor point's own confidence, feature information weight, and probability visibility factor, and using the probability visual information gain score as the collaborative observation utility value.

2. The forest blind spot inspection path planning method based on UAV cooperation according to claim 1, characterized in that, Step S1 includes: Data stream parsing and adaptive histogram equalization are performed on the raw sensor data stream of the aircraft cluster to obtain an enhanced grayscale image stream; Gradient-based primary feature point extraction is performed on the enhanced grayscale image stream to obtain a candidate feature coordinate set; Based on the candidate feature coordinate set, a local neighborhood window is determined in the enhanced grayscale image stream, and local variance statistics and descriptor extraction are performed on the pixel data within the local neighborhood window to obtain a visual feature set and a photometric statistics list. 3.The forest blind spot inspection path planning method based on UAV cooperation according to claim 1, characterized in that, Step S2 includes: Based on the boundary constraints of the prior geometric map and the preset resolution parameters, the three-dimensional coordinates in the visual feature set are discretized, projected, and indexed to obtain a voxel index mapping table. Traverse the voxel index mapping table and perform voxel-level photometric statistical feature aggregation on the feature points inside non-empty voxels to obtain a voxel aggregated data package containing the number of stable feature points and the average photometric variance data. Multidimensional visual confidence calculation is performed on the voxel aggregation data package to obtain the illumination confidence raster.

4. The forest blind spot inspection path planning method based on UAV cooperation according to claim 1, characterized in that, Step S4 includes: The nearest obstacle distance is calculated for the occupied voxels in the illumination confidence raster, and the geometric repulsion cost is calculated by combining the preset safety radius to obtain the geometric distance field. The spatial attributes of voxels in the illumination confidence raster image are determined by using the set of visually degraded regions to determine the region adaptive penalty weights, and the illumination confidence scores of the voxels are inversely proportionally risk-mapped to obtain the visual risk field. The navigation cost field is obtained by multi-source cost weighted fusion of the geometric distance field and the visual risk field.

5. The forest blind spot inspection path planning method based on UAV cooperation according to claim 1, characterized in that, Step S5 includes: Perform a main task path search based on the navigation cost field to obtain the segmented original path; Traverse the original path segmented by marking, match the corresponding cooperative anchor points for the identified blind spot crossing segments according to the anchor point mapping table, and construct a master wingman task queue containing relative temporal constraints to obtain a cooperative waypoint pair sequence. Multi-aircraft spatiotemporal trajectory joint optimization is performed on the cooperative waypoint pair sequence to obtain a visual chain trajectory set.

6. The UAV coordination based forest blind spot inspection path planning method according to claim 1, characterized in that, Step S6 includes: The visual chain trajectory set is parsed and tracked at the low level, and when the state estimation covariance exceeds the limit, extended Kalman filter is used for collaborative correction based on inter-machine relative observation data to obtain the real-time observation data stream. The instantaneous illumination confidence is reassessed on the real-time observation data stream to obtain an instantaneous confidence measurement package that reflects the true illumination conditions at the current location; Based on the instantaneous confidence measurement package, the illumination confidence raster map is numerically corrected online using a recursive Bayesian update strategy and the observation uncertainty coefficient to obtain the updated local confidence map.

7. A forest blind spot inspection route planning system based on UAV collaboration, used to execute the forest blind spot inspection route planning method based on UAV collaboration according to any one of claims 1-6, characterized in that, include: The photometric preprocessing module is used to perform photometric preprocessing on the acquired raw sensor data stream of the drone fleet to obtain a visual feature set and a photometric statistics list. The raw sensor data stream of the drone fleet includes camera image streams, inertial navigation data and illumination sensor readings of each drone. The confidence mapping module is used to perform feature back projection and visual confidence score calculation on a preset three-dimensional voxel grid based on a visual feature set and a photometric statistics list to obtain an illumination confidence raster map. The illumination visibility analysis module is used to perform connected component segmentation and visibility search on the illumination confidence raster map using preset thresholds to obtain a set of visually degraded regions and an anchor point mapping table corresponding to the set of visually degraded regions. The navigation cost field construction module is used to generate a navigation cost field based on the illumination confidence raster map and the set of visually degraded regions; The cooperative escort path solving module is used to solve the multi-aircraft system cooperative escort path from the navigation cost field and anchor point mapping table to obtain a visual chain trajectory set; The swarm cooperative flight control module is used to control the cooperative flight of a swarm of drones based on a visual chain trajectory set.

Images