An autonomous obstacle avoidance control method for a high-voltage transmission line inspection unmanned aerial vehicle

Abstract

The application relates to the technical field of unmanned aerial vehicle obstacle avoidance control, in particular to an autonomous obstacle avoidance control method for a high-voltage transmission line inspection unmanned aerial vehicle, which realizes online identification of newly added Gaussian nodes and real-time updating of an environment model through multi-view visible light data and three-dimensional Gaussian splash model incremental registration, and solves the problem of environment mismatch of a pre-built map. A minimum circumscribed sphere of a newly added obstacle is used to construct an orthogonal plane to divide eight regions, a Gaussian node feature is combined with a distance to quantize an obstacle avoidance risk index, and a direction misjudgment rate is reduced. A guide vector is constructed based on a selected region to decompose a translation direction, a distance difference value is combined to calculate an offset, a control instruction is output, obstacle avoidance safety is ensured, a flight path is automatically cut back, and the inspection autonomy and reliability are improved.

Description

Technical Field

[0001] This application relates to the field of drone obstacle avoidance control technology, specifically to an autonomous obstacle avoidance control method for inspection drones applicable to high-voltage power transmission lines. Background Technology

[0002] Regularly inspecting high-voltage power lines using drones is a crucial measure to ensure the safety of power lines and prevent power accidents. To prevent drones from colliding with obstacles around the lines during inspections, they need to have autonomous obstacle avoidance capabilities to ensure flight safety and the successful completion of the inspection mission.

[0003] Existing autonomous obstacle avoidance control methods for UAVs typically require the use of multiple sensors to measure distances around the UAV, followed by obstacle avoidance via control commands. However, in power transmission line inspection scenarios, due to the need for periodic, multiple missions, the environment around the power transmission line may change with the duration of each inspection, introducing new obstacles. This leads to a mismatch between the pre-built environmental map and the actual obstacles during each inspection. Existing methods fail to adequately consider the near-real-time requirements of 3D reconstruction, resulting in insufficient perception of newly added obstacles in the power transmission line environment and a risk of delayed obstacle avoidance response. Summary of the Invention

[0004] In view of the above, it is necessary to provide an autonomous obstacle avoidance control method for inspection drones of high-voltage transmission lines to solve the above problems.

[0005] One embodiment of this application provides an autonomous obstacle avoidance control method for an inspection drone suitable for high-voltage transmission lines, the method comprising:

[0006] Based on multi-view visible light data collected during the inspection of high-voltage power transmission lines by UAVs, the newly added Gaussian node set and the incrementally updated 3DGS model of the current environment are obtained through the three-dimensional Gaussian splash update algorithm and feature map registration. The newly added obstacles are obtained by clustering the newly added Gaussian node set.

[0007] Construct three reference planes that are mutually orthogonal and all pass through the center of the minimum circumscribed sphere of the newly added obstacle. Divide the minimum circumscribed sphere into eight spatial regions. Analyze the color texture, coordinate scale, and quantity characteristics of all Gaussian nodes in each spatial region to obtain the spatial complexity characteristics of each spatial region. Combine the minimum straight-line distance between the UAV inspection route and the newly added obstacle to determine the obstacle avoidance risk index of each spatial region and select the spatial region as the obstacle avoidance direction based on this index.

[0008] Based on the distance between the newly added obstacle and the inspection route within the selected spatial area, a guidance vector is constructed and decomposed to determine the horizontal and vertical translation directions for obstacle avoidance. Based on the difference between the actual projected distance and the safe distance between the current inspection route of the UAV and the newly added obstacle, the final offset is calculated and control commands are output.

[0009] Preferably, obtaining the newly added Gaussian node set and the incrementally updated current environment 3DGS model specifically involves:

[0010] The visible light data is extracted using a pre-defined deep neural network model, and the visible light data is spatially registered with an existing 3DGS scene model based on the image feature map.

[0011] Calculate the 2D cosine similarity between the registered visible light image and the current 3DGS rendered image, and count the number of effective viewpoints in three-dimensional space that confirm the changes.

[0012] If the 2D cosine similarity is lower than a preset similarity threshold and the number of effective viewpoints is higher than a preset viewpoint threshold, the newly added Gaussian node set and the current environment 3DGS model are output through incremental update optimization; otherwise, the newly added Gaussian node set is an empty set, and the existing 3DGS scene model remains unchanged.

[0013] Preferably, the three reference planes include: a first plane parallel to the ground, a second plane perpendicular to the ground and parallel to the projection direction of the UAV fuselage longitudinal axis on the horizontal plane, and a third plane perpendicular to the ground and perpendicular to the projection direction.

[0014] Preferably, the process of obtaining the spatial complexity features of each spatial region specifically includes:

[0015] For each spatial region, analyze the dispersion of color information of all Gaussian nodes in each channel to determine the texture fluctuation index; analyze the dispersion of scale information of coordinates of all Gaussian nodes in each direction to determine the geometric shape index.

[0016] The mean of the normalized results of the texture fluctuation index and the geometric morphology index of each spatial region is calculated. The mean is then positively fused with the proportion of Gaussian nodes in each spatial region across all spatial regions to obtain the spatial complexity features of each spatial region.

[0017] Preferably, determining the texture fluctuation index specifically involves:

[0018] Calculate the coefficient of variation of the color information of all Gaussian nodes in each channel, and average the obtained coefficients of variation of all channels to obtain the texture fluctuation index.

[0019] Preferably, determining the geometric morphology index specifically involves:

[0020] Calculate the coefficient of variation of the scale information of the coordinates of all Gaussian nodes in each direction, and average the obtained coefficients of variation in all directions to obtain the geometric morphology index.

[0021] Preferably, the specific process for determining the obstacle avoidance risk index for each spatial region is as follows:

[0022] If there are no Gaussian nodes in the current spatial region, their obstacle avoidance risk index is set directly to zero.

[0023] If Gaussian nodes exist, the minimum straight-line distance between the spatial region and the UAV inspection route is obtained, and normalization is performed based on the maximum value of the minimum straight-line distance in all spatial regions. The normalized result is subtracted from the value 1, and the average value is obtained by adding it to the spatial complexity characteristics of the spatial region to obtain the obstacle avoidance risk index.

[0024] Preferably, the obstacle avoidance direction is the spatial region with the lowest obstacle avoidance risk index.

[0025] Preferably, the process of constructing the guiding vector specifically includes:

[0026] If there are no new obstacles in the spatial region corresponding to the obstacle avoidance direction, the vector pointing from the center of the smallest circumscribed sphere to the geometric center of that spatial region will be used as the guiding vector.

[0027] Otherwise, obtain the Gaussian node closest to the UAV's flight path within the spatial region corresponding to the obstacle avoidance direction, and use the opposite direction of the perpendicular vector from that Gaussian node to the UAV's flight path as the guiding vector.

[0028] Preferably, the formula for calculating the final offset is as follows:

[0029] in, This indicates the drone's horizontal offset. This indicates the drone's vertical offset. This represents the magnitude of the horizontal component of the guiding vector. This represents the magnitude of the vertical height component of the guiding vector. Indicates the safe distance for drones. Indicates an additional safety margin; This represents 1 / 2 of the drone's fuselage length. This represents 1 / 2 of the drone's fuselage height, and max() represents the maximum value function.

[0030] To address the problem that existing autonomous obstacle avoidance methods for unmanned aerial vehicles (UAVs) suffer from insufficient real-time performance in 3D reconstruction, leading to delayed perception of newly added obstacles around power transmission lines and increasing the risk of collisions, this application offers the following advantages:

[0031] Firstly, addressing the issue of pre-built maps becoming invalid due to environmental changes (such as tree growth or new illegal constructions) during inspections, this application utilizes multi-view visible light data and incremental feature map registration with the 3DGS model to identify and extract newly added Gaussian node sets online, updating the current environmental model in real time. This mechanism effectively solves the mismatch between pre-built maps and the actual environment, enabling timely and accurate perception of dynamic obstacles around transmission lines.

[0032] Secondly, to address the difficulty in selecting obstacle avoidance directions due to the irregular shapes of newly added obstacles, this application constructs three mutually orthogonal reference planes, using the center of the smallest circumscribed sphere of the newly added obstacle as the origin, dividing the space into eight equal regions. By comprehensively analyzing the color texture, coordinate scale, and quantity distribution characteristics of Gaussian nodes in each region, and combining this with the minimum straight-line distance between the UAV and the obstacle, the obstacle avoidance risk index of each spatial region is accurately quantified. This multi-dimensional assessment of spatial complexity effectively reflects the geometric complexity and collision risk in different directions, significantly reducing the probability of misjudging the obstacle avoidance direction.

[0033] Finally, during the obstacle avoidance execution phase, this application constructs a guidance vector based on the selected safe space area, decomposes it into horizontal and vertical translation directions, and calculates the final offset by combining the difference between the actual projected distance and the safe distance. This control strategy not only ensures that the UAV maintains a safety margin with respect to newly added obstacles during obstacle avoidance, but also automatically and smoothly switches back to the original inspection route after avoiding risks, minimizing interference with the inspection task and comprehensively improving the autonomy and reliability of high-voltage transmission line inspection in complex dynamic environments. Attached Figure Description

[0034] Figure 1 A flowchart of the autonomous obstacle avoidance control method for an inspection drone applicable to high-voltage power transmission lines provided in this application;

[0035] Figure 2 A schematic diagram of the spatial area provided for this application;

[0036] Figure 3 A schematic diagram illustrating the minimum straight-line distance for the drone to pass beside a newly added obstacle, as provided in this application. Detailed Implementation

[0037] This application proposes an autonomous obstacle avoidance control method for UAVs used in the inspection of high-voltage transmission lines, applying the field of UAV obstacle avoidance control technology. (See attached document.) Figure 1 The method includes the following steps:

[0038] Step 1: Based on the multi-view visible light data collected during the inspection of high-voltage power transmission lines by UAVs, the newly added Gaussian node set and the incrementally updated 3DGS model of the current environment are obtained through the 3D Gaussian splash update algorithm and feature map registration. The newly added Gaussian node set is then clustered to obtain the newly added obstacles.

[0039] Prior to real-time data acquisition, a 3D map of the target high-voltage transmission tower scene was constructed based on visible light data from the maiden flight phase. This 3D map is represented by a 3D Gaussian splash model, in which all key points required for detailed inspection (e.g., insulators, fittings, guy wires, etc.) are marked, and a precise mapping from the map coordinate system to the UAV's body coordinate system has been pre-completed. During each subsequent inspection mission, the UAV acquires its initial pose via GPS and IMU modules, and performs visual feature registration by combining real-time image data from the visible light camera with the existing 3D map, thereby obtaining the UAV's precise pose in the map coordinate system.

[0040] During the inspection, the drone uses its onboard rotatable CCD camera to collect multi-view visible light data of the surrounding environment in real time. The acquisition interval is set to 1 second, and the acquisition angles are -15°, 0°, and +15°, with a frontal angle of 0°. Visible light data from three perspectives are collected at each moment and transmitted to the onboard edge computing module via the onboard communication module. On the onboard edge computing module, the real-time collected multi-view visible light data is used as input to the mean filtering algorithm to denoise the visible light data. The denoised multi-view visible light data is then used as input to the CL-Splats algorithm. A preset deep neural network model (DINOv2 model) is used to extract feature maps from the denoised visible light data. Based on these feature maps, the visible light data is spatially registered with an existing 3DGS scene model, outputting a newly added set of Gaussian nodes and an incrementally updated 3DGS model of the current environment.

[0041] Each Gaussian node contains three-dimensional spatial coordinates, color information, and scale information. The color information represents the basic color coefficients of each RGB channel (the color components of each RGB channel under the 0th order spherical harmonic coefficients), reflecting the texture characteristics of the Gaussian node surface. The scale information is the length value of each Gaussian node in the principal axes x, y, and z directions, which is a three-dimensional vector, reflecting the local geometric features covered by the node in space. The number of local optimization iterations is 30.

[0042] After registration, environmental change detection is performed: First, the 2D cosine similarity between the real-time acquired visible light image and the rendered image of the existing 3DGS model at the corresponding viewpoint is calculated. This similarity reflects the degree of difference between the current real environment and the historical map. Second, the number of effective viewpoints confirming changes in the area in 3D space is counted, and 3D change detection is performed. The count of effective viewpoints is used for multi-view cross-validation to eliminate instantaneous noise interference from a single viewpoint caused by illumination reflections, birds, etc. The cosine similarity threshold for 2D change detection and the effective viewpoint threshold for 3D change detection are both configured by the control terminal based on the lighting conditions and hardware performance of the historical inspection environment. In this embodiment, the similarity threshold is set to 0.5, and the viewpoint threshold is set to two viewpoints.

[0043] When the 2D change detection cosine similarity is lower than a preset similarity threshold and the number of effective 3D change detection viewpoints is higher than a preset viewpoint threshold, it indicates the presence of newly added Gaussian nodes. This signifies that significant changes in the physical environment have been cross-confirmed across multiple viewpoints, and thus, a new obstacle is determined to exist. The system uses visible light data as input to 3D Gaussian incremental update algorithms such as CL-Splats (the number of local optimization iterations can be configured, for example, to 30), optimizes, and outputs the set of newly added Gaussian nodes and the incrementally updated 3DGS model of the current environment.

[0044] Otherwise, the newly added Gaussian node set output is empty, and the output 3DGS model remains the existing 3DGS model. The CL-Splats algorithm is a well-known technique in the field of 3D reconstruction, and its specific calculation method will not be elaborated here. If the newly added Gaussian node set output by the CL-Splats algorithm is empty, it indicates that no obvious environmental changes have been detected, and the UAV continues to fly along the original refined inspection route while maintaining the default obstacle avoidance mechanism; if the newly added Gaussian node set is not empty, it indicates that potential obstacles have been detected, and subsequent processing is then performed on this set.

[0045] Considering that the newly added Gaussian node set contains Gaussian nodes for all newly added obstacles, and that the UAV needs to avoid different newly added obstacles during subsequent obstacle avoidance, it is necessary to separate the Gaussian nodes for different newly added obstacles. In one specific implementation of this application, the three-dimensional spatial coordinates of all Gaussian nodes in the newly added Gaussian node set are used as input to the DBSCAN algorithm to obtain one or more clusters. The neighborhood radius is set to 0.4 meters, and the minimum number of sample points is set to 100. Gaussian nodes that are spatially close and have continuous density are clustered into the same cluster, and each cluster is considered a newly added obstacle.

[0046] Step 2: Construct three reference planes that are mutually orthogonal and all pass through the center of the minimum circumscribed sphere of the newly added obstacle. Divide the minimum circumscribed sphere into eight spatial regions. Analyze the color texture, coordinate scale, and quantity characteristics of all Gaussian nodes in each spatial region to obtain the spatial complexity characteristics of each spatial region. Combine the minimum straight-line distance between the UAV inspection route and the newly added obstacle to determine the obstacle avoidance risk index of each spatial region and select the spatial region as the obstacle avoidance direction based on this index.

[0047] During the inspection of high-voltage power lines by drones, any newly added obstacles located at or near the preset flight path may pose a collision risk. Furthermore, the Gaussian splash model does not represent the entire newly added obstacle as a single Gaussian node; instead, it reconstructs it into a point cloud cluster composed of numerous Gaussian nodes based on its spatial geometry, surface texture, and other characteristics. The number of Gaussian nodes corresponding to a newly added obstacle not only reflects the object's size but also reflects its complexity through the differences in the distribution of attributes such as color and scale among the Gaussian nodes. For newly added obstacles, the most directly relevant indicator of collision risk is whether the minimum straight-line distance between the object and the drone is less than the drone's safe distance. Therefore, determining whether a drone will collide with a newly added obstacle requires not only calculating the differences in the distribution of color and scale of the Gaussian nodes but also considering whether the minimum straight-line distance between the object and the drone is less than the safe distance.

[0048] Based on the above analysis, this application first divides the newly added obstacles into spatial regions, analyzes all Gaussian nodes in each spatial region, and reflects the geometric complexity of each spatial region through the dispersion of color and scale distribution; then it analyzes the spatial positional relationship between each spatial region and the UAV, reflecting the degree of impact of each spatial region on the UAV's obstacle avoidance; finally, it combines the two cases to construct the obstacle avoidance risk index for each spatial region.

[0049] Considering that newly added obstacles during high-voltage power line inspections are often irregularly shaped, such as randomly growing trees or haphazardly piled illegal buildings, these obstacles can encroach on different areas to varying degrees. This can lead to drones blindly choosing obstacle avoidance directions, potentially resulting in them selecting directions with high geometric complexity and increasing the risk of collisions. Therefore, in a specific implementation of this application, the q-th newly added obstacle is denoted as L(q), and the centroid of all Gaussian nodes within L(q) is taken as the geometric center of L(q). A minimum circumscribed sphere is constructed with this center as the sphere's center and the Euclidean distance from the sphere's center to the farthest Gaussian node as the radius. Then, the spatial coordinates of all Gaussian nodes within the minimum circumscribed sphere are subtracted from the coordinates of the sphere's center, transforming all Gaussian nodes into a local relative coordinate system with the sphere's center as the origin, facilitating subsequent spatial position analysis.

[0050] Then, with the center of the sphere as the origin, construct three mutually orthogonal planes that all pass through the center of the sphere: the first plane (parallel to the ground), the second plane (perpendicular to the ground and parallel to the projection direction of the UAV's longitudinal axis onto the horizontal plane), and the third plane (perpendicular to the ground and perpendicular to this projection direction). Divide the smallest circumscribed sphere into eight spatial regions: front-up-left, front-up-right, front-down-left, front-down-right, back-up-left, back-up-right, back-down-left, and back-down-right. A schematic diagram of the spatial regions is shown below. Figure 2 As shown, 11 represents the upper hemisphere, 12 represents the lower hemisphere, 13 represents the left hemisphere, and 14 represents the right hemisphere. The black dot in the figure represents the current position of the drone, the directional arrows represent the flight direction of the drone, and 1, 2, ..., 8 represent the markings of eight spatial regions.

[0051] The direction vectors of the eight spatial regions based on the origin are (1,1,1), (-1,1,1), (1,-1,1), (-1,-1,1), (-1,-1,1), (1,-1,-1), (-1,-1,-1), (1,1,-1), (-1,1,-1). The centroid coordinates of each spatial region are the product of the direction vector and half the radius.

[0052] By dividing the space into eight sections, the space around a newly added obstacle can be decomposed into eight independent directional regions. In subsequent analysis, the distribution density, color, and scale of Gaussian nodes in each direction can be effectively analyzed, thereby accurately obtaining the geometric complexity and obstacle avoidance cost of each direction. This enables the UAV to quickly identify collision risks in each direction in a dynamically changing environment, reduce the probability of obstacle avoidance response lag and misjudgment, and improve the UAV's autonomous obstacle avoidance robustness in complex environments.

[0053] The spatial relationship analysis between newly added obstacles and UAVs needs to consider the minimum straight-line distance when the UAV passes by the newly added obstacle. In a specific implementation of this application, the spatial relationship between each spatial region of L(q) and the UAV is further analyzed. The i-th spatial region of L(q) is labeled L(q,i). Taking L(q,i) as an example, the current three-dimensional coordinates of the UAV and the three-dimensional coordinates of the next key point in the preset flight path are obtained from the airborne INS system. A spatial straight line is determined using these two points, and the direction from the UAV's position to the next key point is the UAV's current flight direction. The distances between all Gaussian nodes in L(q,i) and this spatial straight line are calculated, and the minimum value is taken as the minimum straight-line distance of L(q,i). Similarly, the minimum straight-line distances of all spatial regions are calculated, and their minimum values ​​are taken as the minimum straight-line distance of L(q). A schematic diagram of the minimum straight-line distance when the UAV passes by the newly added obstacle is shown below. Figure 3As shown, w1 represents the drone's position, w2 represents the position of L(q), w3 represents the drone's next critical point position, the direction from w1 to w3 is the drone's current flight direction, r1 represents the minimum straight-line distance between w1 and w2, r2 represents the minimum straight-line distance from w2 to the drone's current flight direction, and o represents the intersection of the minimum straight-line distances of L(q) and the drone's current flight path. When the minimum straight-line distance of L(q) is greater than the safe distance, it means that L(q) will not affect the drone's movement, no obstacle avoidance is needed, the calculation ends, and the drone continues to move according to the original settings; when the minimum straight-line distance of L(q) is less than the safe distance for the drone, it means that L(q) has a collision risk.

[0054] Since the number, color distribution, and scale distribution of all Gaussian nodes constituting the newly added obstacle can effectively reflect the complexity of the obstacle itself, and considering the structural differences of the newly added obstacle in different spatial regions of L(q), to reflect this difference, for all Gaussian nodes in L(q,i), the coefficient of variation of the basic color coefficients of the R channel in all their color information is calculated. Similarly, the coefficients of variation of the basic color coefficients of the G and B channels are calculated, and the mean of the coefficients of variation of the basic color coefficients of the RGB three channels is calculated as the texture fluctuation index, reflecting the dispersion of the color information of L(q,i), denoted as . Calculate the coefficient of variation of the length values ​​in the x-direction across all scale information. Similarly, calculate the coefficients of variation of the length values ​​in the y and z directions. Then, calculate the mean of the coefficients of variation of the length values ​​in the x, y, and z directions. This mean is used as a geometric morphology index to reflect the degree of dispersion of the scale information of L(q,i), denoted as […]. ; The larger the value, the more drastic the fluctuation distribution of L(q,i) in optical rendering. In the construction of 3D Gaussian splash scenes, the drastic fluctuation phenomenon in optical rendering usually corresponds to extremely small and irregular objects that are difficult to scan in the physical environment (such as scattered crisscrossing thin wires or overlapping dead branches at the ends). The larger the value, the more irregular the geometry of the newly added obstacle in L(q,i). Since the number of Gaussian nodes also reflects the complexity of the newly added obstacle, combining the number of Gaussian nodes in L(q,i) with... , To construct the spatially complex features of L(q,i), the computational relationship is as follows: ,in Representing the spatial complexity of L(q,i), This represents the number of Gaussian nodes in L(q,i). This represents the total number of Gaussian nodes in L(q). This represents the maximum value of the dispersion of color information across all spatial regions in L(q). This represents the maximum value of the scale information dispersion of all spatial regions in L(q). To represent extremely small positive numbers, to prevent the distribution of color or scale information across all spatial regions in L(q) from being extremely uniform. or The problem of the denominator being zero. The range of its value is [0.0001, 0.001], and its value is 0.0005; , It can be seen as a pair , The results of normalizing the maximum value are shown separately. Because... and The property fluctuations of each Gaussian node are analyzed from two dimensions: color and scale. Therefore, a weighted average is used to... and The normalized results are combined to reflect the degree of texture and geometric non-uniformity within L(q,i); furthermore, since the normalized result of the number of Gaussian nodes in L(q,i) can reflect the spatial scale of L(q,i), the normalized result of the number of nodes is used as an adjustment factor to obtain... . This reflects the spatial complexity of L(q,i). The larger the value, the higher the complexity of L(q,i), the greater the threat to the drone's movement, and the higher the potential flight impact brought about by the spatial complexity of L(q,i).

[0055] Considering that the higher the complexity of each spatial region, the greater its complexity, the lower the success rate of obstacle avoidance in the direction of the UAV during obstacle avoidance. Therefore, an obstacle avoidance risk index for each spatial region is constructed by combining the geometric complexity characteristics and the minimum straight-line distance (taking L(q,i) as an example):

[0056] L(q,i) represents the obstacle avoidance risk index. Let L(q,i) represent the minimum straight-line distance. This represents the maximum value of the minimum straight-line distance across all spatial regions. This represents a preset minimum positive number, set to 0.01 in this embodiment. If there are no Gaussian nodes within L(q,i), it means that L(q,i) has no new obstacle structure information, indicating that the direction is completely open. Open areas will not pose a collision risk to the drone. Set as .

[0057] The obstacle avoidance risk index reflects the overall adverse effect of L(q,i) on UAV obstacle avoidance. The larger the value, the greater the complexity of L(q,i), the greater the obstacle avoidance distance cost, and the less suitable it is as an obstacle avoidance direction for drones.

[0058] This reflects the geometric complexity of L(q,i). This reflects the minimum straight-line distance between L(q,i) and the UAV inspection route. The larger the value, the higher the complexity of L(q,i), and the lower the success rate of the drone in obstacle avoidance. The smaller the value, the smaller the distance between the Gaussian nodes inside L(q,i) and the straight line in space. This means it's more likely to be less than the safe flight distance for the drone, and the longer the obstacle avoidance distance needs to be to keep the drone outside the safe distance. Therefore, and All with They are directly proportional, so we combine them using a weighted average method.

[0059] Similarly, the obstacle avoidance risk index of the eight spatial regions is calculated, and the obstacle avoidance direction of the spatial region with the lowest obstacle avoidance risk index is selected as the obstacle avoidance direction of the UAV. If multiple spatial regions have the same lowest obstacle avoidance risk index, a unique obstacle avoidance direction is selected according to the preset priority order of spatial regions (prioritizing vertically upward spatial regions, followed by lateral regions).

[0060] Step 3: Based on the distance between the newly added obstacle and the inspection route within the selected spatial area, construct the guidance vector and decompose it to determine the horizontal and vertical translation directions for obstacle avoidance. Based on the difference between the actual projected distance and the safe distance between the UAV's current inspection route and the newly added obstacle, calculate the final offset and output the control command.

[0061] If there are no new obstacles in the spatial region corresponding to the obstacle avoidance direction, the vector pointing from the center of the smallest circumscribed sphere to the geometric center of that spatial region is taken as the guiding vector, and the magnitude of the horizontal component of the decomposed guiding vector is... Modulus of vertical height component All values ​​are assigned to 0. If a new obstacle is added to the spatial region corresponding to the obstacle avoidance direction, the Gaussian node closest to the UAV's flight path within that region is obtained. A perpendicular line is drawn from this node to the UAV's flight path, resulting in a perpendicular vector. The opposite direction of this perpendicular vector is used as the guiding vector. If the perpendicular vector is zero, the guiding vector is set to the system's default horizontal orthogonal vector. The obtained guiding vector is decomposed into a horizontal component and a vertical component, and their directions are used as the UAV's translation directions in the horizontal and vertical dimensions, respectively.

[0062] Furthermore, the horizontal and vertical components obtained from the decomposition are used to adjust the drone's offset in the horizontal and vertical directions, respectively:

[0063] Where max() represents the maximum value function. This indicates the drone's horizontal offset. This indicates the drone's vertical offset. This represents the magnitude of the horizontal component of the guiding vector. This represents the magnitude of the vertical height component of the guiding vector. The safe distance for the drone is indicated and obtained from the existing configuration file. This indicates an additional safety margin to prevent insufficient offset. The range of its value is [0.1, 0.5]m, and its value is 0.2m; This represents 1 / 2 of the drone's fuselage length. This represents half the drone's fuselage height; both values ​​are obtained by reading relevant configuration files. If... A value of 0 indicates that there is sufficient safety depth. A value of 0 indicates that there is already a sufficient safe altitude, and no additional translation command will be executed.

[0064] The horizontal and vertical components obtained from the above decomposition, along with their corresponding offsets, are... and As the expected input to the position controller, and combined with the drone's current speed, the obstacle avoidance speed components in the horizontal and vertical directions are calculated separately. For example, if the horizontal direction is left with an offset of 1.2 meters and the vertical direction is upward with an offset of 0.5 meters, the position controller sets the target point to "1.2 meters to the left and 0.5 meters upward." As the drone flies towards this target point, it will simultaneously receive leftward lateral speed commands and upward climb speed commands, thus achieving obstacle avoidance control. When the minimum straight-line distance between the drone and all parts of the newly added obstacle is greater than the safe distance and remains so for more than 1 second, obstacle avoidance is considered complete, and the system automatically switches back to the original refined inspection route to continue performing subsequent inspection tasks according to the preset path.

Claims

1. An autonomous obstacle avoidance control method for UAVs used in the inspection of high-voltage transmission lines, characterized in that: The method includes the following steps: Based on multi-view visible light data collected during the inspection of high-voltage power transmission lines by UAVs, the newly added Gaussian node set and the incrementally updated 3DGS model of the current environment are obtained through the three-dimensional Gaussian splash update algorithm and feature map registration. The newly added obstacles are obtained by clustering the newly added Gaussian node set. Construct three reference planes that are mutually orthogonal and all pass through the center of the minimum circumscribed sphere of the newly added obstacle. Divide the minimum circumscribed sphere into eight spatial regions. Analyze the color texture, coordinate scale, and quantity characteristics of all Gaussian nodes in each spatial region to obtain the spatial complexity characteristics of each spatial region. Combine the minimum straight-line distance between the UAV inspection route and the newly added obstacle to determine the obstacle avoidance risk index of each spatial region and select the spatial region as the obstacle avoidance direction based on this index. Based on the distance between the newly added obstacle and the inspection route within the selected spatial area, a guidance vector is constructed and decomposed to determine the horizontal and vertical translation directions for obstacle avoidance. Based on the difference between the actual projected distance and the safe distance between the current inspection route of the UAV and the newly added obstacle, the final offset is calculated and control commands are output.

2. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 1, characterized in that, The process of obtaining the newly added Gaussian node set and the incrementally updated current environment 3DGS model is as follows: The visible light data is extracted using a pre-defined deep neural network model, and the visible light data is spatially registered with an existing 3DGS scene model based on the image feature map. Calculate the 2D cosine similarity between the registered visible light image and the current 3DGS rendered image, and count the number of effective viewpoints in three-dimensional space that confirm the changes. If the 2D cosine similarity is lower than a preset similarity threshold and the number of effective viewpoints is higher than a preset viewpoint threshold, the newly added Gaussian node set and the current environment 3DGS model are output through incremental update optimization; otherwise, the newly added Gaussian node set is an empty set, and the existing 3DGS scene model remains unchanged.

3. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 1, characterized in that, The three reference planes include: a first plane parallel to the ground, a second plane perpendicular to the ground and parallel to the projection direction of the UAV's longitudinal axis on the horizontal plane, and a third plane perpendicular to the ground and perpendicular to the projection direction.

4. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 1, characterized in that, The process of obtaining the spatial complexity features of each spatial region is as follows: For each spatial region, analyze the dispersion of color information of all Gaussian nodes in each channel to determine the texture fluctuation index; analyze the dispersion of scale information of coordinates of all Gaussian nodes in each direction to determine the geometric shape index. The mean of the normalized results of the texture fluctuation index and the geometric morphology index of each spatial region is calculated. The mean is then positively fused with the proportion of Gaussian nodes in each spatial region across all spatial regions to obtain the spatial complexity features of each spatial region.

5. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 4, characterized in that, The determination of the texture fluctuation index specifically involves: Calculate the coefficient of variation of the color information of all Gaussian nodes in each channel, and average the obtained coefficients of variation of all channels to obtain the texture fluctuation index.

6. The autonomous obstacle avoidance control method for inspection drones of high-voltage transmission lines as described in claim 4, characterized in that, The determination of the geometric shape index specifically refers to: Calculate the coefficient of variation of the scale information of the coordinates of all Gaussian nodes in each direction, and average the obtained coefficients of variation in all directions to obtain the geometric morphology index.

7. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 1, characterized in that, The specific process for determining the obstacle avoidance risk index for each spatial region is as follows: If there are no Gaussian nodes in the current spatial region, their obstacle avoidance risk index is set directly to zero. If Gaussian nodes exist, the minimum straight-line distance between the spatial region and the UAV inspection route is obtained, and normalization is performed based on the maximum value of the minimum straight-line distance in all spatial regions. The obstacle avoidance risk index is obtained by subtracting the normalized result from the value 1, adding it to the spatial complexity characteristics of the region, and then averaging the results.

8. The autonomous obstacle avoidance control method for inspection drones of high-voltage transmission lines as described in claim 1, characterized in that, The obstacle avoidance direction is specifically the spatial region with the lowest obstacle avoidance risk index.

9. The autonomous obstacle avoidance control method for inspection drones applicable to high-voltage transmission lines as described in claim 1, characterized in that, The process of constructing the guiding vector is as follows: If there are no new obstacles in the spatial region corresponding to the obstacle avoidance direction, the vector pointing from the center of the smallest circumscribed sphere to the geometric center of that spatial region will be used as the guiding vector. Otherwise, obtain the Gaussian node closest to the UAV's flight path within the spatial region corresponding to the obstacle avoidance direction, and use the opposite direction of the perpendicular vector from that Gaussian node to the UAV's flight path as the guiding vector.

10. The autonomous obstacle avoidance control method for inspection drones of high-voltage transmission lines as described in claim 1, characterized in that, The formula for calculating the final offset is as follows: in, This indicates the drone's horizontal offset. This indicates the drone's vertical offset. This represents the magnitude of the horizontal component of the guiding vector. This represents the magnitude of the vertical height component of the guiding vector. Indicates the safe distance for drones. Indicates an additional safety margin; This represents 1 / 2 of the drone's fuselage length. This represents 1 / 2 of the drone's fuselage height, and max() represents the maximum value function.

Images