A method and device for alternate cooperative navigation of two unmanned aerial vehicles for long and narrow spaces

Abstract

The application discloses a kind of long and narrow space-oriented dual unmanned aerial vehicle alternate cooperative navigation method and equipment, it is related to robot autonomous navigation, multi-agent cooperative control and multi-sensor fusion technical field.The method comprises: the first unmanned aerial vehicle is set as reference unit, and the second unmanned aerial vehicle is set as unit to be solved;Obtain the ultra-wideband (UWB) ranging value between reference unit and unit to be solved;With the pose of reference unit and the error covariance of reference unit as space reference, state estimation filtering is carried out on unit to be solved to determine the accurate pose information of unit to be solved;Under the condition that the preset alternate condition is met, unit to be solved is converted and set as new reference unit, and the original reference unit is converted into new unit to be solved.In this way, the problems of irreversible divergence of global positioning in long and narrow space and existing vision cooperation easily failing can be solved.

Description

Technical Field

[0001] This application relates to the fields of robot autonomous navigation, multi-agent cooperative control and multi-sensor fusion technology, and in particular to a method and device for alternating cooperative navigation of two unmanned aerial vehicles (UAVs) in narrow spaces. Background Technology

[0002] In existing technologies, most inspection drones employ single-unit local odometer estimation techniques based on vision (such as the VINS algorithm) or lidar (such as the LOAM algorithm). However, narrow spaces, such as tunnels and utility tunnels, are characterized by "extremely simple features and a lack of texture and geometric variation." In such environments, single-unit SLAM is highly susceptible to "corridor degradation," failing to extract effective reference feature points along the tunnel's longitudinal direction (Z-axis). This leads to nonlinear divergence in the accumulated odometer error, ultimately causing severe global positioning drift.

[0003] To eliminate accumulated errors, a simultaneous localization and mapping (SMR) method based on mutual observation between heterogeneous unmanned systems (autonomous vehicles + drones) has been proposed. This approach proposes a concept where heterogeneous robots alternately remain stationary as a reference while another robot moves to explore. However, this approach suffers from the following drawbacks: it heavily relies on visual / laser mutual observation, and its cooperative principle requires the moving robot to be able to clearly see the stationary robot. But in confined spaces, such as mines or old utility tunnels, under harsh conditions with high dust, low illumination, or even complete darkness, visual and laser sensors are prone to failure, leading to momentary interruptions in mutual observation. Summary of the Invention

[0004] This application provides a dual-UAV alternating cooperative navigation method and device for narrow spaces, in order to solve the following technical problem: the problem that global positioning irreversible divergence and existing visual coordination are prone to failure in narrow spaces.

[0005] In a first aspect, embodiments of this application provide a dual-UAV alternating cooperative navigation method for narrow spaces. The method includes: setting a first UAV as a reference unit and a second UAV as a unit to be calculated, wherein the error covariance matrix of the pose estimation of the first UAV is set to a first fixed value approaching zero; obtaining ultra-wideband (UWB) ranging values ​​between the reference unit and the unit to be calculated; using the pose and error covariance of the reference unit as a spatial reference, performing state estimation filtering on the unit to be calculated to determine the precise pose information of the unit to be calculated, wherein the error covariance based on the reference unit is frozen to a first fixed value approaching zero. The first fixed value of zero causes the Kalman gain matrix to be biased toward the state component of the unit to be solved. Through the state constraint filtering algorithm based on asymmetric covariance freezing, the UWB ranging value is converted into an absolute correction amount for the cumulative pose error of the unit to be solved in the narrow space depth direction. Under the condition of satisfying the preset alternation condition, the unit to be solved is converted and set as the new reference unit, and the original reference unit is converted into the new unit to be solved. The accurate pose information of the new reference unit is used as the new spatial reference, and the error covariance matrix of the new reference unit is set to a second fixed value close to zero.

[0006] Secondly, embodiments of this application also provide a dual-UAV alternating cooperative navigation device for narrow spaces, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a dual-UAV alternating cooperative navigation method for narrow spaces as described in the first aspect above.

[0007] Thirdly, embodiments of this application also provide a computer storage medium storing computer-executable instructions, which, when executed, implement a dual-UAV alternating cooperative navigation method for narrow spaces as described in the first aspect above.

[0008] The present application provides a dual-UAV alternating cooperative navigation method and device for narrow spaces, which has the following advantages: This application embodiment utilizes a state constraint filtering algorithm based on asymmetric covariance freezing, which does not rely on mutual visual observation of optical features. Instead, it deeply couples mature UWB technology into the extended Kalman filter architecture of the SLAM backend. Through a leapfrog alternation mechanism of "one stationary unit as the absolute reference (locking the covariance), and one moving unit using UWB strong constraints to correct the Z-axis (eliminating drift)," it mathematically blocks the accumulation of Markov errors, achieving all-weather, centimeter-level accurate positioning under harsh conditions. This overcomes corridor degradation and visual failure, and achieves extremely low drift global positioning without GNSS. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a dual-UAV alternating cooperative navigation method for narrow spaces, provided as an embodiment of this application; Figure 2 This application provides a schematic diagram of a dual-machine positioning process. Figure 3 A schematic diagram illustrating the principle of a dual-machine positioning full-stack collaborative algorithm provided in this application embodiment; Figure 4 A schematic diagram illustrating the physical trajectory deviation comparison of a 30-meter tunnel provided in this application embodiment; Figure 5 A schematic diagram of the absolute trajectory error of the Z-axis and the filtering confidence level in a corridor degradation environment provided in this application embodiment; Figure 6 A schematic diagram comparing the response delay of obstacle avoidance commands when the computing power of a drone is fully loaded, provided as an embodiment of this application; Figure 7a A schematic diagram of a formation topology in a sudden narrow area provided in an embodiment of this application; Figure 7b A schematic diagram illustrating a panoramic coverage assessment provided in an embodiment of this application; Figure 8 This is a schematic diagram of the internal structure of a dual-UAV alternating cooperative navigation device for narrow spaces, provided as an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] In practical applications, Joint Kalman Filtering (EKF) is a state estimation method for multi-sensor or multi-agent systems. Its core idea is to bundle the states of multiple objects to be estimated (such as multiple drones) into a large overall state vector, and use the interrelationships between them (such as relative measurements) for unified estimation and optimization.

[0012] The main steps are as follows: (1) Joint State: The system constructs a joint state vector, which contains the independent states of all agents (such as position, velocity, attitude, etc.). At the same time, a joint error covariance matrix is ​​constructed to describe the uncertainty of each of these states and the correlation between them.

[0013] (2) Unified filtering: In the prediction and update steps of filtering, the joint state and joint covariance matrix are treated as a whole.

[0014] (3) Prediction: Each agent independently predicts its state based on its own motion model (such as IMU data), and its error covariance is also propagated and expanded accordingly.

[0015] (4) Update: When any observation information within the system is obtained (such as UWB ranging between agents, relative visual observation), the observation will update the entire joint state at the same time. This means that an observation can not only correct the state estimate of the directly measured agent, but also indirectly correct the state estimate of other agents through the correlation implied in the covariance matrix.

[0016] The following technical solutions exist in the prior art.

[0017] 1. Traditional SLAM (Simultaneous Localization and Mapping) navigation scheme based on a single UAV: Existing inspection drones mostly employ single-unit local odometer estimation techniques based on vision (such as the VINS algorithm) or lidar (such as the LOAM algorithm). However, tunnel and utility tunnel environments are characterized by "extremely homogeneous features and a lack of texture and geometric variation." In such environments, single-unit SLAM is highly susceptible to "corridor degradation," failing to extract effective reference feature points along the tunnel's longitudinal direction (Z-axis). This leads to nonlinear divergence in accumulated odometer errors, ultimately causing severe global positioning drift. Furthermore, the payload of micro-drones is limited, and their onboard microcomputing platforms often face significant computational bottlenecks when simultaneously processing full point cloud mapping, global positioning, and real-time obstacle avoidance planning. This can easily result in computational delays exceeding 150ms, causing obstacle avoidance lag or even collision hazards during high-speed inspections.

[0018] 2. Fixed formation UAV scheme based on traditional UWB ranging: To overcome the blind spots of a single drone's field of view, some existing technologies have introduced multi-drone formations and equipped the drones with UWB modules. However, in such solutions, the UWB module is only used as a relative reference to measure the distance between two drones, for basic collision avoidance and formation maintenance. Since the SLAM odometry of the two drones drifts in their respective environments, knowing their relative positions does not solve the problem of global coordinate loss, and the system as a whole still suffers from significant Z-axis displacement errors. Furthermore, traditional multi-drone formations are mostly in fixed formations. When construction scaffolding, hanging cables, or ventilation ducts cause sudden narrowing of the cross-section inside the utility tunnel, fixed formations cannot pass through, resulting in extremely poor environmental adaptability.

[0019] 3. A mutual observation scheme based on alternating static states in heterogeneous systems (closest to existing technology): To eliminate accumulated errors, some organizations have proposed a method for simultaneous localization and mapping (SMR) based on mutual observation between heterogeneous unmanned systems (autonomous vehicles + drones). This approach proposes that heterogeneous robots alternately remain stationary as a reference while another robot moves to explore. However, this approach suffers from three unavoidable drawbacks: (1) Heavy reliance on visual / laser mutual observation: The cooperative principle requires that the mobile robot must be able to see the stationary robot. However, in harsh working conditions such as mines and old pipe corridors with high dust, low light or even darkness, visual and laser sensors are very prone to failure, resulting in a momentary interruption of mutual observation.

[0020] (2) The computing power pain point of pure drone swarms is ignored: This solution uses ground unmanned vehicles equipped with high-performance computing devices to process complex data, without taking into account that when pure drones are in formation, both sides face strict computing power and load limitations, and cannot solve the problem of lag in obstacle avoidance computing power caused by high-frequency mapping.

[0021] (3) Lack of flexible spatial passage capability: Since the ground vehicle and the air vehicle are heterogeneous, the problem of formation crowding in narrow air sections is not involved. Therefore, no dynamic formation passage strategy for complex obstacles in the utility tunnel is provided.

[0022] In summary, the existing technology described above has the following problems: 1. To address the problem that global positioning is prone to irreversible divergence and existing visual coordination is easily rendered ineffective under conditions of corridor degradation and harsh working conditions.

[0023] Reason: Real underground long tunnels severely lack geometric and textural features. Due to the inherent Markov property of SLAM algorithms, the odometry error of a single machine in the longitudinal direction (Z-axis) accumulates with distance (i.e., tunnel degradation). Existing "heterogeneous alternating observation" schemes attempting to solve this problem rely on "line-of-sight (LOS) geometric feature matching" using optical / laser methods; however, in harsh conditions such as mine dust or unlit tunnels, the sensors are easily "blinded." Furthermore, conventional multi-machine UWB is only used for distance measurement and cannot constrain the absolute coordinate drift of each sensor.

[0024] Pain points and difficulties: In long-distance GNSS-free environments, whether it is a single unit or a traditional formation, the system cannot maintain high-precision global absolute coordinates, which makes it prone to navigation crashes and getting lost; and in harsh optical environments, the existing visual collaboration mechanism is useless.

[0025] 2. To address the issue of limited computing power in purely aerial micro-formation, which leads to lag in dynamic obstacle avoidance during high-speed inspections.

[0026] The underlying reason: Unlike ground-based unmanned vehicles, inspection drones, due to strict limitations on takeoff weight and endurance, can only carry micro-computing platforms with extremely limited computing power. When the drone needs to simultaneously process high-frequency front-end point cloud feature extraction, global state estimation, dense mapping, and complex local obstacle avoidance path planning, the CPU / GPU is normally under full load, causing computational process blockage.

[0027] Pain Points and Challenges: This computing power bottleneck manifests directly in engineering as system command latency. When a drone is flying at cruising speed and suddenly encounters unmapped random obstacles such as hanging cables or temporary scaffolding, the system simply does not have enough time to avoid them, making catastrophic collisions highly likely.

[0028] 3. Addressing the issues of low throughput in complex spaces and blind spots in single-sided scanning caused by the variable cross-section of the utility tunnel.

[0029] The underlying reason is that tunnels or utility tunnels in actual engineering projects are not ideal regular columns. They are often accompanied by irregular obstacles such as fans, pipeline intersections, and construction barriers, resulting in a sharp reduction in the local usable cross-sectional area. Existing multi-machine collaboration often uses hard-coded fixed geometric formations, lacking the ability to perceive the environment and dynamically reconstruct topological relationships.

[0030] Pain points and difficulties: On the one hand, fixed formations cannot pass through narrow sections and are forced to interrupt the mission and return to base; on the other hand, if a single drone is used to avoid the wall by flying close to it, the field of view (FOV) of the onboard sensor of a single drone is limited, which will cause the other side of the pipe wall to be in a blind spot, and cannot meet the high standard of full coverage inspection requirements.

[0031] To address the aforementioned problems, this application provides a dual-UAV alternating cooperative navigation method for narrow spaces. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0032] Figure 1 This is a flowchart illustrating a dual-UAV alternating cooperative navigation method for narrow spaces, as provided in an embodiment of this application. Figure 1 As shown in the figure, the dual-UAV alternating cooperative navigation method for narrow spaces provided in this application embodiment specifically includes the following steps: Step 101: Set the first UAV as the reference unit and the second UAV as the unit to be solved.

[0033] The error covariance matrix of the first UAV pose estimation is set to a first fixed value that approaches zero.

[0034] Step 102: Obtain the ultra-wideband (UWB) ranging value between the reference unit and the unit to be solved.

[0035] Step 103: Using the pose of the reference unit and the error covariance of the reference unit as a spatial reference, perform state estimation filtering on the unit to be solved to determine the precise pose information of the unit to be solved.

[0036] In this process, the error covariance of the reference unit is frozen to a first fixed value approaching zero, causing the Kalman gain matrix to be biased toward the state component of the unit to be solved. Through a state constraint filtering algorithm based on asymmetric covariance freezing, the UWB ranging value is converted into an absolute correction amount for the cumulative pose error of the unit to be solved in the longitudinal direction of the narrow space.

[0037] It should be noted that the core advantage of the state constraint filtering algorithm based on asymmetric covariance freezing is that it does not rely on external global positioning (such as GPS) and uses relative measurement to suppress cumulative errors. Therefore, it is suitable for various mobile machines that need to work collaboratively in complex environments without GPS or with limited signals. In addition to the aforementioned UAVs, it can also be used for underground pipeline / tunnel inspection robots, mine exploration robots, etc. These machines have autonomous mobility and are equipped with IMU (Inertial Measurement Unit) and UWB. They need to perform multi-machine collaborative positioning, navigation and operation in environments where GPS is denied and in narrow spaces.

[0038] Step 104: Under the condition of satisfying the preset alternation conditions, the unit to be solved is converted and set as the new reference unit, and the original reference unit is converted into the new unit to be solved.

[0039] The precise pose information of the new reference unit serves as a new spatial reference, and the error covariance matrix of the new reference unit is set to a second fixed value that approaches zero.

[0040] This application embodiment utilizes a state constraint filtering algorithm based on asymmetric covariance freezing, which does not rely on mutual visual observation of optical features. Instead, it deeply couples mature UWB technology into the extended Kalman filter architecture of the SLAM backend. Through a leapfrog alternation mechanism of "one stationary unit as the absolute reference (locking the covariance), and one moving unit using UWB strong constraints to correct the Z-axis (eliminating drift)," it mathematically blocks the accumulation of Markov errors, achieving all-weather, centimeter-level accurate positioning under harsh conditions. This overcomes corridor degradation and visual failure, and achieves extremely low drift global positioning without GNSS.

[0041] In one possible implementation, setting the first UAV as the reference unit and the second UAV as the unit to be solved includes: Initialize the known coordinates of the first and second UAVs at the entrance of the narrow space; Wherein, the known coordinate point is the origin of the global relative coordinate system, the first UAV is set to static anchor point mode and hovers in the narrow space, and the second UAV is set to dynamic inspection mode and begins to fly in the narrow space.

[0042] In practical applications, the two drones complete system power-on alignment at a known coordinate point at the entrance of a narrow space, such as a tunnel entrance (either by receiving GNSS signals or by manually placing it on a reference point with known absolute coordinates). The collaborative filtering optimizer initializes the initial position error covariance matrix of the first UAV-A and the second UAV-B to a constant, and this point is then used as the origin of the global relative coordinate system. The positions of all subsequent UAVs are calculated relative to this origin, which fundamentally avoids unknown drift of the entire map globally. Mathematically, the covariance matrix represents the "uncertainty" or "confidence" of the algorithm's position estimation. Setting it to a minimum provides a high-quality starting point for subsequent high-precision filtering calculations. UAV-A is assigned a static anchor mode (Role-Flag=0) and hovers within the tunnel; UAV-B is assigned a dynamic inspection mode (Role-Flag=1) and begins flight inspection along the tunnel direction.

[0043] In one possible implementation, the method further includes: the second UAV in the dynamic inspection mode collects point cloud features of the environment in front and converts the point cloud features into lightweight local obstacle representation information, and sends the local obstacle representation information to the first UAV in the static anchor point mode through an ad hoc network; The first UAV uses a path planning algorithm to generate an obstacle avoidance trajectory based on the local obstacle representation information and transmits it back to the second UAV for execution.

[0044] In practical applications, to avoid CPU overload caused by UAVs simultaneously handling global mapping and local planning, a cross-machine computing power decoupling mechanism can be adopted. The second UAV at the front of the inspection, after its perception module extracts the point cloud features of the environment ahead, no longer performs time-consuming global map stitching. Instead, it downsamples the data and converts it into lightweight local obstacle representation information, such as a lightweight local voxel map or a 3D bounding box. This data is then pushed to the rear UAV in real time via a mesh self-organizing network. After receiving the data, the UAV at the rear, which acts as a static anchor point, calls lightweight algorithms such as the artificial potential field method to generate obstacle avoidance trajectories, achieving low-latency obstacle avoidance.

[0045] In practical applications, to further compress communication bandwidth and reduce back-end computing power, 2D occupancy grid map projection can be used. That is, the second UAV in dynamic inspection mode directly projects the three-dimensional obstacles onto the two-dimensional cross-section of the UAV's safe flight to generate a planar grid. In this way, the first UAV does not need to perform any spatial mapping analysis, but only needs to superimpose the received force vectors into its own flight control attitude calculation, which can also achieve extremely low latency computing power decoupling obstacle avoidance.

[0046] To address the issue of CPU overload and response lag caused by the limited computing power of micro-drone platforms when simultaneously performing dense mapping and obstacle avoidance planning, a mechanism for decoupling master-slave computing power across machines and sharing local voxel maps was constructed. Lightweight obstacle-free path data is transmitted in real time by the forward sensing unit, and the backward inspection unit can directly call lightweight algorithms to generate avoidance trajectories.

[0047] In this way, not only can the bottleneck of computing power in micro airborne systems be alleviated, but the system response latency of high-speed dynamic obstacle avoidance can also be significantly reduced.

[0048] In one possible implementation, the state constraint filtering algorithm based on asymmetric covariance freezing includes: Construct a joint state vector and a joint error covariance matrix that includes the three-dimensional position, velocity, and attitude of the first UAV and the second UAV; By introducing a system logic state parameter, when it is determined that the first UAV is in the static anchor point mode, the error covariance of the first UAV is forcibly frozen to the first fixed value close to zero, so that the error covariance of the first UAV does not accumulate over time. An observation model is established based on the UWB ranging value, and the Kalman gain matrix is ​​solved so that the gain weight of the Kalman gain matrix is ​​unidirectionally biased toward the UWB ranging value. Perform a post-hoc state update. During the update period, keep the three-dimensional coordinates of the first UAV unchanged, perform forced convergence calibration on the Z-axis coordinates of the second UAV that have drifted, and reduce the error covariance of the second UAV.

[0049] In traditional joint Kalman filtering (EKF), the multi-machine error expands together over time. Even with UWB, only the relative distance can be constrained, and the global slippage of the two machines along the tunnel's Z-axis cannot be eliminated. The algorithm described above performs asymmetric reconstruction of the EKF's underlying layer: First, the joint state vector and covariance matrix are constructed. A joint state vector can be built within the edge computing box. ,in These represent the 3D position, velocity, and attitude of UAV-A and UAV-B, respectively. The corresponding joint error covariance matrix ∑ is constructed synchronously, with its main diagonal block matrices being... and .

[0050] Secondly, discrete-time updates constrained by physical state are introduced. In the discrete-time state deduction based on visual odometry, the algorithm introduces a system logical state parameter (Role-Flag). When the UAV-A is determined to be in static anchor mode, the system activates the asymmetric state freezing operator. Intervention is implemented on the state evolution equation. The state update equation of UAV-A is not iterated, and its error covariance matrix ∑A is kept at its minimum value to prevent its displacement error from accumulating over time. The code logic is: if(Role=0):PA=0.0001. UAV-B in flight is affected by the corridor degradation effect, resulting in cumulative movement. Its covariance is divided into blocks. It naturally expands with the distance integral.

[0051] Then, the state observation space and asymmetric Kalman gain are solved to obtain the physical ranging scalar based on UWB hardware. Establish an observation model Solve for the Kalman gain matrix. At that time, the pre-operator has constrained ∑A=0, and the structure of the new covariance matrix produces asymmetric polarization, which causes the weight allocation of the gain matrix K to be unidirectionally biased toward the ranging observation variables of UWB.

[0052] Finally, posterior state update and spatial drift calibration are performed. Based on the aforementioned asymmetric gain matrix K, the posterior state update equation is executed. During the update cycle, the three-dimensional spatial coordinates of UAV-A remain unchanged, while the distorted Z-axis coordinates of UAV-B are subject to a strong unidirectional constraint from the gain matrix, forcing them to converge to the neighborhood of the true value. At the same time, the corresponding uncertain covariance ∑B shrinks significantly after the update, thus completing the global periodic calibration of the corridor degradation drift.

[0053] In one possible implementation, the alternation condition includes at least one of the following: The UWB ranging value between the UAV in the dynamic inspection mode and the UAV in the static anchor point mode reaches a preset distance threshold. The pose estimation error covariance of the UAV in the dynamic inspection mode exceeds the preset covariance threshold.

[0054] In practical applications, to prevent UWB signal disconnection, alternating distance thresholds can be set. When the UAV-B in the dynamic inspection mode reaches the UWB signal distance threshold, or its covariance exceeds the system's preset upper limit, the system instantly triggers a state transition command. The UAV-B stops flying and enters a parking state. The system uses the coordinates calibrated by the UAV-B at the last moment as the static anchor point mode. At the same time, the UAV-A, which was originally in the static anchor point mode, is locked and changes to dynamic inspection mode. After flying past the UAV-B, the distance between the two UAVs is recalculated, and the UAV-A moves forward for inspection.

[0055] In practical applications, navigation also requires acquiring inertial navigation data from both the unit to be calculated and the reference unit. Inertial navigation data refers to the raw data directly measured by the inertial measurement unit (IMU) and preliminarily processed; it is the fundamental information used to calculate the carrier's position, velocity, and attitude. Between the intervals of UWB ranging values ​​(low-frequency, discrete), IMU data (high-frequency, continuous) is the only source of information capable of real-time extrapolation of the UAV's trajectory (position, velocity, and attitude changes), providing input for the time update step of the Kalman filter. The Z-axis drift generated by the UAV in a confined space originates from the accumulated error generated when integrating accelerometer and gyroscope measurements. The algorithm needs to acquire this raw data to accurately model the data and ultimately correct it using UWB observations. Acquiring both the inertial navigation data of the unit to be calculated and the reference unit is necessary because the reference unit (static anchor point) is not an absolutely stationary physical base station, but rather a stationary UAV. To calculate the true relative pose (distance and angle) of the unit to be solved relative to the reference unit, the system must know the reference unit's own attitude (such as yaw angle) and motion trend. Without the reference unit's inertial data, the system cannot accurately establish a relative kinematic model between the two. In the Kalman filter-based fusion algorithm, the passage of time is entirely driven by the inertial measurement unit (IMU). The unit to be solved uses its IMU data to perform continuous integration to calculate its real-time position, providing a predicted state for fusing UWB ranging values.

[0056] The reference unit needs to provide its state calculated based on its own IMU. When the reference unit receives UWB ranging values, the algorithm needs to predict a priori state based on the reference unit's IMU data before it can calculate the error between the UWB measurement and the predicted value, and then complete the state correction and update. During role switching, the original dynamic inspection drone will become a new static anchor point after docking, while the original anchor point will take off to become the new inspection drone. To achieve this seamless switching, the system must always have a complete grasp of the motion state of both drones. If only the inertial data of the unit to be solved is acquired at any time, when the roles are switched, the system will immediately lose attitude awareness of the new anchor point, causing the entire cooperative localization system to collapse.

[0057] In one possible implementation, the method further includes: during the flight of the second UAV in the dynamic inspection mode, obtaining the effective air traffic cross-section width in front of the second UAV; When the effective airspace cross-section width is greater than or equal to the width safety threshold, the first UAV and the second UAV are controlled to maintain a longitudinally connected flight formation; When the effective air traffic cross-section width is less than the width safety threshold, the first UAV and the second UAV are controlled to produce a relative positional offset within the cross-section, thus converting into a stacked flight formation.

[0058] In one possible implementation, the first UAV and the second UAV generate a relative positional offset within a cross-section, converting into a stacked flight formation, including: Based on the sensor field-of-view models of the first UAV and the second UAV, the relative position offset matrix of the two UAVs is solved so that the perception fields of the first UAV and the second UAV can achieve cross coverage in the transitioned overlapping flight formation, eliminating the blind spot on one side. Based on the relative position offset matrix of the two drones and the dynamic constraints of the drones, a continuous and smooth formation change trajectory is generated and executed, so that the first drone and the second drone can complete the change of formation without interrupting the mission.

[0059] In practical applications, to address the "single-sided wall-hugging avoidance scanning blind spot" problem caused by obstacles, an adaptive formation topology controller can be used to operate in real time. The forward-looking perception module of the second UAV-B, in dynamic inspection mode, detects the effective airspace cross-section width directly in front. When the width exceeds a safety threshold, the controller maintains the system in a normal tandem formation to obtain the optimal longitudinal baseline for ranging. When an obstacle is detected that reduces the cross-section width, the controller sends a spatial offset matrix command to the underlying flight control of both UAVs, and the two UAVs smoothly transition to stacked flight formation.

[0060] Specifically, firstly, forward point cloud analysis and effective airspace cross-sectional area extraction are required. The lidar or depth camera of the dynamic forward aircraft (UAV-B) acquires the point cloud of the environment in front in real time. The algorithm extracts point cloud slices within a certain safe look-ahead distance directly in front of the aircraft, projects them onto a 2D cross-section perpendicular to the flight direction, and uses algorithms such as the maximum inscribed rectangle algorithm to remove hanging objects and side wall protrusions in real time, and calculates the maximum safe horizontal airspace width of the current cross-section. and maximum safe vertical navigation altitude Iso-geometric characteristic parameters.

[0061] Then, adaptive topology determination is performed based on multidimensional boundary thresholds. The controller presets the physical safety boundary parameters of the formation configuration, such as the lateral safety threshold for two machines side by side. With the longitudinal safety threshold of stacked flight The extracted spatial geometric parameters can be input into the topological state machine for real-time comparison. (1) If ≥ The state machine maintains its initial state, and the two machines remain in their original, cascaded positions to obtain the optimal collaborative filtering ranging baseline and optimize the accuracy of collaborative navigation.

[0062] (2) If < and ≥ The system is determined to be in a lateral space-constrained state, triggering a state transition command in the state machine, and the target topology is established as a stacked flight state.

[0063] Thus, in the initial state (i.e., when the passage is wide), the target is in tandem formation to obtain the optimal cooperative filtering ranging baseline. When the passage narrows, the target is switched to a stacked formation, which reduces the spatial cross-section required for formation and ensures safe passage.

[0064] Subsequently, the spatial relative offset matrix is ​​dynamically solved. After the state machine undergoes a formation change, the cooperative controller stops issuing absolute path points and instead solves for the two-machine relative position offset matrix that satisfies the target configuration. The core function of this matrix is ​​to ensure that the perception fields of the two sensors overlap, minimizing the overall perception blind spot of the system in any formation. Based on the principle of maximizing the cross-complementary nature of the sensor field of view (FOV), the system dynamically outputs the optimal spatial offset vector that ensures the perception blind spot of a single sensor is covered by the other sensor.

[0065] Finally, trajectory smoothing control based on dynamic constraints can be performed. The underlying flight control module receives the new bias matrix. Instead of abrupt switching, a polynomial trajectory optimization algorithm is introduced. Under the premise of satisfying the dynamic constraints such as the UAV's maximum linear velocity and angular acceleration, a continuous and smooth transition control law is solved and generated. This allows the two UAVs to complete the formation change without interrupting the mission, eliminating blind spots in inspection coverage caused by physical avoidance. In this way, the formation change process is smooth and controllable, avoiding instability or mission interruption caused by violent maneuvers, ultimately achieving continuous inspection without perceptible blind spots.

[0066] To address the sudden reduction in available cross-section due to ventilation equipment within narrow spaces, a flexible formation topology reconfiguration control strategy with cross-section adaptation was introduced. When the system detects that the effective airspace cross-section is below a safe threshold, it can smoothly switch the tandem formation to a stacked formation without interrupting the mission. Through these steps, the system effectively overcomes the throughput limitations caused by abrupt changes in the cross-section of complex utility tunnels and significantly improves the panoramic coverage of point cloud scanning.

[0067] This application also provides a dual-UAV alternating cooperative navigation system for narrow spaces, consisting of two mass-produced standard general-purpose UAVs, defined as the lead UAV-A and the inspection UAV-B. The hardware layer directly utilizes the UAVs' built-in visual cameras and high-frequency IMUs, acquiring front-end visual history data through the official payload development kit (PSDK). Signal-free positioning eliminates the need for expensive and bulky 3D LiDAR, using only a lightweight computing box externally mounted on the top interface of the UAV, embedding a UWB RF transceiver antenna and core algorithm. Addressing the harsh conditions of tunnels where low light and dust cause optical feature failure, the UWB RF module, as a ranging carrier unaffected by environmental optics, can penetrate some dust and output the absolute physical distance between the two UAVs at a fixed frequency (e.g., 50Hz).

[0068] Figure 2 A flowchart of a dual-machine positioning process is shown, specifically, Figure 2 This paper demonstrates the working principle of a dual-UAV collaborative inspection system applied to tunnel environments. The system employs a "one static, one dynamic" role-alternating strategy: the lead UAV A on the left acts as a "static anchor point," indicated by a padlock icon, maintaining a "state locked / absolute reference" mode and providing a stable spatial reference via a UWB radio frequency ranging link. The rear UAV B on the right acts as a "dynamic inspection" unit, equipped with a forward-looking perception module responsible for detecting obstacles ahead and performing inspection tasks along the tunnel's depth. In terms of workflow, based on the high-precision reference established by lead UAV A, rear UAV B uses an asymmetric Kalman filter algorithm to convert the relative distance measured by UWB into an absolute correction for its accumulated error, thus achieving accurate pose calculation. When the flight distance reaches approximately 50 meters or the alternation threshold is triggered, the system issues a "B docks, A takes off" command, causing the two UAVs to switch roles. This cyclical process ensures the continuity and high accuracy of global positioning during tunnel inspections.

[0069] Figure 3 This diagram illustrates the principle of a dual-machine positioning full-stack collaborative algorithm. Specifically, Figure 3 This paper details a closed-loop algorithm for collaborative navigation and positioning of two unmanned aerial vehicles (UAVs, A being a static anchor point and B a dynamic inspection point). Its core logic lies in addressing the problem of eliminating accumulated errors in complex environments such as tunnels through an "asymmetric Kalman filter" and a "role alternation" mechanism. The specific details are as follows: 1. System initialization and role assignment.

[0070] When the system starts, it first acquires the absolute coordinates in the "system starting point baseline initialization" phase and sets the initial joint error covariance matrix of the two UAVs to approach zero, establishing a high-precision initial state. Then, the state machine is activated to assign initial roles (Role-Flag) to the two UAVs, clarifying which one is the static anchor point (Role=0) and which one is the dynamic inspection machine (Role=1).

[0071] 2. Dual-machine parallel operation logic.

[0072] The system is divided into two parallel processing modules: Static anchor point mode (Role=0): Executes the unloading command (hovering or landing), serving as the system's absolute reference point. The most crucial step is to forcibly freeze the covariance decay mechanism, restricting the covariance update equation. This means that the error of anchor point A is artificially "locked," no longer propagating over time, ensuring its stability as an absolute reference frame. Simultaneously, a high-precision visual reference coordinate system ensures the accuracy of the anchor point's position.

[0073] Dynamic Inspection Mode (Role=1): Inspection Unit B primarily relies on Visual Inertial Odometry (VIO) for positioning, including two sub-modules: computational power support and attitude fine-tuning. The system continuously determines whether the "cumulative distance" or "high-precision visual constraint" has reached a preset threshold. If the threshold is not reached, a regular inspection task is executed; if the threshold is reached, an orientation alternation mechanism is triggered, notifying the other party to switch roles and prepare for a role swap.

[0074] Anomaly Handling: If the location is unreliable due to objective factors such as signal obstruction, the system will trigger the mutual inhibition trigger logic and execute the relevant covariance unfreezing to prevent the accumulation and propagation of erroneous data.

[0075] 3. Core Algorithm: Asymmetric Kalman Gain Fusion.

[0076] After obtaining the physical observations from the general UWB hardware ranging system, the system enters the "asymmetric Kalman gain fusion calculation" stage. Because the reference frame covariance approaches zero (the error at anchor point A is extremely small), the calculated gain exhibits an asymmetric tilt. This asymmetric tilt causes the absolute distance constraint of the UWB to be unidirectionally fed to the dynamic inspection unit B, thereby forcibly pulling back and zeroing the drift of B's ​​Z-axis coordinate, and completing the shrinkage and reset of the inspection unit's covariance. This solves the problem of slow convergence caused by the mutual pulling between the two units in traditional filtering.

[0077] 4. Status monitoring and role switching.

[0078] The system continuously monitors whether the relative distance reaches the safety baseline or whether the Z-axis error exceeds the limit. Once an anomaly is detected or a threshold is reached, the system issues a role-switching command, the original dynamic machine stops, and it transforms into a new static anchor point. When switching between old and new roles, the new static machine inherits the high-precision coordinates and minimal covariance after calibration in the last frame, ensuring the continuity and seamless connection of positioning.

[0079] The following example illustrates the specific implementation steps of the above-mentioned dual-UAV alternating cooperative navigation method for narrow spaces.

[0080] (1) Physical system construction based on general-purpose UAVs: In actual engineering deployment, this embodiment selects two standard mass-produced micro UAVs as the execution platform. No modification is required to the underlying mechanics of the UAVs. Only an expansion module containing a UWB radio frequency transceiver antenna and an edge computing board (such as a Raspberry Pi CM4 or Nvidia Jetson Nano) is connected through the PSDK (Payload SDK) interface on the top of the fuselage. The edge computing board reads the binocular visual odometry (VIO) data and depth map output by the UAV flight controller through the local area network interface, and runs the voxel map sharing, asymmetric filtering and flexible formation algorithm proposed in this application in a closed environment. The final motion control command is transmitted back to the general flight controller for execution through the PSDK.

[0081] (2) Computer algorithm simulation and effect evaluation under real tunnel conditions: In order to verify the effectiveness of the above core collaborative algorithm under real harsh conditions, this embodiment constructs a numerical simulation experiment of the algorithm for a 30-meter long corridor degraded tunnel section.

[0082] Simulating real tunnel conditions, the simulation incorporated the actual noise characteristics of the sensors, including the IMU's low-frequency random bias walk, the slip-like abrupt changes in LiDAR / visual point cloud matching features due to the singular tunnel characteristics, and multipath impulse noise interference from UWB signals in enclosed spaces. The alternating distance threshold between the two sensors was set to 15 meters, and the safety tolerance delay threshold was set to 150 ms. The simulation results and evaluation metrics are detailed in the following figures. Figure 4 A schematic diagram showing the comparison of physical trajectory deviations in a 30-meter tunnel is provided. Figure 5 A schematic diagram showing the absolute trajectory error along the Z-axis and the filtering confidence level in the degraded environment of the corridor is presented. Figure 6 This diagram illustrates a comparison of obstacle avoidance command response latency when the drone's computing power is fully utilized. Figure 7a A schematic diagram of the formation topology in a sudden narrow area is shown. Figure 7b A schematic diagram of panoramic coverage assessment is shown.

[0083] The simulation results are analyzed as follows, such as Figure 4As shown, the solid black line (parallel line) represents the absolute true trajectory. In a tunnel lacking environmental characteristics, existing single-machine technology (dashed line) is affected by odometer divergence. The system believes it is flying along the centerline, but the actual physical trajectory has deviated significantly. At 16 meters, it completely penetrates the 0.6-meter solid tunnel wall boundary, resulting in a collision and crash. However, the dual-machine system equipped with the algorithm of this application (solid line moving up and down the aforementioned parallel line), after a long flight, maintains its physical trajectory firmly constrained near the safe centerline with zero deviation, avoiding the risks of getting lost and crashing into the wall.

[0084] like Figure 5 As shown in the figure, the "X" marks represent points where environmental features injected during simulation are lost (i.e., sensor-blinding abrupt change points). Existing single-machine SLAM errors (marked by the X mark) exhibit drastic nonlinear jumps and divergences at these abrupt change points, with a maximum drift of nearly 1.2 meters. In contrast, this system (the broken line below) successfully triggers state transitions at 15 meters, forcibly locking the covariance of the preceding machine's anchor point. Despite the superposition of UWB multipath noise, the 3σ confidence region (shaded band) of the dual-machine state covariance consistently converges within a safe range through the unidirectional pull of the bias Kalman gain, resulting in a global root mean square error (RMSE) of only 0.1 meters, significantly improving positioning accuracy compared to existing single-machine technologies.

[0085] like Figure 6 As shown, the striped area (upper region) represents a high-risk zone for high-speed collisions exceeding 150ms. Existing single-machine technology (upper line) experiences drastic fluctuations in CPU load with environmental complexity when simultaneously performing global mapping and local planning. Latency spikes to around 280ms when encountering sudden features, making it highly susceptible to collisions due to insufficient computing power. This application significantly reduces the global mapping load on the back-end machines by sharing local voxel maps across machines. Figure 6 The system's obstacle avoidance planning delay (lower line) remains extremely stable at around 35ms, and the dynamic obstacle avoidance response speed has been improved by more than 4 times, which is within a safe response time range.

[0086] like Figure 7a As shown, the simulated tunnel experiences a sudden reduction in cross-sectional area at 10-14 meters (construction scaffolding) and 22-25 meters (large ventilation fan). The system's underlying controller responds instantly, changing the control status flag and automatically reconfiguring the dual-machine system from a series connection to a parallel / overlapping connection. For example... Figure 7bAs shown, the dashed line indicates the minimum coverage requirement (90%) for the project. Existing fixed formations or single drones (i.e., some located below 90%) are forced to hug the wall on one side in narrow areas, resulting in a huge blind spot on the other side. The point cloud panoramic coverage instantly drops to the unacceptable range of 60%. In contrast, this system (with coverage consistently above 90%) relies on formation reconstruction and dual-field cross-complementarity to maintain a scanning coverage of over 95% at the narrowest gaps, which is restored after leaving the obstacle area, thus increasing the coverage range of the drone inspection.

[0087] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a dual-UAV alternating cooperative navigation device for narrow spaces, the structure of which is as follows: Figure 8 As shown.

[0088] Figure 8 This is a schematic diagram of the internal structure of a dual-UAV alternating cooperative navigation device for narrow spaces, provided as an embodiment of this application. Figure 8 As shown, the device includes: At least one processor 801; And a memory 802 that is communicatively connected to at least one processor; The memory 802 stores instructions that can be executed by at least one processor. The instructions are executed by at least one processor 801 so that at least one processor 801 can: execute the above-described method for alternating cooperative navigation of two unmanned aerial vehicles in a narrow space.

[0089] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium stores computer-executable instructions, which are configured to execute the aforementioned method for alternating cooperative navigation of two unmanned aerial vehicles (UAVs) in a narrow space.

[0090] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0091] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0097] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0098] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0099] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0100] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A dual-UAV alternating cooperative navigation method for narrow spaces, characterized in that, The method includes: The first UAV is set as the reference unit, and the second UAV is set as the unit to be solved. The error covariance matrix of the pose estimation of the first UAV is set to a first fixed value that approaches zero. Obtain the ultra-wideband (UWB) ranging value between the reference unit and the unit to be solved; Using the pose and error covariance of the reference unit as a spatial reference, state estimation filtering is performed on the unit to be solved to determine the precise pose information of the unit to be solved. Here, the error covariance of the reference unit is frozen to a first fixed value close to zero, causing the Kalman gain matrix to be biased toward the state component of the unit to be solved. Through a state constraint filtering algorithm based on asymmetric covariance freezing, the UWB ranging value is converted into an absolute correction amount for the cumulative pose error of the unit to be solved in the longitudinal direction of the narrow space. Under the condition of satisfying the preset alternation, the unit to be solved is transformed and set as the new reference unit, and the original reference unit is transformed into the new unit to be solved. The accurate pose information of the new reference unit is used as the new spatial reference, and the error covariance matrix of the new reference unit is set to a second fixed value that approaches zero.

2. The method according to claim 1, characterized in that, The step of setting the first UAV as the reference unit and the second UAV as the unit to be solved includes: Initialize the known coordinates of the first and second UAVs at the entrance of the narrow space; Wherein, the known coordinate point is the origin of the global relative coordinate system, the first UAV is set to static anchor point mode and hovers in the narrow space, and the second UAV is set to dynamic inspection mode and begins to fly in the narrow space.

3. The method according to claim 2, characterized in that, The method further includes: the second UAV in the dynamic inspection mode collects point cloud features of the environment in front and converts the point cloud features into lightweight local obstacle representation information, and sends the local obstacle representation information to the first UAV in the static anchor point mode through a self-organizing network; The first UAV uses a path planning algorithm to generate an obstacle avoidance trajectory based on the local obstacle representation information and transmits it back to the second UAV for execution.

4. The method according to claim 2, characterized in that, The state constraint filtering algorithm based on asymmetric covariance freezing includes: Construct a joint state vector and a joint error covariance matrix that includes the three-dimensional position, velocity, and attitude of the first UAV and the second UAV; By introducing a system logic state parameter, when it is determined that the first UAV is in the static anchor point mode, the error covariance of the first UAV is forcibly frozen to the first fixed value close to zero, so that the error covariance of the first UAV does not accumulate over time. An observation model is established based on the UWB ranging values, and the Kalman gain matrix is ​​solved so that the gain weights of the Kalman gain matrix are unidirectionally biased toward the UWB ranging values.

5. The method according to claim 4, characterized in that, The method further includes: Perform a post-hoc state update. During the update period, keep the three-dimensional coordinates of the first UAV unchanged, perform forced convergence calibration on the Z-axis coordinates of the second UAV that have drifted, and reduce the error covariance of the second UAV.

6. The method according to claim 2, characterized in that, The alternation conditions include at least one of the following: The UWB ranging value between the UAV in the dynamic inspection mode and the UAV in the static anchor point mode reaches a preset distance threshold. The pose estimation error covariance of the UAV in the dynamic inspection mode exceeds the preset covariance threshold.

7. The method according to claim 2, characterized in that, The method further includes: During the flight of the second UAV in the dynamic inspection mode, the effective air traffic cross-section width in front of the second UAV is obtained; When the effective airspace cross-section width is greater than or equal to the width safety threshold, the first UAV and the second UAV are controlled to maintain a longitudinally connected flight formation; When the effective air traffic cross-section width is less than the width safety threshold, the first UAV and the second UAV are controlled to produce a relative positional offset within the cross-section, thus converting into a stacked flight formation.

8. The method according to claim 7, characterized in that, The control of the first UAV and the second UAV to generate a relative positional offset within the cross-section, converting them into a stacked flight formation, includes: Based on the sensor field-of-view models of the first UAV and the second UAV, the relative position offset matrix of the two UAVs is solved so that the perception fields of the first UAV and the second UAV can achieve cross coverage in the transitioned overlapping flight formation, eliminating the blind spot on one side. Based on the relative position offset matrix of the two drones and the dynamic constraints of the drones, a continuous and smooth formation change trajectory is generated and executed, so that the first drone and the second drone can complete the change of formation without interrupting the mission.

9. A dual-UAV alternating cooperative navigation device for narrow spaces, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a dual unmanned aerial vehicle (UAV) alternating cooperative navigation method for narrow spaces as described in any one of claims 1-8.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement a dual-UAV alternating cooperative navigation method for narrow spaces as described in any one of claims 1-8.

Images