Unmanned aerial vehicle active avoidance method

Abstract

The invention discloses an active avoidance method for an unmanned aerial vehicle, and relates to the technical field of unmanned aerial vehicles. Comprising the following steps: SA: weather adaptive sensing: obtaining physical characteristics of an external environment through an anti-scattering laser radar group, a super-resolution millimeter wave radar group and an anti-turbulence sound wave positioning group; sB: multi-modal decision fusion: aligning the laser radar data and the millimeter wave radar data, obtaining the confidence interval of each sensor through Gaussian process regression calculation, and determining the weight of the sensor; and SC: hierarchical fault-tolerant control: adjusting a sensor combination according to the visibility of the external environment, and determining a sudden obstacle avoidance track according to the arranged UWB positioning module and the inflatable buffer structure. The sensor combination is automatically switched according to the visibility, and the inflatable buffer structure is started after the emergency mode is triggered, so that the safety of the hovering or landing track of the unmanned aerial vehicle can be ensured, and the obstacle avoidance success rate is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs), and in particular to an active avoidance method for UAVs. Background Art

[0002] With the rapid development of artificial intelligence, automatic control, and sensor technologies, unmanned aerial vehicles (UAVs) are increasingly being used in a variety of fields, including military reconnaissance, logistics and distribution, agricultural plant protection, aerial photography, and emergency rescue. Especially for autonomous flight missions in complex environments, the safety and reliability of UAVs have become key issues in system design.

[0003] During actual flight, drones often face potential threats such as buildings, utility poles, trees, birds, and other aerial obstacles. To ensure flight safety and enhance mission execution, drones need to be able to perceive their surroundings and avoid obstacles in real time. Traditional drones often use passive obstacle avoidance methods, such as relying on operators to manually maneuver around obstacles or using simple infrared / ultrasonic sensors for close-range warnings. However, these methods suffer from slow response, low detection accuracy, and poor adaptability in complex and dynamic environments, making them difficult to meet the demands of high-density, high-speed flight scenarios.

[0004] The Chinese invention patent with publication number CN117742382A discloses a method for avoiding obstacles in a UAV, comprising the following steps: Step 1: Using at least one sensor selected from one or more of a camera, a lidar, a millimeter-wave radar, and an infrared sensor to sense the surrounding environment of the UAV and collect obstacle information; Step 2: By fusing the data collected by the sensor, the collected obstacle information is subjected to spatiotemporal data fusion processing to improve the accuracy and robustness of obstacle detection; Step 3: Based on the fused data, an obstacle avoidance algorithm is used to perform path planning, and the obstacle avoidance algorithm is selected from the dynamic window method; Step 4: Sending the flight instructions calculated by the path planning algorithm to the flight control system of the UAV, the accuracy of obstacle detection and the robustness of the system are improved by adopting spatiotemporal data fusion and adaptive filtering technology, and the dynamic adaptability of data processing is enhanced.

[0005] However, in adverse weather conditions such as rain, snow, fog, haze, and dust, optical / laser sensors experience a sharp drop in signal-to-noise ratio due to Rayleigh scattering, and the density of lidar point clouds decreases, leading to a significant drop in obstacle detection rates and rendering drone perception ineffective. Furthermore, millimeter-wave radars, limited to an angular resolution of 1°-5°, struggle to accurately locate obstacle edges, and acoustic sensors are susceptible to wind noise, resulting in ranging errors. This significantly reduces the effective sampling rate of the point cloud. Consequently, sufficient environmental perception data cannot be obtained to provide stable support for obstacle avoidance decisions, making it difficult to dynamically adjust flight strategies based on real-time environmental changes. This in turn limits the drone's flight safety and mission reliability. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for active avoidance of a UAV to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solutions: a method for active avoidance of a UAV, comprising:

[0008] SA: Weather Adaptive Sensing: Acquires the physical characteristics of the external environment through an anti-scattering lidar group, a super-resolution millimeter-wave radar group, and an anti-turbulence sonar positioning group;

[0009] SB: Multimodal decision fusion: LiDAR and millimeter-wave radar data are aligned and calculated using Gaussian process regression to obtain the confidence interval of each sensor and determine the weight of the sensor.

[0010] SC: Hierarchical Fault Tolerant Control: Adjusts the sensor combination based on the visibility of the external environment, and determines the sudden obstacle avoidance trajectory based on the set UWB positioning module and inflatable buffer structure, including:

[0011] SC1: Set the flight strategy table: According to the preset visibility threshold range, set the sensor combination, specifically:

[0012] When the real-time visibility is less than the lower limit of the preset visibility threshold range, the sensor combination is set to a bionic acoustic sensor and an IMU inertial measurement unit; when the real-time visibility is within the preset visibility threshold range, the sensor combination is set to a super-resolution millimeter-wave radar group and a bionic acoustic sensor; when the real-time visibility is greater than the upper limit of the preset visibility threshold range, the sensor combination is set to an anti-scattering laser radar group and a super-resolution millimeter-wave radar group;

[0013] SC2: Emergency handling: According to the emergency alarm signal, the UAV is switched to hovering mode, and the 3D position of the UAV is determined by the UWB positioning module, and the inflatable buffer structure is activated at the same time;

[0014] SC3: Obstacle Avoidance Path Planning: Based on the current position of the UAV, the starting direction control point, the ending direction control point, and the end point of the path, a trajectory curve is determined, and the cost function of the trajectory curve is obtained. At the same time, all cost functions are compared to determine the minimum cost function. The trajectory curve corresponding to the minimum cost function is the optimal obstacle avoidance path trajectory.

[0015] Furthermore, the physical characteristics of the external environment are obtained, including:

[0016] SA1: Set up an anti-scattering lidar group: Use a dual-wavelength laser transmitter to compensate for rain and snow scattering attenuation and remove raindrop pseudo-targets. It also identifies and filters the characteristic frequency bands of rain and snow, and adjusts the point cloud volatility to a preset range.

[0017] SA2: Set up a super-resolution millimeter-wave radar group: reconstruct sparse echo signals based on the sparse array, sharpen the edges of the millimeter-wave image, and align the lidar data in time and space based on the cross-modal model;

[0018] SA3: Set up the anti-turbulence acoustic wave positioning group: Use the bionic acoustic sensor to obtain sound waves within the preset range, and correct the sound wave propagation time error based on the wind speed vector. Specifically:

[0019]

[0020] in: is the corrected equivalent sound speed, is the speed of sound in a static environment, is the wind speed vector, is the unit vector of the sound wave propagation direction.

[0021] Furthermore, the point cloud volatility is adjusted to a preset range, including:

[0022] SA1.1: Dynamic PRF Adjustment: Adjusting the pulse repetition frequency of the dual-wavelength laser transmitter according to the atmospheric optical thickness to determine the adjusted pulse repetition frequency;

[0023] SA1.2: Scattering compensation: Determine the atmospheric attenuation coefficient based on the laser wavelength and ambient visibility, and compensate the original signal intensity received by the lidar to obtain the compensated laser echo intensity;

[0024] SA1.3: Reflection Filtering: Determine the degree of polarization based on the vertical polarization component strength and the parallel polarization component strength. Compare the degree of polarization with a preset polarization threshold range and filter the echo signals based on the comparison results. Specifically:

[0025] When the polarization degree is within the preset polarization threshold range, the light spot echo signal corresponding to the polarization degree is retained; otherwise, the laser spot echo signal corresponding to the polarization degree is deleted;

[0026] SA1.4: Spectral Analysis: Set the stopband range, passband range, and filter order of the band-stop filter based on the echo characteristics and the motion state of the obstacle, and filter the determined echo characteristics through the band-stop filter;

[0027] SA1.5: Point cloud denoising: Determine the point cloud structure based on the point cloud position and a preset radius. Obtain the overall offset vector of the point cloud based on its 3D coordinates within the point cloud structure. Compare the overall offset vector with a preset offset threshold. Select the point cloud data based on the comparison results. Specifically:

[0028] When the overall offset vector is greater than a preset offset threshold, the point cloud data corresponding to the overall offset vector is deleted; otherwise, the point cloud data corresponding to the overall offset vector is retained.

[0029] Furthermore, according to the preset time window size, the laser echo signal is segmented to obtain a short-term analysis unit, and the power spectrum density of the short-term analysis unit is determined by Fourier transform. At the same time, the power spectrum density is compared with the preset power spectrum threshold range, and the echo characteristics are determined according to the comparison result, specifically:

[0030] When the power spectrum density is greater than the upper limit threshold of the preset power spectrum threshold range, the echo feature corresponding to the power spectrum density is a raindrop signal feature; when the power spectrum density is within the preset power spectrum threshold range, the echo feature corresponding to the power spectrum density is a snowflake signal feature; when the power spectrum density is less than the lower limit threshold of the preset power spectrum threshold range, the corresponding echo feature is an obstacle signal feature.

[0031] Furthermore, each noise point data is corrected according to the noise point data and the overall point cloud data, as follows:

[0032] SA1.5.1: Determine the neighborhood geometric center: Determine the neighborhood geometric center based on the 3D coordinates of all point clouds within the neighborhood.

[0033] SA1.5.2: Noise point movement: Move the noise point data toward the geometric center of the neighborhood to obtain the three-dimensional coordinates of the moved point cloud, specifically:

[0034]

[0035] in: Point Cloud The three-dimensional coordinates after movement, Point Cloud The corresponding neighborhood geometric center, is the three-dimensional coordinate of the i-th point cloud, is the relaxation factor;

[0036] SA1.5.3: Noise point movement: Re-determine the noise point data based on the moved point cloud 3D coordinates, and determine the point cloud volatility based on the number of noise points and the total number of point clouds. Compare the point cloud volatility with a preset volatility threshold, and modify the noise point data based on the comparison result, specifically:

[0037] When the point cloud fluctuation rate is greater than the preset fluctuation threshold, repeat steps SA1.5.1-SA1.5.2 until the point cloud fluctuation rate is no greater than the preset fluctuation threshold; otherwise, the three-dimensional coordinates of the moved point cloud are the three-dimensional coordinates after correction of the noise point data.

[0038] Furthermore, the 30-50kHz Helmholtz resonant cavities are arranged in order from small to large, and an eddy current suppression grid is set at the cavity entrance, and the cavity is filled with porous sound-absorbing material to obtain a bionic acoustic sensor.

[0039] Furthermore, determining the weight of the sensor includes:

[0040] SB1: Spatiotemporal alignment processing: Use the lidar point as the query and the millimeter-wave radar point as the key and value to construct a query-key-value pair, determine the similarity weight between the lidar point and the lidar point, and determine the fused lidar point features based on the similarity weight. Specifically:

[0041]

[0042] in: is the fused lidar point feature, is the attention weight of the lidar point r and the millimeter wave radar point o, is the value vector of millimeter-wave radar point o, is the trainable parameter matrix, is the feature vector of the oth point in the millimeter-wave radar point cloud, is the point index in the millimeter wave radar point cloud;

[0043] SB2: Uncertainty Fusion: Based on the aligned LiDAR point cloud and the MMW radar point cloud, the confidence interval of each data point is determined, and the weight of the sensor is set according to the signal-to-noise ratio of the sensor. Specifically:

[0044]

[0045] in: is the dynamic weight of the sth sensor, is the 1.5th power of the signal-to-noise ratio of the sth sensor, is the base of natural logarithm, is the historical confidence correction factor of the sth sensor, is the slope coefficient of the logistic function, The index of the sensor.

[0046] Furthermore, the inflatable cushioning structure includes but is not limited to an inflatable airbag made of polyurethane-coated nylon.

[0047] Furthermore, the drone's 3D position, remaining battery power, and external environment data are used as inputs to the digital twin model, and the landing trajectory and trajectory score are obtained as outputs. At the same time, all the trajectory scores are compared to determine the maximum trajectory score. The landing trajectory corresponding to the maximum trajectory score is the final landing trajectory.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] First, the present invention automatically switches sensor combinations based on visibility and activates an inflatable cushioning structure when the emergency mode is triggered, thereby ensuring the safety of the drone's hovering or landing trajectory and improving the success rate of obstacle avoidance. At the same time, the determined obstacle avoidance path is optimized through a cost function, which not only ensures the minimum obstacle distance but also reduces energy loss.

[0050] Second, the present invention uses a dual-wavelength lidar to dynamically adjust the pulse repetition frequency and effectively compensates for signal attenuation caused by rain and fog through a scattering compensation algorithm, thereby increasing the point cloud density in low-visibility atmospheric optical thickness. At the same time, a polarization beam splitter and spectrum analysis technology are used to filter out raindrop pseudo-targets, thereby improving the accuracy of obstacle detection.

[0051] Third: The present invention aligns the extracted geometric features of the lidar point cloud with the millimeter-wave radar sparse array super-resolution reconstruction data in time and space, thereby improving the angular resolution. At the same time, through dynamic weighted fusion of the signal-to-noise ratio, the obstacle positioning error can be reduced in rainy and foggy environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the process of the active avoidance method of the UAV in the present invention;

[0053] Figure 2 This is a comparison chart of the penetration of dual-wavelength lasers in rain and fog in the present invention;

[0054] Figure 3 This is a comparison diagram of point cloud density before and after scattering compensation in the present invention;

[0055] Figure 4 This is a simulation diagram of the obstacle avoidance path in the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] Under adverse weather conditions such as rain, snow, fog, haze, and dust, the signal-to-noise ratio of optical / laser sensors will drop sharply due to the Rayleigh scattering effect, and the density of the lidar point cloud will increase, resulting in a significant drop in the obstacle detection rate and the failure of drone perception. At the same time, since the millimeter-wave radar is limited to an angular resolution of 1°-5°, it is difficult to accurately locate the edge of the obstacle, and the acoustic sensor is easily interfered by wind noise and produces ranging errors, which leads to a significant drop in the effective sampling rate of the point cloud. Therefore, when making obstacle avoidance decisions, it is impossible to obtain sufficient environmental perception data for stable support, and it is difficult to dynamically adjust the flight strategy according to real-time environmental changes, thereby limiting the flight safety and mission reliability of the drone. However, this application uses an anti-scattering lidar group, a super-resolution millimeter-wave radar group, and an anti-turbulence acoustic wave positioning group to perform weather adaptive perception, dynamically compensate for sensor attenuation in adverse weather, and simultaneously perform spatiotemporal alignment and weighted fusion of lidar and millimeter-wave radar data, and switch sensor combinations according to environmental visibility. UWB positioning and inflatable buffer structures are used to achieve emergency obstacle avoidance, thereby improving the reliability and safety of drone obstacle avoidance in complex environments such as rain, snow, fog, and haze.

[0058] Example 1

[0059] refer to Figures 1-4 This embodiment provides a method for active avoidance of a drone, which includes the following steps:

[0060] Step SA: Adaptive weather perception. This involves acquiring the optical, electrical, and acoustic physical properties of the external environment through the anti-scattering lidar, super-resolution millimeter-wave radar, and anti-turbulence sonar positioning systems. The details are as follows:

[0061] Step SA1: Set up an anti-scattering lidar group. This involves using a dual-wavelength laser transmitter to compensate for rain and snow scattering attenuation and remove corresponding raindrop pseudo-targets. Simultaneously, the characteristic frequency bands of rain and snow are identified and filtered, and the point cloud volatility is adjusted to a preset range. The details are as follows:

[0062] Step SA1.1: Dynamic PRF adjustment. That is, according to the atmospheric optical thickness, the pulse repetition frequency of the dual-wavelength laser transmitter is adjusted to determine the adjusted pulse repetition frequency. In this embodiment, a dual-wavelength laser transmitter is provided by a 905nm short-pulse laser transmitter and a 1550nm chirped frequency-modulated laser transmitter. The 905nm short pulse is used for short-range high-precision detection, and the 1550nm chirped frequency-modulated laser transmitter is used to penetrate rain and fog. Furthermore, according to the atmospheric optical thickness, the pulse repetition frequency of the dual-wavelength laser transmitter is adjusted in real time, specifically:

[0063]

[0064] in: is the adjusted pulse repetition frequency, is the reference pulse repetition frequency, is the atmospheric optical thickness.

[0065] In the specific implementation, the baseline pulse repetition frequency is 100kHz, and the atmospheric optical depth in rainy and foggy weather is set to 2, so the corresponding adjusted pulse repetition frequency is 130kHz. In other words, the point cloud density is increased.

[0066] Step SA1.2: Scattering compensation. This involves determining the atmospheric attenuation coefficient based on the laser wavelength set in step SA1.1 and the visibility in the harsh environment. The original signal intensity received by the lidar is compensated based on the atmospheric attenuation coefficient to obtain the compensated laser echo intensity. In this embodiment, the atmospheric attenuation coefficient is determined based on the visibility in harsh environments such as rain, fog, and dust, and the laser wavelength of the dual-wavelength laser transmitter. Specifically,

[0067]

[0068] in: is the atmospheric attenuation coefficient, For visibility, is the laser wavelength, is the particle size distribution index.

[0069] In this embodiment, based on the determined atmospheric attenuation coefficient and the detection distance between the laser radar and the target object, the original signal intensity received by the laser radar is compensated by the Mie scattering compensation algorithm to obtain the compensated laser echo intensity, specifically:

[0070]

[0071] in: is the laser echo intensity after Mie scattering compensation, is the original signal strength received by the lidar, is the base of natural logarithm, is the atmospheric attenuation coefficient, is the detection distance between the laser radar and the target.

[0072] In the specific implementation process, the atmospheric attenuation coefficient in moderate rainy weather is set to 0.1m -1 , and the detection distance between the LiDAR and the target is 50 meters. This means that after the Mie scattering compensation algorithm is applied to the original signal intensity received by the LiDAR, only 0.67% of the original signal intensity is retained. This means that data points with low echo intensity can be eliminated.

[0073] Step SA1.3: Reflection filtering. This involves decomposing the echo information of each laser point to obtain the corresponding degree of polarization. This degree of polarization is then compared with a preset polarization threshold range (which can be set based on actual conditions and is not described in detail in this embodiment). The echo signal is then filtered based on the comparison results. Specifically:

[0074] When the obtained polarization degree is within the preset polarization threshold range, the corresponding laser point echo signal is retained. Conversely, when the obtained polarization degree is not within the preset polarization threshold range, the corresponding laser point echo signal is deleted.

[0075] In this embodiment, a polarization beam splitter prism is installed at the laser radar receiving end. That is, the polarization beam splitter prism decomposes the echo signal obtained by the laser radar receiving end to obtain a vertical polarization component and a parallel polarization component. At the same time, a photodetector is used to measure the intensity of the vertical polarization component and the parallel polarization component respectively. Based on the intensity of the vertical polarization component and the parallel polarization component, the corresponding polarization degree is determined, specifically:

[0076]

[0077] in: is the degree of polarization, is the intensity of the vertical polarization component, is the intensity of the parallel polarization component.

[0078] In the specific implementation process, the vertical polarization component intensity of 80W / m was obtained through the photoelectric detector. 2 , the parallel polarization component intensity is 20W / m 2 , the corresponding polarization degree is 0.6. In this embodiment, the preset polarization threshold range is set to [0.1, 0.9]. That is, if the obtained polarization degree is within the preset polarization threshold range, the corresponding laser point echo signal is retained.

[0079] Step SA1.4: Spectral Analysis. This involves segmenting the continuous laser echo signal according to a preset time window size (e.g., a 10ms Hanning window or rectangular window) to obtain corresponding short-term analysis units. Simultaneously, each short-term analysis unit is Fourier transformed to obtain the corresponding power spectral density.

[0080] Specifically, the obtained power spectrum density is compared with a preset power spectrum threshold range (which can be set according to actual conditions, so it is not specifically described in this embodiment), and the corresponding echo characteristics are determined based on the comparison results. Specifically:

[0081] When the obtained power spectrum density is greater than the upper threshold of the preset power spectrum threshold range, the corresponding echo feature is a raindrop signal feature. When the obtained power spectrum density is within the preset power spectrum threshold range, the corresponding echo feature is a snowflake signal feature. When the obtained power spectrum density is less than the lower threshold of the preset power spectrum threshold range, the corresponding echo feature is an obstacle signal feature.

[0082] Furthermore, a corresponding band-stop filter is set according to the determined echo characteristics. Specifically, when the determined echo characteristics are raindrop signal characteristics, the corresponding stopband range is set according to the terminal velocity of the raindrops, specifically:

[0083]

[0084] in: is the Doppler shift of the laser echo reflected by raindrops, is the falling speed of the raindrop, is the laser wavelength, is the frequency fluctuation range.

[0085] Furthermore, when the determined echo feature is a snowflake signal feature, the lower limit value of the stop band is set to the upper limit threshold of the preset power spectrum threshold range according to the preset power spectrum threshold range to avoid excessive interference with the obstacle signal.

[0086] Furthermore, the passband range of the band-stop filter is set according to the movement state of the obstacle. Specifically, when the obstacle is stationary, the passband frequency shift is set to 0. When the obstacle is in motion, the passband frequency shift is set according to the movement speed of the obstacle, specifically:

[0087]

[0088] in: is the Doppler shift, is the radial velocity of the target relative to the lidar, is the laser wavelength.

[0089] Furthermore, according to the stopband range and passband range set by the band-stop filter, the corresponding stopband minimum attenuation, attenuation coefficient, stopband frequency and passband ripple coefficient are obtained, and the order of the filter band is set, specifically:

[0090]

[0091] in: is the minimum order of the filter, is the minimum attenuation in the stop band, is the passband ripple coefficient, is the stop-band attenuation coefficient, is the normalized stopband frequency.

[0092] Step SA1.5: Point cloud denoising. This involves using a kd-tree or octree spatial index to identify other point cloud locations within a preset radius, centering the current point cloud. The current point cloud is then connected to the other point clouds. Each point cloud is treated as a node, and adjacent point clouds are connected via edges to obtain the corresponding point cloud graph structure.

[0093] In the obtained point cloud structure, the overall offset vector of each point cloud is obtained based on the 3D coordinates corresponding to each point cloud. The obtained overall offset vector is compared with a preset offset threshold (which can be set according to the specific actual situation and is not specifically explained in this embodiment), and the point cloud data is filtered based on the comparison result. Specifically:

[0094] When the obtained overall offset vector is greater than the preset offset threshold, the corresponding point cloud data is considered noise point data and is deleted. Conversely, when the obtained overall offset vector is not greater than the preset offset threshold, the corresponding point cloud data is retained.

[0095] In this embodiment, the overall offset vector of each point cloud is obtained according to the three-dimensional coordinates corresponding to each point cloud, specifically:

[0096]

[0097] in: Point Cloud The overall offset vector, is the three-dimensional coordinate of the j-th point cloud, is the three-dimensional coordinate of the i-th point cloud, Point Cloud The neighborhood set of 、 is the index of the point cloud.

[0098] Step SA2: Set up a super-resolution millimeter-wave radar array. This involves reconstructing sparse echo signals based on the configured sparse array and improving angular resolution. Millimeter-wave image edges are also sharpened, and the lidar data is spatially and temporally aligned using a cross-modal model.

[0099] In this example, a sparse uniform rectangular array in the 79 GHz frequency band uses random subsampling to collect 30% of the original data volume. The observation matrix is ​​set as a partial Fourier matrix. The sparse echo signals are reconstructed using an orthogonal matching pursuit algorithm, reducing the data volume while maintaining the required resolution. A motion compensation model is also established based on the drone's flight trajectory, and this model is used to synthesize an equivalent 35x physical aperture.

[0100] In this embodiment, a motion compensation model is established according to the flight trajectory of the drone, specifically:

[0101]

[0102] in: For drones in time The three-dimensional spatial position when is the initial position coordinate of the UAV, is the initial motion speed of the UAV, is the cumulative time from the initial moment, is the acceleration of the UAV.

[0103] Furthermore, the U-Net model is trained using LiDAR point cloud data collected in clear weather as real data. Specifically, the compressed sensing image from the millimeter-wave radar is used as input to the U-Net model, and the output is the sharpened obstacle edges. It is worth noting that in low visibility conditions such as rain, snow, and dust, the millimeter-wave radar obstacle boundaries are annotated using LiDAR data.

[0104] Step SA3: Setting up an anti-turbulence acoustic wave positioning group. That is, using the set bionic resonant cavity to obtain acoustic waves within a preset range, remove wind noise, and correct the acoustic wave propagation time error based on the determined wind speed vector.

[0105] In this embodiment, 30-50 kHz Helmholtz resonators are arranged sequentially according to the frequency gradient to mimic the basilar membrane structure of the bat cochlea. A vortex suppression grille is used at the cavity entrance to reduce airflow noise at wind speeds of 15 m / s. The cavity interior is filled with a porous sound-absorbing material (such as polyurethane foam), resulting in a biomimetic acoustic sensor.

[0106] Furthermore, the 4-6 channel acoustic wave signals acquired by the biomimetic acoustic sensor, including the target echo, wind noise, and body vibration, are filtered and whitened using a 6th-order Butterworth bandpass filter to remove low-frequency body vibration signals and eliminate inter-channel correlation. In this embodiment, the preprocessed acoustic wave signals are iteratively updated based on a preset initialization separation matrix to separate the wind noise signal from the target echo signal.

[0107] Furthermore, based on the wind speed vector measured by the IMU inertial measurement unit, the sound wave propagation time is corrected through the particle filter algorithm, specifically:

[0108]

[0109] in: is the corrected equivalent sound speed, is the speed of sound in a static environment, is the wind speed vector, is the unit vector of the sound wave propagation direction.

[0110] refer to Figure 2 , Figure 2 This is a comparison diagram of the penetration of dual-wavelength laser in rain and fog in this embodiment. Figure 2 It can be seen that the attenuation coefficient of 1550nm (0.06) is lower than that of 905nm (0.1). At the same distance, the 1550nm wavelength has a higher signal strength retention rate, and the 905nm wavelength has a steeper attenuation curve. At the same time, because water droplets and particles in rain and fog scatter and absorb laser light, the shorter 905nm wavelength is more affected by Mie scattering. In other words, in rain and fog, 1550nm laser has better penetration than 905nm, making it suitable for long-distance applications.

[0111] refer to Figure 3 , Figure 3 is a comparison diagram of point cloud density before and after scattering compensation in this embodiment, Figure 3 It can be seen that the original point cloud density (before compensation) decays exponentially with increasing detection distance (at a decay coefficient of -0.05). Specifically, at 10m, the number of point clouds is approximately 1000, while at 100m, the number of point clouds drops to approximately 67. After compensation, the point cloud density still decays with distance, but the decay coefficient is reduced to -0.02. At 100m, the number of point clouds after compensation remains at approximately 135, approximately double the pre-compensation number.

[0112] Step SB: Multimodal decision fusion. This involves aligning the LiDAR data and the millimeter-wave radar data, and using Gaussian process regression to determine the confidence interval for each sensor. The corresponding weights are then adjusted based on the confidence intervals. The details are as follows:

[0113] Step SB1: Spatiotemporal alignment. This involves determining the two timestamps in the millimeter-wave radar point cloud data that have the smallest time difference with the timestamps in the lidar point cloud data, based on the timestamps of the lidar point cloud data and the millimeter-wave radar point cloud data. This means that each frame of lidar point cloud data is aligned with the two millimeter-wave radar point cloud data frames with the smallest timestamp difference.

[0114] In this embodiment, the millimeter-wave radar point cloud data is interpolated according to the UAV flight posture measured by the IMU inertial measurement unit, specifically:

[0115]

[0116] in: is the interpolated millimeter-wave radar point cloud, 、 are two adjacent frames of millimeter-wave radar original point clouds, is the time weighting coefficient, is the timestamp of the lidar data, 、 It is the timestamp of two adjacent frames of millimeter wave radar.

[0117] Furthermore, the PointNet++ algorithm extracts local geometric features (such as surface curvature and normal vectors) from the lidar point cloud data. Using RCS and Doppler velocity filtering, high-confidence point cloud data is selected from the millimeter-wave radar point cloud data. Specifically, a 6DoF transformation matrix is ​​constructed based on the extracted local geometric features and the selected point cloud data. This matrix is ​​then used to align the millimeter-wave radar point cloud data with the lidar coordinate system.

[0118] Furthermore, we use the LiDAR points as queries and the mmWave radar points as keys and values ​​to construct query-key-value pairs and determine the corresponding similarity weights, specifically:

[0119]

[0120] in: is the attention weight of the lidar point r and the millimeter wave radar point o, is the query vector of the lidar point r, is the inversion of the bond vector of the millimeter-wave radar point o, is the dimension of the feature vector.

[0121] In this embodiment, the formula for obtaining the query vector of the laser radar point r and the key vector of the millimeter wave radar point o is specifically:

[0122]

[0123] in: is the query vector of the lidar point r, is the bond vector of millimeter-wave radar point o, 、 is the trainable parameter matrix, is the feature vector of the rth point in the lidar point cloud, is the feature vector of the oth point in the millimeter-wave radar point cloud.

[0124] Specifically, based on the determined similarity weights, the fused lidar point features are determined as follows:

[0125]

[0126] in: is the fused lidar point feature, is the attention weight of the lidar point r and the millimeter wave radar point o, is the value vector of millimeter-wave radar point o, is the trainable parameter matrix, is the feature vector of the oth point in the millimeter-wave radar point cloud, The point index in the millimeter-wave radar point cloud.

[0127] Step SB2: Uncertainty Fusion. This involves determining the confidence interval of each data point using the kernel function set based on the aligned LiDAR point cloud and millimeter-wave radar point cloud obtained in step SB1.

[0128] Furthermore, the weight of each sensor is adjusted according to the signal-to-noise ratio of each sensor, specifically:

[0129]

[0130] in: is the dynamic weight of the sth sensor, is the 1.5th power of the signal-to-noise ratio of the sth sensor, is the base of natural logarithm, is the historical confidence correction factor of the sth sensor, is the slope coefficient of the logistic function, The index of the sensor.

[0131] Step SC: Hierarchical fault-tolerant control. This involves adjusting the corresponding sensor combination based on the visibility of the external environment, and generating the corresponding sudden obstacle avoidance trajectory based on the configured UWB positioning module and inflatable buffer structure. The details are as follows:

[0132] Step SC1: Set the flight strategy table. That is, according to the preset visibility threshold range (which can be set according to the actual situation, so it is not explained in detail in this embodiment), set the sensor combination. Specifically:

[0133] When the real-time visibility is less than the lower limit of the preset visibility threshold range, environmental data is collected through the anti-turbulence acoustic wave positioning group set in step SA3, namely the bionic acoustic sensor and the IMU inertial measurement unit.

[0134] When the real-time visibility is within the preset visibility threshold range, environmental data is collected through the super-resolution millimeter-wave radar group set in step SA2 and the bionic acoustic sensor set in step SA3.

[0135] When the real-time visibility is greater than the upper limit of the preset visibility threshold range, environmental data is collected through the anti-scattering laser radar group set in step SA1 and the super-resolution millimeter-wave radar group set in step SA2.

[0136] In this embodiment, the point cloud attenuation rate of the laser radar is determined based on the number of valid points in the laser radar point cloud data, and the corresponding real-time visibility is determined based on the point cloud attenuation rate of the laser radar and the signal-to-noise ratio of the millimeter wave radar. Specifically,

[0137]

[0138] in: is the visibility estimate after fusion, is the lidar weight, is the visibility calculated by the lidar, Visibility calculated for millimeter-wave radar.

[0139] Furthermore, based on the real-time confidence of the LiDAR and millimeter-wave radar, the LiDAR weight is determined as follows:

[0140]

[0141] in: is the lidar weight, is the real-time confidence of the lidar, is the real-time confidence of millimeter-wave radar.

[0142] Step SC2: Emergency Processing. This involves determining an emergency alarm signal based on the effective point cloud density of the LiDAR and the number of valid targets detected by the millimeter-wave radar. Specifically, an emergency alarm signal is triggered when the effective point cloud density of the LiDAR falls below a preset threshold (e.g., 20%) or when the millimeter-wave radar fails to acquire a valid target for three consecutive frames.

[0143] Furthermore, when the emergency alarm signal is triggered, the drone switches to hover mode and simultaneously deploys its inflatable cushioning structure. Specifically, the cushioning structure in this embodiment is a polyurethane-coated nylon inflatable airbag that is inflated with CO2 gas, reaching 80% of its volume within 0.5 seconds.

[0144] Furthermore, multiple UWB base stations are set up on the ground, with the spacing between adjacent UWB base stations set at 50-100 meters, and each UWB base station can receive TOF data. Specifically, the 3D position of the drone can be determined based on the TOF data obtained.

[0145] In this embodiment, based on the determined drone's 3D position, external environment data, and the drone's remaining battery life, multiple landing trajectories are acquired within the digital twin model, and a trajectory score corresponding to each landing trajectory is determined. Specifically, the trajectory scores corresponding to all landing trajectories are compared, and the maximum trajectory score is determined. The landing trajectory corresponding to this maximum trajectory score is then used as the final landing trajectory.

[0146] Step SC3: Obstacle avoidance path planning. This involves converting the valid point cloud data obtained by the LiDAR, the strongly reflective targets obtained by the millimeter-wave radar, and the positions of the dynamic obstacle bounding boxes obtained by the visual sensor (such as the camera) into the drone's body coordinate system and performing spatiotemporal alignment.

[0147] Furthermore, based on the aligned multi-sensor data, the current position of the drone, the starting direction control point, the ending direction control point, and the end point of the path are determined. Based on the current position of the drone, the starting direction control point, the ending direction control point, and the end point of the path, the corresponding trajectory curve is determined, specifically:

[0148]

[0149] in: For Bezier curves in parameters The 3D coordinates of is the normalized curve parameter, is the current position of the drone, is the starting direction control point, is the ending direction control point, The end point of the path.

[0150] In this embodiment, based on the determined trajectory curves, the cost function corresponding to each trajectory curve is obtained, specifically:

[0151]

[0152] in: is the cost function, is the path length weight coefficient, is the path length, is the obstacle distance weight coefficient, is the minimum distance from the path to the nearest obstacle, is the energy consumption weight coefficient, To estimate energy consumption.

[0153] Specifically, according to the cost function corresponding to each trajectory curve, a minimum cost function is determined, and the trajectory curve corresponding to the minimum cost function is the optimal obstacle avoidance path trajectory.

[0154] refer to Figure 4 , Figure 4 is the obstacle avoidance path simulation diagram in this embodiment, Figure 4 Obstacle 1 is located at (3, 4) with a radius of 1.0, slightly above the middle of the path. Obstacle 2 is located at (5, 2) with a radius of 0.8, slightly below the middle of the path. The generated obstacle avoidance path (solid line) forms an S-shaped curve, controlled by obstacles 1 and 2, naturally bypassing the obstacles while maintaining a safe distance from their edges.

[0155] This embodiment also provides a method for active avoidance of a drone, which uses the above-mentioned method for active avoidance of a drone.

[0156] Example 2

[0157] This embodiment provides a method for active drone avoidance. Its specific implementation is similar to that of Example 1, except that, in step SA1.5, each noise point data is corrected based on the determined noise point data and the overall point cloud data to reduce the point cloud volatility. The present invention is described below with reference to the specific implementation of this embodiment.

[0158] In this embodiment, each noise point data is corrected based on the determined noise point data and the overall point cloud data, specifically as follows:

[0159] Step SA1.5.1: Determine the neighborhood geometric center. That is, based on the 3D coordinates of all point clouds within the neighborhood, determine the corresponding neighborhood geometric center. Specifically:

[0160]

[0161] in: Point Cloud The corresponding neighborhood geometric center, is the three-dimensional coordinate of the j-th point cloud, is the index of the point cloud, Point Cloud The neighborhood set of .

[0162] Step SA1.5.2: Move the noise point. That is, according to the neighborhood geometric center determined in step SA1.5.1, move the noise point to the neighborhood geometric center to obtain the three-dimensional coordinates of the moved point cloud, specifically:

[0163]

[0164] in: Point Cloud The three-dimensional coordinates after movement, Point Cloud The corresponding neighborhood geometric center, is the three-dimensional coordinate of the i-th point cloud, is the relaxation factor.

[0165] Step SA1.5.3: Determine the volatility. That is, based on the 3D coordinates of the point cloud after movement in step SA1.5.2, re-determine the corresponding noise point data, and based on the number of noise points and the total number of point clouds, determine the corresponding point cloud volatility, specifically:

[0166]

[0167] in: is the point cloud volatility, is the number of noise points, is the total number of point clouds.

[0168] Specifically, the obtained point cloud fluctuation rate is compared with a preset fluctuation threshold (which can be set according to the actual situation, so it is not specifically explained in this embodiment), and the noise point data is corrected based on the comparison result. Specifically:

[0169] If the obtained point cloud fluctuation rate exceeds the preset fluctuation threshold, repeat steps SA1.5.1 through SA1.5.2 until the obtained point cloud fluctuation rate is no greater than the preset fluctuation threshold. Conversely, if the obtained point cloud fluctuation rate is no greater than the preset fluctuation threshold, the 3D coordinates of the point cloud after the shift are the 3D coordinates after the noise point data has been corrected.

[0170] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is limited by the accompanying embodiments and their equivalents.

Claims

1. A method for active avoidance of a UAV, characterized in that: Includes: SA: Weather Adaptive Sensing: Acquires the physical characteristics of the external environment through an anti-scattering lidar group, a super-resolution millimeter-wave radar group, and an anti-turbulence sonar positioning group; SB: Multimodal decision fusion: LiDAR and millimeter-wave radar data are aligned and calculated using Gaussian process regression to obtain the confidence interval of each sensor and determine the weight of the sensor. SC: Hierarchical Fault Tolerant Control: Adjusts the sensor combination based on the visibility of the external environment, and determines the sudden obstacle avoidance trajectory based on the set UWB positioning module and inflatable buffer structure, including: SC1: Set the flight strategy table: According to the preset visibility threshold range, set the sensor combination, specifically: When the real-time visibility is less than the lower limit of the preset visibility threshold range, the sensor combination is set to a bionic acoustic sensor and an IMU inertial measurement unit; when the real-time visibility is within the preset visibility threshold range, the sensor combination is set to a super-resolution millimeter-wave radar group and a bionic acoustic sensor; when the real-time visibility is greater than the upper limit of the preset visibility threshold range, the sensor combination is set to an anti-scattering laser radar group and a super-resolution millimeter-wave radar group; SC2: Emergency handling: According to the emergency alarm signal, the UAV is switched to hovering mode, and the 3D position of the UAV is determined by the UWB positioning module, and the inflatable buffer structure is activated at the same time; SC3: Obstacle Avoidance Path Planning: Based on the current position of the UAV, the starting direction control point, the ending direction control point, and the end point of the path, a trajectory curve is determined, and the cost function of the trajectory curve is obtained. At the same time, all cost functions are compared to determine the minimum cost function. The trajectory curve corresponding to the minimum cost function is the optimal obstacle avoidance path trajectory.

2. The method for active avoidance of a UAV according to claim 1, characterized in that: Obtain the physical characteristics of the external environment, including: SA1: Set up an anti-scattering lidar group: Use a dual-wavelength laser transmitter to compensate for rain and snow scattering attenuation and remove raindrop pseudo-targets. It also identifies and filters the characteristic frequency bands of rain and snow, and adjusts the point cloud volatility to a preset range. SA2: Set up a super-resolution millimeter-wave radar group: reconstruct sparse echo signals based on the sparse array, sharpen the edges of the millimeter-wave image, and align the lidar data in time and space based on the cross-modal model; SA3: Set up the anti-turbulence acoustic wave positioning group: Use the bionic acoustic sensor to obtain sound waves within the preset range, and correct the sound wave propagation time error based on the wind speed vector. Specifically: ； in: is the corrected equivalent sound speed, is the speed of sound in a static environment, is the wind speed vector, is the unit vector of the sound wave propagation direction.

3. The method for active avoidance of a UAV according to claim 2, characterized in that: Adjust the point cloud volatility to a preset range, including: SA1.1: Dynamic PRF Adjustment: Adjusting the pulse repetition frequency of the dual-wavelength laser transmitter according to the atmospheric optical thickness to determine the adjusted pulse repetition frequency; SA1.2: Scattering compensation: Determine the atmospheric attenuation coefficient based on the laser wavelength and ambient visibility, and compensate the original signal intensity received by the lidar to obtain the compensated laser echo intensity; SA1.3: Reflection Filtering: Determine the degree of polarization based on the vertical polarization component strength and the parallel polarization component strength. Compare the degree of polarization with a preset polarization threshold range and filter the echo signals based on the comparison results. Specifically: When the polarization degree is within the preset polarization threshold range, the light spot echo signal corresponding to the polarization degree is retained; otherwise, the laser spot echo signal corresponding to the polarization degree is deleted; SA1.4: Spectral Analysis: Set the stopband range, passband range, and filter order of the band-stop filter based on the echo characteristics and the motion state of the obstacle, and filter the determined echo characteristics through the band-stop filter; SA1.5: Point cloud denoising: Determine the point cloud structure based on the point cloud position and a preset radius. Obtain the overall offset vector of the point cloud based on its 3D coordinates within the point cloud structure. Compare the overall offset vector with a preset offset threshold. Select the point cloud data based on the comparison results. Specifically: When the overall offset vector is greater than a preset offset threshold, the point cloud data corresponding to the overall offset vector is deleted; otherwise, the point cloud data corresponding to the overall offset vector is retained.

4. The method for active avoidance of a UAV according to claim 3, characterized in that: According to the preset time window size, the laser echo signal is segmented to obtain a short-term analysis unit, and the power spectrum density of the short-term analysis unit is determined by Fourier transform. At the same time, the power spectrum density is compared with the preset power spectrum threshold range. According to the comparison result, the echo characteristics are determined, specifically: When the power spectrum density is greater than the upper limit threshold of the preset power spectrum threshold range, the echo feature corresponding to the power spectrum density is a raindrop signal feature; when the power spectrum density is within the preset power spectrum threshold range, the echo feature corresponding to the power spectrum density is a snowflake signal feature; when the power spectrum density is less than the lower limit threshold of the preset power spectrum threshold range, the corresponding echo feature is an obstacle signal feature.

5. The method for active avoidance of a UAV according to claim 3, characterized in that: According to the noise point data and the overall point cloud data, each noise point data is corrected as follows: SA1.5.1: Determine the neighborhood geometric center: Determine the neighborhood geometric center based on the 3D coordinates of all point clouds within the neighborhood. SA1.5.2: Noise point movement: Move the noise point data toward the geometric center of the neighborhood to obtain the three-dimensional coordinates of the moved point cloud, specifically: ； in: Point Cloud The three-dimensional coordinates after movement, Point Cloud The corresponding neighborhood geometric center, is the three-dimensional coordinate of the i-th point cloud, is the relaxation factor; SA1.5.3: Noise point movement: Re-determine the noise point data based on the moved point cloud 3D coordinates, and determine the point cloud volatility based on the number of noise points and the total number of point clouds. Compare the point cloud volatility with a preset volatility threshold, and modify the noise point data based on the comparison result, specifically: When the point cloud fluctuation rate is greater than the preset fluctuation threshold, repeat steps SA1.5.1-SA1.5.2 until the point cloud fluctuation rate is no greater than the preset fluctuation threshold; otherwise, the three-dimensional coordinates of the moved point cloud are the three-dimensional coordinates after correction of the noise point data.

6. The method for active avoidance of a UAV according to claim 2, characterized in that: The 30-50kHz Helmholtz resonant cavities are arranged in ascending order, and an eddy current suppression grid is set at the cavity entrance. The cavity is filled with porous sound-absorbing material to obtain a bionic acoustic sensor.

7. The method for active avoidance of a UAV according to claim 1, characterized in that: Determining the weight of the sensor includes: SB1: Spatiotemporal alignment processing: Use the lidar point as the query and the millimeter-wave radar point as the key and value to construct a query-key-value pair, determine the similarity weight between the lidar point and the lidar point, and determine the fused lidar point features based on the similarity weight. Specifically: ； in: is the fused lidar point feature, is the attention weight of the lidar point r and the millimeter wave radar point o, is the value vector of millimeter-wave radar point o, is the trainable parameter matrix, is the feature vector of the oth point in the millimeter-wave radar point cloud, is the point index in the millimeter wave radar point cloud; SB2: Uncertainty Fusion: Based on the aligned LiDAR point cloud and the MMW radar point cloud, the confidence interval of each data point is determined, and the weight of the sensor is set according to the signal-to-noise ratio of the sensor. Specifically: ； in: is the dynamic weight of the sth sensor, is the 1.5th power of the signal-to-noise ratio of the sth sensor, is the base of natural logarithm, is the historical confidence correction factor of the sth sensor, is the slope coefficient of the logistic function, The index of the sensor.

8. The method for active avoidance of a UAV according to claim 1, characterized in that: The inflatable cushioning structure includes but is not limited to an inflatable airbag made of polyurethane-coated nylon.

9. The method for active avoidance of a UAV according to claim 1, characterized in that: The drone's 3D position, remaining battery power, and external environment data are used as input to the digital twin model, and the landing trajectory and trajectory score are obtained as output. At the same time, all the trajectory scores are compared to determine the maximum trajectory score. The landing trajectory corresponding to the maximum trajectory score is the final landing trajectory.

Images