Low-altitude aircraft power line obstacle avoidance detection system based on strobing imaging

Abstract

The application relates to the technical field of unmanned aerial vehicle intelligent obstacle avoidance, and discloses a low-altitude aircraft power line obstacle avoidance detection system based on gating imaging, which dynamically generates an optimal scanning scheme through an adaptive distance layering module, realizes distance-gating imaging through nanosecond-level accurate timing control, extracts power line features through multi-frame fusion and feature enhancement, carries out accurate detection by using a lightweight deep learning network, determines the target position by using time-of-flight ranging and three-dimensional positioning, and finally generates an obstacle avoidance instruction through collision risk assessment and reinforcement learning obstacle avoidance planning. The system enhances image quality by using a weather-adaptive image restoration algorithm under severe weather conditions. The application can realize long-distance accurate detection and real-time obstacle avoidance of power lines in a mountainous low-altitude flight environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent obstacle avoidance technology for unmanned aerial vehicles (UAVs), and more specifically, to a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging. Background Technology

[0002] Low-altitude aircraft face the risk of collision with power lines in applications such as logistics delivery in mountainous areas and power line inspection. Traditional passive optoelectronic obstacle avoidance systems rely on ambient light, and their performance is severely degraded in low visibility conditions such as fog, haze, and rain, with an effective detection range of only 200-300 meters, which is insufficient to provide sufficient obstacle avoidance reaction time for the aircraft. The main problems with existing technologies include: the detection performance of passive optical systems drops sharply under adverse weather conditions, making them unsuitable for the complex environments of mountainous areas; traditional two-dimensional image processing methods struggle to accurately distinguish power lines at different distances, easily leading to false alarms and missed detections; and existing detection networks have a large number of parameters, making real-time processing on embedded platforms difficult.

[0003] To address the aforementioned issues, a gated imaging technique using active laser illumination is employed. Utilizing 1550 nm laser active illumination, which has a lower atmospheric extinction coefficient compared to the visible light band, it offers superior penetration even in hazy conditions. Nanosecond-level precise timing control enables selective imaging of slices at specific distances. By precisely synchronizing laser emission with camera gate selection, background scattering interference is effectively suppressed, achieving three-dimensional imaging along the distance dimension.

[0004] Meanwhile, deep learning and reinforcement learning technologies are further introduced to construct a detection module based on a lightweight network architecture, controlling the number of parameters to meet the real-time processing requirements of the embedded platform. A meteorological adaptive image restoration algorithm is also adopted, dynamically adjusting the image processing strategy according to real-time weather conditions. Through multimodal sensor fusion and physical model constraints, the detection accuracy under harsh conditions is improved. Summary of the Invention

[0005] This invention provides a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging, which solves the technical problems of short detection distance and low detection accuracy under adverse weather conditions in related technologies.

[0006] This invention provides a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging, comprising: The range layering module is used to acquire flight parameters, predict echo intensity, and obtain range slice sequences using an adaptive range layering strategy. The gating imaging module is used to acquire the range slice sequence. It uses a range gating imaging method with precise time control to obtain the range slice image sequence. The electric field line feature enhancement module is used to acquire the distance slice image sequence and uses multi-frame image fusion and electric field line enhancement methods to obtain a list of pixel coordinates of electric field line candidate regions. The lightweight detection module is used to obtain a list of pixel coordinates of candidate regions for power lines and to obtain the power line target detection results using a lightweight detection method. The ranging and positioning module is used to acquire the detection results of power line targets. It uses time-of-flight ranging and three-dimensional spatial positioning methods to obtain the distance and position of the power line targets. The obstacle avoidance decision module is used to obtain the distance and position of the power line target, and to obtain obstacle avoidance command information by using collision risk assessment and obstacle avoidance decision methods.

[0007] In a preferred embodiment, the distance layering module includes: A smart distance slice optimization model is constructed, which combines reinforcement learning strategies and scene adaptation mechanisms to dynamically generate the optimal scanning scheme. The state space is defined to include the current flight speed, echo power distribution, and remaining time budget features. The action space includes the center position, thickness, and pulse accumulation count of each slice. The reward function comprehensively considers detection coverage, signal-to-noise ratio balance, and time efficiency. A deep Q-network is used for decision optimization, and a scene memory module is introduced to cluster and build a library of scenes. During online inference, the corresponding strategy is quickly invoked by scene matching based on the echo power distribution and the current flight status. The output includes the center distance sequence, thickness sequence, pulse accumulation number sequence, and timing control parameter table of the distance slice.

[0008] In a preferred embodiment, the gating imaging module includes: The field-programmable gate array timing control unit operates a high-precision clock to provide a preset nanosecond time resolution, generating laser trigger pulses and camera gating trigger pulses; The laser trigger pulse is a rising edge trigger signal that is transmitted to the laser control board through a shielded cable. At the same time, the internal delay counter is started. When the accumulated value reaches the number of clock cycles corresponding to the preset delay time, a camera gating trigger pulse is generated. The field-programmable gate array has pre-set delay calibration parameters. The deviation between the actual laser emission time and the trigger signal time, and the deviation between the actual camera shutter opening time and the trigger signal time are determined through offline calibration experiments. Compensation is performed during timing generation to ensure that the arrival time of the laser echo coincides precisely with the camera gating window.

[0009] In a preferred embodiment, the power line feature enhancement module includes: A dual-stream encoder structure is constructed to extract multi-scale features from two consecutive frames of images. The feature fusion module uses an attention mechanism to calculate feature correlation. For each position of the feature map of the first frame, the similarity between it and the feature map of the second frame is calculated to construct a correlation tensor. After normalization, the correlation tensor is weighted and summed with the features of the second frame to obtain the alignment features. The fused features are input into the regression network and the parameters of the affine transformation matrix are output. Each frame of image is aligned to the coordinate system of the first frame according to the transformation matrix, and the registration error is less than the preset pixel threshold.

[0010] In a preferred embodiment, the power line feature enhancement module further includes: A deep learning-based background separation and image fusion model is constructed. The generator adopts a U-Net structure. The input is a sequence of registered multi-frame images, and the output is a fused static background image. A temporal attention module is introduced between the encoder and decoder to perform weighted fusion of multi-frame features and calculate the similarity between the features of each frame and the reference frame. The generator loss function includes adversarial loss, reconstruction loss, and temporal consistency loss. During online inference, the generator outputs a fused image and a dynamic region mask. The dynamic regions marked by the mask are filled with surrounding background pixels.

[0011] In a preferred embodiment, the lightweight detection module includes: A lightweight power line detection network was constructed and optimized based on the YOLOv8-nano architecture. The backbone network adopted depthwise separable convolution instead of standard convolution, and strip convolution modules were inserted at key positions of the network to efficiently capture linear features in the horizontal or vertical directions using asymmetric convolution kernels. Knowledge distillation technique is used to optimize network performance. YOLOv8-large is constructed as the teacher network. When training the student network, distillation loss is introduced, which includes feature distillation and logistic distillation. Feature distillation requires that the feature maps of the intermediate layers of the student network be consistent with the corresponding layers of the teacher network. Logical distillation requires that the output probability distribution be consistent with the soft labels of the teacher network.

[0012] In a preferred embodiment, the lightweight detection module further includes: A coordinate attention mechanism is adopted, and a coordinate attention mechanism module is inserted at a key position of the backbone network to adaptively weight the feature maps; Global average pooling is calculated on the feature maps of each channel to obtain channel descriptors, and channel attention weights are generated by learning the dependencies between channels through fully connected layers. Pooling is performed along the horizontal and vertical directions in the spatial dimension to obtain horizontal coordinate encoding and vertical coordinate encoding, and spatial attention weights are generated by learning the importance of spatial location through convolutional layers. Multiply the channel attention weights by the spatial attention weights to obtain a two-dimensional attention map. Multiply the attention map element-wise with the original feature map to achieve feature weighting.

[0013] In a preferred embodiment, the ranging and positioning module includes: The timing control unit reads the laser emission time and camera gating time corresponding to the image frame where the detected power line target is located. The time interval between the two times is calculated to be equal to the flight time of the laser to and from the target. The target distance is equal to the speed of light multiplied by the time interval and then divided by 2. The angle between the target ray and the optical axis is calculated based on the target's position in the image. The actual optical path is equal to the optical axis distance divided by the cosine of the angle. A ranging error correction model is established as a mapping function from pixel position to ranging deviation. The correction amount is obtained by querying the correction model based on the target pixel position. The accurate distance is obtained by adding the correction amount to the coarse distance.

[0014] In a preferred embodiment, the obstacle avoidance decision module includes: Construct an obstacle avoidance trajectory planning model based on reinforcement learning. Define the state space to include the aircraft's current position, velocity, attitude, the position and risk level of the detected power line target, and the destination direction and distance information. Define the action space to include the heading angle adjustment, pitch angle adjustment, velocity adjustment and continuous control variables. The reward function is designed to consider obstacle avoidance safety, trajectory deviation, energy efficiency, and time cost. A deep deterministic policy gradient algorithm is adopted, and a dynamic window method is introduced as a safety constraint layer to verify the feasibility of the actions output by reinforcement learning. Within the dynamic window, forward simulation is performed on each candidate velocity to evaluate the collision risk between the trajectory and obstacles, the target proximity, and the speed smoothness.

[0015] In a preferred embodiment, a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging further includes an image restoration module, specifically comprising: Acquire information from ambient light sensor, temperature and humidity sensor and image statistical features, extract global average gray value, contrast, gradient intensity and edge density statistical features from distance-gated image sequence, construct meteorological recognition neural network to output the probability distribution of each meteorological category; The physical model constraints are embedded in the neural network architecture. The network input is a sequence of multi-distance slice images, and the output is an atmospheric transmittance map and extinction coefficient distribution at each distance. A physical constraint layer is embedded in the network to perform physical consistency verification of the transmittance predicted by the network based on the atmospheric scattering model.

[0016] The beneficial effects of this invention are as follows: The gating imaging technology, which employs 1550 nm active laser illumination and nanosecond-level precise timing control, enables selective imaging of slices at specific distances, effectively suppressing background scattering light interference. Under hazy conditions, the effective detection distance reaches over 850 meters, representing a significant performance improvement compared to traditional passive optoelectronic systems. A lightweight detection network and a reinforcement learning-based obstacle avoidance planning module are constructed, reducing the number of parameters, increasing inference speed on the embedded GPU platform, and reducing the overall system response time, thus meeting the real-time obstacle avoidance requirements of low-altitude aircraft. Attached Figure Description

[0017] Figure 1 This is a block diagram of a low-altitude aircraft power line obstacle avoidance detection system based on gating imaging according to the present invention. Figure 2 This is a flowchart of a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging, according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging, such as... Figures 1 to 2 As shown, it includes the following steps: The range layering module is used to acquire flight parameters, predict echo intensity, and obtain range slice sequences using an adaptive range layering strategy. Based on the real-time flight status of the aircraft and the characteristics of the laser echo, an adaptive range scanning strategy is generated to ensure coverage of key ranges within a limited time window, while maintaining consistent imaging quality across different range slices. This includes the following steps: S11: Obtain the aircraft's speed and braking performance parameters, and use dynamic calculation methods to obtain the minimum safe braking distance and the scanning range boundary; The system acquires the current velocity provided by the aircraft's inertial navigation unit, reads the maximum braking acceleration from the aircraft's dynamics parameter database, and loads the safety margin time from the system configuration file. It calculates the theoretical braking distance required for the aircraft to decelerate from the current velocity to zero using the formula: velocity squared divided by twice the braking acceleration. To account for system response delay and decision time, the product of velocity and safety margin time is added to the theoretical braking distance to obtain the minimum safe braking distance. Based on the minimum safe braking distance, a safety factor of 1.2 is introduced to calculate the near-range scanning boundary, which should be greater than the minimum braking distance to allow for a safety margin. The long-range scanning boundary is set to the system's maximum detection capability of 1500 meters, a value determined by laser power, camera sensitivity, and atmospheric transmission characteristics. The output scanning distance range is then defined.

[0020] S12, obtain system parameters and scanning distance range, and use the echo power prediction algorithm to obtain the laser echo intensity distribution at each distance; The system parameters were read as follows: laser single-pulse energy, divergence angle, and wavelength (1550 nm). The camera receiving lens diameter, quantum efficiency, and typical reflectivity of electric field lines were also read. For electric field line targets at a given distance, the power density of the laser irradiated onto the target surface equals the laser energy divided by the irradiated area, which is determined by the product of distance and the square of the divergence angle. The laser power reflected from the target equals the incident power density multiplied by the target's effective reflective area and reflectivity. The reflected laser undergoes geometric diffusion again during its return journey, and the power reaching the receiving lens equals the reflected power multiplied by the lens area divided by the square of the distance (four times pi). Atmospheric two-way transmission loss was considered, and atmospheric transmittance was introduced. This parameter was obtained from the MODTRAN atmospheric radiative transfer database based on wavelength and visibility. For the 1550 nm band, under hazy conditions with 500 m visibility, the atmospheric extinction coefficient is lower than that of the visible light band, giving this band stronger penetration capabilities. Combining these factors, a functional relationship between distance and echo power was established. Within the scanning distance range, samples were taken at 10-meter intervals, and the predicted echo power at each sampling point was calculated to construct an echo power distribution curve. The mapping table between output distance and echo power provides a basis for subsequent adaptive parameter settings.

[0021] S13, obtain the echo power distribution and camera noise characteristics, and calculate the effective detection range using the minimum detectable power threshold; Obtain the noise equivalent power parameter of the short-wave infrared camera, which characterizes the minimum optical power that the detector can resolve per unit bandwidth. The camera uses a cooled InGaAs focal plane array detector, maintaining an operating temperature of -20 degrees Celsius, effectively suppressing dark current noise. Readout noise, dark current noise, and photon noise parameters of the detector are read, and the total noise power is calculated. A minimum signal-to-noise ratio (SNR) requirement for target detection is set, typically requiring an SNR greater than a first threshold to ensure reliable detection. The minimum SNR is multiplied by the noise power to obtain the minimum detectable echo power threshold. This threshold is compared with the echo power distribution curve obtained in step S12 to identify the range where the echo power is greater than the threshold; this range is the effective detection range of the system. Edge regions where the echo power is close to the threshold are marked as weak signal regions, requiring enhancement measures such as multi-pulse accumulation in subsequent steps to improve the SNR. The effective detection range and the range of weak signal regions are output as constraints for the range stratification strategy.

[0022] S14: Obtain echo power distribution and flight status, and use intelligent distance layering algorithm to obtain the optimal distance slice sequence; A smart distance slicing optimization model is constructed, which combines reinforcement learning strategies and scene adaptation mechanisms to dynamically generate the optimal scanning scheme. The state space is defined to include features such as current flight speed, echo power distribution, and remaining time budget; the action space includes the center position, thickness, and pulse accumulation count of each slice; the reward function comprehensively considers detection coverage, signal-to-noise ratio balance, and time efficiency. A deep Q-network (DQN) is used for decision optimization, consisting of three fully connected layers and employing the ReLU activation function. During the offline training phase, scanning parameters and detection result samples from historical flight missions are collected to construct an experience replay buffer. A scene memory module is introduced to cluster and build a library of typical scenes such as dense power grids in mountainous areas, sparse power grids in suburban areas, and low visibility due to fog and haze. During online inference, based on the echo power distribution and current flight state from step S12, the corresponding strategy is quickly invoked through scene matching, input into the DQN network, and the optimal slice configuration is output. Considering a total system response time budget of 150 milliseconds, deducting 60 milliseconds for processing and decision-making, the image acquisition time is 90 milliseconds, and a single slice imaging takes approximately 9 milliseconds, supporting a maximum of 10 slices. The model optimizes the slice distribution by densely sampling near the minimum safe braking distance and sparsely sampling in distant regions. It outputs the center distance sequence, thickness sequence, pulse accumulation count sequence, and timing control parameter table for 10 distance slices.

[0023] S15: Obtain the range slice sequence and echo power prediction. Using a dynamic gate width and multi-pulse accumulation strategy, obtain the imaging parameter sequence for each slice. Based on the slice thickness obtained in step S14, the camera gate width corresponding to each slice is calculated. The gate width is equal to the optical path time corresponding to the slice thickness, specifically twice the slice thickness divided by the speed of light. For long-distance, weak-signal slices, the signal-to-noise ratio (SNR) of single-pulse imaging is insufficient, requiring multi-pulse accumulation imaging. Based on the echo power prediction in step S12, the SNR of each slice under single-pulse imaging is calculated; for slices with an SNR lower than the target SNR, the required number of accumulation pulses is calculated, where the number of accumulations is equal to the square of the ratio of the target SNR to the single-pulse SNR. To control the total imaging time, a maximum limit is set for the maximum number of accumulations per slice. For slices requiring multi-pulse accumulation, multiple laser pulses are continuously emitted at the same distance delay, and the camera performs multiple gated exposures, stacking multiple frames of images at the pixel level to achieve linear signal accumulation while noise only increases at a square root rate, thereby improving the SNR. Considering the laser's repetition frequency limitation, the interval between adjacent pulses must not be less than the laser capacitor charging time. Output the gating gate width sequence, laser pulse accumulation sequence, and complete timing control parameter table for each distance slice. This parameter table will be loaded into the FPGA distance gating control unit for execution.

[0024] The gating imaging module is used to acquire the range slice sequence. It uses a range gating imaging method with precise time control to obtain the range slice image sequence. By precisely coordinating laser active illumination and gated imaging, selective imaging of slice targets at specific distances is achieved, effectively suppressing background scattering interference from other distance ranges; specifically, the following steps are included: S21, obtain the distance slice parameter table, generate a nanosecond-level trigger signal to obtain the synchronization pulse for laser emission and camera gating; The Field Programmable Gate Array (FPGA) timing control unit receives the distance slice parameter table from step S1, including the delay time, gate width, and pulse count for each slice. The FPGA internally operates a high-precision clock at 100 MHz, providing a time resolution of 10 nanoseconds. Based on the currently scanned distance slice index, the corresponding delay time and gate width parameters are read from the parameter table. The FPGA generates a laser trigger pulse, a rising-edge trigger signal with a pulse width of 20 nanoseconds and a voltage amplitude of 5 volts, transmitted to the laser control board via a shielded cable. Simultaneously, the FPGA starts an internal delay counter, which accumulates in clock cycles. When the accumulated value reaches the number of clock cycles corresponding to the preset delay time, a camera gating trigger pulse is generated. The camera trigger pulse is a TTL level signal; a rising edge indicates the start of exposure, and a falling edge indicates the end of exposure. The pulse width is equal to the gating gate width. To compensate for cable transmission delay and device response delay, the FPGA has pre-set delay calibration parameters. Offline calibration experiments determine the deviation between the actual laser emission time and the trigger signal time, as well as the deviation between the actual camera shutter opening time and the trigger signal time. Compensation is performed during timing generation to ensure that the laser echo arrival time precisely coincides with the camera gating window, achieving sub-nanosecond time synchronization accuracy. The output laser emission trigger pulse and camera gating trigger pulse have a time interval that precisely corresponds to the target distance.

[0025] S22 receives the laser trigger pulse and adopts a collimated beam-expanding illumination optical path design to obtain laser illumination that uniformly covers the camera's field of view; The 1550 nm pulsed laser employs an erbium-doped fiber amplifier architecture, with a laser pulse width of 10 nanoseconds, a single pulse energy of 8 millijoules, and a repetition frequency of 100 Hz. After being collimated by a collimating lens group, the laser is expanded by a beam-expanding lens group with a beam expansion ratio of 5, resulting in an output beam diameter of 25 mm. The expanded laser beam is then homogenized by a microlens array to form a flat-top intensity distribution with a divergence angle of 2 degrees, covering the field of view of the short-wave infrared camera. The optical axis of the laser emitting unit is coaxially aligned with the camera's optical axis, ensuring complete overlap between the illumination and imaging areas. The output is a uniformly collimated laser illumination field that propagates to distant electric field line targets.

[0026] S23, receive the laser illumination field, use diffuse reflection echo propagation to obtain the echo light wave carrying the target spatial information; The laser illumination field propagates to the surface of the target's electric field lines, causing diffuse reflection. The reflected light carries information about the target's spatial location, shape, and surface reflectivity. The reflected light undergoes atmospheric transmission and is attenuated by atmospheric scattering on its return path to the camera. The atmospheric scattering coefficient at the 1550 nm wavelength is lower than that of the visible light band, maintaining high transmittance even under hazy conditions. Upon reaching the camera, the reflected light is collected and focused by the objective lens, which is optimized for achromatic reflection at the 1550 nm wavelength, with a focal length of 50 mm and a relative aperture of F2.0, forming a clear image of the target on the focal plane. The laser echo reflected from the target is then focused onto the surface of the short-wave infrared camera detector.

[0027] S24 receives the echo light wave and the gating trigger pulse, and uses gating-gated exposure imaging to obtain the target image of a slice at a specific distance. The short-wave infrared camera employs a cooled InGaAs focal plane array detector with an array size of 640×512 pixels and a pixel size of 15 micrometers. The detector is maintained at -20 degrees Celsius using thermoelectric cooling, reducing dark current noise to an extremely low level. The detector achieves peak quantum efficiency at 1550 nm and integrates a nanosecond-level response electronic shutter at the front end. After receiving a gating trigger pulse from the FPGA, the shutter control circuit drives the electronic shutter to precisely open and close, with the exposure time equal to the gating width. During the gating window, laser echoes reflected from the target distance slice reach the detector, where photons are absorbed by the InGaAs material, generating a photocurrent. The accumulated capacitance within the pixel integrates this photocurrent. After exposure, the readout circuit converts the charge signal into a 14-bit digital image signal. Because the gating width corresponds only to the optical path time of a specific distance slice, the shutter closes when reflected light from targets at other distances arrives, preventing signal generation. This achieves selective imaging of specific distances, effectively suppressing foreground and background interference. A single-frame distance slice image is output, where the target's position and intensity reflect the spatial distribution of electric field lines within that distance slice.

[0028] S25, acquire the distance slice sequence, and obtain an image sequence covering the entire distance range by image sequence acquisition and storage; The FPGA timing control unit sequentially scans and images 10 distance slices according to the distance slice sequence generated in step S1. For each distance slice, the laser emission and gating imaging process in steps S21 to S24 is repeated. For long-range slices requiring multi-pulse accumulation, multiple imaging operations are performed continuously under the same distance delay parameter. The image processing unit performs pixel-aligned superposition of multiple frames of the same slice, accumulating the gray values ​​of each pixel to achieve signal accumulation. The signal-to-noise ratio of the accumulated image is improved by a factor equal to the square root of the number of accumulations. After imaging a single slice, the FPGA updates the distance slice index, loads the timing parameters of the next slice, and continues the next round of imaging. After the 10 distance slices are scanned sequentially, 10 frames of images are obtained, each frame corresponding to a distance slice, and the images exhibit a spatial distance layering relationship. Image data is transmitted to the image processing unit through the CameraLink high-speed interface, with a transmission bandwidth of hundreds of megabytes per second, ensuring real-time transmission of image data. The image processing unit stores the image sequence in a buffer and records the distance label, timestamp, and imaging parameter metadata corresponding to each frame of image. The output consists of a distance-gated image sequence containing 10 frames and a corresponding distance label sequence, which serves as input data for subsequent image processing.

[0029] The electric field line feature enhancement module is used to acquire the distance slice image sequence and uses multi-frame image fusion and electric field line enhancement methods to obtain a list of pixel coordinates of electric field line candidate regions. The process involves registration, fusion, and feature enhancement of multi-distance slice images to extract linear structural features of electric field lines and suppress background interference and transient noise. Specifically, it includes the following steps: S31, acquire inertial navigation unit data, and use motion compensation prediction algorithm to obtain approximate transformation relationship of continuous frame images; The aircraft undergoes continuous motion during range scanning, resulting in displacement, rotation, and scaling variations between consecutive frames. Angular velocity data from a three-axis gyroscope and linear acceleration data from a three-axis accelerometer are read from the aircraft's inertial navigation unit (INS). The INS data sampling frequency is 200 Hz, and the time synchronization accuracy is aligned with the image acquisition time. The rotation angle changes of the camera attitude, including pitch, yaw, and roll angles, are calculated by integrating the angular velocity within the scanning time window. The translational changes of the camera position are calculated by performing a second integration based on the linear acceleration. Using the camera's pinhole imaging model, the rotation and translation transformations in three-dimensional space are projected onto the image plane to obtain the two-dimensional transformation relationship between images. For small-angle rotations and short-distance translations, the image transformation can be approximated as an affine transformation, including rotation, translation, scaling, and shear components. The affine transformation matrix between consecutive frames is calculated; this matrix is ​​3x3 and contains 6 degrees of freedom. While the transformation relationship predicted by the INS has accumulated errors, it can serve as initial values ​​for subsequent accurate registration, accelerating convergence. Output the sequence of predicted affine transformation matrices between consecutive image frames to provide initial registration values ​​for step S32.

[0030] S32, obtain the image sequence and the predicted transformation matrix, and use a lightweight registration network to obtain the accurate inter-frame transformation matrix; A lightweight image registration network is constructed, predicting inter-frame transformation parameters through end-to-end learning. The network employs a dual-stream encoder structure, extracting multi-scale features from two consecutive frames. Shallow features capture edge details, while deep features encode semantic information. The feature fusion module uses an attention mechanism to calculate feature correlation. For each position in the feature map of the first frame, its similarity to the feature map of the second frame is calculated, constructing a correlation tensor. This correlation tensor is then normalized using softmax and weighted and summed with the features from the second frame to obtain the alignment features. The fused features are input into a regression network consisting of three fully connected layers, outputting six parameters of the affine transformation matrix. Network training employs a strategy combining synthetic and real data. Synthetic data is generated by applying known transformations to real images, while real data comes from historical flight image sequences. Training uses the Adam optimizer with a learning rate of 0.001 and a batch size of 16. The network inference time is approximately 5 milliseconds, achieving sub-pixel level registration accuracy. Each frame image is aligned to the coordinate system of the first frame according to the transformation matrix, outputting a registered image sequence with a registration error of less than 0.3 pixels.

[0031] S33: Obtain the registered image sequence and use intelligent background separation and multi-frame fusion algorithm to obtain high-quality and stable images; A deep learning-based background separation and image fusion model is constructed. The generator adopts a U-Net architecture, taking a sequence of registered multi-frame images as input and outputting a fused static background image. The encoder extracts multi-scale features through convolution and pooling, while the decoder restores image resolution through transposed convolution. A temporal attention module is introduced between the encoder and decoder to perform weighted fusion of multi-frame features, calculate the similarity between each frame's features and the reference frame, and assign greater weight to frames with high similarity, automatically filtering frames containing transient interference. The discriminator adopts a PatchGAN architecture to perform realism judgment on local image patches. The generator's loss function includes adversarial loss, reconstruction loss, and temporal consistency loss. The adversarial loss requires the generated image to deceive the discriminator, the reconstruction loss requires the fused image to be close to the input sequence on average, and the temporal consistency loss requires the static target to remain stable in consecutive frames. The network training adopts an alternating optimization strategy, with training data derived from historical flight image sequences. Training uses the Adam optimizer, with a generator learning rate of 0.0002, a discriminator learning rate of 0.0001, and a batch size of 4. During online inference, the generator outputs a fused image and a dynamic region mask. The dynamic regions marked by the mask are filled with surrounding background pixels. The output is a time-stable fused image, with clear and complete static targets such as power lines. Transient interference is effectively removed, and the image signal-to-noise ratio is improved by several decibels compared to a single frame.

[0032] S34, acquire the fused image, and use linear structuring element opening and closing operations to obtain a morphological image with enhanced linear features; Morphological processing utilizes mathematical morphological operators to extract geometric structural features from the image. A linear structuring element is defined as a long, thin rectangle with a length and width of several pixels, its aspect ratio set to match the scale of the electric field lines in the image. The orientation of the structuring element can be horizontal, vertical, or diagonal, determined based on the dominant direction of the electric field lines in the image. Morphological opening is applied to the fused image obtained in step S33. Opening is a combination of erosion and dilation operations. The erosion operation uses the structuring element to slide across the image. For each position, a high value is output when all pixel values ​​within the area completely covered by the structuring element are higher than a threshold; otherwise, a low value is output. The erosion operation removes isolated noise and burrs smaller than the structuring element size. The dilation operation is the inverse of erosion, using the structuring element to slide across the image. A high value is output at each position as long as high-value pixels exist within the area covered by the structuring element. The dilation operation restores the eroded target contour. The effect of the opening operation is to remove fine noise while preserving the main linear structure. Further application of morphological closing operations, which involve dilation followed by erosion, fills small holes within linear structures and connects broken line segments. Closing operations can connect electric field lines that may appear broken in the image due to occlusion or low contrast. The resulting morphologically enhanced image highlights linear structures and suppresses isolated noise.

[0033] S35, acquire morphological images, and use line detection and connection algorithms to obtain a set of line segments of candidate electric field lines; The Hough transform is a classic method for detecting straight line features in images. For the morphological image obtained in step S34, edge points are extracted as input to the Hough transform. In the Hough parameter space, each edge point in the image space corresponds to a curve in the parameter space; collinear edge points intersect the curves in the parameter space at the same point. A probabilistic Hough transform is used, which reduces computation while maintaining detection performance by randomly sampling edge points. An accumulator array is constructed in the parameter space, where each edge point votes for all points on its corresponding curve in the parameter space, and the accumulator records the number of votes. Accumulator units with high vote counts correspond to straight line features in the image. A vote count threshold is set; accumulator units exceeding the threshold are identified as candidate straight lines. The parameter space is inversely transformed back to the image space to obtain the coordinates of the start and end points of the straight line. A minimum line segment length threshold is set to filter out line segments shorter than this threshold, avoiding false detections of short, interfering line segments. Detected line segments are connected, and it is determined whether the distance between the endpoints of two line segments is less than the connection threshold and whether their directions are consistent. Line segments that meet these conditions are merged to form longer, continuous line segments. Output a set of straight line segments of candidate power lines, each segment represented by the coordinates of its start and end points or its equation.

[0034] S36, obtain the set of line segments, filter them by aspect ratio and direction consistency, and obtain the list of pixel coordinates of candidate regions for electric field lines; Connectivity analysis is performed on the line segment regions detected in step S35. A neighborhood of several pixels is defined around the line segment, and foreground pixels within this neighborhood are extracted. Foreground pixels are defined as pixels with grayscale values ​​higher than the segmentation threshold. Connectivity is marked on the foreground pixels using 4-connectivity or 8-connectivity criteria to group adjacent foreground pixels into the same connected region. The geometric features of each connected region are calculated, including area, perimeter, length and width of the circumscribed rectangle, centroid coordinates, and principal axis direction. The aspect ratio of the connected region is calculated, defined as the longer side of the circumscribed rectangle divided by the shorter side. Since electric field lines are slender targets, their aspect ratio is usually greater than a preset threshold. The principal axis direction of the connected region is calculated using the principal axis inertia method, calculating the second-order moment tensor of the pixel coordinates, with the feature vector corresponding to the principal axis direction. The directional consistency of the connected regions is checked; different segments of the same electric field line should have essentially the same direction. Connected regions with aspect ratios greater than the threshold and satisfying the directional consistency requirement are retained as candidate electric field line regions. Output a list of coordinates of all pixels within the candidate electric field line region, along with the geometric feature parameters of each region. This list serves as the input to the deep learning detection model for candidate regions.

[0035] The lightweight detection module is used to obtain a list of pixel coordinates of candidate regions for power lines and to obtain the power line target detection results using a lightweight detection method. Deep learning models are used to accurately detect and locate candidate regions, improving detection accuracy and reducing false positive rates. The specific steps include: S41. Obtain the training dataset, and use lightweight network design and knowledge distillation to obtain an optimized detection network model; A lightweight power line detection network was constructed and specifically optimized based on the YOLOv8-nano architecture. The backbone network uses depthwise separable convolutions instead of standard convolutions to reduce the number of parameters and computational cost. For the slender target feature of power lines, strip convolutional modules were inserted at key network locations, employing 1×7 or 7×1 asymmetric convolutional kernels to efficiently capture linear features in the horizontal or vertical directions. The neck network uses a feature pyramid structure, upsampling and concatenating feature maps from different levels to achieve multi-scale information fusion. The head network outputs detection results, including bounding box coordinates, target confidence, and class probability. Knowledge distillation was used to optimize network performance, and YOLOv8-large was constructed as the teacher network, with approximately eight times the number of parameters as the student network. The teacher network was trained using a dataset containing tens of thousands of labeled power line images. Distillation loss was introduced during student network training, including feature distillation and logistic distillation. Feature distillation requires that the feature maps of the intermediate layers of the student network be consistent with the corresponding layers of the teacher network, while logistic distillation requires that the output probability distribution be consistent with the soft labels of the teacher network. The total loss function of the student network is a weighted sum of the hard loss for the ground truth labels, the feature distillation loss, and the logistic distillation loss. Through distillation training, the student network achieves over 95% of the detection accuracy of the teacher network with only one-eighth the number of parameters. The output is a lightweight detection network model with parameters controlled within 3.5 megabytes, achieving an inference speed of 35 frames per second on the NVIDIA Jetson Xavier NX platform.

[0036] S42, obtain the backbone network feature map, and use the coordinate attention mechanism to obtain the attention feature map that enhances the features of the slender target; To address the characteristics of slender targets like power lines, a coordinate attention mechanism module is introduced into the network. This module is inserted at several key locations in the backbone network to adaptively weight the feature maps. The coordinate attention mechanism performs feature modeling simultaneously in the channel and spatial dimensions. In the channel dimension, global average pooling is calculated for each channel feature map to obtain channel descriptors. Dependencies between channels are learned through fully connected layers to generate channel attention weights. In the spatial dimension, pooling is performed along the horizontal and vertical directions to obtain horizontal and vertical coordinate codes. The importance of spatial positions is learned through convolutional layers to generate spatial attention weights. The channel attention weights and spatial attention weights are multiplied to obtain a two-dimensional attention map. The attention map is then multiplied element-wise with the original feature map to achieve feature weighting. For slender targets like power lines that extend along a certain direction, coordinate attention can effectively capture their directional features, assigning higher weights to the corresponding direction and enhancing the feature representation of slender targets. The output is an attention-weighted feature map, in which the response intensity of the power lines is enhanced relative to the background.

[0037] S43, obtain feature maps at different levels, and use bottom-up and top-down dual-path fusion to obtain a multi-scale fused feature pyramid; The Feature Pyramid Network fuses feature maps from different levels of the backbone network. The backbone network outputs several feature maps of different resolutions, denoted as low-level features, mid-level features, and high-level features. High-level features have low resolution but rich semantic information, while low-level features have high resolution but weak semantic information. A top-down path is used to upsample high-level features using bilinear interpolation or transposed convolution to increase the feature map resolution to match that of mid-level features. The upsampled high-level features are then concatenated with or element-wise added to mid-level features, fusing information from different levels. The process continues downwards, upsampling the fused features again and fusing them with low-level features. The top-down path injects high-level semantic information into low-level features, enhancing their discriminative power. A bottom-up path is used to progressively reduce the resolution of low-level features through convolution and pooling, fusing them with high-level features and injecting high-resolution spatial detail information into high-level features. After dual-path fusion, feature maps at each level possess both rich semantic information and fine spatial information, which is beneficial for detecting targets at different scales. The output is a multi-scale fused feature pyramid containing several feature maps of different resolutions.

[0038] S44, obtain the feature pyramid, and use high aspect ratio anchor box regression and confidence prediction to obtain the power line bounding box and confidence. The detection head network processes the feature maps of each layer of the feature pyramid. Several anchor boxes are predefined rectangular boxes at each spatial location in the feature map, serving as priors for object detection. For the elongated shape of the power line, high aspect ratio anchor boxes are designed, with aspect ratios set to multiples such as 10, 20, and 30. The anchor boxes are oriented horizontally, vertically, and at several oblique angles to cover various possible orientations of the power line. For each anchor box, the network predicts several parameters, including bounding box regression offset, object confidence, and class probability. The bounding box regression offset represents the adjustment of the anchor box's position and size relative to the true object box, including center point coordinate offset and width / height scaling factors. The object confidence represents the probability that the anchor box contains the object. The class probability represents the probability distribution of the object belonging to various classes; in this application, the classes are power line and background. The network is trained using a loss function that is a weighted sum of classification loss, regression loss, and confidence loss. The classification loss uses cross-entropy, the regression loss uses smoothed L1 loss or IoU loss, and the confidence loss uses binary cross-entropy. The training dataset contains a large number of labeled power line images, covering different scenes, weather conditions, and distance conditions. Data augmentation techniques such as random cropping, rotation, and brightness adjustment are used to increase data diversity and improve the model's generalization ability. After training, the network performs forward inference on the test images and outputs the prediction results for all anchor boxes. The output includes the bounding box coordinates, target confidence, and class probability for each anchor box.

[0039] S45, obtain the anchor box prediction results, and use the non-maximum suppression algorithm to obtain the final list of power line detection bounding boxes; The network output contains a large number of anchor box predictions, many of which point to the same target, indicating duplicate detection. A non-maximum suppression (NMS) algorithm is used to remove duplicate boxes. First, a confidence threshold is set, filtering out anchor boxes with confidence below the threshold and retaining high-confidence candidate boxes. The retained candidate boxes are sorted in descending order of confidence, and the box with the highest confidence is selected as the current detection result. The overlap between this box and other candidate boxes is calculated using the Intersection over Union (IoU) metric, which is equal to the area of ​​the intersection of two boxes divided by the area of ​​their union. Candidate boxes with an IoU greater than the preset NMS threshold are considered duplicate boxes detecting the same target and are deleted. The next highest-confidence box is selected from the remaining candidate boxes, and the above process is repeated until all candidate boxes have been processed. After NMS, only one optimal detection box is retained for each target. The coordinates of the detection boxes are post-processed, converting normalized coordinates to image pixel coordinates. The center point coordinates of the detection box are calculated as the target's location point. Output the power line target detection results, including the bounding box coordinates, center point coordinates, and confidence score of each target. These results are used for subsequent distance measurement and localization.

[0040] S46, obtain the student network model, and use INT8 quantization technology to obtain an optimized embedded deployment model; To further improve the model's inference speed and reduce resource consumption, the student network was quantized using INT8. The network weights and activation values ​​were converted from 32-bit floating-point to 8-bit integers, reducing the model size to one-quarter and increasing inference speed several times over. The quantization process employed post-training quantization or quantization-aware training, calibrating quantization parameters to minimize accuracy loss. The quantized model was deployed on an embedded GPU platform such as the NVIDIA Jetson series, utilizing the TensorRT inference acceleration framework, achieving inference speeds exceeding 30 frames per second. The optimized embedded deployment model was output, meeting real-time requirements.

[0041] The ranging and positioning module is used to acquire the detection results of power line targets. It uses time-of-flight ranging and three-dimensional spatial positioning methods to obtain the distance and position of the power line targets. The detection results of the image plane are converted into three-dimensional spatial coordinates to achieve accurate ranging and positioning of power line targets; specifically, the following steps are included: S51: Obtain the detection results and time-series records, and use flight time calculation to obtain the approximate distance to the target; The timing control unit in step S2 reads the laser emission time and camera gating time corresponding to the image frame where the detected power line target is located. The laser emission time is recorded by the FPGA timing unit at the same time as triggering the laser, and the gating time is the time when the FPGA sends the gating trigger signal. The time interval between the two times is calculated, which is equal to the flight time of the laser to and from the target. According to the principle of the constancy of the speed of light, the laser propagation distance is equal to the speed of light multiplied by the propagation time. Since the laser travels twice the distance to and from the target, the target distance is equal to the speed of light multiplied by the time interval and then divided by 2. The speed of light is taken as 299,792,458 meters per second. Since gating imaging corresponds to a distance slice rather than a single distance point, the actual target distance is located near the center distance of the slice, and the center distance of the slice is used as the initial value of the coarse distance. Considering the accuracy limitations of the system timing control, cable transmission delay, device response delay, etc., the coarse distance has a certain error. The coarse distance value of the target is output, which serves as the initial estimate for subsequent precise ranging.

[0042] S52, obtain the pixel coordinates of the detection result, and use distortion correction and coordinate transformation to obtain the normalized camera coordinates of the target; The intrinsic parameter matrix is ​​read from the camera calibration file. This matrix contains parameters such as focal length, principal point coordinates, and pixel size. Distortion coefficients, including radial and tangential distortion coefficients, are read. Radial distortion is caused by the lens shape, resulting in barrel or pincushion distortion at the image edges; tangential distortion is caused by misalignment between the lens and sensor installation, resulting in asymmetric image distortion. For the pixel coordinates of the power line target center point detected in step S4, distortion correction is first performed. According to the distortion model, the relationship between pixel coordinates and ideal coordinates is described by a polynomial. The radial distortion component is proportional to the power of the distance from the pixel to the principal point, while the tangential distortion component is related to the cross term of the pixel coordinates. The distorted pixel coordinates are mapped back to the ideal pixel coordinates using iterative solving or lookup table methods. The ideal pixel coordinates are then converted to normalized camera coordinates. Normalized coordinates are defined as pixel coordinates minus the principal point coordinates divided by the focal length. The physical meaning of normalized coordinates is the projection of the target ray's direction vector in the camera coordinate system onto the unit focal length plane. Normalized coordinates are dimensionless, eliminating the influence of camera intrinsic parameters. Output the normalized camera coordinates of the target, which represent the direction of the target relative to the camera's optical center.

[0043] S53, obtains coarse distance and normalized coordinates, and uses the principle of triangulation to obtain the three-dimensional position of the target in the camera coordinate system; According to the pinhole camera model, there is a proportional relationship between the target's 3D coordinates and normalized coordinates. The target's 3D coordinates in the camera coordinate system are equal to the normalized coordinates multiplied by the target's depth distance. The depth distance is the straight-line distance from the target to the camera's optical center, approximately equal to the coarse distance obtained in step S51. Specifically, the horizontal component of the normalized coordinates is multiplied by the depth distance to obtain the horizontal component of the 3D coordinates, and the vertical component of the normalized coordinates is multiplied by the depth distance to obtain the vertical component of the 3D coordinates. The depth component is directly equal to the depth distance. The camera coordinate system is defined as follows: the optical center is the origin, the optical axis is the positive Z-axis, the horizontal direction of the image is the X-axis, and the vertical direction of the image is the Y-axis. The output is the target's 3D coordinates in the camera coordinate system, representing its position relative to the camera's optical center.

[0044] S54: Obtain the target pixel position and rough distance, and use the distance correction model to obtain the precisely corrected target distance; The time-of-flight distance obtained in step S51 assumes the laser propagates along the optical axis. However, when the target is located at the edge of the field of view, the actual optical path length between the laser and the target is greater than the optical axis distance. Based on the target's position in the image, the angle between the target ray and the optical axis is calculated. This angle is equal to the arctangent of the normalized coordinate vector's modulus. The actual optical path is equal to the optical axis distance divided by the cosine of the angle. For small angles, the increase in optical path is half the distance multiplied by the square of the angle. Furthermore, the non-uniformity of refractive index at the lens edge and manufacturing errors also introduce ranging bias. A ranging error correction model is established, which is a mapping function from pixel position to ranging bias, obtained through offline calibration experiments. The calibration experiment is conducted on a target board at a known distance, fitting the deviation between the measured distance and the true distance as a polynomial function of the pixel position. During online ranging, the correction model is queried based on the target pixel position to obtain the correction amount. The coarse distance is then added to the correction amount to obtain the precise distance. The corrected ranging accuracy reaches 0.3 meters within a distance of 1000 meters, meeting obstacle avoidance requirements. Outputs the precisely corrected target distance, which is used to update the depth component of the 3D coordinates.

[0045] S55: Obtain the three-dimensional position of the camera coordinate system, and use coordinate system transformation to obtain the position of the target in the aircraft body coordinate system; The current pitch, yaw, and roll angles are read from the aircraft's attitude and heading reference system. These attitude angles describe the rotational relationship between the aircraft's body coordinate system and the navigation coordinate system. The extrinsic parameter matrix of the camera coordinate system relative to the aircraft's body coordinate system is read from the system installation calibration file. This extrinsic parameter includes a rotation matrix and a translation vector. The rotation matrix describes the axial relationship between the two coordinate systems, and the translation vector describes the position of the camera's optical center in the body coordinate system. The target camera coordinates obtained in step S53 are transformed to the body coordinate system using the rotation matrix, and then the translation vector is added to obtain the target's position in the aircraft's body coordinate system. The body coordinate system is defined as follows: the aircraft's center of mass is the origin, the nose direction is the positive X-axis, the right wing direction is the positive Y-axis, and vertically downwards is the positive Z-axis. The three-dimensional position coordinates of the target in the aircraft's body coordinate system and the Euclidean distance from the target to the aircraft's center of mass are output. This information is directly used by the flight control system.

[0046] S56, obtain the target's position in the body coordinate system and GPS data, and use geographic coordinate transformation to obtain the target's geographical location in the world coordinate system; The aircraft reads its current location (longitude, latitude, and altitude) from its GPS receiver. The GPS position corresponds to the aircraft's center of mass in the WGS84 global geodetic coordinate system. The target's relative position in the body coordinate system is transformed to the navigation coordinate system using an attitude rotation matrix. The navigation coordinate system is a northeast-northeast geodetic coordinate system, with the X-axis pointing north, the Y-axis pointing east, and the Z-axis pointing to the Earth's center. The relative position in the navigation coordinate system is added to the aircraft's GPS position to obtain the target's absolute geographical location. For addition operations in the geodetic coordinate system, the Earth's curvature must be considered, using a local tangent plane approximation or a geodetic coordinate transformation formula. The target's longitude, latitude, and altitude are output. This information can be used for updating the power line geographic information database and for information sharing during multi-aircraft cooperative obstacle avoidance.

[0047] The obstacle avoidance decision module is used to obtain the distance and position of the power line target, and to obtain obstacle avoidance command information by using collision risk assessment and obstacle avoidance decision methods; Based on the detected power line locations, the collision risk is assessed, an obstacle avoidance decision is generated, and sent to the aircraft control system via a standard interface; specifically, the following steps are included: S61: Obtain the target's three-dimensional position and flight trajectory, and use the minimum approach distance prediction algorithm to obtain collision risk parameters; Step S5 retrieves the three-dimensional position coordinates of each detected electric field line target in the aircraft's coordinate system. The current velocity vector, comprising three directional components, is read from the aircraft's navigation system. The planned flight trajectory, consisting of several waypoints, is read from the flight management system, each containing its position coordinates and expected arrival time. Assuming the aircraft moves at a constant velocity in a straight line along the current velocity vector direction, the minimum approach distance between the aircraft trajectory and each electric field line target is calculated. For point targets, the minimum approach distance equals the perpendicular distance from the target's position to the straight line of the flight trajectory, calculated using the point-to-line distance formula. For linear targets such as electric field lines, the extension direction of the electric field lines must be considered; the electric field lines are modeled as spatial straight line segments, and the shortest distance between two skew lines is calculated. The time of reaching the minimum approach distance is calculated, which is equal to the current time plus the distance from the current position to the nearest point divided by the velocity. The collision time is calculated, defined as the remaining time to reach the minimum approach distance. Collision risk parameters, including the minimum approach distance, collision time, and nearest point coordinates, are output for each target.

[0048] S62, obtain collision risk parameters, and use multi-factor comprehensive scoring to obtain the collision risk level; A collision risk assessment model is established, comprehensively considering factors such as minimum approach distance, collision time, target relative speed, and aircraft maneuverability. A risk scoring function is defined, where the risk score is inversely proportional to both the minimum approach distance and the collision time. Targets with a minimum approach distance less than a safe distance threshold are assigned a high-risk score; the safe distance threshold is set based on aircraft size and obstacle avoidance margin. Targets with a collision time less than a safe time threshold are also assigned a high-risk score; the safe time threshold is set based on aircraft braking capability and decision response time. The risk scores are normalized to a range of 0 to 1, and risks are categorized into several levels based on the score: low risk, medium risk, high risk, and emergency risk. Low risk indicates the target is far away and there is sufficient time for collision, requiring no obstacle avoidance action; medium risk indicates the target is approaching but there is still room for obstacle avoidance, suggesting course adjustment; high risk indicates the target is very close and collision time is imminent, requiring immediate obstacle avoidance; emergency risk indicates an imminent collision, requiring emergency braking or emergency maneuvering. Risk assessments are performed on all detected power line targets, and the target with the highest risk level is selected as the primary threat target. The risk score and risk level of each target, as well as the identifier of the primary threat target, are output.

[0049] S63: Obtain the target location and risk level, and use an intelligent obstacle avoidance trajectory planning algorithm to obtain the optimal obstacle avoidance maneuver command; A reinforcement learning-based obstacle avoidance trajectory planning model is constructed. The state space is defined to include the aircraft's current position, velocity, attitude, the location and risk level of detected power line targets, and the direction and distance to the destination. The action space is defined to include continuous control variables such as yaw, pitch, and velocity adjustments. The yaw adjustment range is -30° to +30°, the pitch adjustment range is -15° to +15°, and the velocity adjustment range is -5 m / s to +5 m / s. A reward function is designed, comprehensively considering obstacle avoidance safety, trajectory deviation, energy efficiency, and time cost. The basic reward is a positive reward for successfully avoiding an obstacle and a negative reward for colliding with an obstacle; the trajectory deviation penalty is the distance between the current position and the planned route multiplied by a penalty coefficient; the energy penalty is the square of the maneuver acceleration multiplied by an energy consumption coefficient. A deep deterministic policy gradient (DDPG) algorithm is adopted, including an Actor network and a Critic network. The Actor network outputs the optimal action based on the current state, and the Critic network evaluates the Q-value of the state-action pair. The Actor network employs a 3-layer perceptron structure, with the output layer using a tanh activation function to constrain actions within a reasonable range. The Critic network takes the concatenation of the state vector and action vector as input and outputs a Q-value scalar. Training utilizes an experience replay mechanism, storing interactive samples in a replay buffer. A Dynamic Window Method (DWA) is introduced as a safety constraint layer to verify the feasibility of actions output by reinforcement learning. The DWA calculates the achievable velocity space for the next time step based on the aircraft's current velocity and acceleration constraints. Forward simulations are performed on each candidate velocity within the dynamic window to evaluate the collision risk between the trajectory and obstacles, target proximity, and velocity smoothness, selecting the velocity with the highest overall score as the final action. During training, a flight scenario including power line obstacles is constructed in a simulation environment, with randomized scenario parameters. During online inference, the current state vector is constructed based on the target position and risk level obtained in steps S61 and S62, input into the trained Actor network, and outputs a suggested obstacle avoidance action. This action is then input into the DWA safety constraint layer for verification and correction to obtain the final action. The output includes suggested obstacle avoidance command parameters such as heading angle, pitch angle, and velocity adjustment.

[0050] S64 acquires obstacle avoidance decision information, encapsulates it using ARINC429 standard messages, and obtains flight control interface communication data packets; The obstacle avoidance decision information generated in steps S61 to S63 is encapsulated into a standard avionics data format. The ARINC429 protocol is a widely used unidirectional data bus standard in the aviation industry. Data is transmitted in 32-bit words, each containing fields such as tag, data, symbol state matrix, and parity check. Data tags for obstacle avoidance information are defined, with different tag numbers assigned to different information types, such as target distance tags, risk level tags, and avoidance heading tags. Floating-point data is converted to ARINC429 BNR binary numbers or BCD binary-encoded decimal numbers. The symbol state matrix field is set to identify the validity, functional status, and hardware status of the data. Parity bits are calculated to ensure the integrity of data transmission. Multiple 32-bit words are combined into data packets and output according to a predetermined frame structure and transmission frequency. The data packets are sent to the flight control computer through dual redundant channels. Redundant channels improve system reliability; if one channel fails, the other continues to operate. Standard-format obstacle avoidance data packets are output and continuously sent at a frequency of several times per second to ensure the flight control system receives obstacle avoidance information in real time.

[0051] S65 receives obstacle avoidance data packets, adopts a hierarchical decision-making mechanism, and obtains execution instructions for automatic obstacle avoidance or manual takeover. After receiving the obstacle avoidance data packet, the aircraft control system determines the execution strategy based on the current flight mode and risk level. Flight modes include manual, semi-automatic, and fully automatic. In manual mode, the system only issues visual and auditory warnings to the pilot, displaying the location of power lines and collision risks; the pilot then decides whether to take obstacle avoidance action. In semi-automatic mode, for low- and medium-risk targets, the system issues warning suggestions; for high- and emergency-risk targets, the system automatically executes obstacle avoidance maneuvers and notifies the pilot, who can take over control at any time. In fully automatic mode, the system automatically executes obstacle avoidance for all risk levels without manual intervention. The obstacle avoidance execution module generates aircraft control commands based on the suggested heading, including control surface deflection angles, engine thrust, and desired flight attitude values. These commands are sent to the flight control law module for closed-loop control. The system includes a manual takeover interface; the pilot can exit automatic obstacle avoidance mode and resume manual control at any time via the joystick or takeover button, ensuring final decision-making authority remains with the pilot. Obstacle avoidance execution status information is output, and an obstacle avoidance event log is recorded, including detection time, target parameters, decision-making process, and execution results, for subsequent analysis and system optimization.

[0052] S66: Obtain the detection results and enhanced images, and use augmented reality annotation technology to obtain the operator monitoring video output; The system provides a video output interface for operators or pilots to monitor the obstacle avoidance system's operational status in real time. The shortwave infrared image acquired in step S2 or the enhanced image from step S6 is used as the background video. Visual annotations of the detection results are overlaid on the video image. For power line targets detected in step S4, colored rectangles are drawn at their bounding boxes, with the rectangle colors set according to risk level: green for low risk, yellow for medium risk, orange for high risk, and red for emergency risk. Text information, including target distance, collision time, and risk level, is annotated next to the target. Vector arrows indicating the aircraft's current heading and suggested avoidance heading are drawn to display the obstacle avoidance maneuver direction. System operational status information, including imaging mode, weather conditions, processing frame rate, and system health status, is displayed above the video. The video output format supports HDMI or SDI standard interfaces, with a resolution of HD or higher and a frame rate of 30 frames per second, meeting real-time monitoring requirements. The video can be recorded and stored as a file for post-event review and training. The output monitoring video with augmented reality annotations provides intuitive obstacle avoidance situational awareness.

[0053] In one embodiment of the present invention, to address the image degradation problem under severe weather conditions such as fog and rain, adaptive image enhancement and restoration are also required to improve the detectability of power lines. A low-altitude aircraft power line obstacle avoidance detection system based on gated imaging further includes: an image restoration module, used to acquire a distance-gated image sequence and employ a weather-adaptive image restoration algorithm to obtain an enhanced image; specifically as follows: S71 acquires sensor data and image statistical features, and uses a meteorological intelligent recognition algorithm to obtain weather category and intensity level; A multimodal meteorological recognition system was constructed, integrating ambient light sensors, temperature and humidity sensors, and image statistical features. The ambient light sensor measures ambient illuminance in the visible and near-infrared bands, with a sampling frequency of 10 Hz. The temperature and humidity sensor measures ambient temperature and relative humidity; humidity information plays a crucial role in distinguishing between haze and rainfall. Multidimensional statistical features were extracted from the distance-gated image sequence obtained in step S2, including global average grayscale value, contrast, gradient strength, and edge density. Contrast is defined as the standard deviation of image grayscale divided by the average value; gradient strength is the average response of the image after passing through the Sobel operator; and edge density is the proportion of edge pixels to total pixels. A meteorological recognition neural network was constructed, with the network input being a multimodal feature vector containing illuminance values, temperature, humidity, and image statistical features. The network architecture adopted a multilayer perceptron (MLP) structure, containing three fully connected layers, each using the ReLU activation function and Dropout regularization. The output layer used the softmax activation function to output the probability distribution of each meteorological category. Meteorological categories were defined as sunny, light fog, moderate fog, heavy fog, light rain, moderate rain, and heavy rain. The network training data comes from meteorological samples accumulated from historical flight missions, with a dataset size of tens of thousands of samples. Training uses the cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, and a batch size of 64. During online inference, the system collects sensor data and image features in real time, inputs them into the trained network, and outputs the current meteorological category label, intensity level, and confidence score.

[0054] S72, obtain image sequences and meteorological categories, and use an atmospheric parameter estimation network to obtain the distribution of transmittance and scattering coefficient; An atmospheric parameter estimation network is constructed, embedding physical model constraints into the neural network architecture. The network input is a sequence of multi-distance sliced ​​images obtained in step S2, and the output is the atmospheric transmittance map and extinction coefficient distribution at each distance. The network architecture adopts an encoder-decoder structure. The encoder extracts features from the input image sequence, and the decoder generates the transmittance map. The encoder uses a 3D convolutional neural network, treating multiple frames of images as spatiotemporal volumes. The 3D convolution simultaneously captures spatial features and distance dimension variations. The decoder uses transposed convolution for upsampling to restore the original image resolution. A physical constraint layer is embedded in the network to perform physical consistency verification of the transmittance predicted by the network based on the atmospheric scattering model. The atmospheric scattering model is expressed as: the intensity of the observed image equals the scene radiance multiplied by the transmittance plus atmospheric light multiplied by 1 minus the transmittance. The physical constraint layer embeds this model as a differentiable computation graph into the network, reconstructs the observed image based on the predicted transmittance and atmospheric light, and calculates the difference between the reconstructed image and the actual observed image as the physical consistency loss. The total network loss function includes transmittance supervision loss, physical consistency loss, and smoothing regularization loss. The network training employs a strategy combining synthetic and real data, using the Adam optimizer with a learning rate of 0.0001 and a batch size of 8. During online inference, the network performs forward propagation on the input image sequence, with a single inference time of approximately 15 milliseconds. The output includes transmittance maps and extinction coefficient distributions at various distances, with a transmittance estimation error of less than 5%.

[0055] S73, acquire the distance-gated image sequence, and use local minimum filtering to obtain the dark channel image of the scene; The dark channel prior is based on the statistical regularity of natural images: in non-sky regions, at least one color channel's pixel value approaches zero. For short-wave infrared single-channel images, the dark channel is defined as the minimum pixel value within a local image patch. For each frame in the range-gated image sequence obtained in step S2, a local window of several pixels multiplied by several pixels is set at each pixel position. The minimum value of all pixels within the window is calculated, and this minimum value is assigned to the center pixel, resulting in the dark channel image. The dark channel reflects the degree of atmospheric light scattering; in fog-free areas, the dark channel value is close to zero, while in foggy areas, the dark channel value is higher. Based on the dark channel image, the global atmospheric light intensity is estimated by selecting the top percentage of brightest pixels in the dark channel image and finding the maximum gray value corresponding to these pixel positions in the original image; this value is considered the atmospheric light intensity. The dark channel image and the global atmospheric light intensity are output for accurate transmittance estimation.

[0056] S74: Obtain the dark channel image and atmospheric light intensity, and use transmittance inversion calculation to obtain the transmittance map of the scene; According to the atmospheric scattering physics model, the intensity of the observed image equals the scene's true radiance multiplied by the transmittance, plus atmospheric light multiplied by 1 minus the transmittance. This model is modified and applied to the dark channel. Assuming the scene's true radiance in the dark channel approaches zero, the intensity of the observed image in the dark channel equals atmospheric light multiplied by 1 minus the transmittance. This leads to the inverse relationship where the transmittance equals 1 minus the observed dark channel intensity divided by atmospheric light. For the dark channel image and atmospheric light intensity obtained in step S73, the transmittance at each pixel location is calculated. To avoid excessive dehazing leading to image distortion, a retention coefficient is introduced in the transmittance calculation, and the transmittance is multiplied by this coefficient for attenuation. For pixels with excessively low transmittance, a lower threshold is set to avoid division by zero errors and noise amplification. The transmittance map of the scene is output, and this map is fused with the transmittance estimation result from step S72, using a weighted average or selecting an estimate with higher confidence.

[0057] S75, acquire the transmittance map and distance-gated image sequence, and use guided filtering to keep the edges smooth to obtain a refined transmittance map; The transmittance map obtained in step S74 is calculated within a local window, resulting in block artifacts and insufficient precision at target edges. Guided filtering is used to refine the transmittance map. Guided filtering uses an image from the distance-gated image sequence as a guide image to smooth the edges of the transmittance map. The core idea of ​​guided filtering is that the output image has a linear relationship with the guide image within the local window. The linear coefficients are solved by minimizing the difference between the output and input while maintaining the linear relationship. Guided filtering performs strong smoothing in flat areas and maintains transitions in edge areas to avoid edge blurring. The window radius and regularization parameter of the guided filter are set; the window radius controls the degree of smoothing, and the regularization parameter controls the dependence on the guide image. Guided filtering is applied to the transmittance map from step S74, using the distance-gated image sequence from step S2 as a guide, to output a refined transmittance map. This map is precisely aligned at target edges and is smooth internally without block artifacts.

[0058] S76, obtain the refined transmittance map and atmospheric light intensity, and use the inverse operation of the atmospheric scattering model to obtain the dehazed restored image; Based on the inverse operation of the atmospheric scattering model, the true radiance of the scene is equal to the observed image minus atmospheric light multiplied by 1 minus transmittance, then divided by the transmittance. For each frame in the distance-gated image sequence obtained in step S2, the refined transmittance map from step S75 and the atmospheric light intensity from step S73 are used to calculate the restored image pixel by pixel. The restoration formula is: the restored image equals the original image minus atmospheric light divided by the transmittance, plus an offset. The offset is used to adjust the dynamic range of the restored image to maintain a consistent format with the original image. For pixels with transmittance close to zero, the division operation will amplify noise; therefore, truncation is performed by setting a lower limit for transmittance. The contrast of the restored image is improved, and details of distant targets are restored. The dehazed restored image sequence is output, with enhanced contrast between electric field lines and the background, which is beneficial for subsequent detection.

[0059] S77, obtain the restored image sequence, and use contrast-limited adaptive histogram equalization to obtain the final enhanced image; The restored image sequence from step S76 undergoes further contrast enhancement. A contrast-limited adaptive histogram equalization algorithm is employed. This algorithm divides the image into several sub-blocks, performs histogram equalization independently on each sub-block, and smoothly transitions the sub-block boundaries using bilinear interpolation. Histogram equalization redistributes gray levels, making the image histogram more uniformly distributed, thereby improving contrast. To avoid excessive noise amplification, a contrast limit is introduced, cropping portions of the histogram exceeding the limit height and redistributing the cropped pixels to other gray levels. The contrast limit threshold is adaptively adjusted based on the image noise level; a low threshold is used for images with high noise, and a high threshold is used for images with low noise. The processed image maintains overall brightness while improving local contrast and enhancing the edge sharpness of electric field lines. The final enhanced image sequence is output, replacing the distance-gated image sequence from step S2 as input to step S3 for subsequent processing.

[0060] In one embodiment of the present invention, an application example of a low-altitude aircraft power line obstacle avoidance detection system based on gated imaging is provided: This example focuses on the application of power line obstacle avoidance for logistics drones in mountainous areas. During a particular flight mission, the logistics drone took off from a suburban distribution center, bound for a mountain village. The flight path traversed several valleys, passing multiple high-voltage power lines. The drone maintained an altitude of 120 meters and a speed of 70 kilometers per hour. During the flight, localized fog reduced visibility to approximately 600 meters, posing a challenge to traditional photoelectric systems. The power line obstacle avoidance detection system of this invention operates in real-time, continuously performing distance-gated scanning imaging of the area ahead.

[0061] Adaptive distance stratification and precise timing control: At the 10-minute mark of flight, the system activates the adaptive distance stratification module. Based on the aircraft's current speed of 70 km / h (approximately 19.4 m / s) and maximum braking acceleration of 3 m / s², the minimum safe braking distance is calculated to be approximately 63 meters. The system sets the near-range scanning boundary at 75 meters (with a safety factor of 1.2) and the far-range scanning boundary at 1500 meters. Based on laser echo power prediction, the system automatically generates 10 distance slices within the range of 75 meters to 1500 meters, with center distances of 100 meters, 200 meters, 350 meters, 550 meters, 800 meters, 1050 meters, 1250 meters, 1350 meters, 1450 meters, and 1500 meters, respectively. The system's precise timing control module generates nanosecond-level trigger pulses, and the FPGA timing unit operates at a 100 MHz clock frequency, achieving a time resolution of 10 nanoseconds, ensuring precise synchronization between laser emission and camera gating.

[0062] Multi-frame fusion and electric field line feature enhancement: At the 15-minute mark of flight, the system acquired 10 range slice images. The multi-frame fusion module first reads the aircraft's angular velocity and linear acceleration data from the inertial navigation unit and calculates motion compensation parameters between consecutive frames. A lightweight registration network performs precise image registration, with the registration error controlled within 0.3 pixels. The background separation and fusion module uses a U-Net generator and a PatchGAN discriminator to fuse the registered multi-frame images, outputting a temporally stable fused image. The morphological processing module applies linear structuring element opening and closing operations to enhance the linear features of the electric field lines. Hough transform detects straight segments of candidate electric field lines, and the connected component analysis module filters candidate regions based on aspect ratio and directional consistency.

[0063] Lightweight deep learning detection: The system's lightweight detection module is based on the YOLOv8-nano architecture. This module employs depthwise separable convolutions and strip convolutions to efficiently capture the slender features of electric field lines. A coordinate attention mechanism simultaneously models features in both channel and spatial dimensions, enhancing the feature representation of slender targets. A feature pyramid network enables multi-scale fusion, with high aspect ratio anchor boxes (10x, 20x, 30x) covering various orientations of the electric field lines. Non-maximum suppression removes duplicate detection boxes, outputting the final detection results. After INT8 quantization, the model achieves an inference speed of 35 frames per second on the NVIDIA Jetson Xavier NX platform.

[0064] Time-of-flight ranging and 3D positioning: The system reads the laser emission time and camera gating time from the FPGA timing unit to calculate the time of flight. Based on the principle of the constancy of the speed of light, the target distance equals the speed of light multiplied by the time interval and then divided by 2. The intrinsic parameter matrix and distortion coefficients are read from the camera calibration file, and the pixel coordinates of the detected target center point are distorted and converted to normalized camera coordinates. Based on the pinhole imaging model, the normalized coordinates are multiplied by the depth distance to obtain the target's 3D position in the camera coordinate system. The distance correction model is adjusted based on the target pixel position, achieving a ranging accuracy of 0.3 meters within a 1000-meter range. Through coordinate system transformation, the target position is converted to the aircraft's coordinate system, and then further converted to the world coordinate system using GPS data.

[0065] Weather-adaptive image restoration: At the 22-minute mark of flight, the drone entered an area with dense fog. The system's multimodal weather recognition module, integrating ambient light, temperature and humidity sensors, and image statistical features, accurately identified the current weather conditions as moderate fog / haze, with a visibility of approximately 550 meters and a confidence level of 0.89. The atmospheric parameter estimation network used 3D convolution to process multi-distance slice images, outputting transmittance maps at each distance. The dark channel prior algorithm calculated local minima and estimated atmospheric light intensity. Transmittance inversion and guided filtering refined the transmittance. The atmospheric scattering model inverse operation restored the defogging image, and contrast-limited adaptive histogram equalization was used for final enhancement.

[0066] Collision risk assessment and obstacle avoidance decision-making: The system's obstacle avoidance decision-making module calculates the minimum approach distance and collision time based on the detected power line positions and flight trajectories. A multi-factor comprehensive scoring model considers factors such as distance, time, and relative speed, classifying risks into four levels: low risk, medium risk, high risk, and emergency risk. The reinforcement learning-based obstacle avoidance trajectory planning module uses the DDPG algorithm to output the optimal heading angle, pitch angle, and speed adjustment. The Dynamic Window Method (DWA) is used as a safety constraint layer for feasibility verification. Obstacle avoidance decision information is encapsulated in ARINC429 standard messages and sent to the flight control system at a frequency of several times per second. The flight control system performs obstacle avoidance maneuvers in semi-automatic mode, and the pilot can take over control at any time.

[0067] Specific detection results and performance indicators: At the 15-minute mark of flight, the system detected a high-voltage power line spanning a valley approximately 950 meters ahead. The system's adaptive distance layering module increased the slice density near 950 meters based on the current scene characteristics. Examples of the acquired data are shown in Table 1. Table 1: Detected target parameters of power lines

[0068] The system completes the entire process of image acquisition, processing, target localization, and risk assessment within 145 milliseconds. Based on the risk assessment results, the position and collision risk of each power line target have been accurately assessed. The system's intelligent obstacle avoidance planning module generates the optimal avoidance strategy for each target. Taking target 1 as an example, the power line is located 18 meters to the right of the current flight path. If the current heading is maintained, the aircraft will approach the power line in 48 seconds, posing a collision risk. The system recommends adjusting the heading to the right by 5 degrees, increasing the minimum approach distance between the aircraft and the power line to over 45 meters under the new heading, meeting the safety margin. The obstacle avoidance decision information is sent to the flight control system via the ARINC429 interface. The flight control system executes the heading adjustment in semi-automatic mode, and the UAV smoothly bypasses the power line and continues flying along the planned flight path. The monitoring video shows that the power lines are clearly marked in the shortwave infrared image, using different colors according to the risk level: green for low risk, yellow for medium risk, orange for high risk, and red for emergency risk.

[0069] Continuing flight for 22 minutes, the drone entered an area of ​​denser fog, further reducing visibility. The system accurately identified the weather conditions as moderate fog / haze, with visibility of approximately 550 meters and a confidence level of 0.89. Through intelligent defogging processing, image contrast was improved by nearly 3 times, the signal-to-noise ratio for power line detection was improved by approximately 13 dB, and the transmittance estimation accuracy reached over 96%. On the enhanced image, the system successfully detected another power line (target 2) 1150 meters ahead. Early detection of the power line provided the aircraft with ample time to avoid it, allowing for smooth obstacle avoidance decisions and ensuring flight safety.

[0070] Throughout the flight, the system detected a total of 7 power line targets with a 100% accuracy rate and no missed or false detections. Even in light fog, the effective detection range remained above 850 meters, superior to traditional passive photoelectric systems (200-300 meters).

[0071] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A low-altitude aerial vehicle power line obstacle avoidance detection system based on gated imaging, characterized in that, include: The range layering module is used to acquire flight parameters, predict echo intensity, and obtain range slice sequences using an adaptive range layering strategy. The gating imaging module is used to acquire the range slice sequence. It uses a range gating imaging method with precise time control to obtain the range slice image sequence. The electric field line feature enhancement module is used to acquire the distance slice image sequence and uses multi-frame image fusion and electric field line enhancement methods to obtain a list of pixel coordinates of electric field line candidate regions. The lightweight detection module is used to obtain a list of pixel coordinates of candidate regions for power lines and to obtain the power line target detection results using a lightweight detection method. The ranging and positioning module is used to acquire the detection results of power line targets. It uses time-of-flight ranging and three-dimensional spatial positioning methods to obtain the distance and position of the power line targets. The obstacle avoidance decision module is used to obtain the distance and position of the power line target, and to obtain obstacle avoidance command information by using collision risk assessment and obstacle avoidance decision methods.

2. The low-altitude aircraft power line obstacle avoidance detection system based on gated imaging according to claim 1, characterized in that, The distance layering module includes: A smart distance slice optimization model is constructed, which combines reinforcement learning strategies and scene adaptation mechanisms to dynamically generate the optimal scanning scheme. The state space is defined to include the current flight speed, echo power distribution, and remaining time budget features. The action space includes the center position, thickness, and pulse accumulation count of each slice. The reward function comprehensively considers detection coverage, signal-to-noise ratio balance, and time efficiency. A deep Q-network is used for decision optimization, and a scene memory module is introduced to cluster and build a library of scenes. During online inference, the corresponding strategy is quickly invoked by scene matching based on the echo power distribution and the current flight status. The output includes the center distance sequence, thickness sequence, pulse accumulation number sequence, and timing control parameter table of the distance slice.

3. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The gating imaging module includes: The field-programmable gate array timing control unit operates a high-precision clock to provide a preset nanosecond time resolution, generating laser trigger pulses and camera gating trigger pulses; The laser trigger pulse is a rising edge trigger signal that is transmitted to the laser control board through a shielded cable. At the same time, the internal delay counter is started. When the accumulated value reaches the number of clock cycles corresponding to the preset delay time, a camera gating trigger pulse is generated. The field-programmable gate array has pre-set delay calibration parameters. The deviation between the actual laser emission time and the trigger signal time, and the deviation between the actual camera shutter opening time and the trigger signal time are determined through offline calibration experiments. Compensation is performed during timing generation to ensure that the arrival time of the laser echo coincides precisely with the camera gating window.

4. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The power line feature enhancement module includes: A dual-stream encoder structure is constructed to extract multi-scale features from two consecutive frames of images. The feature fusion module uses an attention mechanism to calculate feature correlation. For each position of the feature map of the first frame, the similarity between it and the feature map of the second frame is calculated to construct a correlation tensor. After normalization, the correlation tensor is weighted and summed with the features of the second frame to obtain the alignment features. The fused features are input into the regression network and the parameters of the affine transformation matrix are output. Each frame of image is aligned to the coordinate system of the first frame according to the transformation matrix, and the registration error is less than the preset pixel threshold.

5. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The power line feature enhancement module also includes: A deep learning-based background separation and image fusion model is constructed. The generator adopts a U-Net structure. The input is a sequence of registered multi-frame images, and the output is a fused static background image. A temporal attention module is introduced between the encoder and decoder to perform weighted fusion of multi-frame features and calculate the similarity between the features of each frame and the reference frame. The generator loss function includes adversarial loss, reconstruction loss, and temporal consistency loss. During online inference, the generator outputs a fused image and a dynamic region mask. The dynamic regions marked by the mask are filled with surrounding background pixels.

6. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The lightweight detection module includes: A lightweight power line detection network was constructed and optimized based on the YOLOv8-nano architecture. The backbone network adopted depthwise separable convolution instead of standard convolution, and strip convolution modules were inserted at key positions of the network to efficiently capture linear features in the horizontal or vertical directions using asymmetric convolution kernels. Knowledge distillation technique is used to optimize network performance. YOLOv8-large is constructed as the teacher network. When training the student network, distillation loss is introduced, which includes feature distillation and logistic distillation. Feature distillation requires that the feature maps of the intermediate layers of the student network be consistent with the corresponding layers of the teacher network. Logical distillation requires that the output probability distribution be consistent with the soft labels of the teacher network.

7. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The lightweight detection module also includes: A coordinate attention mechanism is adopted, and a coordinate attention mechanism module is inserted at a key position of the backbone network to adaptively weight the feature maps; Global average pooling is calculated on the feature maps of each channel to obtain channel descriptors, and channel attention weights are generated by learning the dependencies between channels through fully connected layers. Pooling is performed along the horizontal and vertical directions in the spatial dimension to obtain horizontal coordinate encoding and vertical coordinate encoding, and spatial attention weights are generated by learning the importance of spatial location through convolutional layers. Multiply the channel attention weights by the spatial attention weights to obtain a two-dimensional attention map. Multiply the attention map element-wise with the original feature map to achieve feature weighting.

8. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The ranging and positioning module includes: The timing control unit reads the laser emission time and camera gating time corresponding to the image frame where the detected power line target is located. The time interval between the two times is calculated to be equal to the flight time of the laser to and from the target. The target distance is equal to the speed of light multiplied by the time interval and then divided by 2. The angle between the target ray and the optical axis is calculated based on the target's position in the image. The actual optical path is equal to the optical axis distance divided by the cosine of the angle. A ranging error correction model is established as a mapping function from pixel position to ranging deviation. The correction amount is obtained by querying the correction model based on the target pixel position. The accurate distance is obtained by adding the correction amount to the coarse distance.

9. The low-altitude aerial vehicle power line obstacle detection system based on gated imaging according to claim 1, wherein, The obstacle avoidance decision module includes: Construct an obstacle avoidance trajectory planning model based on reinforcement learning. Define the state space to include the aircraft's current position, velocity, attitude, the position and risk level of the detected power line target, and the destination direction and distance information. Define the action space to include the heading angle adjustment, pitch angle adjustment, velocity adjustment and continuous control variables. The reward function is designed to consider obstacle avoidance safety, trajectory deviation, energy efficiency, and time cost. A deep deterministic policy gradient algorithm is adopted, and a dynamic window method is introduced as a safety constraint layer to verify the feasibility of the actions output by reinforcement learning. Within the dynamic window, forward simulation is performed on each candidate velocity to evaluate the collision risk between the trajectory and obstacles, the target proximity, and the speed smoothness.

10. A low-altitude aircraft power line obstacle avoidance detection system based on gated imaging according to claim 1, characterized in that, It also includes an image restoration module, which includes: Acquire information from ambient light sensor, temperature and humidity sensor and image statistical features, extract global average gray value, contrast, gradient intensity and edge density statistical features from distance-gated image sequence, construct meteorological recognition neural network to output the probability distribution of each meteorological category; The physical model constraints are embedded in the neural network architecture. The network input is a sequence of multi-distance slice images, and the output is an atmospheric transmittance map and extinction coefficient distribution at each distance. A physical constraint layer is embedded in the network to perform physical consistency verification of the transmittance predicted by the network based on the atmospheric scattering model.

Images