An optical flow-based low-altitude speed measurement lightweight method for aircrafts

Abstract

The application discloses a kind of low-altitude speed measurement lightweight methods of aircraft based on optical flow, belong to computer vision and aircraft navigation control technical field, the method is collected ground image by downward-looking monocular camera, after down-sampling, bilinear interpolation scale normalization, in the image center region uniform extraction fixed size feature block, using normalized cross correlation (NCC) block matching search is carried out, after weighted fusion and outlier rejection, it is obtained robust pixel displacement, combined with pinhole camera model, height above ground level and IMU attitude data solution and compensation horizontal velocity, while establishing speed-height feasible region constraint, adaptive adjustment block matching search range, the present application is extremely low in calculation, without initialization, low-texture scene robustness is strong, it can be realized ≥10Hz Real-time operation in embedded platform, it is suitable for aircraft 0.2m-1.5m ultra-low altitude take-off stage horizontal velocity accurate estimation.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and specifically to a lightweight method for low-altitude speed measurement of aircraft based on optical flow. Background Technology

[0002] Currently, the methods widely used in the field of visual velocity measurement of aircraft mainly fall into the following three categories.

[0003] Optical flow velocity measurement modules in flight control systems, such as the PX4 flight control system, estimate velocity based on dedicated optical flow sensors or downward-looking cameras. A global pixel displacement estimation strategy with fixed parameters is employed to calculate the average pixel motion over the entire image or a fixed area. Simple rotation compensation is then performed using angular velocity measurements from a gyroscope to remove pixel motion components caused by the aircraft's own rotation. Finally, the pixel displacement is converted into linear velocity using altitude information provided by ultrasonic or laser rangefinders.

[0004] VIO methods, such as the optimized VINS class and the filter-based MSCKF method.

[0005] The former technical process is as follows: First, corner features are extracted from the image using detection operators such as FAST or ORB, and inter-frame feature correspondences are established through optical flow tracing or descriptor matching. Then, acceleration and angular velocity are integrated using an IMU pre-integration model to obtain inter-frame relative motion constraints. In the backend, a sliding window nonlinear optimization framework is used to jointly construct a cost function from reprojection errors and IMU pre-integration residuals. The Gauss-Newton or Levenberg-Marquardt algorithm iteratively solves for state variables such as camera pose, IMU bias, and 3D coordinates of feature points. To control the computational scale, older frames outside the window are marginalized, and their information is compressed into prior constraints. The entire system also requires a dedicated initialization process to estimate gravity direction, scale factor, and IMU bias.

[0006] The latter differs from the former by maintaining multiple camera pose states (rather than the 3D coordinates of feature points) within a sliding window in the extended Kalman filter's state vector. It utilizes an IMU motion model as the filter's prediction step, integrating the IMU-measured acceleration and angular velocity through continuous-time motion equations to predict the current state and its covariance matrix. When a new image frame arrives, multi-state geometric constraints are established for the tracked feature points across multiple camera poses. By marginalizing and eliminating the feature point coordinates from the observation equations, a residual equation related only to the camera pose is constructed as the filter's update step. In the update phase, MSCKF performs triangulation on each feature point to obtain an initial estimate of its 3D coordinates. Subsequently, it calculates the Jacobian matrix and eliminates the feature point parameters from the observation model through QR decomposition or null projection, ultimately obtaining a compressed observation residual used for EKF state updates. When the number of camera poses in the sliding window exceeds a set limit, the oldest camera state needs to be pruned, and the covariance matrix updated accordingly.

[0007] One drawback of the VIO method is its extremely high computational cost and resource consumption on the chip. Optimized VINS systems require multiple computationally intensive steps, including feature point extraction, descriptor computation or optical flow tracking, IMU pre-integration, nonlinear graph optimization, and edge detection, with single-frame processing latency typically in the tens to hundreds of milliseconds range. While filter-based MSCKF avoids nonlinear iterative optimization, it still requires feature point extraction and tracking, multi-state Jacobian matrix computation, QR decomposition or null projection, and maintenance and updating of the high-dimensional covariance matrix. As the number of camera states and tracked features within the sliding window increases, the matrix computation load of the EKF update step also increases significantly. Both methods have high processor computing power requirements and resource consumption, making it difficult to achieve high-frequency (≥10Hz) real-time operation simultaneously with other complex tasks on the embedded computing platforms commonly used in aircraft.

[0008] Another drawback of VIO is the complexity and time-consuming initialization process. Most VINS methods require a dedicated initialization phase to estimate parameters such as gravity direction, scale factor, and IMU bias. Similarly, MSCKF requires accumulating sufficient visual-inertial observations during the startup phase to complete state initialization and covariance matrix convergence. This process typically requires sufficient motion excitation from the aircraft, taking several seconds or even longer. It cannot complete initialization and provide reliable velocity estimates in time within the 0.2m to 1.5m altitude window at takeoff. Both methods rely on feature point extraction and tracking at the front end. When the ground texture at ultra-low altitudes is simple, feature points are sparse or unevenly distributed, making feature extraction and tracking prone to failure, leading to system degradation or even complete loss. Furthermore, the most crucial information during takeoff is horizontal velocity information. Using these two large-scale six-degree-of-freedom methods to solve the velocity estimation problem results in significant computational waste and redundant system complexity.

[0009] The PX4 optical flow module is primarily used in vertical takeoff procedures. It employs global average pixel displacement estimation, which has limited matching accuracy in scenarios with uneven ground texture complexity. Its search strategy is simple, lacking the ability to dynamically adjust the search range based on flight speed and altitude; its attitude compensation mechanism is also rudimentary, relying solely on gyroscope angular velocity for rotational compensation, failing to fully utilize extrinsic parameter calibration information between the IMU and camera to correct for changes in effective observation altitude caused by pitch and roll angles; and it lacks theoretical feasible domain analysis of horizontal speed, altitude, and search range, thus failing to provide clear speed-altitude constraints for the flight control system's takeoff strategy.

[0010] Therefore, it is necessary to invent a lightweight method for low-altitude speed measurement of aircraft based on optical flow to solve the above problems. Summary of the Invention

[0011] The purpose of this invention is to provide a lightweight method for low-altitude speed measurement of aircraft based on optical flow, so as to solve the problems in the above-mentioned technology.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a lightweight method for low-altitude velocity measurement of aircraft based on optical flow, comprising the following steps: S1. Image downsampling preprocessing; S2, Bilinear interpolation scale normalization; S3. Extraction of uniform feature blocks in the central region; S4, Normalized Cross-Correlation NCC Block Matching Search; S5, NCC weighted pixel displacement estimation and outlier removal; S6. Velocity estimation based on pinhole camera model; S7, IMU attitude compensation; S8. Velocity-height feasible region constraint and adaptive search range adjustment.

[0013] Preferably, a sequence of ground images is continuously acquired using a downward-facing monocular camera, with each adjacent frame designated as a reference frame. and the current frame The original image is downsampled, reducing the resolution from 1280×1080 to something like 640×540.

[0014] Preferably, to ensure that the reference frame and the current frame have consistent scale during matching, bilinear interpolation is performed on the two frames. For any pixel position in the target image... Map back to the original image coordinates = , = ,make , , Then the pixel values ​​of the four neighboring points are: Therefore, the interpolation result is: This ensures that even if there are slight changes in focal length or differences in image scaling between two frames, pixel-level scale consistency can still be maintained.

[0015] Preferably, in the central region of the downsampled image, N rectangular feature blocks of size (21×21) pixels are selected at uniform intervals. The purpose of selecting the central region is to avoid areas with large distortion at the image edges and improve matching reliability. The selection of feature blocks does not depend on any corner point or feature point detection operators (such as FAST, Shi-Tomasi, ORB, etc.), but directly samples uniformly on the image in a fixed grid manner, which fundamentally avoids the computational overhead of feature point detection and the problem of insufficient feature points in low-texture scenes.

[0016] Preferably, the coordinates in the reference frame are template blocks The pixel value , It is a template block The average of all pixel values ​​is then normalized to the cross-correlation value. It can be calculated using the following formula: Traverse all candidate positions within the search range, making The position where the maximum value is obtained is the best matching position for that feature block. , It refers to the pixel displacement at that position.

[0017] Preferably, for the matching results of all N feature blocks, outlier removal is performed first: an NCC threshold is set. Feature block matching results with a maximum NCC value lower than the threshold are removed. The remaining matching results are then analyzed according to the normalized cross-correlation value. The weighted average is taken as the pixel displacement.

[0018] Preferably, using the calibrated IMU-camera extrinsic parameters, the pitch angle and roll angle are introduced to calculate the tilt compensation factor: tilt_factor=cos(pitch)×cos(roll), and the speed correction is completed based on the effective height e_height=height / tilt_factor.

[0019] Preferably, based on the pinhole camera model and the assumption that the camera is mounted vertically downwards, pixel displacement is converted into the actual horizontal velocity of the aircraft. Using the camera's focal length parameter, we can convert the pixel movement into the actual angular change of the camera's field of view, resulting in the displacement conversion from pixel to angle: The angle change itself only indicates how much the camera's field of view has changed. By knowing the time interval between two frames, the angular velocity can be calculated. To obtain the actual speed (linear velocity) of the camera, the height of the camera above the ground is needed. Therefore, the formula for calculating the speed is: The focal length parameter converts pixel coordinates into physical angles, specifically the principal point coordinates. (cx, cy) Calculate the pixel offset relative to the image center, and correct the effect of lens distortion on pixel coordinates using distortion parameters.

[0020] Preferably, the camera observes a flat ground, but in reality, the aircraft will inevitably have a tilted attitude during flight. By calibrating the IMU-camera extrinsic parameters, IMU attitude compensation can be incorporated into the calculation to improve the accuracy of velocity estimation. When the aircraft rotates, feature points fixed on the ground will generate additional motion in the image; the compensation amount at this time is angular velocity × linear velocity × time. When the camera tilts, velocity compensation needs to be made for pitch and roll angles relative to camera height. : is the speed of visual computing, Laser altimeter parameters Pitch angle, : Roll angle, then compensation factor Effective height Speed ​​compensation amount .

[0021] Preferably, to ensure the real-time performance of the velocity measurement method under certain computational resources and the reliability of the aircraft's visual global navigation method, the optical flow velocity measurement method can establish a constraint relationship between horizontal flight speed and flight altitude within a search range of [missing information]. ± Pixels (downsampled image coordinates), corresponding original image search range ± Pixels, inter-frame time interval The camera focal length is Flight altitude is Under these conditions, in order for the search range to cover the actual pixel displacement, the following must be satisfied: That is, the horizontal speed of the aircraft With height The following constraints must be met: This constraint provides a clear basis for the takeoff strategy of the flight control system: at a given altitude, the horizontal speed of the aircraft should not exceed the above-mentioned upper limit; conversely, at a given speed, the aircraft should reach a certain altitude to ensure the reliability of the speed measurement results. When using visual feature tracking methods for navigation, the speed control of the aircraft can also be determined using this speed measurement tool with reference to the search window size. Based on the above constraints, the search range required for the next frame can be dynamically calculated from the velocity estimate obtained in the previous frame and the current altitude. When the aircraft is moving slowly or at a high altitude, the search range is automatically narrowed to reduce computational load; when the speed increases or the altitude decreases, the search range is automatically expanded to ensure a high success rate of matching.

[0022] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. Ultra-low computational overhead, meeting the real-time requirements of embedded systems, and will not compete for resources when running on a spacecraft computing platform with limited resources and other resource-intensive programs.

[0023] 2. No feature point extraction is required, and it is robust in low-texture scenes.

[0024] 3. No initialization process is required; speed can be output instantly upon takeoff.

[0025] 4. It can be used specifically for horizontal velocity measurement during oblique takeoff, and has a complete IMU attitude compensation model to meet special flight conditions.

[0026] 5. Establish speed-altitude feasible domain constraints, adaptively adjust, and assist visual navigation methods in constraining flight control settings for flight speed. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the optical flow velocimetry method of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0029] This invention provides, for example Figure 1 The illustrated lightweight method for low-altitude velocity measurement of aircraft based on optical flow includes the following steps: S1. Image downsampling preprocessing; S2, Bilinear interpolation scale normalization; S3. Extraction of uniform feature blocks in the central region; S4, Normalized Cross-Correlation NCC Block Matching Search; S5, NCC weighted pixel displacement estimation and outlier removal; S6. Velocity estimation based on pinhole camera model; S7, IMU attitude compensation; S8. Velocity-height feasible region constraint and adaptive search range adjustment.

[0030] Specifically, a series of ground images are continuously acquired using a monocular camera installed in a downward-facing configuration, with each adjacent frame designated as a reference frame. and the current frame The original image is downsampled, reducing the resolution from 1280×1080 to, for example, 640×540. To ensure that the reference frame and the current frame have consistent scale during matching, bilinear interpolation is performed on the two images. For any pixel position on the target image... Map back to the original image coordinates = , = ,make , , Then the pixel values ​​of the four neighboring points are: Therefore, the interpolation result is: This ensures pixel-level scale consistency even with minor focal length changes or image scaling differences between two frame captures. In the central region of the downsampled image, N rectangular feature blocks of size (21×21) pixels are selected at uniform intervals. Selecting the central region avoids areas with significant image edge distortion, improving matching reliability. Feature block selection does not rely on any corner points or feature point detection operators (such as FAST, Shi-Tomasi, ORB, etc.), but rather directly samples uniformly on the image using a fixed grid, fundamentally avoiding the computational overhead of feature point detection and the problem of insufficient feature points in low-texture scenes. The coordinates in the reference frame are... template blocks The pixel value , It is a template block The average of all pixel values ​​is then normalized to the cross-correlation value. It can be calculated using the following formula: Traverse all candidate positions within the search range, making The position where the maximum value is obtained is the best matching position for that feature block. , This refers to the pixel displacement at that location. For the matching results of all N feature blocks, outlier removal is first performed: an NCC threshold is set. Feature block matching results with a maximum NCC value lower than the threshold are removed. The remaining matching results are then analyzed according to the normalized cross-correlation value. The weighted average is taken as the pixel displacement. Using the calibrated IMU-camera extrinsic parameters, the pitch angle and roll angle are introduced to calculate the tilt compensation factor: tilt_factor=cos(pitch)×cos(roll), and the speed correction is completed based on the effective height e_height=height / tilt_factor. When the UAV cruises at a constant speed of 5 m / s, the speed error measured by this method is ≤0.2 m / s, and the total time for single-frame speed measurement is ≤4 ms. Compared with the traditional optical flow speed measurement method, the computational complexity is reduced by more than 90%. It can run stably on low-computing-power processors of small UAVs, meet the real-time speed measurement requirements of low-altitude cruise, and still maintain high speed measurement accuracy under slight attitude fluctuations.

[0031] Specifically, based on the pinhole camera model and the assumption that the camera is mounted vertically downwards, pixel displacement is converted into the actual horizontal velocity of the aircraft. Using the camera's focal length parameter, we can convert the pixel movement into the actual angular change of the camera's viewpoint. The resulting displacement conversion from pixel to angle is as follows: The angle change itself only indicates how much the camera's field of view has changed. By knowing the time interval between two frames, the angular velocity can be calculated. To obtain the actual speed (linear velocity) of the camera, the height of the camera above the ground is needed. Therefore, the formula for calculating the speed is: The focal length parameter converts pixel coordinates into physical angles, specifically the principal point coordinates. (cx, cy) The calculation of pixel offset relative to the image center, distortion parameter correction of the effect of lens distortion on pixel coordinates, and the camera's observation of a flat ground, all contribute to improving the accuracy of velocity estimation. While the camera observes a flat ground, the aircraft will inevitably have a tilted attitude during flight. By using calibrated IMU-camera extrinsic parameters, IMU attitude compensation can be incorporated into the calculation. When the aircraft rotates, feature points fixed to the ground will generate additional motion in the image; the compensation amount at this time is angular velocity × linear velocity × time. When the camera tilts, velocity compensation needs to be made for pitch and roll angles relative to camera height. : is the speed of visual computing, Laser altimeter parameters Pitch angle, : Roll angle, then compensation factor Effective height Speed ​​compensation amount .

[0032] To ensure the real-time performance of the velocity measurement method under certain computational resources and the reliability of the aircraft's visual global navigation method, the optical flow velocity measurement method can establish a constraint relationship between horizontal flight speed and flight altitude within a search range. ± Pixels (downsampled image coordinates), corresponding original image search range ± Pixels, inter-frame time interval The camera focal length is Flight altitude is Under these conditions, in order for the search range to cover the actual pixel displacement, the following must be satisfied: That is, the horizontal speed of the aircraft With height The following constraints must be met: This constraint provides a clear basis for the takeoff strategy of the flight control system: at a given altitude, the horizontal speed of the aircraft should not exceed the above upper limit; conversely, at a given speed, the aircraft should reach a certain altitude to ensure the reliability of the speed measurement results. When using visual feature tracking for navigation, the speed control of the aircraft can also be determined by this speed measurement tool with reference to the search window size. Based on the above constraint, the search range required for the next frame can be dynamically calculated from the speed estimate obtained from the previous frame and the current altitude. Specifically, when the aircraft is slow or at a high altitude, the search range is automatically narrowed to reduce the computational load; when the speed increases or the altitude decreases, the search range is automatically expanded to ensure a high success rate of matching. When a fixed-wing aircraft flies at a constant speed of 10 m / s, the speed error measured by this method is ≤0.3 m / s, and the total time for single-frame velocity measurement is ≤5 ms. It can operate stably on the lightweight flight control system of fixed-wing aircraft. It can still maintain stable velocity measurement performance in scenarios with uneven textures, slight changes in lighting, and attitude fluctuations in the field. Compared with traditional optical flow velocity measurement methods, it does not require complex feature point detection and iterative optimization, making it more suitable for low-altitude reconnaissance applications of small fixed-wing aircraft with limited resources.

[0033] Working principle of this invention: Refer to the instruction manual appendix Figure 1 When using this invention, the original camera image has a high resolution (1280×1080), and direct processing would significantly consume the aircraft's embedded computing resources. Therefore, the first step is to reduce the resolution to 640×540 through image downsampling preprocessing. This reduces the amount of data and subsequent computational load while preserving key ground texture information, ensuring real-time performance. To address the slight focal length changes and image scaling differences that may exist between two camera frames, a bilinear interpolation algorithm is introduced to achieve scale normalization. The target image pixels are mapped back to the original image coordinates, and the interpolation result is calculated by weighting the four neighboring pixels to ensure that the pixel-level scale is consistent when the reference frame matches the current frame. This eliminates the matching offset error caused by scale differences from the source. Traditional feature point detection operators such as FAST, ORB, and Shi-Tomasi are abandoned to avoid the pain points of missing feature points and high feature detection computation overhead in low-texture ground scenes. A fixed grid uniform sampling method is adopted, selecting N 21×21 pixels in the center region of the downsampled image. The core advantages of the rectangular feature blocks are: large distortion coefficients in image edge regions, effective avoidance of edge distortion in the central region, improved matching reliability, no feature point detection step, significantly reduced computation, and stable distribution of uniformly sampled feature blocks, adaptable to various low-altitude ground scenes (such as flat roads, grass, and weakly textured sites), ensuring sufficient and stable matching samples. The Normalized Cross-Correlation (NCC) algorithm is used to complete accurate feature block matching. Using the feature block of the reference frame as a template, the candidate positions are traversed within the search range of the current frame, and the NCC correlation value between the template and the candidate region is calculated. The position corresponding to the maximum correlation value is the best matching point. The pixel displacement of a single feature block is obtained through coordinate difference. To further improve displacement accuracy, an NCC threshold of 0.6 is set to remove abnormal results with low matching confidence. The remaining effective matching results are weighted and averaged according to the NCC correlation value to obtain a globally robust pixel displacement value, eliminating displacement deviation caused by local mismatches and small ground disturbances, and achieving accurate pixel-level displacement estimation. Based on the pinhole camera imaging model and the assumption of vertical downward-looking camera installation, this method converts image pixel displacement into the physical horizontal velocity of the aircraft, completing the mapping from visual quantities to physical quantities. First, pixel offset is converted into a change in viewing angle using camera intrinsic parameters (focal length, principal point coordinates). Then, angular velocity is calculated based on the inter-frame time interval. Finally, relying on the aircraft's altitude, the angular velocity is converted into horizontal linear velocity. Specifically, the camera focal length enables the conversion from pixel coordinates to physical angles, the principal point coordinates calibrate the image center offset, and distortion parameters correct inherent lens distortion, ensuring a precise correspondence between pixel displacement and actual physical displacement. During low-altitude flight, the aircraft inevitably experiences pitch and roll attitude tilt, which can cause the camera's downward-looking viewing angle to shift, directly resulting in errors in altitude measurement and velocity calculation. This method calibrates the IMU and camera extrinsic parameters to obtain real-time pitch and roll angles, and calculates the tilt compensation factor tilt_factor=c. The formula os(pitch)×cos(roll) is used to derive the effective height e_height = actual height / tilt compensation factor. Based on the effective height, attitude correction is performed on the visual velocity measurement results to compensate for velocity deviation caused by attitude tilt and eliminate additional motion errors in the image caused by aircraft rotation. This significantly improves the velocity measurement accuracy under tilted flight conditions. To balance real-time performance and velocity measurement reliability, a mathematical constraint relationship between horizontal velocity and flight altitude is established. The correlation formulas for search range, velocity, altitude, focal length, and inter-frame interval are clarified, and the upper limit of the aircraft's horizontal velocity is limited, providing a basis for safe flight strategies for the flight control system. At the same time, based on the velocity of the previous frame and the current altitude, the matching search range for the next frame is dynamically calculated: when the aircraft is flying at low speed and high altitude, the search range is automatically reduced to reduce the amount of computation; when flying at high speed and low altitude, the search range is automatically expanded to ensure the matching success rate. This achieves an adaptive balance between computing power consumption and velocity measurement accuracy, adapting to different flight conditions.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow, characterized in that, Includes the following steps: S1. Image downsampling preprocessing; S2, Bilinear interpolation scale normalization; S3. Extraction of uniform feature blocks in the central region; S4, Normalized Cross-Correlation NCC Block Matching Search; S5, NCC weighted pixel displacement estimation and outlier removal; S6. Velocity estimation based on pinhole camera model; S7, IMU attitude compensation; S8. Velocity-height feasible region constraint and adaptive search range adjustment.

2. The lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: A series of ground images were continuously acquired using a downward-facing monocular camera, with each adjacent frame designated as a reference frame. and the current frame The original image is downsampled, reducing the resolution from 1280×1080 to something like 640×540.

3. The lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: To ensure that the reference frame and the current frame have consistent scale during matching, bilinear interpolation is performed on both images. For any pixel position in the target image... Map back to the original image coordinates = , = ,make , , Then the pixel values ​​of the four neighboring points are: Therefore, the interpolation result is: This ensures that even if there are slight changes in focal length or differences in image scaling between two frames, pixel-level scale consistency can still be maintained.

4. The lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: In the central region of the downsampled image, N rectangular feature blocks of size (21×21) pixels are selected at uniform intervals. The purpose of selecting the central region is to avoid areas with large distortion at the image edges and improve matching reliability. The selection of feature blocks does not depend on any corner point or feature point detection operators (such as FAST, Shi-Tomasi, ORB, etc.), but directly samples uniformly on the image in a fixed grid manner, which fundamentally avoids the computational overhead of feature point detection and the problem of insufficient feature points in low-texture scenes.

5. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: The coordinates in the reference frame are template blocks The pixel value , It is a template block The average of all pixel values ​​is then normalized to the cross-correlation value. It can be calculated using the following formula: Traverse all candidate positions within the search range, making The position where the maximum value is obtained is the best matching position for that feature block. , It refers to the pixel displacement at that position.

6. The lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: For the matching results of all N feature blocks, outlier removal is first performed: an NCC threshold is set. Feature block matching results with a maximum NCC value lower than the threshold are removed. The remaining matching results are then analyzed according to the normalized cross-correlation value. The weighted average is taken as the pixel displacement.

7. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: Using the calibrated IMU-camera extrinsic parameters, the pitch angle and roll angle are introduced to calculate the tilt compensation factor: tilt_factor=cos(pitch)×cos(roll), and the speed correction is completed based on the effective height e_height=height / tilt_factor.

8. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: Based on the pinhole camera model and the assumption that the camera is mounted vertically downwards, pixel displacement is converted into the actual horizontal velocity of the aircraft. Using the camera's focal length parameter, we can convert the pixel movement into the actual angular change of the camera's viewpoint. The resulting displacement conversion from pixel to angle is as follows: The angle change itself only indicates how much the camera's field of view has changed. By knowing the time interval between two frames, the angular velocity can be calculated. To obtain the actual speed (linear velocity) of the camera, the height of the camera above the ground is needed. Therefore, the formula for calculating the speed is: The focal length parameter converts pixel coordinates into physical angles, specifically the principal point coordinates. (cx, cy) Calculate the pixel offset relative to the image center, and correct the effect of lens distortion on pixel coordinates using distortion parameters.

9. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: The camera observes a flat ground, but in reality, the aircraft will inevitably have a tilted attitude during flight. By calibrating the IMU-camera extrinsic parameters, IMU attitude compensation can be incorporated into the calculation to improve the accuracy of velocity estimation. When the aircraft rotates, feature points fixed on the ground will generate additional motion in the image. The compensation amount at this time is angular velocity × linear velocity × time. When the camera tilts, velocity compensation needs to be made for the pitch and roll angles relative to the camera height. : is the speed of visual computing, Laser altimeter parameters Pitch angle, : Roll angle, then compensation factor Effective height Speed ​​compensation amount .

10. A lightweight method for low-altitude velocity measurement of aircraft based on optical flow according to claim 1, characterized in that: To ensure the real-time performance of the velocity measurement method under certain computational resources and the reliability of the aircraft's visual global navigation method, the optical flow velocity measurement method can establish a constraint relationship between horizontal flight speed and flight altitude within a search range. ± Pixels (downsampled image coordinates), corresponding original image search range ± Pixels, inter-frame time interval The camera focal length is Flight altitude is Under these conditions, in order for the search range to cover the actual pixel displacement, the following must be satisfied: That is, the horizontal speed of the aircraft With height The following constraints must be met: This constraint provides a clear basis for the takeoff strategy of the flight control system: at a given altitude, the horizontal speed of the aircraft should not exceed the above-mentioned upper limit; conversely, at a given speed, the aircraft should reach a certain altitude to ensure the reliability of the speed measurement results. When using visual feature tracking methods for navigation, the speed control of the aircraft can also be determined using this speed measurement tool with reference to the search window size. Based on the above constraints, the search range required for the next frame can be dynamically calculated from the velocity estimate obtained in the previous frame and the current altitude. When the aircraft is moving slowly or at a high altitude, the search range is automatically narrowed to reduce computational load; when the speed increases or the altitude decreases, the search range is automatically expanded to ensure a high success rate of matching.

Images