An unmanned aerial vehicle sampling landing positioning method based on a staple tracking algorithm

Abstract

The application discloses a kind of unmanned plane sampling landing positioning method based on Staple tracking algorithm, by abandoning single vision-inertial integrated navigation scheme, Staple target tracking algorithm is combined with binocular vision three-dimensional reconstruction, the long-term running drift error problem of traditional integrated navigation is solved, while grey image, pure color sampling target optimization Staple position calculation logic is aimed at, weighted centroid is replaced by single point maximum value, target positioning accuracy is significantly improved, can satisfy the high-precision sampling landing requirement of unmanned plane from 2m height, by laser height data pre-correction target scale, stage reinitialization tracking model, dynamically adjust ORB algorithm parameter and image cropping strategy, effectively deal with scale, illumination interference, while increasing right eye coordinate deviation check, three-dimensional height tail filter, coordinate smoothing and multiple robustness design, substantially reduce the tracking failure risk caused by image degradation, feature missing.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and specifically to a UAV sampling landing and positioning method based on the Staple tracking algorithm. Background Technology

[0002] Currently, the mainstream positioning scheme for UAVs and deep space exploration vehicles is the vision-inertial integrated navigation method. This scheme tightly couples two types of sensors: inertial measurement unit (IMU) and vision camera. The IMU outputs angular velocity and acceleration data at high frequency and calculates the motion state of the device through integration. It has high positioning accuracy in the short term, but the error will continue to accumulate over time. The vision system relies on the principle of multi-view geometry to extract features such as corner points of the image to calculate the relative pose. There is no cumulative error, but the data output frequency is low and it is easily affected by changes in lighting and scale. Existing vision-inertial integrated navigation fuses data from two types of sensors through frameworks such as extended Kalman filtering and nonlinear optimization. It uses IMU to compensate for insufficient visual frame rate and image blurring, and relies on vision to suppress IMU integral drift, thus outputting six-degree-of-freedom continuous pose. However, after long-term operation, this scheme still produces significant drift errors in heading angle and position, which cannot meet the requirements of high-precision sampling and landing. In sampling missions such as those involving Mars spacecraft, UAVs are required to land precisely from a height of 2 meters above the sample. The drift error of traditional vision-inertial navigation systems can directly cause the landing point to deviate from the target sampling area, resulting in the failure of the rock sampling mission.

[0003] Therefore, it is necessary to invent a UAV sampling landing and positioning method based on the Staple tracking algorithm to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a UAV sampling landing and positioning method based on the Staple tracking algorithm to solve the problems in the above-mentioned technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a UAV sampling landing and positioning method based on the Staple tracking algorithm, comprising the following steps: S1. Data Subscription and Acquisition: The UAV flies to a height of 2m above the sample to be sampled, obtains the two-dimensional coordinates of the sample in the coordinate system of the left eye camera image, and simultaneously subscribes to the data nodes of the binocular camera, navigation module and laser altimeter, and continuously receives binocular images, navigation pose and laser altitude data to complete the synchronous reception of multi-source data. S2. Staple algorithm initialization in the first frame: Staple algorithm initialization is completed based on the first frame left eye image, sample center point and initial scale of 80×80 pixels, and position estimation, scale estimation and model update operations are performed in sequence. S3. ORB feature matching of the first frame binocular image: crop the local area of ​​the left and right eye images of the first frame, use the ORB algorithm to extract and match key points to obtain an initial set of key point pairs; S4. Right eye sample coordinate calculation: The RHO algorithm is used in combination with the key point pair set to solve the homography matrix between the left and right eye images. The sample coordinates in the first frame right eye image are obtained by mapping the homography matrix. If the number of effective feature points is less than 4, the fault is reported to the flight control module and the drone is controlled to land directly. S5. Key point pair screening: The truncation method is used to screen the matching key point pairs and retain the 12 key point pairs that are closest to the center point of the sample. S6. 3D Coordinate Calculation: Distortion and triangulation are performed on the selected key points to obtain the 3D coordinates of the key points. Then, height outliers are removed by the truncated mean method, and the average height is calculated. Combined with the average height of the in-camera participants, the 2D coordinates of the sample are back-projected to the camera coordinate system to obtain the 3D coordinates of the sample. S7. Coordinate system transformation: Combining the installation matrix of the camera and IMU and the UAV navigation pose data, the three-dimensional coordinates of the sample in the camera coordinate system are transformed into three-dimensional coordinates in the takeoff airport coordinate system. S8. Subsequent Frame Scale Pre-correction: For each subsequent frame, the target scale of the Staple algorithm is pre-corrected using the laser height data of the current frame and the previous frame. S9. Frame-by-frame target tracking and dynamic model processing: Staple target tracking is performed using the corrected scale. When the laser height drops to 1.1m and 0.7m for the first time, the Staple model is re-initialized at a scale of 120×120 pixels. The position of the right eye sample is inferred by combining the motion vectors of the left and right eye images. The Staple model update and binocular feature matching are completed in parallel through multi-threading. The image cropping range and ORB algorithm parameters are dynamically adjusted according to the number of feature points. S10, Robustness Verification and Landing Control: The calculated coordinates of the right eye sample are checked for deviation. The key point screening, 3D calculation, coordinate system transformation and coordinate smoothing process are repeated. When the sample height drops to 0.2m, or the number of valid feature points in 3 frames is less than 10, the positioning calculation is terminated and the UAV is controlled to switch to landing mode to complete the landing.

[0006] Preferably, the position estimation step of the Staple algorithm in S2 specifically includes: The search area is delineated and scaled based on the historical target location and scale. HOG features of the search area are extracted, and a template score map is calculated by combining Fourier transform and filter template. Simultaneously, a pixel-level probability map is calculated based on the foreground and background color histograms to generate a color score map. The template score map and the color score map are weighted and fused with fixed weights of 0.7 and 0.3 respectively to obtain the final confidence map. The region containing the maximum value of the confidence map is selected, and the weighted centroid of pixels within an 11×11 pixel range with a response value not lower than 90% of the maximum value is calculated as the target center of the current frame. If the maximum confidence value is less than 0.2, tracking failure is determined, the flight control is reported, and a go-around and resampling are performed.

[0007] Preferably, the scaling step of the Staple algorithm in S2 specifically includes: A scale pyramid is constructed based on the current target center and historical scales, with the scale level set to 33 and the scale factor step size set to 1.02. HOG features are extracted from image patches of different scales to form a scale-feature matrix. The response values ​​of each scale are calculated by combining the scale filter and Fourier transform, and the scale corresponding to the maximum response value is selected as the current target scale.

[0008] Preferably, the model update step of the Staple algorithm in S2 is as follows: The HOG template filter, color histogram model, and scale filter are updated online using an exponential moving average method, with the template learning rate set to 0.01, and the color model and scale learning rates both set to 0.04. During the algorithm initialization phase, the learning rates for all three types are set to 1, and the initial filter model is generated completely from the first frame of data.

[0009] Preferably, step S3 specifically involves: cropping a 250×250 pixel region centered on the sample center in the left eye image of the first frame and a 300×300 pixel region centered on the sample center in the right eye image; performing ORB binary feature matching on the two cropped images; restoring the matched key points to the original image coordinate system; and obtaining a set of key point pairs.

[0010] Preferably, the S8 mesoscale pre-correction formula is: s′t 1=st 1×htht 1, where st 1 represents the target scale of the previous frame, ht 1. ht represents the laser height data of the previous frame and the current frame, respectively, and s′t represents the laser height data of the current frame. 1 represents the corrected target scale.

[0011] Preferably, the rule for dynamically adjusting the ORB algorithm and image cropping in S9 is as follows: When the number of keypoint pairs is less than 10 for the first time, the FAST threshold of the ORB algorithm is lowered to increase the number of corner points; when the number of keypoint pairs is less than 10 for the second time, the cropping area of ​​the binocular image is uniformly adjusted to 360×360 pixels, and the cropped image is downsampled by 4 times; if the number of keypoint pairs is less than 4, the 3D coordinate calculation of the current frame is skipped, and the next frame image is processed directly.

[0012] Preferably, the robustness verification method for the right eye sample coordinates in S10 is as follows: Calculate the Euclidean distance between the right eye sample coordinates obtained from the homography matrix solution and the inferred coordinates of the motion vector. If this distance is greater than 25 / ht pixels (ht is the laser height in the current frame), then use the inferred coordinates instead of the coordinates obtained from the homography matrix solution.

[0013] Preferably, the truncated mean method processing rule in S6 is as follows: The height data of all matching key points are extracted to form a set. The highest 20% and lowest 20% of the height values ​​in the set are removed. The average value of the remaining heights is calculated as the estimated height of the sample. If no valid height data is available after processing, a fault is reported and the drone is controlled to land directly.

[0014] Preferably, in step S10, the exponential moving average method is used to smooth the three-dimensional coordinates of the sample in the takeoff airport coordinate system. The smoothing calculation formula is: Pwt=α Pwo(t)+(1 α) Pwt 1, where the smoothing coefficient α is 0.7, Pwo(t) is the original solution coordinate of the current frame, and Pwt 1 represents the smoothed coordinates of the previous frame, and Pwt represents the final smoothed coordinates of the current frame.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. Positioning accuracy has been greatly improved. This invention abandons the single vision-inertial integrated navigation scheme and combines the Staple target tracking algorithm with binocular vision 3D reconstruction, which solves the drift error problem that has plagued traditional integrated navigation for a long time. At the same time, it optimizes the Staple position calculation logic for grayscale images and solid color sampled targets, and uses weighted centroids to replace the maximum value of a single point, which significantly improves the target positioning accuracy and can meet the high-precision sampling and landing requirements of UAVs starting from a height of 2m.

[0016] 2. Strong resistance to interference and environmental adaptability To address the issues of drastic changes in target scale, dimming of lighting, and decreased image contrast during the descent of a drone, this invention effectively addresses scale and lighting interference by pre-correcting the target scale using laser altitude data, re-initializing the tracking model in stages, and dynamically adjusting ORB algorithm parameters and image cropping strategies. Simultaneously, it incorporates multiple robust designs, including right eye coordinate deviation verification, 3D height truncation filtering, and coordinate smoothing, significantly reducing the risk of tracking failure caused by image degradation and feature loss.

[0017] 3. Excellent real-time performance and high operating efficiency. This invention uses Fourier transform to accelerate the Staple algorithm's related operations, and uses multi-threaded parallel execution of model update and feature matching tasks. While ensuring computational accuracy, it improves the overall algorithm's frame rate and meets the real-time requirements of the UAV's dynamic landing process.

[0018] 4. Wide range of applications This invention can be used not only for ground-based industrial and agricultural sampling drones, but also for rock sampling operations of extraterrestrial deep space exploration equipment such as Mars spacecraft, expanding the application scenarios of computer vision and drone positioning technology. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the sample detection and positioning process of the present invention; Figure 2 This is a schematic diagram of the overall process flow of the present invention. Detailed Implementation

[0020] 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.

[0021] This invention provides, for example Figure 1-2 The method for UAV sampling landing and localization based on the Staple tracking algorithm, as shown, includes the following steps: S1. Data Subscription and Acquisition: The UAV flies to a height of 2m above the sample to be sampled, obtains the two-dimensional coordinates of the sample in the coordinate system of the left eye camera image, and simultaneously subscribes to the data nodes of the binocular camera, navigation module and laser altimeter, and continuously receives binocular images, navigation pose and laser altitude data to complete the synchronous reception of multi-source data. S2. Staple algorithm initialization in the first frame: Staple algorithm initialization is completed based on the first frame left eye image, sample center point and initial scale of 80×80 pixels, and position estimation, scale estimation and model update operations are performed in sequence. S3. ORB feature matching of the first frame binocular image: crop the local area of ​​the left and right eye images of the first frame, use the ORB algorithm to extract and match key points to obtain an initial set of key point pairs; S4. Right eye sample coordinate calculation: The RHO algorithm is used in combination with the key point pair set to solve the homography matrix between the left and right eye images. The sample coordinates in the first frame right eye image are obtained by mapping the homography matrix. If the number of effective feature points is less than 4, the fault is reported to the flight control module and the drone is controlled to land directly. S5. Key point pair screening: The truncation method is used to screen the matching key point pairs and retain the 12 key point pairs that are closest to the center point of the sample. S6. 3D Coordinate Calculation: Distortion and triangulation are performed on the selected key points to obtain the 3D coordinates of the key points. Then, height outliers are removed by the truncated mean method, and the average height is calculated. Combined with the average height of the in-camera participants, the 2D coordinates of the sample are back-projected to the camera coordinate system to obtain the 3D coordinates of the sample. S7. Coordinate system transformation: Combining the installation matrix of the camera and IMU and the UAV navigation pose data, the three-dimensional coordinates of the sample in the camera coordinate system are transformed into three-dimensional coordinates in the takeoff airport coordinate system. S8. Subsequent Frame Scale Pre-correction: For each subsequent frame, the target scale of the Staple algorithm is pre-corrected using the laser height data of the current frame and the previous frame. S9. Frame-by-frame target tracking and dynamic model processing: Staple target tracking is performed using the corrected scale. When the laser height drops to 1.1m and 0.7m for the first time, the Staple model is re-initialized at a scale of 120×120 pixels. The position of the right eye sample is inferred by combining the motion vectors of the left and right eye images. The Staple model update and binocular feature matching are completed in parallel through multi-threading. The image cropping range and ORB algorithm parameters are dynamically adjusted according to the number of feature points. S10, Robustness Verification and Landing Control: The calculated coordinates of the right eye sample are checked for deviation. The key point screening, 3D calculation, coordinate system transformation and coordinate smoothing process are repeated. When the sample height drops to 0.2m, or the number of valid feature points in 3 frames is less than 10, the positioning calculation is terminated and the UAV is controlled to switch to landing mode to complete the landing.

[0022] The position estimation step of the Staple algorithm in S2 is as follows: The search area is delineated and scaled based on the historical target position and scale; HOG features of the search area are extracted; a template score map is calculated by combining Fourier transform and filter template; simultaneously, a pixel-level probability map is calculated based on the foreground and background color histograms to generate a color score map; the template score map and the color score map are weighted and fused with fixed weights of 0.7 and 0.3 to obtain the final confidence map; the region containing the maximum value of the confidence map is selected, and the weighted centroids of pixels within an 11×11 pixel range with response values ​​not lower than 90% of the maximum value are calculated as the target center of the current frame; if the maximum confidence value is less than 0.2, tracking failure is determined, the flight control is reported, and a re-flight and secondary sampling are performed. The scale estimation step of the Staple algorithm in S2 is as follows: A scale pyramid is constructed based on the current target center and historical scale, with the scale level set to 33 and the scale factor step size set to 1.02; different scale image patches are processed. HOG features are extracted to construct a scale-feature matrix. The scale filter and Fourier transform are combined to calculate the response values ​​at each scale. The scale corresponding to the maximum response value is selected as the current target scale. Specifically, the model update step of the Staple algorithm in S2 involves: online updates of the HOG template filter, color histogram model, and scale filter using an exponential moving average method, with the template learning rate set to 0.01, and the color model and scale learning rates both set to 0.04. During algorithm initialization, the learning rates for all three types are set to 1, and the initial filter model is generated completely from the first frame data. Specifically, in S3, a 250×250 pixel region centered on the sample center is cropped from the left eye image of the first frame, and a 300×300 pixel region centered on the sample center is cropped from the right eye image. ORB binary feature matching is performed on the two cropped images, and the matched keypoints are restored to the original image coordinate system to obtain a set of keypoint pairs. The scale pre-correction formula in S8 is: s′t 1=st 1×htht 1, where st 1 represents the target scale of the previous frame, ht 1. ht represents the laser height data of the previous frame and the current frame, respectively, and s′t represents the laser height data of the current frame. 1 represents the corrected target scale. The rules for dynamically adjusting the ORB algorithm and image cropping in S9 are as follows: When the number of keypoint pairs is less than 10 for the first time, the FAST threshold of the ORB algorithm is lowered to increase the number of corner points; when the number of keypoint pairs is less than 10 for the second time, the cropping area of ​​the binocular image is uniformly adjusted to 360×360 pixels, and the cropped image is downsampled by 4 times; if the number of keypoint pairs is less than 4, the calculation of the three-dimensional coordinates of the current frame is skipped, and the next frame image is processed directly. The robustness verification method of the right eye sample coordinates in S10 is: calculate the Euclidean distance between the right eye sample coordinates calculated by the homography matrix and the motion vector inferred coordinates. If the distance is greater than 25... For pixels / ht (where ht is the laser height in the current frame), inferred coordinates are used instead of the homography matrix to calculate the coordinates. The truncated mean method in S6 is as follows: extract the height data of all matching key points to form a set, remove the highest 20% and lowest 20% of height values ​​from the set, and calculate the average of the remaining heights as the estimated height of the sample. If no valid height data is available after processing, a fault is reported and the UAV is controlled to land directly. In S10, the exponential moving average method is used to smooth the three-dimensional coordinates of the sample in the takeoff airport coordinate system. The smoothing calculation formula is: Pwt = α Pwo(t)+(1 α) Pwt 1, where the smoothing coefficient α is 0.7, Pwo(t) is the original solution coordinate of the current frame, and Pwt 1 represents the smoothed coordinates of the previous frame, and Pwt represents the final smoothed coordinates of the current frame; Specifically, this embodiment describes a Mars spacecraft rock sampling and landing scenario. The UAV initially hovers 2m directly above the rock sample, equipped with a binocular camera, IMU, and laser altimeter. The method of this invention is used to complete the entire positioning and landing process. The specific steps are as follows: The UAV hovers 2m above the target, acquires the two-dimensional coordinates of the rock in the left-eye image, and simultaneously connects the data links of the binocular camera, navigation, and laser altimeter to complete data synchronization. The first frame initializes the Staple algorithm at an 80×80 pixel scale, sequentially performing position, scale estimation, and model update. The calculated maximum confidence score is 0.45 (greater than 0.2). The weighted centroid is calculated using effective pixels within an 11×11 region to determine the center of the rock target. The left and right eye images are cropped, and the ORB algorithm is run to complete feature matching, obtaining 28 keypoint pairs. The RHO algorithm is used to solve the homography matrix, resulting in 22 sets of inliers (greater than 4). The right-eye rock coordinates are calculated normally, and the distance between the inliers and the sample center is calculated. The 12 keypoint pairs with the smallest distances are selected. Distortion removal and triangulation were performed to obtain the 3D coordinates. After removing the top 20% and bottom 20% of the data in terms of height, the average height was calculated. Combined with the camera intrinsic parameters and back-projection, the 3D coordinates of the rock in the camera coordinate system were obtained. Combined with the camera-IMU mounting matrix and UAV navigation data, the coordinate system was transformed to obtain the rock coordinates in the world coordinate system. The UAV began to descend. In each frame, the scale of the Staple target was corrected using the laser height of the previous and next frames. When the laser height first dropped to 1.1m and 0.7m, the Staple model was reinitialized at a scale of 120×120 pixels. The position of the right eye was calculated based on the left eye motion vector. Multi-threaded model updates and feature matching were completed. In the later stages of descent, the lighting weakened, and the number of key points was insufficient. The FAST threshold was lowered, the clipping area was expanded, and downsampling was performed in sequence to ensure that feature matching was normal. The right eye coordinate deviation was checked, and the 3D coordinates were smoothed using an exponential moving average. When the laser height dropped to 0.2m, the positioning calculation was stopped, the UAV switched to landing mode, and landed accurately on the rock sampling area, completing the mission.

[0023] Working principle of this invention: Refer to the instruction manual appendix Figure 1-2When using this invention, the method first uses the improved Staple target tracking algorithm as the core, combining binocular vision 3D reconstruction, multi-sensor data fusion, dynamic strategy control, and multi-level coordinate transformation technology to carry out UAV sampling, landing, and positioning. After the UAV arrives at a position 2 meters above the sampling target, it simultaneously collects multi-source data such as binocular images, navigation pose, and laser altitude. Staple algorithm initialization is then completed. This algorithm calculates the target position by fusing HOG features and color features, and achieves adaptive target scale estimation based on the scale pyramid. Simultaneously, an exponential moving average strategy is used to iteratively update the model. Furthermore, weighted centroid positioning is used to improve tracking accuracy for grayscale images and solid-color targets. A re-entry mechanism can be triggered when tracking fails. Subsequently, the ORB algorithm is used to complete binocular image feature matching, and the RHO algorithm is used to solve the homography matrix, combined with threshold verification to suppress errors. Finally, distortion correction is performed. The system calculates the target's 3D coordinates using triangulation and truncated mean filtering. It then uses the camera and IMU mounting matrix and navigation pose to transform the camera coordinate system to the takeoff airport coordinate system. The coordinate data is smoothed using exponential moving average. During the UAV's descent, the tracking scale is corrected in real-time using laser altitude data. The Staple model is reinitialized at designated altitude nodes to prevent the target from overflowing the frame. Simultaneously, the ORB algorithm parameters and image cropping range are dynamically adjusted to address issues such as dimming lighting and feature reduction. Multi-threaded operation ensures the algorithm's real-time performance. The system continuously loops through the positioning process and guides the UAV towards the target. When the target altitude drops to 0.2 meters or the accumulated feature count is insufficient, positioning is terminated and the UAV is controlled to land. This overall solution effectively overcomes the drift defects of traditional vision-inertial navigation systems, enabling high-precision sampling and landing of UAVs in environments with varying scales and lighting conditions.

[0024] 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 sampling landing and localization method for unmanned aerial vehicles (UAVs) based on the Staple tracking algorithm, characterized in that, Includes the following steps: S1. Data Subscription and Acquisition: The UAV flies to a height of 2m above the sample to be sampled, obtains the two-dimensional coordinates of the sample in the coordinate system of the left eye camera image, and simultaneously subscribes to the data nodes of the binocular camera, navigation module and laser altimeter, and continuously receives binocular images, navigation pose and laser altitude data to complete the synchronous reception of multi-source data. S2. Staple algorithm initialization in the first frame: Staple algorithm initialization is completed based on the first frame left eye image, sample center point and initial scale of 80×80 pixels, and position estimation, scale estimation and model update operations are performed in sequence. S3. ORB feature matching of the first frame binocular image: crop the local area of ​​the left and right eye images of the first frame, use the ORB algorithm to extract and match key points to obtain an initial set of key point pairs; S4. Right eye sample coordinate calculation: The RHO algorithm is used in combination with the key point pair set to solve the homography matrix between the left and right eye images. The sample coordinates in the first frame right eye image are obtained by mapping the homography matrix. If the number of effective feature points is less than 4, the fault is reported to the flight control module and the drone is controlled to land directly. S5. Key point pair screening: The truncation method is used to screen the matching key point pairs and retain the 12 key point pairs that are closest to the center point of the sample. S6. 3D Coordinate Calculation: Distortion and triangulation are performed on the selected key points to obtain the 3D coordinates of the key points. Then, height outliers are removed by the truncated mean method, and the average height is calculated. Combined with the average height of the in-camera participants, the 2D coordinates of the sample are back-projected to the camera coordinate system to obtain the 3D coordinates of the sample. S7. Coordinate system transformation: Combining the installation matrix of the camera and IMU and the UAV navigation pose data, the three-dimensional coordinates of the sample in the camera coordinate system are transformed into three-dimensional coordinates in the takeoff airport coordinate system. S8. Subsequent Frame Scale Pre-correction: For each subsequent frame, the target scale of the Staple algorithm is pre-corrected using the laser height data of the current frame and the previous frame. S9. Frame-by-frame target tracking and dynamic model processing: Staple target tracking is performed using the corrected scale. When the laser height drops to 1.1m and 0.7m for the first time, the Staple model is re-initialized at a scale of 120×120 pixels. The position of the right eye sample is inferred by combining the motion vectors of the left and right eye images. The Staple model update and binocular feature matching are completed in parallel through multi-threading. The image cropping range and ORB algorithm parameters are dynamically adjusted according to the number of feature points. S10, Robustness Verification and Landing Control: The calculated coordinates of the right eye sample are checked for deviation. The key point screening, 3D calculation, coordinate system transformation and coordinate smoothing process are repeated. When the sample height drops to 0.2m, or the number of valid feature points in 3 frames is less than 10, the positioning calculation is terminated and the UAV is controlled to switch to landing mode to complete the landing.

2. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The position estimation steps of the Staple algorithm in S2 are as follows: The search area is delineated and scaled based on the historical target location and scale. HOG features of the search area are extracted, and a template score map is calculated by combining Fourier transform and filter template. Simultaneously, a pixel-level probability map is calculated based on the foreground and background color histograms to generate a color score map. The template score map and the color score map are weighted and fused with fixed weights of 0.7 and 0.3 respectively to obtain the final confidence map. The region containing the maximum value of the confidence map is selected, and the weighted centroid of pixels within an 11×11 pixel range with a response value not lower than 90% of the maximum value is calculated as the target center of the current frame. If the maximum confidence value is less than 0.2, tracking failure is determined, the flight control is reported, and a go-around and resampling are performed.

3. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The specific scaling steps of the Staple algorithm in S2 are as follows: A scale pyramid is constructed based on the current target center and historical scales, with the scale level set to 33 and the scale factor step size set to 1.

02. HOG features are extracted from image patches of different scales to form a scale-feature matrix. The response values ​​of each scale are calculated by combining the scale filter and Fourier transform, and the scale corresponding to the maximum response value is selected as the current target scale.

4. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The specific model update steps of the Staple algorithm in S2 are as follows: The HOG template filter, color histogram model, and scale filter are updated online using an exponential moving average method, with the template learning rate set to 0.01, and the color model and scale learning rates both set to 0.

04. During the algorithm initialization phase, the learning rates for all three types are set to 1, and the initial filter model is generated completely from the first frame of data.

5. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: Specifically, S3 involves cropping a 250×250 pixel region centered on the sample center in the left eye image of the first frame and a 300×300 pixel region centered on the sample center in the right eye image. ORB binary feature matching is then performed on the two cropped images, and the matched key points are restored to the original image coordinate system to obtain a set of key point pairs.

6. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The S8 mesoscale pre-correction formula is: s′t 1=st 1×htht 1, where st 1 represents the target scale of the previous frame, ht 1. ht represents the laser height data of the previous frame and the current frame, respectively, and s′t represents the laser height data of the current frame. 1 represents the corrected target scale.

7. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The rules for dynamically adjusting the ORB algorithm and image cropping in S9 are as follows: When the number of keypoint pairs is less than 10 for the first time, the FAST threshold of the ORB algorithm is lowered to increase the number of corner points; when the number of keypoint pairs is less than 10 for the second time, the cropping area of ​​the binocular image is uniformly adjusted to 360×360 pixels, and the cropped image is downsampled by 4 times; if the number of keypoint pairs is less than 4, the 3D coordinate calculation of the current frame is skipped, and the next frame image is processed directly.

8. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The robustness verification method for the right eye sample coordinates in S10 is as follows: Calculate the Euclidean distance between the right eye sample coordinates obtained from the homography matrix solution and the inferred coordinates of the motion vector. If this distance is greater than 25 / ht pixels (ht is the laser height in the current frame), then use the inferred coordinates instead of the coordinates obtained from the homography matrix solution.

9. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: The processing rule for the truncated mean method in S6 is as follows: The height data of all matching key points are extracted to form a set. The highest 20% and lowest 20% of the height values ​​in the set are removed. The average value of the remaining heights is calculated as the estimated height of the sample. If no valid height data is available after processing, a fault is reported and the drone is controlled to land directly.

10. The UAV sampling landing and positioning method based on the Staple tracking algorithm according to claim 1, characterized in that: In step S10, the exponential moving average method is used to smooth the three-dimensional coordinates of the sample in the takeoff airport coordinate system. The smoothing calculation formula is: Pwt=α Pwo(t)+(1 α) Pwt 1, where the smoothing coefficient α is 0.7, Pwo(t) is the original solution coordinate of the current frame, and Pwt 1 represents the smoothed coordinates of the previous frame, and Pwt represents the final smoothed coordinates of the current frame.

Images