Unmanned aerial vehicle pose estimation method and system based on fixed marker visual inspection
Optimizing pose estimation through deep learning model and erosion point algorithm, the problem of low accuracy and susceptibility to noise interference in traditional methods is solved, and pose estimation with higher accuracy and stability is achieved, suitable for complex scenarios and Drone flight.
Patent Information
- Application Number
- CN202411020512.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-15
AI Technical Summary
Traditional pose estimation algorithms rely on image feature points, have low accuracy and are susceptible to noise interference, especially in areas with poor occlusion, poor lighting and unobtrusive textures, and are difficult to deal with complex scenes.
Using a deep learning-based method, the airport runway is targeted detection using the YOLOv3 model, the runway corner coordinates are extracted, and the pose estimation is optimized through the erosion point algorithm and the orthogonal iteration algorithm to replace the traditional PnP algorithm.
It improves the accuracy and stability of pose estimation, enhances generalization ability, and can accurately extract feature points in complex scenarios, reducing dependence on sensor costs.
Smart Images

Figure CN118967809A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of posture estimation in the field of computer vision, and in particular to a method and system for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers. Background Art
[0002] The pose estimation algorithm has a wide range of application scenarios in traffic navigation, mechanical production and military fields. In the scenario of autonomous driving, in addition to the need to detect vehicles and pedestrians in road traffic, it is also necessary to calculate the position and posture information of the target, so as to estimate the movement direction and trajectory of the target object, and provide a basis for the next driving behavior. On the production line, the robotic arm also needs to predict the relative pose between the robotic arm and the object when picking up and delivering objects. In the military field, pose estimation can predict enemy weapons and equipment, so as to respond in a targeted manner. For example, by calculating the pitch and rotation angles of the enemy's gun barrel, the enemy's strike range can be predicted; or based on the known enemy position, how to carry out precise strikes, etc.
[0003] Traditional pose estimation algorithms can generally be divided into three categories: feature point-based methods, template matching-based algorithms, and three-dimensional description operator-based algorithms.
[0004] The pose estimation algorithm based on feature correspondence first identifies feature information on the image, and then matches the features with the target object in the world coordinate system to calculate the camera pose. The feature point correspondence method requires the image coordinates of at least three points and their corresponding coordinates in the world coordinate system. Through perspective projection and other methods, the pose information of the camera based on the world coordinate system is calculated. This type of problem is usually called the PnP problem. PnP is a classic algorithm for solving pose problems, but it requires iterative calculations, has cumbersome steps, and is inefficient. In order to solve the problems of the PnP algorithm, many optimization algorithms have been proposed, such as EPnP (Efficient PnP), RPnP (Robust PnP) and UPnP (Universal PnP). Based on PnP, these algorithms further improve accuracy and efficiency through non-iterative algorithms or optimization of rotation matrix parameters.
[0005] Algorithms based on template matching need to model the target object and generate templates in different postures. This type of work usually uses a virtual camera to shoot the target object, thereby obtaining models of the target object at different angles. Template matching methods need to build models based on specific target objects, and the usage scenarios are different. Hinterstoisser built the LINEMOD dataset and used a depth camera to automatically generate templates from 3D models, but the effect is poor for objects with occlusion.
[0006] The main difference between 3D feature descriptors and 2D image operators is that they use the depth information of the image, so their main research object is data objects such as point clouds that carry 3D information of the target object. Rust et al. proposed the viewpoint histogram feature descriptor (VFH), which completed the extraction of geometric features and the description of pose information in 3D space. Unlike VFH, which extracts image features from a global perspective, Drost et al. designed a set of local feature descriptors (PPFs) to extract the distance and normal angle between point pairs, and perform pose calculations based on matching point pairs.
[0007] Traditional pose estimation algorithms rely heavily on feature points in the image, and the feature points obtained by image processing operators are usually not very accurate. At the same time, such algorithms cannot obtain accurate feature points in areas where feature points are blocked, lighting conditions are poor, or texture features are not obvious, and do not have good generalization.
[0008] With the progress of deep network research, pose estimation methods based on deep learning have become a current research hotspot. Compared with traditional algorithms, deep networks do not require manual design of specific operators, but use the model's expressive power to adaptively extract feature points, resulting in more accurate results. At the same time, it also solves the problem of poor generalization performance of template matching. In terms of feature extraction, deep networks can be roughly divided into two categories: target detection models and semantic segmentation models.
[0009] The target detection model predicts the position of the target object in the image to extract feature points. R-CNN is one of the earliest deep learning detection algorithms, and it is the first to apply convolutional neural networks (CNN) to target detection tasks. R-CNN outputs the location and classification information of the target object in two stages. On this basis, Fast R-CNN and Faster R-CNN have improved on R-CNN, further improving the inference speed of the R-CNN model while ensuring the accuracy of target detection.
[0010] The YOLO model is another major category in the target detection model. Its characteristic is that the image is divided into grid cells, each of which is responsible for predicting one or more bounding boxes. Each grid cell also predicts the location of the bounding box and the class probability of the target. The emergence of the YOLO model enables the feature point extraction algorithm based on target detection to achieve higher efficiency.
[0011] In addition to the object detection model, the semantic segmentation model directly identifies the location information of the object points in the image by classifying the pixels. This type of model is represented by U-Net, which realizes the transmission of image features at different resolutions through the skip connection mechanism.
[0012] Although current deep learning models are able to fully extract image features, these models cannot be directly used for pose estimation and need to be further integrated with pose estimation algorithms. Summary of the invention
[0013] The present invention provides a method and system for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers, which can improve the stability and accuracy of the posture estimation results.
[0014] A method for estimating the posture of a drone based on visual detection of fixed markers comprises the following steps:
[0015] (1) Build a deep model and use airport runway images for target detection training;
[0016] (2) Use the deep model for inference, input the airport runway image taken by the onboard camera, and output the coordinates of the runway corner points;
[0017] (3) Obtain corner points with high confidence through post-processing filtering and screening;
[0018] (4) Use the vanishing point algorithm to calculate the current posture state;
[0019] (5) Use the orthogonal iterative algorithm to optimize the posture state and obtain the final posture result of the airborne camera.
[0020] The method proposed in this paper combines the high-precision detection capability of deep learning models with the simplicity and operating efficiency of traditional algorithms. In the feature point prediction stage, it does not rely on specific image processing operators, but extracts the feature points of target objects through target detection algorithms. It has stronger generalization, can accurately extract the feature points of objects in different forms, and can handle more complex scenarios.
[0021] In step (1), the deep model uses the YOLOv3 framework, takes the airport runway as the target object, and uses the four corner points of the runway as feature points. The training data is the airport runway images at different flight altitudes. Through training, the model is able to perform target detection on the airport runway and output the coordinates of the four corner points of the runway.
[0022] The specific process of step (2) is as follows:
[0023] Use the deep model trained in step (1) and input the airport runway image taken by the onboard camera to detect the four corner points of the runway and obtain their coordinates and their confidence levels respectively.
[0024] The specific process of step (3) is as follows:
[0025] (3-1) Sort the output corner points according to their confidence, and retain three candidate results for each corner point;
[0026] (3-2) Setting a confidence threshold and removing candidate results below the threshold;
[0027] (3-3) For the remaining corner points, convex polygon detection is performed to remove unreasonable detection results; through the post-processing process, a set of corner point detection results are finally obtained based on the confidence level. Taking the lower right corner point as the reference, the coordinates of the four corner points are expressed as (x1, y1), (x2, y2), (x3, y3), and (x4, y4) in counterclockwise order.
[0028] In step (4), the vanishing point algorithm is based on two runway points and two runway edges as constraints, thereby calculating the preliminary pose of the airborne camera.
[0029] In step (5), the orthogonal iterative algorithm uses the straight line information of the runway and fully utilizes the overall position state of the runway to improve the accuracy and stability of the translation matrix. Finally, the position and posture result of the current airborne camera is obtained by converting the displacement matrix and the rotation matrix.
[0030] A drone pose estimation system based on visual detection of fixed markers includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned drone pose estimation method.
[0031] A UAV posture estimation system based on visual detection of fixed markers, comprising:
[0032] Visual detection module: Use the YOLOv3 deep model to detect airport runways and extract corner features;
[0033] Pose estimation module: Based on the obtained corner point coordinates, the camera pose is solved through the vanishing point algorithm and orthogonal iteration algorithm.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] 1. The present invention makes up for the shortcomings of existing image detection operators such as low precision, poor generalization, and susceptibility to noise interference, and improves the accuracy and robustness of key point detection.
[0036] 2. The present invention uses the vanishing point algorithm and the orthogonal iterative algorithm to replace the traditional PnP algorithm, thereby ensuring the stability and accuracy of the pose estimation results.
[0037] 3. The present invention combines the traditional pose estimation algorithm with the deep learning algorithm, which improves the generalization ability and accuracy of the algorithm while ensuring the operation efficiency of the algorithm. The target detection of the deep model is generalizable and can handle various complex scenes.
[0038] 4. The present invention can be used in the fields of aircraft positioning, yaw prediction, etc. during the take-off, landing and taxiing phases of civil UAVs to reduce the cost of using sensors such as GPS and IMU. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 This is a flow chart of a method for estimating the posture of a UAV based on visual detection of fixed markers according to the present invention;
[0040] Figure 2 Schematic diagram of the principle of the vanishing point algorithm in the present invention;
[0041] Figure 3 It is a schematic diagram of the principle of the orthogonal iterative algorithm in the present invention;
[0042] Figure 4 This is a flowchart of the model training in the present invention;
[0043] Figure 5 Schematic diagram of corner points extracted from airport runway images. DETAILED DESCRIPTION
[0044] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.
[0045] The present invention uses a deep learning model to identify the feature points of the target object, and then integrates it with the existing traditional pose estimation algorithm to produce a pose estimation system with higher accuracy. The target object image is obtained through an onboard camera sensor, and the camera pose information is finally output.
[0046] The present invention strengthens the existing pose estimation algorithm from two aspects. First, the feature extraction algorithm based on deep learning has higher accuracy and generalization than the traditional image operator, and can significantly improve the accuracy of feature point extraction. Secondly, the existing pose estimation algorithm is optimized, and the coordinate mapping algorithm such as the vanishing point is used to replace the traditional PnP algorithm, further improving the accuracy and operation efficiency of the algorithm.
[0047] like Figure 1 As shown, a method for estimating the posture of a UAV based on visual detection of fixed markers includes the following steps:
[0048] Step (1): build a deep model and use runway images for target detection training.
[0049] The deep model is trained using an image detection dataset. For the drone application scenario, the airport runway is used as the target object, and the four corner points of the runway are used as feature points. Unlike the traditional target detection model, which first generates the target object classification result and then generates the anchor box, the deep model here mainly functions to identify the location of feature points. Therefore, the runway corner points are directly predicted, and the specific location of the runway is enclosed in the form of a convex polygon using the obtained corner point coordinates.
[0050] The dataset contains camera images of airport runways at different flight altitudes. The deep model needs to accurately detect the coordinates of the corner points of airport runways of different shapes and sizes. When two nearby corner points cannot be observed, it is also necessary to identify the intersection of the runway boundary line and the image boundary in the image.
[0051] In order to balance the accuracy of prediction and the inference speed of the model, this algorithm uses the currently more mature YOLOv3 model to implement feature point detection function.
[0052] Step (2), use the deep model to infer and output the coordinates of the runway corner points.
[0053] Use the depth model obtained in step (1) to input the airport runway image, perform target detection, and output the coordinate information and confidence of the four corner points of the runway. At this time, the runway corner points are not accurate enough and need to be filtered through the post-processing module to obtain the final corner point coordinates.
[0054] Step (3), through post-processing filtering and screening, obtain corner points with higher confidence.
[0055] In order to obtain a detection result with a high confidence level, it is necessary to filter the corner points generated in step (2) during post-processing, which can be roughly divided into the following three stages:
[0056] (3-1) Sort the output corner points according to their confidence, and retain three candidate results for each corner point
[0057] (3-2) Set the confidence threshold and remove candidate results below the threshold
[0058] (3-3) Perform convex polygon detection on possible corner points to remove unreasonable detection results
[0059] Through the post-processing process, we finally get a set of corner point detection results based on the confidence level. Taking the lower right corner point as the reference, the coordinates of the four corner points are expressed as (x1, y1), (x2, y2), (x3, y3), and (x4, y4) in counterclockwise order.
[0060] Step (4), use the vanishing point algorithm to calculate the current posture state.
[0061] The vanishing point algorithm is used to implement the coordinate mapping function for the corner point coordinates obtained in step (3). The pose obtained by PnP solution is based on the idea of optimization, which minimizes the reprojection error of the feature points of the target object on the canvas relative to the world coordinates. This scheme gives equal importance to each feature point of the target object. In fact, due to the different camera perspectives, the prediction effects of different feature points of the target object are often different. For airport runways, during the take-off and landing phases of drones, the two near corner points are usually clearer than the two far corner points, and the prediction results are more accurate. In order to solve the errors caused by the traditional PnP algorithm in the coordinate mapping stage, the vanishing point algorithm is used instead.
[0062] The vanishing point algorithm can estimate the camera pose based on two points and two runway sidelines as constraints. The algorithm can make full use of these clear features and improve the accuracy of pose estimation by selectively selecting the closer corner points as the main calculation objects, which makes the algorithm more adaptable and stable in actual scenes. At the same time, the calculation process of the vanishing point algorithm is relatively simple, and it can usually run quickly in applications with high real-time requirements.
[0063] like Figure 2 As shown, two side lines, l3 and the crossing threshold line l2 have been detected. P2 and P3 are two corner points of the runway, and the runway width is L. Runway coordinate system o w x w y w z w The starting point is at the center of runway P2P3, and z w The axis is defined as the centerline of the runway, x w Axis right, y w The camera frame is fixed on the plane with the axis down. The origin is at the projection center, also called the pinhole. The horizontal axis of the image is u, the vertical axis is v, and the camera coordinate system is o. c x c y c z c Rigidly fixed to the aircraft, z c is the optical axis, x c ,y c The direction is consistent with u and v. A point P i Projection onto the image plane p i (u i ,v i ) on the depth k i is the optical center and P i Consider a pinhole camera model, so the corresponding camera coordinates are shown in formula (1).
[0064]
[0065] Among them, the variable
[0066] According to camera imaging theory, the parallel lines l1 and l2 will be at the vanishing point p(u p ,v p ) intersect, and in addition o c p / / l1 / / l3.
[0067] Through the above analysis, we can get the following three geometric constraints:
[0068] ①l1⊥l2②|P2P3|=L③l1 / / l3
[0069] Assuming the pose of the camera relative to the drone is known, we only look for the pose of the camera relative to the runway. This part is based on the visible threshold line.
[0070] First, after inferring the depths of P2 and P3 from the three constraints above, we can get the camera coordinates and camera coordinates of the three lines. Secondly, the attitude of the camera relative to the runway is determined by the corresponding three lines. Finally, the position of the camera is obtained using the coplanarity equation.
[0071] The vector l2 can be obtained by express: l1 is parallel to From constraint (1), we can get:
[0072]
[0073] The k2 and k3 relationships are:
[0074] k3=sk2(3)where
[0075] Constraint (2) determines the following formula:
[0076]
[0077] Replacing (3) with (4) yields a quadratic equation for k2, which can be solved as follows:
[0078]
[0079] According to the previous section, k2 has been calculated, the camera coordinates P i c and camera vector The corresponding l2 in the camera frame can be determined. The vector Equivalent to The orthonormal basis in the camera coordinate system is With the runway coordinate system It is the expression of the same vector in different coordinate systems. The rotation matrix R is determined by the equation RB = A, which can be transformed into the following least squares problem:
[0080] min||A-RB|| 2 s R T R=I3 (6)
[0081] The constrained least squares problem can be solved using Umeyama. Let the singular value decomposition UDV T It's AB T (UU T =VV T =I,D=diag(d i ),d1≥d2≥d3≥0), when the rank (AB T )≥2, the rotation matrix can be determined:
[0082] R=USV T (7)
[0083] In the current case, the rotation matrix R is known, and the point P in the camera coordinate system i c and point P in the runway coordinate system i , solve the translation matrix T of the runway origin relative to the camera origin, which can be obtained by the following formula:
[0084] P i c =RP i +T (8)
[0085] Step (4) uses the correspondence between any two points on the runway in the camera coordinate system and the world coordinate system to solve the translation matrix of the camera relative to the airport origin. This method is simple and direct, but the problem that comes with it is that it only uses a few points on the runway as the reference to solve the problem, and the information utilization is not perfect.
[0086] Step (5), use the orthogonal iterative algorithm to optimize the posture state and obtain the final posture result.
[0087] Formula (8) is improved and solved by using an orthogonal iterative algorithm. This algorithm uses the straight line information of the runway and the overall position state of the runway to improve the accuracy and stability of the translation matrix.
[0088] Depend on Figure 3 As shown, c x c y c z c is the camera coordinate system with the origin at the camera focus, and z c The axis coincides with the optical axis, and z cpoints to the front of the camera. I represents the normalized image plane vector, which is parallel to the o z on the point c Axis. O-XYZ is the spatial coordinate system. Both coordinate systems are right-handed coordinate systems.
[0089] Consider a space line I i , in spatial coordinates, this line is characterized by its unit direction vector d i and an arbitrary point p i Therefore, regarding L i Any point can be represented by λd i +p i Indicates that λ is a scalar, let l i is the normalized image vector L i .
[0090] The image plane I, represented by a in the camera coordinate system i x+b i y+c i =0, where n i =(a i ,b i ,c i ) t is a unit vector. L i , l i and c On the same plane (with N i denoted), called the interpretation plane.
[0091] Note that d i and p i Can be expressed in camera coordinates alone as Rd i and Rp i +t, because L i and c P i All in N i On, it gives:
[0092]
[0093] Formula (9) is the basic function for pose estimation from two-dimensional to three-dimensional line correspondence.
[0094] The basis function (9) states that any space line and its image line are in the same plane containing the projection center. Most algorithms are based on this form of basis functions. However, from Euclidean space points, a new form of basis equation is derived.
[0095] Let c i is online i The closest point on I o Point (see Figure 3 ), then the straight line o c d i Perpendicular to l i . Assume i ,m i The straight lines l i and c d i The unit direction vectors of , which can be easily calculated as:
[0096]
[0097] Note i ,m i} is a two-dimensional Euclidean space N i If ε is N i Any vector that is:
[0098]
[0099] definition Replace ε=Rd i and ε=Rp i + t, we can get:
[0100]
[0101] Formula (12) is called the new form of the basic function. It can be easily verified that So we get:
[0102]
[0103] So the new form of the basic function is equivalent to formula (9).
[0104] The OI algorithm is extended to straight line correspondence, and a displacement estimation method is derived:
[0105]
[0106] The orthogonal iterative algorithm improves the accuracy and stability of the vanishing point solution. Finally, the current camera's position and posture can be obtained through simple conversion of the displacement matrix and the rotation matrix.
[0107] The present invention uses a deep learning algorithm to replace the traditional image processing operator to achieve high-precision feature point extraction of the target object, and then uses the improved coordinate mapping algorithm to solve and obtain the current posture information.
[0108] The present invention improves the existing pose estimation algorithm in two aspects. First, it uses a deep learning algorithm to detect key points, which solves the defects of traditional image processing operators, such as fixed templates, poor accuracy, and susceptibility to noise interference. Second, it uses the vanishing point algorithm and orthogonal iteration algorithm to replace the existing PnP algorithm, making the algorithm more robust and further improving accuracy.
[0109] To verify the effect of the present invention, a certain runway is taken as an example to perform posture estimation using the present invention. Figure 4 The training process of the algorithm model is demonstrated.
[0110] S01, after completing the environment configuration, use YOLOv3 as the deep learning model. Among them, use Darknet-53 as the backbone network, YOLOv3Neck as the neck network, and YOLORunwayHead as the detection head. After the construction is completed, initialize the weights.
[0111] S02, use the airport runway image to train the model. Use the trained model to predict the runway image and extract the corner point information. The results are as follows: Figure 5 shown.
[0112] S03, the obtained detection results are post-processed and input into the pose estimation module, and the current camera pose result is solved by the vanishing point algorithm and the orthogonal iteration algorithm.
[0113] The airport runway camera images at different flight altitudes were collected, including 2,383 images in total, covering different stages of flight. The corner prediction accuracy was used to measure the algorithm performance of the visual detection module, and the results are shown in Table 1 below:
[0114] Table 1
[0115] Average corner error value (pixel) Accuracy Corner 1 [2.32897215,1.99478364] 98.6% Corner 2 [5.21634743,1.98813884] 98.6% Corner 3 [3.04829136,1.37873625] 95.7% Corner 4 [2.12986092,1.27265906] 99.2% average value [3.18,1.66] 98.0%
[0116] The corner point error value is in pixels and is divided into two dimensions: horizontal and vertical. The results show that the average corner point detection error of the visual detection module is within five pixels, and the average accuracy reaches 98.0%, which can meet the accuracy requirements of the posture detection module.
[0117] Using airport runway images at different heights for pose estimation, the results are shown in Table 2 below:
[0118] Table 2
[0119]
[0120] Among them, ft is the unit of feet. At different flight altitudes, the posture estimation module outputs high positioning accuracy and attitude estimation accuracy, which is sufficient to meet normal flight requirements.
[0121] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers, characterized in that: The following steps are involved: (1) Build a deep model and use airport runway images for target detection training; (2) Use the deep model for inference, input the airport runway image taken by the onboard camera, and output the coordinates of the runway corner points; (3) Obtain corner points with high confidence through post-processing filtering and screening; (4) Use the vanishing point algorithm to calculate the current posture state; (5) Use the orthogonal iterative algorithm to optimize the posture state and obtain the final posture result of the airborne camera.
2. The method for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers according to claim 1, characterized in that: In step (1), the deep model uses the YOLOv3 framework, takes the airport runway as the target object, and uses the four corner points of the runway as feature points. The training data is the airport runway images at different flight altitudes. Through training, the model is able to perform target detection on the airport runway and output the coordinates of the four corner points of the runway.
3. The method for estimating the position and posture of an unmanned aerial vehicle based on visual detection of fixed markers according to claim 1, characterized in that: The specific process of step (2) is as follows: Use the deep model trained in step (1) and input the airport runway image taken by the onboard camera to detect the four corner points of the runway and obtain their coordinates and their confidence levels respectively.
4. The method for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers according to claim 1, characterized in that: The specific process of step (3) is as follows: (3-1) Sort the output corner points according to their confidence, and retain three candidate results for each corner point; (3-2) Setting a confidence threshold and removing candidate results below the threshold; (3-3) For the remaining corner points, perform convex polygon detection to remove unreasonable detection results; Through the post-processing process, we finally get a set of corner point detection results based on the confidence level. Taking the lower right corner point as the reference, the coordinates of the four corner points are expressed as (x1, y1), (x2, y2), (x3, y3), and (x4, y4) in counterclockwise order.
5. The method for estimating the posture of an unmanned aerial vehicle based on visual detection of fixed markers according to claim 1, characterized in that: In step (4), the vanishing point algorithm is based on two runway points and two runway edges as constraints, thereby calculating the preliminary pose of the airborne camera.
6. The method for estimating the position and posture of an unmanned aerial vehicle based on visual detection of fixed markers according to claim 1, characterized in that: In step (5), the orthogonal iterative algorithm uses the straight line information of the runway and makes full use of the overall position state of the runway to improve the accuracy and stability of the translation matrix. Finally, the position and posture result of the current airborne camera is obtained by converting the displacement matrix and the rotation matrix.
7. A drone posture estimation system based on visual detection of fixed markers, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the drone pose estimation method described in any one of claims 1 to 6.
8. A drone posture estimation system based on visual detection of fixed markers, characterized in that: include: Visual detection module: Use the YOLOv3 deep model to detect airport runways and extract corner features; Pose estimation module: Based on the obtained corner point coordinates, the camera pose is solved through the vanishing point algorithm and orthogonal iteration algorithm.
Citation Information
Patent Citations
Fixed wing aircraft vision-auxiliary landing navigation method under low visibility
CN109341700A
Fast and steady monocular camera pose estimation algorithm
CN113436249A
Airport runway corner detection method and device, computer equipment and storage medium
CN115019179A
Unmanned aerial vehicle autonomous landing guiding method based on vision
CN115793677A
Unmanned aerial vehicle autonomous landing method and device based on visual image guidance
CN117636284A
