An Unmanned Aerial Vehicle Full-Automatic Landing Method Based on Visual Scene Target Fusion Recognition in a Denied Environment

Abstract

The present invention discloses a fully automatic landing method for an unmanned aerial vehicle (UAV) based on visual scene target fusion recognition in a denial environment. When the UAV performs a cruise inspection mission and enters the denial environment, visual scene target information is obtained in real time. Based on the visual scene target coordinates, a mathematical model for UAV autonomous positioning is established to calculate the current coordinates of the UAV. The current coordinates of the UAV are used as the reference navigation coordinates, and the nearest alternate landing point is used as the end point, and it is navigated to the nearest alternate landing point through trajectory planning. The visual scene information of the landing target target is obtained, the recognition and positioning of the target target are completed according to the visual scene information of the target target, and the UAV is controlled to land fully automatically according to the positioning result. The present invention realizes the fusion of infrared images to solve the problem of low visibility and UAV autonomous positioning under complex imaging, and avoids the safety problem of UAV out of control in the denial environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of UAV control, and particularly relates to an automatic UAV landing method based on visual target fusion recognition in a denial environment. Background Art

[0002] As an airborne mobile surveillance platform that can achieve rapid transfer, the unmanned aerial vehicle (UAV) has strong mobility and a wide monitoring range. At present, the development of UAVs tends to be fully intelligent and fully autonomous, and is widely used in cruise inspection tasks. The positioning accuracy of differential GPS aircraft can reach below 10 cm, but it is easily affected by the environment. The denial environment refers to the situation where when a UAV flies in a city or indoors, due to obstacles such as building clusters, forests, and walls, satellite signals are lost and unavailable, or signal interference causes GPS to be unable to provide accurate positioning information. Cruise inspection tasks are often carried out in urban areas, where the terrain environment of the landing area is complex. In a denial environment, the positioning accuracy is insufficient and the real-time performance is poor. If the landing attitude conflicts with the landing terrain, there are serious safety hazards such as the tilt and rollover of the aircraft body, which may cause property and personnel damage, and it is difficult to meet the fully automatic and safe landing of UAVs in a denial environment.

[0003] In the prior art, in a denial environment, the safe landing of UAVs is generally achieved by manual intervention in the landing process. Due to the large operation area of cruise inspection tasks, it is often difficult for the pilot to arrive at the scene in time to intervene in the landing process. In addition, absolute position known base stations can be set up on the ground to calculate the relative distance and angle for positioning. Although this method can enable the UAV to calculate its own coordinates, multiple base stations need to be set up in advance, with high hardware requirements and high costs, and there are significant equipment limitations. There are also some solutions to guide the UAV to land through infrared. However, since these solutions require entering the infrared guidance range in advance, and the generation of the denial environment is not predictable. Accurate three-dimensional positioning requires two or more infrared beacons, and it is difficult to achieve full infrared coverage of the entire operation area. Moreover, infrared guidance requires manual intervention to ensure flight safety, and it is difficult to ensure the fully automatic and safe landing of UAVs. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an automatic UAV landing method based on visual target fusion recognition in a denial environment in view of the above-mentioned deficiencies of the prior art. The geodetic coordinates of the pre-collected target are obtained through visual target fusion recognition, and then the precise coordinates of the UAV are calculated through the inner triangle centroid pose estimation algorithm to achieve the fully automatic and safe landing of the UAV, and the fusion of infrared images is used to solve the problem of UAV autonomous positioning in low visibility and complex imaging, avoiding the safety problem of UAV out of control in a denial environment.

[0005] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] An automatic UAV landing method based on visual target fusion recognition in a denied environment, comprising:

[0007] Step 1: When the UAV executes a cruise inspection mission and enters a denied environment, obtain visual target information in real time;

[0008] Step 2: Based on the visual target coordinates, establish a mathematical model for UAV autonomous positioning and solve the current coordinates of the UAV;

[0009] Step 3: Use the current coordinates of the UAV as the reference navigation coordinates and the nearest alternate landing point as the end point, and navigate to the nearest alternate landing point through trajectory planning;

[0010] Step 4: Obtain the visual information of the landing target target, complete the recognition and positioning of the target target according to the visual information of the target target, and control the UAV to land automatically according to the positioning result.

[0011] To optimize the above technical solution, the specific measures taken also include:

[0012] In the above steps 1 and 4, the visual information is obtained by improving the TY-YOLO recognition algorithm. The specific improvement and training methods of the improved TY-YOLO recognition algorithm include:

[0013] 1) For the unique aerial perspective of the UAV, the visible angle of the object changes, and the basic network is specifically trained;

[0014] Add a 52×52 prediction branch at the 8-fold downsampling of the shallow network structure of the convolutional neural network, and fuse the shallow and deep features through splicing;

[0015] 2) Add the distance between the predicted box and the center coordinates of the real box to the loss function. By incorporating the coordinate error between the center points into the standard of loss convergence, minimize the real error between the center points during training;

[0016] 3) Compress the recognition model for recognition speed. Specifically:

[0017] Replace the original convolutional layer with an ES module;

[0018] For the convolutional neural network, retain the first 9 convolutional layers, perform the above replacement on the 10th, 11th, and 13th convolutional layers, and remove the pooling layer between the two layers;

[0019] 4) Prune and compress the recognition model through channel pruning;

[0020] 5) Add camera infrared images during training, and fuse infrared imaging to achieve visual target recognition in low-visibility and complex imaging environments.

[0021] The above-mentioned ES module consists of an S compression layer and an E expansion layer. The S compression layer consists of several 1×1 convolutional kernels, and the E expansion layer consists of several convolutional kernels with sizes of 1 and 3. Taking cin as the number of input channels, cout as the number of output channels, S as the number of input channels of the S compression layer, E1 and E2 as the number of input channels of the E expansion layer respectively, and the convolutional kernel is represented by m, the formula for calculating the number of parameters of the ES module is:

[0022] N = (c in ×m s 2 +1) × S + (S × m E1 2 +1) × E 1 +(S × m E2 2 +1) × E 2 .

[0023] The above-mentioned pruning and compression of the recognition model through channel pruning is as follows:

[0024] Divide the channel importance attributes to perform sparse training on the pre-trained model;

[0025] Set the channel threshold, and perform channel pruning on the channels of the sparse model whose importance is lower than the threshold;

[0026] Fine-tune the pruned sparse model, and after initializing the network, train it until the loss function curve converges again;

[0027] If the accuracy of the model drops to the accuracy threshold, the channel pruning ends. If it cannot be achieved, readjust the pruning ratio and prune again.

[0028] The above-mentioned step 2 calculates the current coordinates of the UAV through the inner triangle centroid positioning algorithm.

[0029] The above-mentioned step 4 is specifically: obtain the landing target target through the improved TY-YOLO recognition algorithm. After obtaining the target view information, perform image point positioning on the recognized target prediction box, obtain the image point coordinates of the center point of the target recognition, calculate the relative offset of the image point coordinate center, and adjust the UAV gimbal axis to make the target target always within the view range;

[0030] Based on the view information of the target target and the image point coordinates, establish a mathematical model for estimating the relative pose of the UAV target, complete the estimation of the relative pose of the UAV and the target, calculate the relative attitude and adjust the yaw attitude and pitch attitude of the UAV, and then calculate the relative coordinate offset of the UAV;

[0031] Further calculations are performed to obtain the heading of the UAV relative to the target, correct the landing error, and adjust the UAV center and the target center to be on the same Z-axis. When the error is 0, the UAV lands, completing the automatic landing process of the UAV.

[0032] The present invention has the following beneficial effects:

[0033] The present invention improves the traditional deep learning model, proposes an improved TY-YOLO recognition algorithm suitable for UAV on-board, and supplements infrared images during training, effectively solving the problem of visual target recognition in low-visibility complex imaging environments. By improving the network structure and pruning the model, the speed of target recognition is increased, better adapting to the application scenarios where the UAV needs real-time recognition. By integrating the visual target recognition technology, through improving the TY-YOLO recognition algorithm, while maintaining the accuracy, the recognition speed is greatly improved, enabling real-time and accurate visual target recognition, and supplementing camera infrared images during training, effectively solving the problem of visual target recognition in low-visibility complex imaging environments.

[0034] The present invention obtains the current coordinates of the UAV through visual target relative pose estimation, uses them as the reference coordinates, and takes the target coordinates of the next alternate landing point as the end point, shortening the distance through trajectory planning in a denied environment. After entering the visual range of the target, image point positioning is performed on the target, the center point image coordinates of the target are obtained, and the relative pose difference between the UAV and the target is calculated through the pose estimation model, adjusting the pan-tilt axis and the UAV attitude to achieve fully automatic and safe landing. It can achieve centimeter-level fully automatic and precise landing of the UAV, effectively improving safety and reliability, solving the problem of fully automatic and safe landing in the UAV cruise inspection task, and avoiding the out-of-control safety problem in the cruise inspection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a schematic diagram of the UAV cruise process;

[0036] Figure 2 is a schematic diagram of the landing process;

[0037] Figure 3 is a schematic diagram of the improved prediction branch;

[0038] Figure 4 is a schematic diagram of the main steps of channel pruning;

[0039] Figure 5 is a schematic diagram of UAV target recognition. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following further describes the embodiments of the present invention in detail with reference to the drawings.

[0041] As Figure 1As shown in the figure, a fully automatic landing method for an unmanned aerial vehicle (UAV) based on visual target fusion recognition in a denied environment according to the present invention includes:

[0042] Step 1: When the UAV is performing a cruise inspection mission and enters a denied environment, the visual target information is obtained in real time by improving the TY-YOLO recognition algorithm;

[0043] That is, when the UAV is performing a cruise inspection mission, the satellite signal is lost, and the UAV enters a denied environment. At this time, the UAV starts the visual target recognition algorithm to fuse the UAV vision-assisted positioning, and the visual target information in the current area is detected in real time by improving the TY-YOLO recognition algorithm.

[0044] Step 2: Based on the visual target coordinates, establish a UAV autonomous positioning mathematical model and solve the current coordinates of the UAV;

[0045] Step 3: Take the current coordinates of the UAV as the reference navigation coordinates and the nearest alternate landing point as the end point, and navigate to the nearest alternate landing point through trajectory planning;

[0046] Steps 2-3 are:

[0047] For the obtained visual target information, establish a UAV autonomous positioning mathematical model in combination with the target coordinates, and solve the current coordinate information of the UAV;

[0048] Raise the altitude to the safety threshold for obstacle and terrain avoidance, and navigate to the nearest UAV emergency alternate landing point through UAV trajectory planning, gradually shortening the distance to the alternate landing point;

[0049] Step 4: Obtain the visual information of the landing target target, complete the recognition and positioning of the target target according to the visual information of the target target, and control the UAV to land fully automatically according to the positioning result. As Figure 2 shown, specifically:

[0050] Obtain the landing target target by improving the TY-YOLO recognition algorithm. After obtaining the target visual information, perform image point positioning on the recognized target prediction box, obtain the image point coordinates of the center point of the target recognition, calculate the relative offset of the image point coordinate center, and adjust the UAV pan-tilt axis to make the target target always within the visual range;

[0051] Based on the visual information of the target target and the image point coordinates, establish a UAV target relative pose estimation mathematical model, complete the relative pose estimation of the UAV and the target, calculate the relative attitude and adjust the yaw attitude and pitch attitude of the UAV, and then calculate the relative coordinate offset of the UAV;

[0052] Further calculations are performed to obtain the heading of the drone relative to the target, correct the landing error, and adjust the drone center and the target center to be on the same Z-axis. When the error is 0, the drone executes a landing to complete the automatic landing process of the drone.

[0053] The specific introduction is as follows:

[0054] In steps 1 and 4, the improved TY-YOLO recognition algorithm is used to obtain visual information. The specific improvement and training methods of the improved TY-YOLO recognition algorithm include:

[0055] Compared with the traditional recognition algorithm, the improved TY-YOLO recognition algorithm for visual target recognition is improved according to the specificity of the drone recognition task, fuses features, optimizes the network structure, and replaces some convolutional layers through the ES module. Without sacrificing much recognition accuracy, it greatly reduces the computational overhead and improves the recognition speed, better adapting to the on-board recognition task. At the same time, a loss function based on the improvement of the center coordinates is proposed to further improve the accuracy of the drone landing center coordinates, better adapting to the fully automatic and safe landing task of the drone. Channel pruning is performed during training and infrared images are supplemented to effectively solve the visual target recognition in low-visibility and complex imaging environments.

[0056] 1.1 Since the drone recognition task has high requirements for real-time performance and accuracy, the traditional target recognition algorithm is improved. According to the actual flight characteristics of the drone, the pixel ratio of the target during the landing process changes greatly, and the training and learning of the convolutional neural network are divided into different levels. Therefore, a prediction branch is added to the original algorithm to improve the recognition accuracy of the drone for small ground targets and optimize the recognition algorithm. Specifically:

[0057] First, due to the unique aerial perspective of the drone, the visual angle of the object changes, and specific training of the basic network is required.

[0058] In addition, when the drone takes pictures of the ground recognition target in the air, the object distance is much greater than the image distance, the pixels of the target in the image account for a small proportion, and during the landing process, the pixel ratio of the target gradually increases. The shallow layer of the convolutional neural network mainly extracts features composed of simple shapes such as edges, and the deep convolutional layer is responsible for extracting abstract and complex target contours and other features. Compared with the shallow layer, it has more accurate position information, and the semantic information features of the deep layer are more obvious. To improve the influence caused by the change of the target scale, a 52×52 prediction branch is added at the 8-fold downsampling of the shallow network structure, as shown in the appendix Figure 3 shown. By splicing and fusing shallow and deep features, the recognition effect when the pixel ratio of the target is small is improved, and not much computational overhead is added.

[0059] 1.2 Since the present invention has relatively high requirements for the positioning accuracy of the target, it is necessary to estimate the relative pose of the UAV through the image point coordinates. Therefore, the loss function is improved, and the accuracy of image point positioning is incorporated into the convergence criterion. The distance between the center coordinates of the predicted box and the true box is added to the loss function. By incorporating the coordinate error between the center points into the loss convergence criterion, the true error between the center points is minimized during training. As shown in the following formula, where Al and Bl represent the center point coordinates, and d is the maximum distance between the two boxes.

[0060] Loss TY = 1 - IoU + o 2 (Al, Bl) / d 2

[0061] 1.3 Since the UAV is affected by the load and cannot carry large computing devices when performing on-board recognition tasks, and the UAV task has relatively high requirements for the recognition speed, the recognition model is compressed.

[0062] 1.3.1 Replace the original convolutional layer. The formula for calculating the number of parameters of the original convolutional layer is:

[0063] N conv = c out ×(c in ×m 2 + 1)

[0064] where m is the size of the convolutional kernel, c in is the number of input channels, and c out is the number of output channels.

[0065] 1.3.2 The ES module consists of an S compression layer and an E expansion layer. The S compression layer consists of several 1×1 convolutional kernels, and the E expansion layer consists of several convolutional kernels with sizes of 1 and 3. With c in as the number of input channels, c out as the number of output channels, S as the number of input channels of the S compression layer, E1 and E2 as the number of input channels of the E expansion layer respectively, and the convolutional kernel is represented by m. Then the calculation of the number of parameters of the ES module is:

[0066] N = (c in ×m s 2 + 1)×S + (S×m E1 2 + 1)×E 1 +(S×m E2 2 + 1)×E 2

[0067] 1.3.3 Assume that the number of input channels is 64, the number of output channels is 128, the convolutional kernel size is 3×3, and the expansion compression ratio is 4. Calculating separately according to the above formulas 1.3.1 and 1.3.2, we can get: (64×3 2 +1)×128 = 73856, 16×(64×1 2 +1)+64×(16×1 2 +1)+64×(16×3 2 +1) = 11408. Through calculation, it can be obtained that the number of parameters after replacement is reduced by 62448, and the reduction ratio is 84.5%, which greatly reduces the computational cost.

[0068] 1.3.4 According to the characteristics of feature extraction of convolutional neural networks, while avoiding too much parameter compression from affecting the recognition accuracy, retain the first 9 convolutional layers, replace the 10th, 11th, and 13th layers, and remove the pooling layers between the two layers. The overall number of parameters is reduced by about 67%, the recognition accuracy of the final target is lost by 3.75%, and the recognition speed is increased by 7.9 FPS.

[0069] 1.4 Since the application scenario of the UAV recognition part of the present invention is mainly the fully automatic landing process of the UAV. During this process, it is required to realize the recognition of the target in real time and quickly, calculate the relative pose of the UAV, obtain the coordinate offset, and there are high requirements for the recognition speed of the target. Therefore, the model is slightly compressed through channel pruning. The specific pruning process is as shown in the appendix Figure 4 as follows.

[0070] 1.4.1 Divide the channel importance attributes to perform sparse training on the pre-trained model.

[0071] 1.4.2 Set the channel threshold, and prune the channels with channel importance lower than the threshold in the sparse model.

[0072] 1.4.3 Fine-tune the pruned sparse model, and train it until the loss function curve converges again after initializing the network.

[0073] 1.4.4 If the accuracy of the pruned model after fine-tuning drops to the accuracy threshold, the channel pruning ends. If it cannot be achieved, readjust the pruning ratio and prune again.

[0074] 1.4.5 Add camera infrared images during training to fuse infrared imaging to solve the visual target recognition in low visibility and complex imaging environments.

[0075] In step 1, when the UAV is performing a cruise inspection task, due to interference from the environment, etc., it loses the GPS signal and enters a denied environment. At this time, the improved TY-YOLO recognition algorithm is used to recognize the visual target.

[0076] In Step 2, after detecting the target visual scene information, obtain the image point coordinates of the visual scene target, establish a relative pose estimation model of the UAV to calculate the current coordinates of the UAV, and realize the autonomous positioning of the UAV in the denial environment, specifically as follows:

[0077] 2.1 According to the obtained image point coordinates of the visual scene target, set the rotation matrix as R and the translation matrix as T:

[0078]

[0079] 2.2 The coordinates of the visual scene target in the camera coordinate system are:

[0080]

[0081] Among them, R G R N and R B are the rotation matrices for coordinate transformation. T G is the origin of the UAV geodetic coordinate system in the reference coordinate system, X C and X G are the coordinates of the target in the camera coordinate system and the geodetic rectangular coordinate system respectively is the UAV coordinate information, and (φ, γ, θ) is the UAV attitude information. The azimuth and elevation angles of the camera are represented by (α, β).

[0082] 2.3 According to the geometric relationship of the perspective projection optical path, we can obtain

[0083]

[0084] Among them, f is the focal length of the camera, x 0 , y 0 are the interior orientation elements of the image, x and y are the image plane coordinates of the visual scene target image point a in the image, X A , Y A , Z A are the coordinates of the target in the reference coordinate system, and (X S , Y S , Z S ) is the optical center coordinate of the camera.

[0085] 2.4 According to the coordinate transformation formula and the collinearity equation, taking the geodetic rectangular coordinates (X A , Y A , Z A ) of point A, the current coordinates of the UAV, as the unknowns, the following formula is derived:

[0086]

[0087] 2.5 Based on the principle of binocular vision positioning, three sets of coordinate points of point A in the reference coordinate system can be obtained:

[0088]

[0089] Among them

[0090]

[0091]

[0092]

[0093] 2.6 When observing and positioning the visual scene target, due to the positioning error of the visual scene target itself, the image point positioning error, etc., there may be a situation where the extension line of the imaging optical axis and the image point of the visual scene target cannot intersect at one point, resulting in a large positioning error. To solve the interference of redundant information and improve the positioning accuracy, the error is reduced through the inner triangle centroid positioning algorithm.

[0094] Assume that the true coordinates of the UAV at point A in the reference system are (X A , Y A , Z A ), and the calculated UAV coordinates are (X Ai , Y Ai , Z Ai ). Then, if you want to obtain the most accurate positioning coordinates, it should satisfy:

[0095]

[0096] According to the inner triangle centroid algorithm, the sum of the squares of the deviations of each variable value from its average is equal to the minimum value. Therefore, the optimal estimate of the UAV coordinate A is:

[0097]

[0098] 2.7 Calculate the current coordinates of the UAV.

[0099] In step 3, raise the flight altitude of the UAV to the safety threshold for obstacle and terrain avoidance;

[0100] Take the current coordinates of the UAV as the reference navigation coordinates and the coordinates of the alternate landing point as the target coordinates, and navigate to the target alternate landing point of the UAV through trajectory planning.

[0101] The specific implementation process of step 4 is as follows:

[0102] Gradually shorten the distance to the target target. After obtaining the visual scene information of the target target, identify the target data and perform image point positioning on the identified target prediction box.

[0103] Obtain the image point coordinates of the center point of the target recognition, output the center point coordinates and calculate the relative image point coordinate offset. As Figure 5 shown.Figure 5 The box in it represents the target prediction box, and the data in the lower right corner represents the image point coordinates of the center point of the current target.

[0104] Combined with the UAV IMU data, the pose estimation of the UAV relative to the target is obtained by solving the above formulas simultaneously, and further the heading of the UAV relative to the target is calculated.

[0105] Based on the visual information of the target and the image point coordinates, a mathematical model for UAV pose estimation is established, the attitude is calculated, and the yaw attitude and pitch attitude of the UAV are adjusted, and then the relative coordinate offset of the UAV is calculated. According to the relative pose estimation of the UAV, the current coordinate offset (X e , Y e )

[0106] According to the data obtained in the previous step, the coordinate error of the UAV is corrected, and the UAV coordinates are adjusted so that the center of the UAV and the center of the target are on the same Z-axis, that is, (X e , Y e ) is zeroed.

[0107] Continuously correct the coordinate error until the error is stable at 0, (X e , Y e ) is zeroed, the UAV landing is executed, and the automatic landing process of the UAV is completed.

[0108] The above is only the preferred implementation mode of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A fully automatic landing method for unmanned aerial vehicles based on visual target fusion recognition in a denied environment, characterized in that, it includes: Step 1: When the unmanned aerial vehicle executes a cruise inspection mission and enters a denied environment, real-time visual target information is obtained; Step 2: Based on the visual target coordinates, an autonomous positioning mathematical model of the unmanned aerial vehicle is established to calculate the current coordinates of the unmanned aerial vehicle; Step 3: Using the current coordinates of the unmanned aerial vehicle as the reference navigation coordinates and the nearest alternate landing point as the end point, navigate to the nearest alternate landing point through trajectory planning; Step 4: Obtain the visual information of the landing target target, complete the recognition and positioning of the target target according to the visual information of the target target, and control the fully automatic landing of the unmanned aerial vehicle according to the positioning result; In Steps 1 and 4, the improved TY-YOLO recognition algorithm is used to obtain visual information. The specific improvement and training methods of the improved TY-YOLO recognition algorithm include: 1), For the unique aerial perspective of the unmanned aerial vehicle, the visual angle of the object changes, and specific training is carried out on the basic network; Add a 52×52 prediction branch at the 8-fold downsampling of the shallow network structure of the convolutional neural network, and fuse the shallow and deep features through splicing; 2), Add the distance between the predicted box and the center coordinates of the true box to the loss function. By incorporating the coordinate error between the center points into the standard of loss convergence, minimize the true error between the center points during training; 3), Compress the recognition model for recognition speed. Specifically: Replace the original convolutional layer with the ES module; For the convolutional neural network, retain the first 9 convolutional layers, perform the above replacement on the 10th, 11th, and 13th convolutional layers, and remove the pooling layer between the two layers; 4), Prune and compress the recognition model through channel pruning; 5), Add camera infrared images during training, and fuse infrared imaging to achieve visual target recognition in low-visibility and complex imaging environments.

2. The fully automatic landing method for unmanned aerial vehicles based on visual target fusion recognition in a denied environment according to claim 1, characterized in that, The ES module is composed of an S compression layer and an E expansion layer. The S compression layer is composed of several 1×1 convolutional kernels, and the E expansion layer is composed of several convolutional kernels with sizes of 1 and 3. Taking cin as the number of input channels, cout as the number of output channels, S as the number of input channels of the S compression layer, E1 and E2 as the number of input channels of the E expansion layer respectively, and the convolutional kernel is represented by m. The parameter calculation formula of the ES module is: N = (c in × m s 2 + 1) × S + (S × m E1 2 + 1) × E 1 + (S × m E2 2 + 1) × E 2 。 3. The fully automatic landing method for unmanned aerial vehicles based on visual target fusion recognition in a denied environment according to claim 1, characterized in that, The pruning and compression of the recognition model through channel pruning is as follows: Divide the channel importance attributes to perform sparse training on the pre-trained model; Set the channel threshold, and perform channel pruning on the channels of the sparse model whose importance is lower than the threshold; Fine-tune the pruned sparse model, and train it to the loss function curve to converge again after initializing the network; If the accuracy of the model drops to the accuracy threshold, the channel pruning ends. If it cannot be achieved, readjust the pruning ratio and prune again.

4. A fully automatic landing method for an unmanned aerial vehicle (UAV) based on visual target fusion recognition in a denial environment according to claim 1, characterized in that, in step 2, the current coordinates of the UAV are calculated by the inner triangle centroid positioning algorithm.

5. A fully automatic landing method for an unmanned aerial vehicle (UAV) based on visual target fusion recognition in a denial environment according to claim 1, characterized in that, step 4 is specifically: the landing target is obtained by improving the TY-YOLO recognition algorithm. After obtaining the visual information of the target, the image point positioning of the predicted box of the recognized target is carried out, the image point coordinates of the center of the target recognition are obtained, the relative offset of the image point coordinate center is calculated, and the axial direction of the UAV gimbal is adjusted so that the target is always within the visual range; a mathematical model for estimating the relative pose of the UAV and the target is established based on the visual information of the target and the image point coordinates, the relative pose of the UAV and the target is estimated, the relative attitude is calculated, and the yaw attitude and pitch attitude of the UAV are adjusted, and then the relative coordinate offset of the UAV is calculated; further calculate the heading of the UAV relative to the target, correct the landing error, adjust the center of the UAV and the center of the target to be on the same Z-axis. When the error is 0, execute the landing of the UAV to complete the automatic landing process of the UAV.

Images