A quadrotor unmanned aerial vehicle pose estimation method based on a Re-LSTM network

Abstract

The application discloses a quadrotor unmanned aerial vehicle pose estimation method based on a Re-LSTM network, and steps of the method comprise the following steps: 1, collecting the pose of the quadrotor unmanned aerial vehicle at the current time t by using an inertial measurement unit (IMU), a depth camera and a GPS module respectively, and performing data preprocessing and fusion; 2, establishing a Re-LSTM network and processing the fused pose data to obtain the pose estimation value of the quadrotor unmanned aerial vehicle at the current time t; 3, constructing a loss function and training the Re-LSTM network to obtain an optimal pose estimation model, thereby predicting the current pose of the quadrotor unmanned aerial vehicle. The application can accurately and timely estimate the current pose of the quadrotor unmanned aerial vehicle, thereby guaranteeing flight safety and significantly improving control accuracy and task execution efficiency.

Description

Technical Field

[0001] This invention belongs to the field of multi-sensor fusion and deep learning, and more specifically, it is a pose estimation method for quadcopter UAVs based on Re-LSTM networks. Background Technology

[0002] Drones have been widely used in various fields worldwide, especially in agriculture (such as crop planting and pesticide spraying), industry (such as equipment maintenance and inspection in hazardous environments), and the military (such as site reconnaissance). However, current deep learning-based drone pose estimation methods often require the integration of data from multiple sensors. How to effectively integrate heterogeneous data remains a challenge. Furthermore, in complex environments, such as densely populated urban areas and forests, and under conditions of varying lighting, sensor data may lack sufficient accuracy and robustness. Therefore, even after data fusion, better algorithms and data support are still needed to improve the accuracy and stability of pose estimation. Thus, improving the accuracy of drone pose estimation after integrating multi-source data remains a challenge.

[0003] Existing pose estimation methods rely on neural network models that have low accuracy in estimating the attitude of quadcopter drones in outdoor scenarios. This leads to the remote control room misjudging the current flight status of the drone and issuing inappropriate commands, resulting in damage to the drone. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a Re-LSTM network-based quadrotor drone pose estimation method. The aim is to optimize the LSTM network structure and train the network to obtain a more accurate model for quadrotor drone pose estimation, thereby enabling real-time and accurate output of the quadrotor drone's current attitude. This allows a remote workstation to accurately determine the quadrotor drone's attitude in real time and issue timely commands, assisting the drone in efficiently completing its tasks.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention provides a pose estimation method for a quadrotor UAV based on a Re-LSTM network, characterized by the following steps:

[0007] Step 1: Collect data from the quadcopter UAV at the current time t and preprocess it to obtain the pose feature E of the quadcopter UAV at the current time t. t_cur ;

[0008] Step 2: Establish a Re-LSTM network and perform E... t_cur The process is performed to obtain the predicted pose feature at the current time t output by the r-th output unit. r,t_pred ;

[0009] Step 3: Construct the loss function Loss t_pose A Re-LSTM network is trained to obtain a trained quadcopter drone pose estimation model, which is used to estimate the current pose of the quadcopter drone in real time.

[0010] The method for estimating the pose of a quadcopter UAV based on a Re-LSTM network described in this invention is characterized in that step 1 includes the following steps:

[0011] Step 1.1: Collect the attitude [q] of the quadcopter UAV at the current time t. 0_t ,q 1_t ,q 2_t ,q 3_t ] T and position coordinates [x 1_t ,y 1_t ,z 1_t ] T And the depth image of the object observed by the quadcopter UAV at the current time t [dep] 0_t ,img 0_t ] T Where T represents transpose, dep 0_t The image represents the depth pixel value of an object at the current time t. 0_t The color pixel value representing the depth image of the object at the current time t;

[0012] Step 1.2: Use equation (1-1) to [q] 0_t ,q 1_t ,q 2_t ,q 3_t ] T Data preprocessing is performed to obtain the preprocessed attitude [q] of the quadcopter UAV at the current time t. 0_t_pro ,q 1_t_pro ,q 2_t_pro ,q 3_t_pro ] T

[0013]

[0014] In equation (1-1): α1, β1, γ1 are the yaw angle, pitch angle, and roll angle of the quadcopter UAV at the current time t, respectively, Δt represents the time difference, and Δt∈[0,1];

[0015] Step 1.3: Use equation (1-2) to [dep 0_t ,img 0_t ] T Perform data preprocessing to obtain the preprocessed object depth image [dep] at the current time t. 0_t_pro,img 0_t_pro ] T :

[0016]

[0017] In equation (1-2): f represents the focal length of the depth camera. The depth image resolution of the depth camera along the x-axis direction. Let C1 and C2 be the depth image resolution of the depth camera along the y-axis, where C1 and C2 are unequal constants and neither is zero.

[0018] Step 1.4: Use equation (1-3) to obtain the pose feature E of the quadcopter UAV at the current time t. t_cur :

[0019]

[0020] In equation (1-3): ε is the zero bias error of the observation process, τ is the observation noise, G is the state transition matrix, and we have:

[0021]

[0022] The Re-LSTM network in step 2 includes, in sequence: Input module, Exact module, Forget module, Update module, Hidden layer module, and Output module.

[0023] Step 2.1: The input module consists of P basic units connected in parallel;

[0024] The pose feature E t_cur Each basic unit of the input module is processed separately to obtain the basic pose feature set E at the current time t. t_pre ={E p,t_pre |p=1,2,..P}, where E p,t_pre This represents the basic pose feature output by the p-th basic unit at the current time t;

[0025] Step 2.2: The feature extraction module Exact consists of U parallel extraction units, where any u-th extraction unit is composed of k1 convolutional layers Conv and k2 pooling layers Pool in sequence.

[0026] E t_pre In the input feature extraction module Exact, the input features are first processed by k1 convolutional layers Conv of U extraction units to obtain the coarse pose feature set E at the current time t. t_Exact ={E u,t_Exact |u=1,2,..U}, where Eu,t_Exact This represents the coarse pose feature output by the u-th extraction unit at the current time t;

[0027] E t_Exact After processing by k2 pooling layers of u extraction units, the standard pose feature set E at the current time t is obtained. t_Exact ={E u,t_Pool |u=1,2,..U}, where E u,t_Pool This represents the standard pose feature of the current time t output by the u-th extraction unit;

[0028] Step 2.3: The Forget module consists of k3 feature compensation branches connected in parallel;

[0029] E u,t_pool The input is fed into k3 feature compensation branches, and the u-th compensated pose feature E at the current time t is generated using equation (2-1). u,t_Fus Thus, the compensated pose feature set E at the current time t is obtained. t_Fus ={E u,t_Fus |u=1,2,..U}:

[0030]

[0031] In equation (2-1), E u,i,t_Fus W represents the u-th compensated pose feature output at time t of the i-th feature compensation branch. i b is the weight of the i-th feature compensation branch. i It is the bias of the i-th feature compensation branch, and W i b i All are not 0; k represents the number of the feature compensation branch, and k3 represents the total number of feature compensation branches;

[0032] Step 2.4: The Update module consists of O parallel update units, where any o-th update unit is composed of g3 upsampling layers in sequence;

[0033] E t_Fus The samples are input into the O update units of the Update module for upsampling processing, resulting in the precise pose feature set E at the current time t. t_Pci ={E o,t_Pci |o=1,2..O}, where E o,t_Pci This represents the precise pose feature at time t output by the o-th update unit;

[0034] Step 2.5: The Hidden layer module consists of O hidden units connected in parallel, where any O-th hidden unit consists of an activation layer;

[0035] E t_Pci The inputs are processed into the O hidden units of the Hidden layer module, and the activation pose feature set E at the current time t is obtained using equation (2-2). t_Act ={E o,t_Act |o=1,2..O}:

[0036]

[0037] In equation (2-2), E o,t_Act This represents the activation pose feature output by the o-th hidden unit at the current time t;

[0038] Step 2.6: The Output module consists of R parallel output units, where any r-th output unit is composed of g4 upsampling layers and 1 fully connected layer in sequence.

[0039] E t_Act After processing the R output units in the Output module, the predicted pose feature set at the current time t is obtained. t_pred ={pose r,t_pred |r=1,2..R}, where pose r,t_pred This represents the predicted pose feature of the r-th output unit at the current time t.

[0040] Step 3 includes the following steps:

[0041] Step 3.1: Construct the pose prediction loss function Loss for the quadcopter UAV at the current time t using equation (3-1). t_pose :

[0042]

[0043] In equation (3-1): The depth pixel value dep represents the preprocessed object depth image at the current time t. 0_t_pro standard deviation The image represents the color pixel value of the preprocessed object depth image at the current time t. 0_t_pro Standard deviation, σ dep0_t_pro The depth pixel value dep represents the preprocessed object depth image at the current time t. 0_t_pro The variance, σ img0_t_pro The image represents the depth pixel value of the preprocessed object depth image at the current time t. 0_t_pro The variance of , |||| is the L2 norm; d1 and d2 represent the measurement noise contained in the depth pixels of the preprocessed object depth image at the current time t, and the measurement noise contained in the color pixels of the preprocessed object depth image at the current time t, respectively.

[0044] Step 3.2: Iteratively train the Re-LSTM network using the SGD optimizer and calculate the total loss function Loss. t_pose Adjust the network parameters until the total loss function is lost. t_pose The process continues until convergence, thus obtaining a trained quadcopter drone pose estimation model, which is used to estimate the current pose of the quadcopter drone in real time.

[0045] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the pose estimation method of the quadcopter unmanned aerial vehicle, and the processor is configured to execute the program stored in the memory.

[0046] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the quadcopter UAV pose estimation method.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. The feature extraction module Exact in the Re-LSTM network of this invention is composed of U parallel basic units, where any u-th basic unit consists of k1 convolutional layers (Conv) and k2 pooling layers (Pool) in sequence. This overcomes the deficiency of insufficient feature extraction by the input gate in LSTM networks, reduces the computational pressure on the Forget module during feature selection, and improves the overall network efficiency. In real-world scenarios, drones may face various dynamic environments, such as wind speed changes and airflow disturbances. The feature extraction module Exact can learn the changing patterns of historical pose data to extract features adapted to dynamic environments, thereby improving the robustness of pose estimation.

[0049] 2. The Forget module in the Re-LSTM network of this invention consists of g3 parallel feature compensation branches. This module can dynamically determine which branch's features to combine or ablate, preventing interference from too much old information, reducing the data computation burden, and overcoming the shortcomings of the original LSTM network that only eliminates features without combining them and has excessive algorithm consumption. It avoids the loss of important features, thus allowing the network to focus more on extracting sensor data that is beneficial for estimating UAV pose, thereby improving the accuracy of UAV pose estimation.

[0050] 3. The Hidden layer module in the Re-LSTM network of this invention consists of 0 parallel activation layers. This module can flexibly select the features contained in the time-stamped sensor data sequence, while overcoming the influence of various dynamic environments, such as airflow, wind speed changes, sensor noise, etc., capturing the temporal relationship between different sensors, discarding unimportant features or retaining and strengthening the extraction of related features, thereby better understanding the attitude changes of the UAV and more accurately predicting the future attitude characteristics of the UAV, thus improving the success rate of the UAV's autonomous navigation and autonomous obstacle avoidance. Attached Figure Description

[0051] Figure 1 This is an architecture diagram of the quadrotor UAV pose estimation method based on Re-LSTM network in this invention;

[0052] Figure 2 This is a flowchart of the quadrotor UAV pose estimation method based on Re-LSTM network in this invention;

[0053] Figure 3 This is a diagram of the Re-LSTM network structure in this invention. Detailed Implementation

[0054] In this embodiment, a quadcopter UAV pose estimation method based on a Re-LSTM network is described. (See [link to relevant documentation]). Figure 1 and Figure 2 Specifically, it includes the following steps:

[0055] Step 1: The data processing module collects and processes the pose data of the quadcopter UAV.

[0056] Step 1.1: Collect the attitude [q] of the quadcopter UAV at the current time t. 0_t ,q 1_t ,q 2_t ,q 3_t ] T and position coordinates [x 1_t ,y 1_t ,z 1_t ] T And the depth image of the object observed by the quadcopter UAV at the current time t [dep] 0_t ,img 0_t ] T Where T represents transpose, dep 0_t The image represents the depth pixel value of an object at the current time t. 0_t The color pixel value representing the depth image of the object at the current time t;

[0057] In this embodiment, system noise inherent in the inertial measurement unit (IMU) and the influence of magnetic fields during measurement can cause errors in the UAV's positioning. Inappropriate hardware parameter settings and inherent system noise in the depth camera can also lead to errors in the generated depth image, causing positioning errors for the UAV. Furthermore, GPS signal instability can cause deviations in the UAV's positioning. Therefore, preprocessing of the UAV pose data acquired by the IMU, depth camera, and GPS is necessary to improve the quality of the data source and ensure more accurate and reliable results in subsequent UAV pose estimation applications.

[0058] Step 1.2: Use equation (1-1) to [q] 0_t ,q 1_t ,q 2_t ,q 3_t ] T Data preprocessing is performed to obtain the preprocessed attitude [q] of the quadcopter UAV at the current time t. 0_t_pro ,q 1_t_pro ,q 2_t_pro ,q 3_t_pro ] T :

[0059]

[0060] In equation (1-1): α1, β1, γ1 are the yaw angle, pitch angle, and roll angle of the quadcopter UAV at the current time t, respectively, Δt represents the time difference, and Δt∈[0,1].

[0061] Step 1.3: Use equation (1-2) to [dep 0_t ,img 0_t ] T Perform data preprocessing to obtain the preprocessed object depth image [dep] at the current time t. 0_t_pro ,img 0_t_pro ] T :

[0062]

[0063] In equation (1-2): f represents the focal length of the depth camera. The depth image resolution of the depth camera along the x-axis direction. Let C1 and C2 be the depth image resolution of the depth camera along the y-axis, where C1 and C2 are unequal constants and neither is zero.

[0064] Step 1.4: Use equation (1-3) to obtain the pose feature E of the quadcopter UAV at the current time t. t_cur :

[0065]

[0066] In equation (1-3): ε is the zero bias error of the observation process, τ is the observation noise, G is the state transition matrix, and we have:

[0067]

[0068] Step 2, as follows Figure 3 As shown, a Re-LSTM network is established, which includes, in sequence: Input module, Exact module, Forget module, Update module, Hidden layer module, and Output module.

[0069] Step 2.1: The input module consists of P basic units connected in parallel;

[0070] The pose feature E t_cur Each basic unit of the input module is processed to obtain the basic pose feature set E at the current time t. t_pre ={E p,t_pre |p=1,2,..P}, where E p,t_pre This represents the basic pose feature output by the p-th basic unit at the current time t;

[0071] Step 2.2: The feature extraction module Exact consists of U parallel extraction units, where any u-th extraction unit is composed of k1 convolutional layers (Conv) and k2 pooling layers (Pool) in sequence. In this embodiment, u is 4, k1 is 2, and k2 is 2. The coarsely estimated current pose E of the quadcopter UAV is... t_cur After entering the convolutional layer, the input features are first decomposed into several pose features, thus allowing for the initial extraction of the input features E. t_cur The pose characteristics in the middle.

[0072] E t_pre In the input feature extraction module Exact, the input features are first processed by k1 convolutional layers Conv of U extraction units to obtain the coarse pose feature set E at the current time t. t_Exact ={E u,t_Exact |u=1,2,..U}, where E u,t_Exact This represents the coarse pose feature output by the u-th extraction unit at the current time t;

[0073] In this embodiment, any convolutional layer Conv is determined by T n It consists of several filters, each with a size of 3*3, and each filter contains 3*L... n There are 1 convolutional kernel, and the stride of each kernel is 1.

[0074] Et_Exact After processing by k2 pooling layers of u extraction units, the standard pose feature set E at the current time t is obtained. t_Exact ={E u,t_Pool |u=1,2,..U}, where E u,t_Pool Let represent the standard pose feature at the current time t output by the u-th extraction unit; in this embodiment, k2 is 2, and any pooling layer sequentially performs downsampling processing on the input coarse feature set, thereby changing the dimension of the pose feature to be processed and reducing the computational complexity for subsequent network layers.

[0075] Step 2.3: The Forget module consists of k3 feature compensation branches connected in parallel;

[0076] E u,t_pool The input is fed into k3 feature compensation branches, and the u-th compensated pose feature E at the current time t is generated using equation (2-1). u,t_Fus Thus, the compensated pose feature set E at the current time t is obtained. t_Fus ={E u,t_Fus |u=1,2,..U}:

[0077]

[0078] In equation (2-1), E u,i,t_Fus W represents the u-th compensated pose feature output at time t of the i-th feature compensation branch. i b is the weight of the i-th feature compensation branch. i It is the bias of the i-th feature compensation branch, and W i b i All are not 0; k represents the number of the feature compensation branch, and the numbers of the k3 feature compensation branches are 1, 2, 3...k,k+1,k3 in sequence.

[0079] In this embodiment, k3 is set to 4, thereby enabling the Forget module to effectively capture or discard the standard pose features of the UAV at the current time t and fuse them, thus enhancing the extraction of effective pose features.

[0080] Step 2.4: The Update module consists of O parallel update units, where any O-th update unit is composed of g3 upsampling layers in sequence. In this embodiment, O is 8 and g3 is 2, which allows for faster and more efficient integration of the pose features transmitted by the forgetting module.

[0081] E t_Fus The samples are input into the O update units of the Update module for upsampling processing, resulting in the precise pose feature set E at the current time t. t_Pci ={Eo,t_Pci |o=1,2..O}, where E o,t_Pci This represents the precise pose feature at time t output by the o-th update unit;

[0082] Step 2.5: The Hidden layer module consists of O basic units connected in parallel. Each of the Oth basic units is an activation layer; in this embodiment, O is set to 8, which allows more unextracted or ignored information to be added to the compensation features, thereby improving the accuracy of the compensation features in predicting the UAV's pose at the current time t.

[0083] E t_Pci The inputs are processed into the O hidden units of the Hidden layer module, and the activation pose feature set E at the current time t is obtained using equation (2-2). t_Act ={E o,t_Act |o=1,2..O}:

[0084]

[0085] In equation (2-2), E o,t_Act This represents the activation pose feature of the output of the o-th hidden unit at the current time t.

[0086] Step 2.6: The Output module consists of R parallel output units, where any r-th output unit is composed of g4 upsampling layers and 1 fully connected layer in sequence; in this embodiment, R is 6 and g4 is 3.

[0087] E t_Act After processing the R output units in the Output module, the predicted pose feature set at the current time t is obtained. t_pred ={pose r,t_pred |r=1,2..R}, where pose r,t_pred This represents the predicted pose feature of the r-th output unit at the current time t.

[0088] The size of any fully connected layer in the Output module is 3*3*L. n This allows for the full extraction and utilization of the pose information hidden in each activation feature, which can be used to assist in predicting the pose of the UAV at the current time t.

[0089] Step 3: Construct the loss function Loss t_pose And train the Re-LSTM network;

[0090] Step 3.1: Construct the pose prediction loss function Loss for the quadcopter UAV at the current time t using equation (3-1). t_pose :

[0091]

[0092] In equation (3-1): The depth pixel value dep represents the preprocessed object depth image at the current time t. 0_t_pro standard deviation The image represents the color pixel value of the preprocessed object depth image at the current time t. 0_t_pro standard deviation The depth pixel value dep represents the preprocessed object depth image at the current time t. 0_t_pro variance The image represents the depth pixel value of the preprocessed object depth image at the current time t. 0_t_pro The variance of , |||| is the L2 norm; d1 and d2 represent the measurement noise contained in the depth pixels of the preprocessed object depth image at the current time t, and the measurement noise contained in the color pixels of the preprocessed object depth image at the current time t, respectively.

[0093] Step 3.2: Iteratively train the Re-LSTM network using the SGD optimizer, ensuring that parameter updates are unaffected by gradient transformations, thereby efficiently calculating the loss function. t_pose And adjust the network parameters until the total loss function is lost. t_pose The training continues until convergence, resulting in a well-trained quadcopter drone pose estimation model, which is used to estimate the current pose of the quadcopter drone in real time.

[0094] In this embodiment, the SGD optimizer uses randomness quality to update the parameters of the Re-LSTM network, thereby avoiding getting trapped in local optima. Furthermore, the random selection of features in each iteration allows the SGD optimizer to quickly find the optimal solution.

[0095] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0096] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for pose estimation of a quadrotor UAV based on a Re-LSTM network, characterized in that, Includes the following steps: Step 1: Collect data from the quadcopter UAV at the current time t and preprocess it to obtain the pose features of the quadcopter UAV at the current time t. ; Step 2: Establish a Re-LSTM network and perform... The process is performed to obtain the predicted pose features at the current time t output by the r-th output unit. ; The Re-LSTM network comprises, in sequence: an input module (Input), a feature extraction module (Exact), a forget module (Forget), an update module (Update), a hidden layer module (Hidden), and an output module (Output). Step 2.1: The input module consists of parallel... It consists of several basic units; pose features Each basic unit of the input module is processed separately to obtain the basic pose feature set at the current time t. ,in, This represents the basic pose feature output by the p-th basic unit at the current time t; Step 2.2: The feature extraction module Exact consists of parallel... The extraction unit consists of k1 convolutional layers (Conv) and k2 pooling layers (Pool) in sequence. Will In the input feature extraction module Exact, the features are first processed by k1 convolutional layers Conv of U extraction units to obtain the coarse pose feature set at the current time t. ,in, This represents the coarse pose feature output by the u-th extraction unit at the current time t; After processing by k2 pooling layers of u extraction units, the standard pose feature set at time t is obtained. ,in, This represents the standard pose feature of the current time t output by the u-th extraction unit; Step 2.3: The Forget module consists of k3 feature compensation branches connected in parallel; Will The input is fed into k3 feature compensation branches, and the u-th compensated pose feature at the current time t is generated using equation (2-1). Thus, the set of compensated pose features at the current time t is obtained. : (2-1) In equation (2-1), W i It is the first The weights of each feature compensation branch, b i It is the bias of the i-th feature compensation branch, and W i b i All are not 0; k represents the number of the feature compensation branch, and k3 represents the total number of feature compensation branches; Step 2.4: The Update module consists of O parallel update units, where any o-th update unit is composed of g3 upsampling layers in sequence; Will The samples are input into the O update units of the Update module for upsampling processing to obtain the precise pose feature set at the current time t. ,in, This represents the precise pose feature at time t output by the o-th update unit; Step 2.5: The Hidden layer module consists of O hidden units connected in parallel, where any O-th hidden unit consists of an activation layer; Will The inputs are processed into the O hidden units of the Hidden layer module, and the set of activation pose features at the current time t is obtained using equation (2-2). : (2-2) In equation (2-2), This represents the activation pose feature output by the o-th hidden unit at the current time t; Step 2.6: The Output module consists of R parallel output units, where any r-th output unit is composed of g4 upsampling layers and 1 fully connected layer in sequence. Will After processing by R output units in the Output module, the predicted pose feature set at the current time t is obtained. ,in, This represents the predicted pose feature of the r-th output unit at the current time t. Step 3: Construct the loss function A Re-LSTM network is trained to obtain a trained quadcopter drone pose estimation model, which is used to estimate the current pose of the quadcopter drone in real time.

2. The method for pose estimation of a quadrotor UAV based on a Re-LSTM network according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Collect the attitude of the quadcopter UAV at the current time t. and position coordinates And the depth image of the object observed by the quadcopter drone at the current time t. Where T represents transpose, This represents the depth pixel value of the object's depth image at the current time t. The color pixel value representing the depth image of the object at the current time t; Step 1.2, using equation (1-1) to... Data preprocessing is performed to obtain the preprocessed attitude of the quadcopter UAV at the current time t. : (1-1) In equation (1-1): Let yaw angle, pitch angle and roll angle of the quadcopter UAV at the current time t, where Δt represents the time difference and Δt∈[0,1]; Step 1.3, using equation (1-2) to... Perform data preprocessing to obtain the preprocessed object depth image at the current time t. : (1-2) In equation (1-2): f represents the focal length of the depth camera. The depth image resolution of the depth camera along the x-axis direction. Let be the depth image resolution of the depth camera along the y-axis. , These are distinct constants, and none of them are zero; Step 1.4: Use equation (1-3) to obtain the pose characteristics of the quadcopter UAV at the current time t. : (1-3) In equation (1-3): For the zero bias error of the observation process, To observe the noise, Let be the state transition matrix, and we have: (1-4)。 3. The method for pose estimation of a quadrotor UAV based on a Re-LSTM network according to claim 2, characterized in that, Step 3 includes the following steps: Step 3.1: Construct the pose prediction loss function of the quadcopter UAV at the current time t using equation (3-1). : (3-1) In equation (3-1): This represents the depth pixel value of the preprocessed object depth image at the current time t. standard deviation This represents the color pixel value of the preprocessed object depth image at the current time t. standard deviation This represents the depth pixel value of the preprocessed object depth image at the current time t. variance This represents the depth pixel value of the preprocessed object depth image at the current time t. The variance, ||| is the L2 norm; , These represent the measurement noise contained in the depth pixels of the preprocessed object depth image at the current time t, and the measurement noise contained in the color pixels of the preprocessed object depth image at the current time t, respectively. Step 3.2: Iteratively train the Re-LSTM network using the SGD optimizer and calculate the total loss function. Adjust the network parameters until the total loss function is reached. The process continues until convergence, thus obtaining a trained quadcopter drone pose estimation model, which is used to estimate the current pose of the quadcopter drone in real time.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the pose estimation method of any one of claims 1-3 for a quadcopter UAV, and the processor is configured to execute the programs stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the quadcopter UAV pose estimation method according to any one of claims 1-3.

Images