A method for eliminating low-quality UAV inspection photos by integrating support vector machine

Abstract

The present invention discloses a method for eliminating low-quality inspection photos taken by drones that integrates a support vector machine. The method includes the following processes: labeling and preprocessing the collected inspection photos to form a training set; training a deep residual network and a support vector machine classifier in combination with the training set; collecting inspection photos to be tested and preprocessing them, then inputting the inspection photos to be tested into the deep residual network to obtain feature vectors, then inputting the feature vectors into the support vector machine classifier to determine the labels of the inspection photos to be tested and then performing classification detection; eliminating low-quality inspection photos and outputting the attribute information of the low-quality inspection photos. The present invention extracts feature vectors through a deep residual network, and then distinguishes the inspection photos through a support vector machine classifier. The process is simple and effective, and improves the accuracy and efficiency of eliminating low-quality inspection photos.

Description

Technical Field

[0001] The present invention relates to a method for eliminating low-quality inspection photos in the field of image processing, and in particular to a method for eliminating low-quality inspection photos of unmanned aerial vehicles (UAVs) by integrating a support vector machine. Background Art

[0002] With the rapid development of the national economy, the scale of my country's power grid construction continues to expand. Furthermore, transmission line equipment operates in complex and changing outdoor environments year-round, exposing itself to various defects and hidden dangers that can affect safe operation. Therefore, to ensure the safe and reliable operation of transmission lines, line maintenance personnel are required to regularly inspect transmission lines to promptly identify and eliminate defects that could affect normal operation.

[0003] Traditional inspection methods rely on inspectors to inspect transmission and distribution lines visually or by climbing towers. Inspectors travel along the line corridors, then, upon reaching the poles, use binoculars and handheld infrared detection instruments to survey the towers from the bottom up. Alternatively, they ascend the towers and use equipment such as electric field detectors. This inspection method is particularly challenging in harsh conditions, such as those across rivers or in mountainous areas. Not only is it labor-intensive and challenging, but the uneven quality of personnel also leads to frequent missed and false detections, resulting in extremely low inspection efficiency. Therefore, it is imperative to effectively leverage artificial intelligence theories and technologies to achieve intelligent operations and maintenance.

[0004] The use of drone inspections significantly improves the efficiency of on-site inspections of power transmission lines. This refined inspection process involves the acquisition and classification of massive amounts of photo data. During this acquisition process, factors such as the drone equipment and the environment can result in some low-quality inspection photos, such as overexposure, darkness, and blurriness. To prevent defects from being missed due to the use of low-quality photos, which can lead to untimely repairs and accidents, it is crucial to promptly identify and remove low-quality inspection photos from the database and retake clearer photos of the locations. Summary of the Invention

[0005] To eliminate low-quality inspection photos collected by drone inspections, such as those that are overexposed or underexposed, this paper provides a method for eliminating low-quality drone inspection photos by integrating a support vector machine. This method extracts feature vectors from inspection photos by building a deep residual network. These feature vectors are then processed using a support vector machine algorithm to achieve classification and detection of the inspection photos. Finally, the identified low-quality inspection photos are eliminated, and their attribute information is output.

[0006] The technical solution adopted in the present invention is:

[0007] The method comprises the following steps:

[0008] Step 1: A training set is composed of pre-processed and labeled inspection photos and their labels. The inspection photos are obtained by fixed-point collection of tower parts by drones. The labeled inspection photos include low-quality inspection photos and normal inspection photos. The low-quality inspection photos are inspection photos with an average pixel value less than 50 or an average pixel value greater than 200, and the normal inspection photos are inspection photos with an average pixel value not less than 50 and not greater than 200. The parts of the tower include the tower head, tower base sign, large and small side line corridors, left and right side ground wires, left and right upper phase parts, left and right middle phase parts, and left and right lower phase parts.

[0009] Step 2: Train the deep residual network and support vector machine classifier based on the training set;

[0010] Step 3: Collect and preprocess the inspection photos to be tested. Then, input the preprocessed inspection photos to the deep residual network to obtain the feature vector of the inspection photos to be tested. Then, input the feature vector of the inspection photos to be tested into the support vector machine classifier to determine the label of the inspection photos to be tested, and then classify the inspection photos to determine whether they are low-quality inspection photos or normal inspection photos;

[0011] Step 4: Eliminate the low-quality inspection photos detected in step 3 and output the attribute information of the low-quality inspection photos.

[0012] The preprocessing is to perform image cropping and image rotation on the inspection photos that have been marked with labels.

[0013] The step 2 is specifically as follows:

[0014] 3.1. The inspection photos in the training set are sequentially fed into a deep residual network. The weights of the deep residual network are iteratively updated using a mutual entropy loss function based on backpropagation to minimize the loss function. Once the loss function converges, the feature vector output by the deep residual network is fed into a support vector machine classifier.

[0015] 3.2. Combined with the labels of the inspection photos, the feature vectors of the input support vector machine classifier are transformed using the Gaussian kernel function. The Gaussian kernel function K() is as follows:

[0016]

[0017] Among them, X i represents the feature vector of the i-th inspection photo in the training set, and i∈[1,m], m represents the number of inspection photos in the training set; y i represents the label of the i-th inspection photo in the training set; γ represents the hyperparameter of the Gaussian kernel function, which takes a value of 1;

[0018] The eigenvector after transformation by the Gaussian kernel function is:

[0019]

[0020] Among them, X' i is the feature vector of the i-th inspection photo after transformation by Gaussian kernel function;

[0021] 3.3. Combine the labels of the inspection photos and the eigenvectors converted by the Gaussian kernel function obtained in step 3.2, and use the coordinate descent algorithm to iteratively solve the separating hyperplane equation to determine the optimal parameters of the separating hyperplane equation, and then obtain the optimal classification function. The optimal parameters of the separating hyperplane equation include the eigenvector α of the inspection photo corresponding to the minimum solution of the separating hyperplane equation. * and the average bias vector b * ;

[0022] The separating hyperplane equation is specifically:

[0023] ω T X' i +b=0

[0024] Where ω is the parameter matrix; ω T is the transposed matrix of the parameter matrix ω; b is the initial bias vector;

[0025] The parameter matrix is ​​specifically determined by the following formula:

[0026]

[0027] The constraints on the optimal parameters of the separating hyperplane equation are specifically:

[0028]

[0029] Where, X j represents the feature vector of the jth inspection photo in the training set, and j∈[1,m]; y j is the label of the jth inspection photo in the training set; C is the penalty coefficient, which is 50; min is the minimization function; st is the constraint function;

[0030] The support vector machine classifier is composed of the above iteratively solved separating hyperplane equation and the optimal classification function, which is specifically:

[0031]

[0032]

[0033] Where, α * b is the eigenvector of the inspection photo corresponding to the minimum solution of the separating hyperplane equation;* is the average bias vector; X is the feature vector of the inspection photo to be tested; y is the label of the inspection photo to be tested; f() is the optimal classification function, and f(X) = 1 or -1. If f(X) = 1, the inspection photo to be tested is a low-quality inspection photo. If f(X) = -1, the inspection photo to be tested is a normal inspection photo.

[0034] Among them, the average bias vector b * It is determined by the following formula:

[0035]

[0036] Where s is the number of eigenvectors that satisfy the constraints, and 0 <s<m,b s is the bias vector of the eigenvector that satisfies the constraints;

[0037] Among them, the bias vector b of the feature vector that meets the constraint condition s It is determined by the following formula:

[0038]

[0039] Where, X s is the eigenvector that satisfies the constraints, y s The label of the inspection photo corresponding to the feature vector that meets the constraints.

[0040] The deep residual network in step 2 is mainly composed of a first network stacking module, a second network stacking module, a third network stacking module, a fourth network stacking module and a fifth network stacking module; the first network stacking module, the second network stacking module, the third network stacking module, the fourth network stacking module and the fifth network stacking module are connected in series in sequence;

[0041] The first network stacking module and the second network stacking module have the same structure, both including two network layers, one addition layer and one pooling layer, and the two network layers, one addition layer and one pooling layer are connected in series in sequence, and each network layer is mainly composed of a convolution layer, a batch normalization layer and an activation function connected in series in sequence; a skip link structure is used inside the first network stacking module and the second network stacking module to fuse the inspection photos, and the skip link structure is to input the input of each network stacking module and the output of the second network layer of the stacking module into the addition layer at the same time for pixel addition processing, and then input them into the pooling layer for pooling processing;

[0042] The third network stacking module is mainly composed of three network layers, a 1×1 convolution layer and a pooling layer, and the three network layers, the 1×1 convolution layer and the pooling layer of the third network stacking module are connected in series in sequence, and each network layer of the third network stacking module has the same network layer structure as the first network stacking module;

[0043] The fourth network stacking module and the fifth network stacking module have the same structure, both including three network layers, a 1×1 convolution layer and a pooling layer, and each network layer in the fourth network stacking module and the fifth network stacking module is mainly composed of an asymmetric convolution layer, a batch normalization layer and an activation function connected in series in sequence.

[0044] The specific steps of step 4 are: automatically eliminating the detected low-quality inspection photos through software tools, and saving the low-quality inspection photos to a low-quality photo folder, then reading the attribute information of the low-quality inspection photos in the low-quality photo folder, and automatically organizing the attribute information of the low-quality inspection photos into an Excel table.

[0045] The attribute information includes the tower number of the inspection tower, the name of the shooting location, the longitude and latitude information and the altitude when the drone was shooting, the drone's yaw angle, the drone's pitch angle, the drone's roll angle, the gimbal's yaw angle, the gimbal's pitch angle, and the gimbal's roll angle. The beneficial effects of the present invention are:

[0046] This paper optimizes and improves a deep residual network to build an optimized deep residual network. This network extracts the feature vectors of inspection photos, processes them using a support vector machine, removes low-quality inspection photos, and outputs their attribute information. Compared to existing methods for classifying low-quality inspection photos, the proposed method is simple and effective, with high accuracy and processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 Schematic diagram of the overall structure of the deep residual network combined with the support vector machine of the present invention;

[0049] Figure 3 Schematic diagram of an Excel table according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, the method includes the following steps:

[0052] Step 1: The training set consists of pre-processed and labeled inspection photos and their labels. The inspection photos are obtained by drones from fixed-point acquisition of tower parts. The labeled inspection photos include low-quality inspection photos and normal inspection photos. Low-quality inspection photos are inspection photos with an average pixel value less than 50 or an average pixel value greater than 200, and normal inspection photos are inspection photos with an average pixel value of not less than 50 and not greater than 200. Among them, photos with an average pixel value less than 50 are underexposed, and photos with an average pixel value greater than 200 are overexposed. The parts of the tower include the tower head, tower base, line corridors on the large and small sides, left and right ground wires, upper and lower parts on the left and right sides, middle parts on the left and right sides, and lower parts on the left and right sides.

[0053] Step 2: Train the deep residual network and support vector machine classifier based on the training set;

[0054] Step 3: Collect and preprocess the inspection photos to be tested. Then, input the preprocessed inspection photos to the deep residual network to obtain the feature vector of the inspection photos to be tested. Then, input the feature vector of the inspection photos to be tested into the support vector machine classifier to determine the label of the inspection photos to be tested, and then classify the inspection photos to determine whether they are low-quality inspection photos or normal inspection photos;

[0055] Step 4: Eliminate the low-quality inspection photos detected in step 3 and output the attribute information of the low-quality inspection photos.

[0056] In step 1, 200 inspection photos were collected on cloudy days with insufficient light, sunny days with abundant light, and sunny days with normal light. Preprocessing consisted of cropping and rotating the labeled inspection photos. Cropping maintained the aspect ratio of the photos, using the center of the photos as the reference point to crop a region of the photos with a ratio of 0.6 or 0.8. Image rotation involved rotating the photos 90° or 180° clockwise, resulting in a total of 3,000 preprocessed inspection photos.

[0057] Step 2 is as follows:

[0058] 3.1. Inspection photos from the training set were sequentially fed into a deep residual network. The network's weights were iteratively updated using a mutual entropy loss function based on the feature vectors output by the network through backpropagation to minimize the loss function. After the loss function converged, the feature vectors output by the network were fed into a support vector machine classifier. The trained deep residual network outputs 160 feature vectors, each with a dimension of 169.

[0059] 3.2. Combined with the labels of the inspection photos, the feature vectors input into the support vector machine classifier are transformed using the Gaussian kernel function to convert the nonlinearly separable feature vectors into linearly separable feature vectors. The Gaussian kernel function K(X i ,y i ) are as follows:

[0060]

[0061] Among them, X i represents the feature vector of the i-th inspection photo in the training set, and i∈[1,m], m represents the number of inspection photos; y i represents the label of the i-th inspection photo in the training set; γ represents the hyperparameter of the Gaussian kernel function, which takes a value of 1;

[0062] The eigenvector after transformation by the Gaussian kernel function is:

[0063]

[0064] Among them, X' i is the feature vector of the i-th inspection photo after transformation by Gaussian kernel function;

[0065] 3.3. Combine the labels of the inspection photos and the eigenvectors converted by the Gaussian kernel function obtained in step 3.2, and use the coordinate descent algorithm to iteratively solve the separating hyperplane equation to determine the optimal parameters of the separating hyperplane equation, and then obtain the optimal classification function. The optimal parameters of the separating hyperplane equation include the eigenvector α of the inspection photo corresponding to the minimum solution of the separating hyperplane equation. * and the average bias vector b * ;

[0066] The separating hyperplane equation is specifically:

[0067] ω T X' i +b=0

[0068] Where ω is the parameter matrix; ω T is the transposed matrix of the parameter matrix ω; b is the initial bias vector; and the parameter matrix is ​​specifically determined by the following formula:

[0069]

[0070] The constraints on the optimal parameters of the separating hyperplane equation are specifically:

[0071]

[0072] Where, X j represents the feature vector of the jth inspection photo in the training set, and j∈[1,m]; yj is the label of the jth inspection photo in the training set; C is the penalty coefficient, which is 50; min is the minimization function; st is the constraint function;

[0073] The support vector machine classifier is composed of the above iteratively solved separation hyperplane equation and the optimal classification function, which is specifically:

[0074]

[0075]

[0076] Where, α * b is the eigenvector of the inspection photo corresponding to the minimum solution of the separating hyperplane equation; * is the average bias vector; X is the feature vector of the inspection photo to be tested; y is the label of the inspection photo to be tested; f() is the optimal classification function, and f(X) = 1 or -1. If f(X) = 1, the inspection photo to be tested is a low-quality inspection photo. If f(X) = -1, the inspection photo to be tested is a normal inspection photo.

[0077] Average bias vector b * It is determined by the following formula:

[0078]

[0079] Where S is the number of eigenvectors that satisfy the constraints, and 0 <s<m,b s is the bias vector of the eigenvector that satisfies the constraints;

[0080] The bias vector b of the eigenvector that satisfies the constraints s It is determined by the following formula:

[0081]

[0082] Where, X s is the eigenvector that satisfies the constraints, y s The label of the inspection photo corresponding to the feature vector that meets the constraints.

[0083] The binary classification problem of the inspection photo feature vectors is converted into the problem of solving the separating hyperplane equation based on the inspection photo feature vectors.

[0084] The deep residual network in step 2 is mainly composed of a first network stacking module, a second network stacking module, a third network stacking module, a fourth network stacking module and a fifth network stacking module; the first network stacking module, the second network stacking module, the third network stacking module, the fourth network stacking module and the fifth network stacking module are connected in series in sequence;

[0085] The first network stacking module and the second network stacking module have the same structure, both including two network layers, one addition layer and one pooling layer, and the two network layers, one addition layer and one pooling layer are connected in series in sequence, and each network layer is mainly composed of a convolution layer, a batch normalization layer and an activation function connected in series in sequence; a skip link structure is used inside the first network stacking module and the second network stacking module to fuse the inspection photos, and the skip link structure is to combine the input of each network stacking module with the output of the second network layer of the stacking module, that is, the output processed by the two network layers in sequence, and input them into the addition layer for pixel addition processing, and then input them into the pooling layer for pooling processing;

[0086] The third network stacking module mainly consists of three network layers, one 1×1 convolution layer and one pooling layer, and the three network layers, one 1×1 convolution layer and one pooling layer of the third network stacking module are connected in series in sequence, and each network layer of the third network stacking module has the same network layer structure as the first network stacking module;

[0087] The fourth network stacking module and the fifth network stacking module have the same structure, both including three network layers, a 1×1 convolution layer and a pooling layer, and each network layer in the fourth network stacking module and the fifth network stacking module is mainly composed of an asymmetric convolution layer, a batch normalization layer and an activation function connected in series.

[0088] like Figure 2 As shown in the figure, the deep residual network and the support vector machine classifier are connected in series, and the deep residual network combined with the support vector machine algorithm is used to classify and judge the pre-processed inspection photos.

[0089] Step 5 is specifically as follows: by using software tools, specifically using the os.remove() function in the Python program to automatically remove the low-quality inspection photos detected, and save the low-quality inspection photos to the low-quality photo folder, and then read the attribute information of the low-quality inspection photos in the low-quality photo folder, and automatically organize the attribute information of the low-quality inspection photos into an Excel table, such as Figure 3 shown.

[0090] The attribute information includes the tower number of the tower where the inspection photo was taken, the name of the shooting location, the latitude and longitude information and altitude when the drone was taking the photo, the drone's yaw angle, drone's pitch angle, drone's roll angle, gimbal yaw angle, gimbal pitch angle and gimbal roll angle.

[0091] This paper introduces a skip link structure into the deep residual network to integrate the features of inspection photos, building an improved deep residual network model. This model can make the characteristics of the inspection photo feature vector more distinct. By integrating the support vector machine algorithm to discriminate the feature vectors of inspection photos, the network training speed is greatly accelerated, shortening the training time by 1 / 4 compared to the original network model. In addition, the training accuracy rate is increased by nearly 8% when tested on the test set. Compared with the original network model, the method of the present invention improves the accuracy of calculation while reducing the amount of calculation.

[0092] Finally, it should be noted that the above embodiment and the proposed method are merely representative examples of the present invention. Obviously, the technical solution of the present invention is not limited to the above embodiment and the proposed control method, and many variations are possible. A person skilled in the art will be able to directly derive or make various improvements based on the present disclosure, and such improvements should also be within the scope of protection of the claims of this application.

Claims

1. A method for eliminating low-quality UAV inspection photos by integrating support vector machines, characterized by: The method comprises the following steps: Step 1: A training set is composed of pre-processed and labeled inspection photos and their labels. The inspection photos are obtained by fixed-point collection of tower parts by drones. The labeled inspection photos include low-quality inspection photos and normal inspection photos. The low-quality inspection photos are inspection photos with an average pixel value less than 50 or an average pixel value greater than 200, and the normal inspection photos are inspection photos with an average pixel value not less than 50 and not greater than 200. The parts of the tower include the tower head, tower base sign, large and small side line corridors, left and right side ground wires, left and right upper phase parts, left and right middle phase parts, and left and right lower phase parts. Step 2: Train the deep residual network and support vector machine classifier based on the training set; Step 3: Collect and preprocess the inspection photos to be tested. Then, input the preprocessed inspection photos to the deep residual network to obtain the feature vector of the inspection photos to be tested. Then, input the feature vector of the inspection photos to be tested into the support vector machine classifier to determine the label of the inspection photos to be tested, and then classify the inspection photos to determine whether they are low-quality inspection photos or normal inspection photos; Step 4: Eliminate the low-quality inspection photos detected in step 3 and output the attribute information of the low-quality inspection photos; The deep residual network in step 2 is mainly composed of a first network stacking module, a second network stacking module, a third network stacking module, a fourth network stacking module and a fifth network stacking module; the first network stacking module, the second network stacking module, the third network stacking module, the fourth network stacking module and the fifth network stacking module are connected in series in sequence; The first network stacking module and the second network stacking module have the same structure, both including two network layers, one addition layer and one pooling layer, and the two network layers, one addition layer and one pooling layer are connected in series in sequence, and each network layer is mainly composed of a convolution layer, a batch normalization layer and an activation function connected in series in sequence; a skip link structure is used inside the first network stacking module and the second network stacking module to fuse the inspection photos, and the skip link structure is to input the input of each network stacking module and the output of the second network layer of the stacking module into the addition layer at the same time for pixel addition processing, and then input them into the pooling layer for pooling processing; The third network stacking module is mainly composed of three network layers, a 1×1 convolution layer and a pooling layer, and the three network layers, the 1×1 convolution layer and the pooling layer of the third network stacking module are connected in series in sequence, and each network layer of the third network stacking module has the same network layer structure as the first network stacking module; The fourth network stacking module and the fifth network stacking module have the same structure, both including three network layers, a 1×1 convolution layer and a pooling layer, and each network layer in the fourth network stacking module and the fifth network stacking module is mainly composed of an asymmetric convolution layer, a batch normalization layer and an activation function connected in series in sequence.

2. The method for eliminating low-quality UAV inspection photos by integrating support vector machine according to claim 1 is characterized by: The preprocessing is to perform image cropping and image rotation on the inspection photos that have been marked with labels.

3. The method for eliminating low-quality UAV inspection photos by integrating support vector machine according to claim 1 is characterized by: The step 2 is specifically as follows: 3.

1. The inspection photos in the training set are sequentially fed into a deep residual network. The weights of the deep residual network are iteratively updated using a mutual entropy loss function based on backpropagation to minimize the loss function. Once the loss function converges, the feature vector output by the deep residual network is fed into a support vector machine classifier. 3.

2. Combined with the labels of the inspection photos, the feature vectors of the input support vector machine classifier are transformed using the Gaussian kernel function. The Gaussian kernel function K() is as follows: Among them, X i represents the feature vector of the i-th inspection photo in the training set, and i∈[1,m], m represents the number of inspection photos in the training set; y i represents the label of the i-th inspection photo in the training set; γ represents the hyperparameter of the Gaussian kernel function, which takes a value of 1; The eigenvector after transformation by the Gaussian kernel function is: Among them, X' i is the feature vector of the i-th inspection photo after transformation by Gaussian kernel function; 3.

3. Combine the labels of the inspection photos and the eigenvectors converted by the Gaussian kernel function obtained in step 3.2, and use the coordinate descent algorithm to iteratively solve the separating hyperplane equation to determine the optimal parameters of the separating hyperplane equation, and then obtain the optimal classification function. The optimal parameters of the separating hyperplane equation include the eigenvector α of the inspection photo corresponding to the minimum solution of the separating hyperplane equation. * and the average bias vector b * ; The separating hyperplane equation is specifically: oh T X' i +b=0 Where ω is the parameter matrix; ω T is the transposed matrix of the parameter matrix ω; b is the initial bias vector; The parameter matrix is ​​specifically determined by the following formula: The constraints on the optimal parameters of the separating hyperplane equation are specifically: Where, X j represents the feature vector of the jth inspection photo in the training set, and j∈[1,m]; y j is the label of the jth inspection photo in the training set; C is the penalty coefficient, which is 50; min is the minimization function; st is the constraint function; The support vector machine classifier is composed of the above iteratively solved separating hyperplane equation and the optimal classification function, which is specifically: Where, α * b is the eigenvector of the inspection photo corresponding to the minimum solution of the separating hyperplane equation; * is the average bias vector; X is the feature vector of the inspection photo to be tested; y is the label of the inspection photo to be tested; f() is the optimal classification function, and f(X) = 1 or -1. If f(X) = 1, the inspection photo to be tested is a low-quality inspection photo. If f(X) = -1, the inspection photo to be tested is a normal inspection photo. Among them, the average bias vector b * It is determined by the following formula: Where s is the number of eigenvectors that satisfy the constraints, and 0 <s<m,b s is the bias vector of the eigenvector that satisfies the constraints; Among them, the bias vector b of the feature vector that meets the constraint condition s It is determined by the following formula: Where, X s is the eigenvector that satisfies the constraints, y s The label of the inspection photo corresponding to the feature vector that meets the constraints.

4. The method for eliminating low-quality UAV inspection photos by integrating support vector machines according to claim 1 is characterized by: The specific steps of step 4 are: automatically eliminating the detected low-quality inspection photos through software tools, and saving the low-quality inspection photos to a low-quality photo folder, then reading the attribute information of the low-quality inspection photos in the low-quality photo folder, and automatically organizing the attribute information of the low-quality inspection photos into an Excel table.

5. The method for eliminating low-quality UAV inspection photos by integrating support vector machine according to claim 4 is characterized by: The attribute information includes the tower number of the tower where the inspection photo was taken, the name of the shooting location, the latitude and longitude information and altitude when the drone was shooting, the drone yaw angle, the drone pitch angle, the drone roll angle, the gimbal yaw angle, the gimbal pitch angle and the gimbal roll angle.

Images