Unmanned aerial vehicle aerial photography parameter dynamic optimization method based on AI model

Abstract

The invention belongs to the technical field of unmanned aerial vehicle aerial photography, and particularly discloses an unmanned aerial vehicle aerial photography parameter dynamic optimization method based on an AI model, and the method comprises the following steps: obtaining a current waypoint coordinate and a photographing point coordinate of an unmanned aerial vehicle, and calculating the actual distances between the two, including a linear distance, a horizontal distance and a vertical distance; an input feature is constructed, the actual distance is combined with an environment feature and a target feature, the environment feature comprises illumination intensity and a weather type, and the target feature comprises a shooting target type and a target size ratio; processing the input features by using a pre-trained AI model, and outputting optimal zoom multiple, pan-tilt pitch angle and exposure parameters; and dynamically adjusting the aerial photography parameters of the unmanned aerial vehicle according to the optimal parameters. The unmanned aerial vehicle aerial image quality can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle aerial photography, and more particularly to an unmanned aerial vehicle aerial photography parameter dynamic optimization method based on an AI model. BACKGROUND

[0002] With the popularization of unmanned aerial vehicle aerial photography technology, the requirement for aerial image quality is increasingly improved. The aerial image quality directly depends on the rationality of the shooting parameters. The traditional unmanned aerial vehicle needs to preset the zoom ratio, gimbal pitch angle, and exposure amount and the like manually in advance. When the unmanned aerial vehicle flight trajectory is complex and the distance between the aerial point and the shooting point dynamically changes, manual operation is difficult to adapt in real time, which easily leads to problems such as overexposure, underexposure, target composition deviation, and detail blur in the shooting picture. Some unmanned aerial vehicles use a fixed parameter adjustment algorithm, which only makes a simple linear adjustment according to single height or distance information. The zoom, gimbal pitch angle, and exposure amount are coupled, for example, when the zoom ratio increases, the light quantity will decrease, and the exposure amount needs to be adjusted synchronously to ensure the picture brightness. The prior art often adjusts a single parameter, ignoring the synergistic effect between parameters, which leads to poor overall shooting effect.

[0003] Therefore, how to improve the unmanned aerial vehicle aerial image quality is a problem to be solved at present. SUMMARY

[0004] In view of the defects of the prior art, the purpose of the application is to provide an unmanned aerial vehicle aerial photography parameter dynamic optimization method based on an AI model, which can improve the unmanned aerial vehicle aerial image quality.

[0005] To achieve the above purpose, in a first aspect, the application provides an unmanned aerial vehicle aerial photography parameter dynamic optimization method based on an AI model, comprising the following steps: S10, obtaining the current aerial point coordinates and shooting point coordinates of the unmanned aerial vehicle, calculating the actual distance therebetween, including the straight-line distance, horizontal distance, and vertical distance; S20, constructing input features, combining the actual distance with environmental features and target features, the environmental features including the light intensity and weather type, and the target features including the target type and target size proportion; S30, processing the input features by using a pre-trained AI model, and outputting the optimal zoom ratio, gimbal pitch angle, and exposure amount parameters; S40, dynamically adjusting the aerial photography parameters of the unmanned aerial vehicle according to the optimal parameters, the zoom ratio determining the initial zoom ratio according to the straight-line distance and correcting it in combination with the target size proportion, the gimbal pitch angle calculating the initial pitch angle according to the horizontal distance and the vertical distance and adjusting it in combination with the target type, and the exposure amount calculating the basic exposure value according to the light intensity and correcting it in combination with the zoom ratio and the weather type.

[0006] As a further preferred, in step S40, the initial zoom factor f init The calculation formula is: f init =k×D, where k is the distance-zoom ratio coefficient, and D is the straight-line distance. In the zoom factor correction, if the target size ratio is less than 10%, increase 10%-20% based on the initial zoom factor f init If the target size ratio is greater than 50%, decrease 10%-20%.

[0007] As a further preferred, in step S40, the correction of the zoom factor also includes hardware limiting the final value to be within the physical zoom range of the UAV camera.

[0008] As a further preferred, in step S40, the initial tilt angle The calculation formula is: =arctan(Dv / Dh), where Dh is the horizontal distance and Dv is the vertical distance. In the gimbal tilt angle adjustment, if the building facade is shot, the tilt angle is corrected to -10°-0°, and if the terrain is shot, the tilt angle is corrected to -30°- -60°.

[0009] As a further preferred, in step S40, the adjustment of the gimbal tilt angle also includes setting the angle change rate threshold to not more than 5° / s to avoid the gimbal rotating quickly causing the picture to shake.

[0010] As a further preferred, in step S40, the calculation formula of the base exposure value is: = , where L is the light intensity. In the exposure amount correction, the exposure amount value is increased by 0.3 for every 1x zoom increase; meanwhile, it is adjusted according to the weather type.

[0011] As a further preferred, in step S40, the correction of the exposure amount also includes decomposing the corrected exposure value into a combination of shutter speed, ISO sensitivity, and aperture, prioritizing the shutter speed, followed by the ISO, and fixing the aperture to the maximum aperture.

[0012] As a further preferred, in step S30, the AI model adopts a CNN-LSTM hybrid neural network structure, which includes: Convolutional layer for extracting spatial information in environmental features and target features; Long short-term memory layer for processing time sequence features with varying distances over time; Fully connected layer for outputting the optimal parameters.

[0013] As a further preferred, in step S30, the training data set of the AI model contains aerial photography data samples under different distances, different environments and different targets, each sample is labeled with input features and artificial labeled optimal parameters, the mean square error loss function is used to calculate the error between the model output parameters and the artificial labeled optimal parameters, and the Adam optimizer is used for training optimization.

[0014] In a second aspect, the present application provides an AI model-based dynamic optimization system for aerial photography parameters of a UAV, comprising: A waypoint-shooting point distance calculation module is configured to obtain the current waypoint coordinates and shooting point coordinates of the UAV, and calculate the actual distance between the two, including straight-line distance, horizontal distance and vertical distance. An AI parameter optimization model is configured to build input features, and then use a pre-trained AI model to process the input features to output optimal zoom ratio, gimbal pitch angle and exposure parameters; the input features include actual distance and environmental features, target features, the environmental features include light intensity and weather type, and the target features include target type and target size proportion A parameter execution control module is configured to dynamically adjust the aerial photography parameters of the UAV according to the optimal parameters, the zoom ratio is determined according to the initial zoom ratio of the straight-line distance and combined with the target size proportion for correction, the gimbal pitch angle is calculated according to the initial pitch angle of the horizontal distance and the vertical distance and combined with the target type for adjustment, and the exposure is calculated according to the basic exposure value of the light intensity and combined with the zoom ratio and the weather type for correction. The AI model-based dynamic optimization method and system for aerial photography parameters of a UAV provided by the present application have the following effects: by calculating the various distance information between the waypoint and the shooting point in real time, combining the environmental light, weather conditions and target type and size, and using the trained AI model to realize the collaborative optimization and dynamic adjustment of the zoom ratio, gimbal pitch angle and exposure, the defects of traditional methods that rely only on single height or distance information for linear adjustment and ignore the coupling relationship between parameters are overcome, thereby effectively avoiding problems such as overexposure, underexposure, composition shift and detail blur, and significantly improving the overall quality and adaptability of aerial images. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the AI model-based dynamic optimization method for aerial photography parameters of a UAV provided by the present application; Figure 2 is a parameter adjustment flowchart provided by an embodiment of the present application; Figure 3 is an architecture diagram of the AI model-based dynamic optimization system for aerial photography parameters of a UAV provided by an embodiment of the present application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] like Figure 1 As shown, this application provides a method for dynamic optimization of UAV aerial photography parameters based on an AI model, including steps S10 to S40, which are detailed below: Step S10: Obtain the current waypoint coordinates of the drone and the coordinates of the shooting point, and calculate the actual distance between them, including straight-line distance, horizontal distance, and vertical distance. This step is used to provide core input features for the AI ​​model. In step S10, coordinate acquisition: the three-dimensional coordinates of the current waypoint can be obtained in real time through the high-precision GPS module mounted on the UAV. ), Coordinates of the shooting point ( This information can be obtained through the real-time target detection module.

[0018] Distance Calculation: The linear distance D between two points can be calculated using the sampled spatial distance formula. The formula is: D = ; The formula for calculating the horizontal distance Dh is: Dh = ; The formula for calculating the vertical distance Dv is: Dv = | | is used for the precise calculation of subsequent gimbal pitch angle and zoom parameters.

[0019] Step S20: Construct input features by combining actual distance with environmental and target features to form a multi-dimensional feature vector that comprehensively reflects the real-time status of the shooting scene.

[0020] Specifically, the environmental features provided in this application include light intensity and weather type, and the target features include the type of target being photographed and the target size ratio.

[0021] Step S30: The input features are processed using a pre-trained AI model to output the optimal zoom magnification, gimbal tilt angle, and exposure parameters.

[0022] In step S30, the AI ​​model may adopt a CNN-LSTM hybrid neural network structure, which includes: a convolutional layer for extracting spatial information from environmental features and target features; a long short-term memory layer for processing temporal features of distance changing over time; and a fully connected layer for outputting the optimal parameters.

[0023] Step S40: Dynamically adjust the aerial photography parameters of the UAV according to the optimal parameters.

[0024] The zoom ratio is initially determined based on the straight-line distance and then adjusted according to the target size ratio. If the target size ratio is low, the zoom is increased to emphasize the subject; if the ratio is high, the zoom is decreased to preserve background information. The gimbal tilt angle is calculated based on the horizontal and vertical distances and adjusted according to the type of target being photographed. For example, when photographing buildings, the gimbal is raised to ensure a frontal view; when photographing terrain, the tilt is adjusted to capture the terrain features. The exposure is calculated based on the light intensity and then compensated for using the zoom ratio and weather type to adapt to changes in light intake and ambient lighting conditions.

[0025] Based on the same inventive concept, this application also provides a dynamic optimization system for UAV aerial photography parameters based on an AI model, including a waypoint-shooting point distance calculation module, an AI parameter optimization model, and a parameter execution control module.

[0026] The waypoint-shooting point distance calculation module is used to obtain the current waypoint coordinates and shooting point coordinates of the UAV, and calculate the actual distance between the two, including straight-line distance, horizontal distance and vertical distance.

[0027] The AI ​​parameter optimization model is used to construct input features, and then the pre-trained AI model is used to process the input features to output the optimal zoom magnification, gimbal tilt angle and exposure parameters.

[0028] In this application, the input features include actual distance and environmental features, and target features. The environmental features include light intensity and weather type, and the target features include the type of target being photographed and the target size ratio.

[0029] The parameter execution control module is used to dynamically adjust the aerial photography parameters of the UAV according to the optimal parameters. The zoom magnification is determined based on the straight-line distance and corrected according to the target size ratio. The gimbal pitch angle is calculated based on the horizontal and vertical distances and adjusted according to the type of shooting target. The exposure is calculated based on the light intensity and corrected according to the zoom magnification and weather type. It should be noted that the functions of each module provided in this application can be found in the steps described in the aforementioned optimization method, and will not be repeated here.

[0030] The AI-based method and system for dynamic optimization of drone aerial photography parameters provided in this application have the following effects: By calculating various distance information between waypoints and shooting points in real time, and combining multi-dimensional features such as ambient lighting, weather conditions, and target type and size, the trained AI model is used to achieve coordinated optimization and dynamic adjustment of zoom magnification, gimbal pitch angle, and exposure. This overcomes the shortcomings of traditional methods that rely solely on single altitude or distance information for linear adjustment and ignore the coupling relationship between parameters, thereby effectively avoiding problems such as overexposure, underexposure, composition shift, and blurry details, and significantly improving the overall quality and adaptability of aerial images.

[0031] In one embodiment, the technical solution to achieve the above objective can be as follows: This embodiment provides a dynamic optimization system for UAV aerial photography parameters based on an AI model, which includes three parts: a waypoint-shooting point distance calculation module, an AI parameter optimization model, and a parameter execution control module.

[0032] The waypoint-shooting point distance calculation module provided in this embodiment is used to obtain the current waypoint coordinates and shooting point coordinates of the UAV in real time, and calculate the actual distance between the two, providing core input features for the AI ​​model. Coordinate acquisition: The three-dimensional coordinates of the current waypoint are obtained through a high-precision GPS module mounted on the UAV. ), Coordinates of the shooting point ( This information can be obtained through the real-time target detection module. Distance Calculation: The actual straight-line distance D is calculated using the spatial distance formula, which is: D= Horizontal distance is calculated separately. Dh= and vertical distance Dv=∣ | is used for the precise calculation of subsequent gimbal pitch angle and zoom parameters.

[0033] The AI ​​parameter optimization model provided in this embodiment is used to construct input features, and then the pre-trained AI model is used to process the input features to output the optimal zoom magnification, gimbal tilt angle and exposure parameters.

[0034] In this embodiment, the AI ​​model can adopt a CNN-LSTM hybrid neural network structure, taking "distance features + environmental features + target features" as input, and outputting the optimal zoom magnification, gimbal tilt angle, and exposure parameters. The model training and inference process is as follows: (1) Input feature construction: Key features: straight-line distance D from waypoint to shooting point, horizontal distance Dh, vertical distance Dv; Environmental characteristics: Light intensity and weather type obtained through the light sensor mounted on the drone; Target characteristics: The type of the target being photographed and the percentage of its size, obtained through the image recognition module. (2) Model structure design: Convolutional layers: Three convolutional kernels are used to extract spatial information from environmental and target features; Long Short-Term Memory Layer: Two LSTM layers are set up to process the temporal features of distance changing over time, avoiding frequent fluctuations in parameter adjustments; Fully connected layer: Set 3 output nodes to correspond to the zoom magnification (f), gimbal pitch angle (θ), and exposure (E) respectively. The output range is limited by the activation function. (3) Model training: Dataset construction: Aerial photography data was collected at different distances, in different environments, and with different targets, totaling 100,000+ samples. Each sample was labeled with "input features + manually labeled optimal parameters". Loss function: The mean squared error loss function is used to calculate the error between the model output parameters and the manually labeled optimal parameters. The formula is as follows: Loss=1 / 3[ + + ] Training optimization: The Adam optimizer is used, with the learning rate initially set to 0.001 and decreasing by 10% every 50 epochs. Training stops when the loss function value is below 0.01. The model inference time is ≤100ms, which meets the real-time requirements.

[0035] The parameter execution control module provided in this embodiment is used to perform secondary calibration based on hardware characteristics after the AI ​​model outputs parameters, such as... Figure 2 As shown, to ensure the parameters are executable, the optimization logic is as follows: (1) Zoom magnification optimization: Basic mapping: The initial zoom magnification is determined based on the straight-line distance D, using the formula f. init =k×D, where k is the distance-zoom ratio factor and D is the straight-line distance; AI Correction: The AI ​​model corrects f based on the target size ratio. init If the target percentage is less than 10%, then in f init Increase by 10%-20% on the basis; if it exceeds 50%, decrease by 10%-20%. Hardware limitations: The final zoom level must be limited to the physical zoom range of the drone camera; if it exceeds the range, the maximum value will be used. (2) Gimbal pitch angle optimization: Initial angle calculation: Calculate the initial pitch angle based on the horizontal distance Dh and the vertical distance Dv, using the following formula: =arctan(Dv / Dh); AI Correction: Adjusted according to the type of shooting target. When shooting building facades, θ is corrected to -10°~0° (to ensure frontal shooting). When shooting terrain, θ is corrected to -30°~-60°. Smooth control: Set the angle change rate threshold to ≤5° / s to avoid image jitter caused by rapid gimbal rotation.

[0036] (3) Exposure optimization: Basic Exposure Calculation: Calculate the basic EV value based on the illumination intensity L, using the following formula: = ; AI Correction: Combined with zoom magnification f-correction, because the amount of light entering the camera decreases as the zoom increases, the EV value increases by 0.3 for every 1x increase in zoom; it also adjusts according to weather type; Parameter decomposition: Decompose the EV value into a combination of shutter speed, ISO sensitivity, and aperture. Prioritize adjusting the shutter speed, then adjust the ISO, and keep the aperture at its maximum to ensure sufficient light intake.

[0037] The beneficial effects of this embodiment are as follows: it can automate the adjustment of drone parameters without human intervention; the AI ​​model can dynamically adjust parameters according to distance to adapt to complex flight trajectories and improve aerial photography efficiency; through multi-parameter collaborative optimization, it can solve the problems of overexposure, underexposure, and composition shift, and improve the clarity and color reproduction of aerial images by more than 30%; combined with environmental and target characteristics, it can adapt to a variety of scenarios and cope with high-speed drone flight scenarios.

[0038] The following describes specific embodiments, such as Figure 3 The present application will be described as shown.

[0039] (a) Hardware Architecture Unmanned Aerial Vehicle (UAV) Platform: A multi-rotor UAV equipped with high-precision GPS and an IMU (Inertial Measurement Unit); Camera and gimbal: Features an optical zoom camera and a three-axis stabilized gimbal; Sensor module: integrates a light sensor and an image acquisition module; Computing module: Equipped with an edge computing chip, used for running AI model inference and parameter calculation, with a computing power of ≥21 TOPS. (II) Software Architecture Data Acquisition Layer: Real-time acquisition of UAV waypoint coordinates, shooting point coordinates, light intensity, and image data; Preprocess the collected data. Model inference layer: Load the trained CNN-LSTM model; read the input features every 100ms, input them into the model for inference, and output the zoom level, gimbal tilt angle, and exposure parameters. Parameter control layer: Convert the parameters output by the AI ​​model into hardware control commands; The system monitors the parameter execution status in real time, obtains the actual zoom magnification through camera feedback, compares it with the target value, and resends the control command if the error exceeds 5%. (III) Implementation Steps (1) Model training phase: Collected over 100,000 sample data, covering different distances, lighting conditions, and target types; A CNN-LSTM model was built using the PyTorch framework, trained until the loss function value was below 0.01, converted to ONNX format, and deployed to an edge computing chip. (2) Flight preparation phase: Preset the shooting point coordinates in the drone ground station software and adjust the threshold by setting parameters; The drone is launched, GPS positioning and sensor calibration are completed, and the AI ​​model is loaded onto the edge computing chip.

[0040] (3) Aerial photography execution phase: The drone flies along a preset route, calculating the distance between the current waypoint and the shooting point every 100ms; The image acquisition module identifies the type and size ratio of the target being photographed, and the light sensor obtains the real-time light intensity. These features are then input into the AI ​​model. The AI ​​model outputs the optimal parameters, and the parameter control layer sends instructions to the camera and gimbal to achieve dynamic adjustment of zoom, pitch angle, and exposure. If the target is a moving target, the target detection module updates the coordinates of the shooting point in real time and repeats the above parameter adjustment process. Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for dynamic optimization of drone aerial photography parameters based on an AI model, characterized in that, Includes the following steps: S10: Obtain the current waypoint coordinates and shooting point coordinates of the drone, and calculate the actual distance between them, including straight-line distance, horizontal distance and vertical distance; S20, construct input features by combining the actual distance with environmental features and target features. The environmental features include light intensity and weather type, and the target features include the type of target being photographed and the target size ratio. S30: The input features are processed using a pre-trained AI model to output the optimal zoom magnification, gimbal tilt angle, and exposure parameters. S40, dynamically adjust the aerial photography parameters of the UAV according to the optimal parameters. The zoom factor is determined based on the straight-line distance and corrected in combination with the target size ratio. The gimbal pitch angle is calculated based on the horizontal and vertical distances and adjusted in combination with the type of shooting target. The exposure is calculated based on the light intensity and corrected in combination with the zoom factor and weather type.

2. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1, characterized in that, In step S40, the initial zoom factor f init The calculation formula is: f init =k×D, where k is the distance-zoom ratio factor and D is the straight-line distance; In zoom magnification correction, if the target size accounts for less than 10%, then at the initial zoom magnification f... init Increase by 10% to 20% on the basis; if the target size accounts for more than 50%, decrease by 10% to 20%.

3. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1 or 2, characterized in that, In step S40, the correction of the zoom factor also includes hardware limiting the final value to keep it within the physical zoom range of the drone camera.

4. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1, characterized in that, In step S40, the initial pitch angle The calculation formula is: =arctan(Dv / Dh), where Dh is the horizontal distance and Dv is the vertical distance; When adjusting the tilt angle of the gimbal, if shooting a building facade, the tilt angle should be corrected to -10° to 0°, and if shooting terrain, the tilt angle should be corrected to -30° to -60°.

5. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1 or 4, characterized in that, In step S40, the adjustment of the gimbal pitch angle also includes setting an angle change rate threshold of no more than 5° / s to avoid rapid gimbal rotation causing image jitter.

6. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1, characterized in that, In step S40, the base exposure value The calculation formula is: = , where L is the light intensity; In exposure correction, the exposure value increases by 0.3 for every 1x increase in zoom; it is also adjusted according to the weather type.

7. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1 or 6, characterized in that, In step S40, the exposure correction also includes decomposing the corrected exposure value into a combination of shutter speed, ISO sensitivity, and aperture, prioritizing the adjustment of shutter speed, then adjusting ISO, and fixing the aperture to the maximum aperture.

8. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1, characterized in that, In step S30, the AI ​​model adopts a CNN-LSTM hybrid neural network structure, which includes: Convolutional layers are used to extract spatial information from environmental and target features; Long Short-Term Memory (LSTM) layer is used to process temporal features where distance changes over time; A fully connected layer is used to output the optimal parameters.

9. The method for dynamic optimization of UAV aerial photography parameters as described in claim 1, characterized in that, In step S30, the training dataset of the AI ​​model contains aerial photography data samples under different distances, environments, and targets. Each sample is labeled with input features and manually labeled optimal parameters. The mean squared error loss function is used to calculate the error between the model output parameters and the manually labeled optimal parameters, and the Adam optimizer is used for training optimization.

10. A dynamic optimization system for drone aerial photography parameters based on an AI model, characterized in that, include: The waypoint-shooting point distance calculation module is used to obtain the current waypoint coordinates and shooting point coordinates of the UAV, and calculate the actual distance between the two, including straight-line distance, horizontal distance and vertical distance; An AI parameter optimization model is used to construct input features. Then, a pre-trained AI model processes these input features to output optimal zoom magnification, gimbal tilt angle, and exposure parameters. The input features include actual distance and environmental features, as well as target features. Environmental features include light intensity and weather type, while target features include the type of target being photographed and the target size percentage. The parameter execution control module is used to dynamically adjust the aerial photography parameters of the UAV according to the optimal parameters. The zoom magnification is determined based on the straight-line distance and corrected in combination with the target size ratio. The gimbal pitch angle is calculated based on the horizontal and vertical distances and adjusted in combination with the type of shooting target. The exposure is calculated based on the light intensity and corrected in combination with the zoom magnification and weather type.

Images