Unmanned aerial vehicle take-off and landing parking apron

By adopting deep learning technology and positioning marking modules in the take-off and landing apron, the precise positioning and safe take-off and landing of the drone are achieved, solving the problem of inaccurate positioning in the existing technology and improving the launch of the drone Reduced safety and efficiency.

CN118918499APending Publication Date: 2024-11-08GUANGDONG CHENGJIN TECH CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410950085.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing drone take-off and landing aprons lack accuracy and intelligence in positioning and navigation, and are susceptible to human and environmental factors, resulting in inaccurate positioning and even accidents.

Method used

A drone take-off and landing apron was designed, including a physical layer, a perception layer and a maintenance layer. The physical layer uses durable, anti-slip and waterproof materials. The perception layer is accurately positioned through positioning marking modules, machine learning modules and apron identification modules. The machine learning module uses deep learning technology to train image recognition models, identify the contours and positioning marks of the apron, and provide accurate landing coordinate information. The maintenance layer promptly detects and deals with obstacles and surface problems through the abnormal detection module.

Benefits of technology

It improves the accuracy and safety of the take-off and landing of drones, reduces the impact of human factors and environmental factors on positioning, improves operating efficiency, reduces the error rate of human operations, and promptly ensures the safe take-off and landing of drones.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918499A_ABST
    Figure CN118918499A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of unmanned aerial vehicle take-off and landing positioning, and discloses an unmanned aerial vehicle take-off and landing parking apron which comprises a physical layer, a sensing layer and a maintenance layer. The physical layer comprises a parking apron body, and the parking apron body is made of a durable material, so that the stability and durability of the parking apron are ensured; the sensing layer realizes accurate positioning and intelligent navigation of the unmanned aerial vehicle through a positioning marking module, a machine learning module and a parking apron identification module, and effectively avoids the influence of human and environmental factors on positioning; the maintenance layer monitors the state of the parking apron in real time through an image receiving module, an anomaly detection module and a maintenance notification module, timely finds and processes the problems of obstacles, surface damage or pollution and the like, and ensures the safety and availability of the parking apron; through the intelligent technology, the accuracy and safety of taking-off and landing of the unmanned aerial vehicle are remarkably improved, the accident risk is reduced, meanwhile, the maintenance efficiency of the parking apron is improved, and powerful support is provided for wide application of the unmanned aerial vehicle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) take-off and landing positioning, and in particular to an unmanned aerial vehicle (UAV) take-off and landing apron. Background Art

[0002] With the rapid development and widespread application of drone technology, drone landing aprons are important infrastructure for drone take-off and landing. Their performance and intelligence level directly affect the safe take-off and landing of drones and their operating efficiency. Existing aprons lack accuracy and intelligence in drone positioning and navigation. They usually rely on simple physical markings or manual operations to assist drone positioning. This method is not only inefficient, but also easily affected by human and environmental factors, resulting in inaccurate positioning and even accidents. Therefore, we propose a drone landing apron. Summary of the invention

[0003] The purpose of the present invention is to provide a drone take-off and landing apron to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a drone take-off and landing apron, comprising a physical layer, a perception layer and a maintenance layer;

[0005] The physical layer includes a helipad body for providing a take-off and landing platform for drones, and the helipad body is made of durable, non-slip and waterproof materials;

[0006] The perception layer includes a positioning marking module, a machine learning module and an apron recognition module;

[0007] The positioning mark module is used to set a contour mark and a positioning mark with a unique identifier on the apron body, the contour mark is used to indicate whether the current apron of the drone is available, and the positioning mark is used to provide accurate landing coordinate information to the drone;

[0008] The machine learning module uses deep learning technology to train an image recognition model, and the image recognition model includes an apron contour recognition model and an apron positioning mark recognition model. The contour recognition model is used to determine whether the current apron can land, and the positioning mark recognition model is used to obtain the landing coordinate information set by the positioning mark module on the apron body;

[0009] The apron recognition module is used to capture the image of the apron and obtain the apron that can be landed and its coordinate information; the apron recognition module further includes an image acquisition unit, an image preprocessing unit and an image analysis unit, the image acquisition unit is responsible for capturing the image of the apron in real time; the image preprocessing unit is used to perform denoising and enhancement processing on the captured image; the image analysis unit is used to pass the image output by the image preprocessing unit to the machine learning module to obtain the apron that can be landed and its coordinate information; after obtaining the coordinate information, the drone lands at the specified location according to the coordinates;

[0010] The maintenance layer includes an image receiving module, an abnormality detection module and a maintenance notification module;

[0011] The image receiving module is responsible for receiving the image data captured by the apron identification module; the anomaly detection module further analyzes and processes the received images, including obstacle detection and apron surface condition assessment, to determine whether there are obstacles, surface damage or pollution on the apron; the maintenance notification module is responsible for generating maintenance notifications and sending the notifications to relevant personnel or systems, so as to arrange staff to perform maintenance and repair of the apron in a timely manner.

[0012] Preferably, the positioning mark adopts a symbol design that is easy to identify and distinguish, including the use of a specific color, shape or pattern combination, and the outline mark and the positioning mark are provided with night navigation indicator lights on the edges of the mark symbols.

[0013] Preferably, the positioning mark set by the positioning mark module on the apron body is the longitude and latitude values ​​of the location of the apron.

[0014] Preferably, the specific steps of the machine learning module using deep learning technology to train the image recognition model include:

[0015] Step 1: Data Preparation

[0016] Step 1.1: Collect an image dataset of the apron through field photography. The image dataset contains images of the apron at different angles, different lighting conditions, and different backgrounds.

[0017] Step 1.2: Accurately label the collected apron image data to indicate the outline and positioning marks of the apron;

[0018] Step 1.3: Divide the labeled image data into a training set and a validation set. The training set is used to train the deep learning model, and the validation set is used to evaluate the performance of the model.

[0019] Step 2: Build a deep learning model

[0020] Step 2.1: Use convolutional neural networks to build contour recognition models and positioning marker recognition models. Both the contour recognition model and the positioning marker recognition model contain multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is responsible for extracting features from the image, the pooling layer is used to reduce the dimension and complexity of the features, and the fully connected layer maps the learned features to the final classification results.

[0021] Step 2.2: Initialize the model’s weights and bias parameters;

[0022] Step 3: Train the Deep Learning Model

[0023] Step 3.1: Train the model using the training set;

[0024] Step 3.2: Input the image data into the model through the forward propagation algorithm and calculate the output of each layer;

[0025] Step 3.3: Use the set cross entropy loss function to calculate the difference between the model prediction and the true annotation;

[0026] Step 3.4: Update the weights and biases of the model based on the gradient information of the loss function through the back-propagation algorithm and the gradient descent optimizer;

[0027] Step 3.5: Perform training iteratively for multiple times until the model converges or reaches the preset number of training rounds.

[0028] Step 4: Evaluate and optimize the model:

[0029] Step 4.1: Evaluate the trained model using the validation set.

[0030] Step 4.2: Calculate the model’s accuracy, recall, and F1 score evaluation metrics;

[0031] Step 4.3: Optimize the model based on the evaluation results, including adjusting the network structure, integrating data enhancement techniques, and modifying hyperparameters.

[0032] Preferably, the contour recognition model is a binary classification model for determining whether the apron is available for landing, and the cross entropy loss function used in its training process is:

[0033]

[0034] Among them, y i Represents the true label of the i-th sample, with a value of 0 or 1. Represents the model's predicted probability for the i-th sample, N is the total number of samples, and this function is used to calculate the difference between the predicted probability distribution and the true label distribution.

[0035] Preferably, the image acquisition unit is arranged at the bottom of the drone, and the image acquisition unit comprises:

[0036] A high-definition camera, with its lens facing the apron, is used to capture high-definition images of the apron;

[0037] An image sensor, connected to the high-definition camera, for converting an optical image captured by the high-definition camera into an electrical signal;

[0038] An image processing circuit is connected to the image sensor and is used to receive the electrical signal of the image sensor and convert it into digital image data;

[0039] An anti-shake device is used to keep the HD camera stable during the flight of the drone;

[0040] The data transmission interface is used to transmit the digital image data to the image analysis unit and the image receiving module for further processing.

[0041] Preferably, the image preprocessing unit uses Gaussian filtering to eliminate noise in the image by replacing the value of each pixel in the image with a weighted average of the pixel values ​​in its neighborhood:

[0042]

[0043] Where G(x,y) is the value of the Gaussian function at the two-dimensional coordinate (x,y) of the image, x and y are the coordinates in two-dimensional space, σ is the standard deviation of the Gaussian distribution, e is the base of the natural logarithm, and π is the circumference of a circle;

[0044] The weighted average grayscale value of the surrounding pixels is calculated for each pixel of the image through the calculation rule of G(x,y), and this value is used to replace the grayscale value of the current pixel.

[0045] Preferably, the image preprocessing unit performs image enhancement in the following manner:

[0046] The Sobel detection operator is used to perform convolution operations on the image and calculate the gradient strength and direction of each pixel. Each pixel in the image is traversed and its gradient strength is compared with the gradient strength of the adjacent pixels in its gradient direction. If the current pixel is not a local maximum, its gradient strength is set to zero, thereby eliminating unmarked pixels and reducing marking artifacts.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The positioning mark module sets unique contour marks and positioning marks on the apron body, providing accurate landing coordinate information for the drone. Compared with traditional physical marks or manual operation methods, the positioning mark of the present invention is more accurate and stable, effectively reducing the impact of human and environmental factors on positioning.

[0049] The machine learning module uses deep learning technology to train image recognition models, which can automatically determine whether the current apron can be landed and obtain accurate landing coordinate information. Through intelligent recognition methods, it not only improves operational efficiency but also reduces the error rate of manual operations.

[0050] The anomaly detection module of the maintenance layer can detect obstacles and assess the surface condition of the apron, and promptly detect and deal with obstacles, surface damage or pollution on the apron, etc., which helps to ensure the safe takeoff and landing of drones and reduce the risk of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the overall structure diagram of the present invention;

[0052] Figure 2 Diagram of the steps for building and training an image recognition model;

[0053] Figure 3 A diagram of the steps for image enhancement to be implemented by the image preprocessing unit;

[0054] Figure 4 This is an example picture of the apron body. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See also Figure 1-4 ,The present invention provides a technical solution: a UAV take-off and landing apron, including a physical layer, a perception layer and a maintenance layer.

[0057] The physical layer mainly includes a helipad body for providing a take-off and landing platform for drones. The helipad body is made of durable, non-slip and waterproof materials to ensure that the helipad can maintain stable performance under various weather and environmental conditions, providing a safe and reliable take-off and landing environment for drones.

[0058] The perception layer is the core part of the intelligent and precise positioning of the apron, which includes a positioning marking module, a machine learning module and an apron identification module.

[0059] The positioning mark module is specifically used to set special contour marks and positioning marks on the apron body. These marks have unique identifications and are easy for drones to identify. Among them, the main function of the contour mark is to indicate to the drone whether the current apron is available, so as to avoid the drone taking off and landing on an unavailable or dangerous apron, while the positioning mark is used to provide the drone with accurate landing coordinate information to help the drone land accurately and safely at the designated location.

[0060] The machine learning module uses deep learning technology to train the image recognition model, which further includes a helipad contour recognition model and a helipad positioning mark recognition model. The main task of the contour recognition model is to determine whether the current helipad is suitable for drone landing, while the positioning mark recognition model is used to accurately obtain the landing coordinate information set by the positioning mark module on the helipad body.

[0061] The apron recognition module has the main function of capturing the image of the apron and obtaining the apron and its coordinate information from it. The apron recognition module further includes an image acquisition unit, an image preprocessing unit and an image analysis unit. The image acquisition unit is responsible for capturing the image of the apron in real time, and the image preprocessing unit performs denoising and enhancement processing on the captured image to improve the quality of the image. Finally, the image analysis unit passes the preprocessed image to the machine learning module, and accurately obtains the apron and its coordinate information that can be landed through the trained image recognition model. After obtaining this information, the drone can land accurately at the specified location according to the provided coordinates.

[0062] The maintenance layer is mainly responsible for the daily maintenance and abnormality detection of the apron, and includes an image receiving module, an abnormality detection module and a maintenance notification module.

[0063] The image receiving module is responsible for receiving the image data captured by the apron recognition module, and these data will be used for subsequent anomaly detection.

[0064] The anomaly detection module further analyzes and processes the received images, including obstacle detection to ensure that there are no obstacles on the apron that may affect the take-off and landing of the drone; it also includes an apron surface condition assessment, such as detecting whether the apron surface is damaged or contaminated, which may affect the safe take-off and landing of the drone.

[0065] Once an abnormality is detected, the maintenance notification module will generate a maintenance notification and send the notification to relevant personnel or systems, ensuring that problems with the apron can be discovered and handled in a timely manner, thereby ensuring the safe take-off and landing of drones and operational efficiency.

[0066] The present invention will be further described below in conjunction with Examples 1 to 3:

[0067] Embodiment 1:

[0068] See attached Figure 4 The positioning mark adopts a specific symbol design, which is easy for the drone to recognize the image. The positioning mark uses high-contrast, easy-to-distinguish colors, such as bright red or blue as the main color of the mark to ensure that it can be clearly identified under various lighting conditions. In addition to color, the shape and pattern of the contour mark are designed with specific geometric shapes, such as the H-shaped symbol design. These shapes and patterns have a high degree of distinction in image recognition, which helps the drone to accurately lock the landing coordinates.

[0069] The size of the apron is moderate, neither too large to affect the usable area of ​​the apron, nor too small to make it difficult to identify. At the same time, the layout of multiple positioning marks on the apron has been carefully planned to ensure that the drone can identify at least one clear mark from different angles.

[0070] In order to enhance visibility at night or in low-light conditions, contour marks and positioning marks are equipped with night flight indicator lights on the edges of the mark symbols. These indicator lights use low-power, high-brightness light sources such as LEDs, and use specific flashing frequencies or color changes to enhance recognition. The setting of indicator lights not only improves the safety of drone takeoff and landing at night, but also facilitates manual visual confirmation.

[0071] The positioning mark module sets the positioning mark on the apron body, which is specifically the longitude and latitude values ​​of the apron location, such as longitude: 113.426793, latitude: 22.526470. This setting method can accurately identify the geographical location of the apron, facilitating accurate navigation and landing of drones or other aircraft.

[0072] Embodiment 2:

[0073] The specific steps of the machine learning module using deep learning technology to train the image recognition model include:

[0074] Step 1: Data Preparation

[0075] Step 1.1: Collect a dataset of images of the apron by field shooting

[0076] Select multiple different aprons as data collection points to ensure data diversity and generalization. Shoot at different time periods (such as morning, noon, and evening) to capture images of the apron under different lighting conditions. Shoot the apron from multiple angles, including front, side, and top views, to simulate the various perspectives that the drone may encounter in actual flight. Shoot in different weather conditions, such as sunny, cloudy, and rainy days, to enhance the robustness of the model. Use a high-resolution camera or the drone's own camera to shoot to ensure image quality.

[0077] Step 1.2: Accurately label the collected apron image data

[0078] The apron outline in each image is accurately depicted to ensure the accuracy and completeness of the outline. The positioning mark is accurately marked inside the apron outline, including the center point, shape and size of the mark. Other interference factors (such as shadows, debris, etc.) that may exist in the image are also marked or removed as necessary.

[0079] Step 1.3: Divide the labeled image data into training set and validation set

[0080] According to the total amount and diversity of the data, the ratio of the training set and the validation set is reasonably divided to ensure that the data distribution in the training set and the validation set is similar to avoid model deviation during the training process. At the same time, the training set and the validation set are randomly shuffled to enhance the generalization ability of the model.

[0081] Step 2: Build a deep learning model

[0082] Step 2.1: Use convolutional neural network to build contour recognition model and positioning mark recognition model

[0083] Choose a network architecture: According to the specific needs of apron image recognition, choose a suitable convolutional neural network architecture as the basis, such as VGGNet or ResNet.

[0084] Design model structure: For contour recognition model and positioning marker recognition model, similar network structures can be designed, but they can also be fine-tuned according to the characteristics of each task. The model will contain multiple convolutional layers to extract features from the image; pooling layers to reduce the dimension of features and enhance the robustness of the model; and fully connected layers to map the learned features to the final classification or regression results.

[0085] Step 2.2: Initialize the model’s weights and bias parameters

[0086] Weight initialization: Initialize the model's weight parameters randomly before training begins.

[0087] Bias initialization: Similar to weights, bias parameters also need to be initialized. Initialize the bias to 0 or a small constant to avoid introducing too much bias at the beginning of training.

[0088] Step 3: Train the Deep Learning Model

[0089] Step 3.1: Train the model using the training set

[0090] Data loading and preprocessing: First, load the training dataset divided in step 1.3. For each image, perform preprocessing operations such as scaling and normalization to ensure that the input data meets the requirements of the model.

[0091] Step 3.2: Input the image data into the model through the forward propagation algorithm and calculate the output of each layer

[0092] Forward propagation: The preprocessed image data is used as input and forward propagated through the model. In this process, the data passes through each layer of the model in turn. Each layer processes the input data according to its parameters and activation function and generates output.

[0093] Step 3.3: Using the set cross entropy loss function, calculate the difference between the model prediction and the true annotation

[0094] Loss function calculation: In the output layer of the model, the cross entropy loss function is used to calculate the difference between the model prediction result and the true annotation. This loss value is used in the subsequent back propagation and optimization process.

[0095] Step 3.4: Update the model’s weights and biases based on the gradient information of the loss function through the back-propagation algorithm and gradient descent optimizer

[0096] Back propagation and optimization: Based on the error calculated by the loss function, the gradient of each layer parameter is calculated through the back propagation algorithm. Then, the Adam gradient descent optimizer is used to update the weight and bias parameters of the model based on these gradient information. This process minimizes the loss function and makes the model's prediction results closer to the true annotation.

[0097] Step 3.5: Iterate the training multiple times until the model converges or reaches the preset number of training rounds

[0098] Iterative training: Repeat steps 3.2 to 3.4, i.e., forward propagation, loss calculation, back propagation, and parameter update, until the model's performance on the validation set reaches the preset standard (such as accuracy and loss value), or reaches the preset maximum number of training rounds. During the training process, the model's weights and bias parameters can be saved periodically for subsequent evaluation and use.

[0099] Step 4: Evaluate and optimize the model

[0100] Step 4.1: Evaluate the trained model using the validation set

[0101] Data loading and preprocessing: First, load the validation dataset divided in step 1.3. Similar to the training dataset, perform necessary preprocessing operations on each image in the validation dataset to ensure the consistency and accuracy of the input data.

[0102] Model evaluation: The preprocessed validation data is input into the trained model, and the model’s prediction results are obtained through forward propagation. These prediction results are compared with the true annotations in the validation dataset to evaluate the performance of the model.

[0103] Step 4.2: Calculate the model’s precision, recall, and F1 score evaluation metrics

[0104] Evaluation indicator calculation: Based on the model's prediction results and true annotations, the model's accuracy, recall, and F1 score evaluation indicators are calculated. These indicators are used to comprehensively reflect the model's performance in identifying apron contours and locating marks.

[0105] Precision: The ratio of samples predicted to be positive and actually positive to all samples predicted to be positive, reflecting the accuracy of the model's prediction of positive samples.

[0106] Recall: The ratio of samples predicted to be positive and actually positive to all samples actually positive, which reflects the ability of the model to find all positive samples.

[0107] F1 score: The harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model.

[0108] Step 4.3: Optimize the model based on the evaluation results

[0109] Model optimization strategy: Based on the evaluation indicators calculated in step 4.2, analyze the shortcomings of the model in identifying the apron contour and positioning the markers. In response to these shortcomings, formulate corresponding optimization strategies, such as adjusting the network structure, adding data enhancement techniques, or modifying hyperparameters.

[0110] Adjust the network structure: Based on the evaluation results, you can try to increase or decrease the number of convolutional layers, pooling layers, or fully connected layers to adjust the complexity and learning ability of the model.

[0111] Data enhancement: By introducing more data enhancement techniques (such as rotation, scaling, translation, etc.), the generalization ability of the model is increased, so that it can better cope with changes in various practical scenarios.

[0112] Modify hyperparameters: Adjust hyperparameters such as learning rate, batch size, optimizer, etc. to find a model configuration that is more suitable for the current task.

[0113] Iterative optimization and verification: After implementing the optimization strategy, retrain the model and verify it using the validation data set. Through multiple iterative optimization and verification processes, the performance of the model is gradually improved until a satisfactory recognition effect is achieved.

[0114] Here is a brief code example of building and training an image recognition model using Python as the programming language:

[0115] #Import TensorFlow library and related modules

[0116] import tensorflow as tf

[0117] from tensorflow.keras.models import Model

[0118] from tensorflow.keras.layers import Input,Conv2D,MaxPooling2D,Flatten,Dense

[0119] from tensorflow.keras.optimizers import Adam

[0120] #Define the function to build a deep learning model

[0121] def build_model(input_shape):

[0122] #Create an input layer and specify the shape of the input

[0123] inputs=Input(shape=input_shape)

[0124] #Add a convolutional layer, using 32 3x3 convolution kernels, the activation function is ReLU, and the edge padding is the same to maintain the spatial size

[0125] x=Conv2D(32,(3,3),activation='relu',padding='same')(inputs)

[0126] #Add the maximum pooling layer with a pooling window of 2x2

[0127] x=MaxPooling2D(pool_size=(2,2))(x)

[0128] #Flattening operation, flattening the multi-dimensional input into one dimension so that it can be connected to the fully connected layer

[0129] x=Flatten()(x)

[0130] #Add a fully connected layer with 128 neurons and ReLU activation function

[0131] x=Dense(128,activation='relu')(x)

[0132] #Add output layer, use softmax activation function for multi-classification, num_classes is the total number of categories

[0133] predictions=Dense(num_classes,activation='softmax')(x)

[0134] #Build the model, specify input and output

[0135] model=Model(inputs=inputs,outputs=predictions)

[0136] #Return the constructed model

[0137] return model

[0138] #Set model parameters

[0139] input_shape = (256, 256, 3) #The size of the input image is 256x256 pixels, 3 color channels

[0140] num_classes = 2 # Assume that the classification task has 2 categories, such as whether the apron can land

[0141] #Call function to build model

[0142] model=build_model(input_shape)

[0143] #Compile the model, specify the optimizer, loss function and evaluation index, use the Adam optimizer, set the learning rate to 0.0001, and the evaluation index is accuracy

[0144] model.compile(optimizer=Adam(learning_rate=0.0001),loss='categorical_crossentropy',metrics=['accuracy'])

[0145] #X_train is the preprocessed training image data, y_train is the corresponding label, the shape of X_train is (number of samples, image height, image width, number of color channels) y_train should be one-hot encoded, and the shape is (number of samples, number of categories)

[0146] #Use the fit method to train the model, specify the training data, batch size, and number of training rounds (epochs). The validation_split parameter means that 20% of the training data is divided as a validation set

[0147] history=model.fit(X_train,y_train,batch_size=32,epochs=10,validation_split=0.2)

[0148] TensorFlow in the above code is an open source machine learning library that is widely used in deep learning and other machine learning applications. It allows developers to build and train neural network models and provides a large number of tools and APIs to simplify and accelerate the process.

[0149] The contour recognition model is a binary classification model that determines whether the apron can be landed. The cross entropy loss function used in its training process is:

[0150]

[0151] Among them, y i Represents the true label of the i-th sample, with a value of 0 or 1, representing two possible categories. It represents the predicted probability of the model for the i-th sample, the probability value is between 0 and 1, N is the total number of samples, and this function is used to calculate the difference between the predicted probability distribution and the true label distribution.

[0152] When i = 1, the loss function The term is 0, so the loss mainly focuses on That is, the logarithm of the probability that the model predicts the positive class. If is close to 1, the loss is small; if If it is close to 0, the loss is large.

[0153] When i = 0, the loss function The item is 0, and the loss is mainly concerned with That is, the logarithm of the probability that the model predicts a negative class. If is close to 0, the loss is small; if If it is close to 1, the loss is large.

[0154] For all N samples, their cross entropy losses are calculated separately and summed up. Finally, the average cross entropy loss is obtained by dividing by the number of samples N. This average value is used as the loss of the entire model on the current batch of data for subsequent gradient calculations and model updates.

[0155] By calculating the cross entropy loss function value, we can use optimization algorithms such as gradient descent to update the model parameters to minimize the loss function, thereby improving the accuracy of the model in the apron contour and positioning mark recognition tasks.

[0156] Embodiment 3:

[0157] The image acquisition unit is specifically set at the bottom of the drone to capture high-definition images of the apron. The following are the specific implementations of the image acquisition unit:

[0158] HD Camera: Choose a HD camera with high resolution and wide-angle field of view, and ensure that its lens is facing the apron to capture HD images of the apron. The camera should have autofocus and light correction functions to adapt to image capture under different lighting conditions.

[0159] Image sensor: The image sensor is closely connected to the high-definition camera and is responsible for converting the optical image captured by the camera into an electrical signal. High-performance CMOS or CCD sensors are selected to ensure high fidelity and sensitivity of image conversion.

[0160] Image processing circuit: The image processing circuit is connected to the image sensor, receives the electrical signal from the sensor, and converts it into digital image data through an analog-to-digital converter.

[0161] Anti-shake device: The anti-shake device is designed as a mechanical or electronic anti-shake system to reduce image blur caused by vibration or airflow interference during the flight of the drone. The motion state of the drone is monitored in real time through components such as gyroscopes and accelerometers, and the position of the camera or the reading speed of the image sensor are dynamically adjusted to maintain image stability.

[0162] Data transmission interface: The data transmission interface is designed as a high-speed, stable data transmission channel that supports wireless or wired connection. The interface should be compatible with multiple data transmission protocols to ensure that digital image data can be smoothly transmitted to the image analysis unit and image receiving module for further processing.

[0163] Implementation steps:

[0164] i) Installation and commissioning:

[0165] Install the HD camera at a suitable position on the bottom of the drone, making sure the lens is facing the landing pad and the field of view is unobstructed. Connect the image sensor, image processing circuit, and anti-shake device, and perform overall debugging to ensure that all components are working properly.

[0166] ii) Image capture and conversion:

[0167] When the drone flies near the apron, the high-definition camera begins to capture high-definition images of the apron. The image sensor converts the captured optical image into an electrical signal and transmits it to the image processing circuit.

[0168] iii) Image processing and transmission: After receiving the electrical signal, the image processing circuit generates digital image data through analog-to-digital conversion and applies image processing algorithms to optimize image quality. The processed digital image data is transmitted to the image analysis unit and image receiving module in real time through the data transmission interface.

[0169] iv) Stability guarantee:

[0170] During the flight of the drone, the anti-shake device continues to work to reduce the impact of vibration on image quality.

[0171] The image preprocessing unit uses Gaussian filtering to eliminate noise in the image by replacing the value of each pixel in the image with the weighted average of the pixel values ​​in its neighborhood:

[0172]

[0173] Where G(x,y) is the value of the Gaussian function at the two-dimensional coordinate (x,y) of the image, x and y are the coordinates in two-dimensional space, σ is the standard deviation of the Gaussian distribution, e is the base of the natural logarithm, and π is the circumference of a circle;

[0174] The weighted average grayscale value of the surrounding pixels is calculated for each pixel of the image through the calculation rule of G(x,y), and this value is used to replace the grayscale value of the current pixel.

[0175] The standard deviation determines the smoothness of the Gaussian filter, and the Gaussian kernel size determines the size of the image area covered by the filter. If the Gaussian kernel size is 3x3 and the standard deviation σ=1.0, the following is an example of the calculation result:

[0176]

[0177] Each element in the matrix G is a weight value. These weights are calculated according to the G(x,y) function and normalized so that the sum of all weights is 1. Applying the Gaussian kernel to each pixel of the image can obtain the pixel values ​​after filtering of all pixels.

[0178] The image preprocessing unit performs image enhancement in the following manner: performing a convolution operation on the image using a Sobel detection operator to calculate the gradient strength and direction of each pixel; traversing each pixel in the image, comparing its gradient strength with the gradient strength of adjacent pixels in its gradient direction, and if the current pixel is not a local maximum, setting its gradient strength to zero, thereby eliminating non-marked pixels and reducing marked artifacts.

[0179] The specific steps of image enhancement are:

[0180] Step 1. Define the Sobel operator: The Sobel operator contains two sets of 3x3 convolution kernels, one set for detecting horizontal edges and the other for detecting vertical edges.

[0181] Step 2: Convolution operation: Use these two sets of convolution kernels to perform convolution operation on the input image and calculate the gradient strength of each pixel in the horizontal and vertical directions respectively.

[0182] Step 3: Calculate the gradient strength and direction: Based on the convolution result, calculate the total gradient strength (obtained by the square root of the sum of the squares of the horizontal and vertical gradient strengths) and gradient direction (obtained by calculating the ratio of the horizontal and vertical gradient strengths using the inverse tangent function) of each pixel.

[0183] Step 4, non-local maximum suppression: traverse each pixel in the processed image, for each pixel, find the gradient strength of the neighboring pixels along its gradient direction, compare the gradient strength of the current pixel with the gradient strength of the neighboring pixels in its gradient direction, if the gradient strength of the current pixel is not the local maximum in its gradient direction, then set the gradient strength of the pixel to zero. The purpose of this step is to eliminate those pixels that are not edges or markers, thereby reducing noise and artifacts in the image.

[0184] Step 5, output result: After the above processing, an enhanced image is output, in which non-marked pixels are effectively eliminated and marked artifacts are significantly reduced.

[0185] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0186] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A drone take-off and landing apron, characterized by: Includes physical layer, perception layer and maintenance layer; The physical layer includes a helipad body for providing a take-off and landing platform for drones, and the helipad body is made of durable, non-slip and waterproof materials; The perception layer includes a positioning marking module, a machine learning module and an apron recognition module; The positioning mark module is used to set a contour mark and a positioning mark with a unique identifier on the apron body, the contour mark is used to indicate whether the drone can land on the current apron, and the positioning mark is used to provide the drone with accurate landing coordinate information; The machine learning module uses deep learning technology to train an image recognition model, and the image recognition model includes an apron contour recognition model and an apron positioning mark recognition model. The contour recognition model is used to determine whether the current apron can land, and the positioning mark recognition model is used to obtain the landing coordinate information set by the positioning mark module on the apron body; The apron recognition module is used to capture the image of the apron and obtain the apron that can be landed and its coordinate information; the apron recognition module further includes an image acquisition unit, an image preprocessing unit and an image analysis unit, the image acquisition unit is responsible for capturing the image of the apron in real time; the image preprocessing unit is used to perform denoising and enhancement processing on the captured image; the image analysis unit is used to pass the image output by the image preprocessing unit to the machine learning module to obtain the apron that can be landed and its coordinate information; after obtaining the coordinate information, the drone lands at the specified location according to the coordinates; The maintenance layer includes an image receiving module, an abnormality detection module and a maintenance notification module; The image receiving module is responsible for receiving the image data captured by the apron recognition module; The anomaly detection module further analyzes and processes the received images, including obstacle detection and apron surface condition assessment, to determine whether there are obstacles, surface damage or pollution on the apron; the maintenance notification module is responsible for generating maintenance notifications and sending the notifications to relevant personnel or systems, so as to arrange staff to maintain and repair the apron in a timely manner.

2. The UAV take-off and landing apron according to claim 1, characterized in that: The positioning mark adopts a symbol design that is easy to identify and distinguish, including the use of a specific color, shape or pattern combination, and the outline mark and the positioning mark are both provided with night navigation indicator lights on the edges of the mark symbols.

3. The UAV take-off and landing apron according to claim 2, characterized in that: The positioning mark set by the positioning mark module on the apron body is the longitude and latitude values ​​of the location of the apron.

4. The UAV take-off and landing apron according to claim 3, characterized in that: The specific steps of the machine learning module using deep learning technology to train the image recognition model include: Step 1: Data Preparation Step 1.1: Collect an image dataset of the apron through field photography. The image dataset contains images of the apron at different angles, different lighting conditions, and different backgrounds. Step 1.2: Accurately label the collected apron image data to indicate the outline and positioning marks of the apron; Step 1.3: Divide the labeled image data into a training set and a validation set. The training set is used to train the deep learning model, and the validation set is used to evaluate the performance of the model. Step 2: Build a deep learning model Step 2.1: Use convolutional neural networks to build contour recognition models and positioning marker recognition models. Both the contour recognition model and the positioning marker recognition model contain multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is responsible for extracting features from the image, the pooling layer is used to reduce the dimension and complexity of the features, and the fully connected layer maps the learned features to the final classification results. Step 2.2: Initialize the model’s weights and bias parameters; Step 3: Train the Deep Learning Model Step 3.1: Train the model using the training set; Step 3.2: Input the image data into the model through the forward propagation algorithm and calculate the output of each layer; Step 3.3: Use the set cross entropy loss function to calculate the difference between the model prediction and the true annotation; Step 3.4: Update the weights and biases of the model based on the gradient information of the loss function through the back-propagation algorithm and the gradient descent optimizer; Step 3.5: Perform training iteratively for multiple times until the model converges or reaches the preset number of training rounds. Step 4: Evaluate and optimize the model: Step 4.1: Evaluate the trained model using the validation set. Step 4.2: Calculate the model’s accuracy, recall, and F1 score evaluation metrics; Step 4.3: Optimize the model based on the evaluation results, including adjusting the network structure, integrating data enhancement techniques, and modifying hyperparameters.

5. The UAV take-off and landing apron according to claim 4, characterized in that: The contour recognition model is a binary classification model that determines whether the apron can be landed. The cross entropy loss function used in its training process is: Among them, y i Represents the true label of the i-th sample, with a value of 0 or 1. Represents the model's predicted probability for the i-th sample, N is the total number of samples, and this function is used to calculate the difference between the predicted probability distribution and the true label distribution.

6. The UAV take-off and landing apron according to claim 5, characterized in that: The image acquisition unit is arranged at the bottom of the drone, and the image acquisition unit comprises: A high-definition camera, with its lens facing the apron, is used to capture high-definition images of the apron; An image sensor, connected to the high-definition camera, for converting an optical image captured by the high-definition camera into an electrical signal; An image processing circuit is connected to the image sensor and is used to receive the electrical signal of the image sensor and convert it into digital image data; An anti-shake device is used to keep the HD camera stable during the flight of the drone; The data transmission interface is used to transmit the digital image data to the image analysis unit and the image receiving module for further processing.

7. The UAV take-off and landing apron according to claim 6, characterized in that: The image preprocessing unit uses Gaussian filtering to eliminate noise in the image by replacing the value of each pixel in the image with the weighted average of the pixel values ​​in its neighborhood: Where G(χ,y) is the value of the Gaussian function at the two-dimensional coordinate (x,y) of the image, x and y are the coordinates in two-dimensional space, σ is the standard deviation of the Gaussian distribution, e is the base of the natural logarithm, and π is the circumference of a circle; The weighted average grayscale value of the surrounding pixels is calculated for each pixel of the image through the calculation rule of G(×,y), and this value is used to replace the grayscale value of the current pixel.

8. The UAV take-off and landing apron according to claim 7, characterized in that: The image preprocessing unit performs image enhancement in the following manner: The Sobel detection operator is used to perform convolution operations on the image and calculate the gradient strength and direction of each pixel. Each pixel in the image is traversed and its gradient strength is compared with the gradient strength of the adjacent pixels in its gradient direction. If the current pixel is not a local maximum, its gradient strength is set to zero, thereby eliminating unmarked pixels and reducing marking artifacts.

Citation Information

Cited By

  • Column switch intelligent early warning and emergency disposal system and method

    CN120912762A