1 of 38

UAV FOR SPRAYING APPLICATIONS WITH SIMULATIONS IN GAZEBO

UNDER THE GUIDANCE OF

Dr. Ashish Kumar Dhara

Assistant Professor

Department of Electrical Engineering

NIT Durgapur

A PRESENTATION BY

Deepayan Pal (19EE8003)

Rishab Dugar (19EE8016)

Sai Charan Allu (19EE8033)

Vemula Rahul (19EE8067)

National Institute of Technology, Durgapur

Department of Electrical Engineering

2 of 38

INTRODUCTION

Agriculture is a vital industry, and the demand for food is increasing globally.
The sector is also a significant contributor to India's GDP, accounting for around 17% of the country's total GDP.

3 of 38

CONCERNS

Farmers face several challenges, including labor shortages, soil degradation, and environmental concerns.
Erratic weather patterns and unpredictable rainfall affect crop yields
Crop failures and financial losses for farmers
Extreme weather events such as floods and droughts disrupt the agricultural cycle
Adverse health effects faced by farmers due to manual fertilization.

4 of 38

UAV’S

Among these adverse effects we choose Health effects faced by farmers to help come up with a solution.
Unmanned Aerial Vehicles (UAVs) or drones can be used for various agricultural applications, including fertilization.
The use of UAVs for fertilization can reduce labor costs, increase efficiency, and minimize environmental impact by reducing fertilizer runoff and waste moreover it can help automate fertilization in turn which reduces health issues.

5 of 38

EXISTING DRONE’S ISSUES

Major Issues of the present-day drone technology:
Limited range and coverage area, which can reduce efficiency and effectiveness for some applications
Limited support for some applications and software, which can reduce versatility and compatibility.
High prices, which cannot be afforded by farmers.

6 of 38

Raspberry Pi and its Interfacing with UAV

7 of 38

RASPBERRY PI

Raspberry Pi is a credit card-sized computer that can be used in various applications, including drone control.
Drone interfacing with Raspberry Pi enables users to control the drone remotely through a computer or mobile device.
Raspberry Pi is an open-source platform, meaning that its design, schematics, and source code are freely available for anyone to modify and use, making it a popular choice for DIY projects.

8 of 38

WHY RASPBERRY PI FOR DRONE CONTROL?

Low-cost computer that offers impressive performance for its price.
Versatile and compatible with various software and hardware.
Easy to set up and use, even for beginners.
Growing demand for low-cost and versatile microprocessors for various applications.
Potential for expanding into new markets and applications beyond drone control.

9 of 38

INTERFACING OF RASPBERRY PI MODEL WITH DRONE

Drones and Raspberry Pi’s are cool on their own, but if brought together, they can form a very powerful tool. They give a drone the power of a computer and the ability to go anywhere. Maybe robots are finally taking over!
Python has a wide scope of functionality for artificial intelligence. For example, the programming libraries TensorFlow, scikit-learn, and Keras are all primarily based on Python. Therefore, it can be used to control and train RPI-enabled drones.
Cameras can provide a unique video feed offering users a window into their drones’ perspectives. A camera can be simple or can be mounted on a gimbal to be used as a first-person view camera (FPV) on higher-end drones. FPV cameras are also available as plug-and-play solutions but can still be difficult for a beginner to master.

11 of 38

PROPOSED WORKFLOW

12 of 38

Semantic Segmentation :

Land boundary Identification

13 of 38

WHY DO WE NEED LAND BOUNDARY IDENTIFICATION?

The drone needs to spray within the land borders. So, it becomes crucial to detect land borders.
We use image processing algorithms to detect land borders. There are two methods: Classical Image Processing and Deep Neural Network algorithms.

14 of 38

WHY DO CLASSICAL IMAGE PROCESSING TECHNIQUE FAILS TO GIVE GOOD RESULTS?

Classical image processing models typically rely on handcrafted features and rules to extract relevant information from images.

Handcrafted features are pre-defined by experts and may not capture all the important patterns and variations in the image data.

Handcrafted rules are also inflexible and may not adapt well to different types of images and tasks.

On the other hand, deep neural network (DNN) algorithms can automatically learn features and rules from the image data, without the need for expert knowledge.

15 of 38

DEEP NEURAL NETWORKS FOR LAND BORDER IDENTIFICATION

Deep neural networks (DNNs) have shown promising results in land border identification by learning complex features and patterns from images.
DNNs are trained using a supervised learning approach, where the network learns to associate input data with output labels.
DNNs can also be integrated with unmanned aerial vehicles (UAVs) for real-time border monitoring and identification.

16 of 38

DEEP NEURAL NETWORKS FOR SEMANTIC SEGMENTATION

Semantic segmentation is a computer vision task that involves classifying every pixel in an image into a specific category or class.
Deep neural networks (DNNs) have shown promising results in semantic segmentation by learning high-level features and context from images.
DNNs for semantic segmentation typically use an encoder-decoder architecture, where the encoder extracts high-level features and the decoder generates the segmentation map.
The encoder can use various convolutional neural network (CNN) architectures, such as U-Net, R2U-Net, or Attention R2U-Net, to extract features.

17 of 38

CONTD.

We have used different FCNN algorithms in our drone model like:
UNET: "U-Net" (a convolutional neural network architecture for image segmentation)
R2UNET: "Residual Refinement U-Net" (a modified version of U-Net with residual connections for improved performance)
A-UNET: "Attention U-Net" (a modified version of U-Net that incorporates attention mechanisms to focus on important features during segmentation)
A-R2UNET: "Attention Residual Refinement U-Net" (a combination of A-UNET and R2UNET, which includes both attention mechanisms and residual connections for improved performance in image segmentation tasks)

18 of 38

FCNN

FCNN stands for Fully Convolutional Neural Network, a type of deep learning algorithm commonly used for image processing and segmentation tasks.
FCNN is particularly useful for tasks that require pixel-level segmentation, such as identifying objects or regions within an image.
FCNNs use convolutional layers to extract features from the input image and then use transposed convolutions to up sample the features and generate the output segmentation mask.

19 of 38

CONTD.

Model	Description	Advantages
UNET	An encoder-decoder architecture with skip connections	Good performance on small datasets, fast training and inference
R2UNET	UNET with recurrent connections	Better performance on sequential data
Attention-UNET	UNET with attention mechanism	Improved performance on fine details and object boundaries
Attention-RUNET	RUNET with attention mechanism	Improved performance on sequential data with fine details

20 of 38

COMPARISON BETWEEN DIFFERENT SEGMENTATION TECHNIQUES

21 of 38

Coverage Path Planning & Visualization

22 of 38

COVERAGE PATH PLANNING (CPP)

Coverage path planning is the process of designing a path for a drone to follow in order to cover a specific area or region.
The objective is to ensure efficient and effective coverage while minimizing overlap and wasted resources.
The resulting path must be feasible for the drone to execute and the drone must be capable of accurately following the path while avoiding obstacles and maintaining safe flight conditions.

23 of 38

ALGORITHMS USED FOR CPP

GRID-BASED ALGORITHMS

These algorithms divide the area to be covered into a grid of cells and plan a path that covers each cell only once.
Examples of grid-based algorithms include the Wavefront algorithm and the D* Lite algorithm.

POTENTIAL FIELD ALGORITHMS

These algorithms use a virtual force field to guide the robot through the area to be covered. The robot is attracted to unexplored areas and repelled from obstacles.
Examples of potential field algorithms include the Artificial Potential Field algorithm and the Vector Field Histogram algorithm.

24 of 38

ALGORITHMS USED FOR CPP

RANDOMIZED ALGORITHMS

These algorithms generate a random path through the area to be covered, making adjustments to the path as necessary to avoid obstacles and ensure complete coverage.
Examples of randomized algorithms include the Rapidly-exploring Random Tree algorithm and the Monte Carlo Localization algorithm.

OPTIMIZATION ALGORITHMS

These algorithms optimize a pre-defined cost function to find the most efficient path for covering the area.
Examples of optimization algorithms include the Traveling Salesman Problem algorithm and the Ant Colony Optimization algorithm.

25 of 38

SIMULATION OF CPP

A sample simulation of coverage path planning of a virtual drone is shown in the figure below which was done in gazebo software.�

26 of 38

INTRODUCTION TO THE GAZEBO SOFTWARE

Gazebo is an open-source software tool for simulating and testing robotics systems. It is a physics-based 3D simulator that allows users to create virtual environments and objects to test robotic systems and algorithms in a safe and controlled environment.
Gazebo can be used for a wide range of robotics applications, including robot design, testing of control algorithms, sensor integration, and multi-robot coordination. It has a large community of developers and users who contribute to the software and share their simulation models and environments.

27 of 38

WHY GAZEBO SOFTWARE?

Realistic simulation: Gazebo provides a realistic simulation environment that mimics real-world physics. It simulates the interaction of the drone with the environment, including wind, gravity, and obstacles, making it an ideal tool for testing drone algorithms before actual deployment.
Open-source: Gazebo is an open-source tool, which means that anyone can use, modify and contribute to its development. This has made it a popular choice among researchers, hobbyists, and professionals.

28 of 38

WHY GAZEBO SOFTWARE?

Integration with ROS: Gazebo integrates with the Robot Operating System (ROS), which is an open-source robotics middleware platform.
Customization: Gazebo allows users to customize the simulation environment to match their specific needs.
Cost-effective: Using Gazebo for drone simulation is cost-effective as it eliminates the need for expensive hardware and reduces the risk of damaging the drone during testing.

29 of 38

SIMULATIONS IN GAZEBO SOFTWARE

Here are some simulations of a virtual drone made in the gazebo simulator.�

30 of 38

INTRODUCTION TO THE SPRAYING SYSTEM

Spraying systems use a variety of nozzles that differ in efficiency, performance, and technical specifications.
The spraying system includes both the spraying material, such as pesticides or fertilizers, and a spraying nozzle.
The controller is the second component, which regulates the sprayer's nozzle. A pressure pump is also an essential component of the sprinkler system, which propels the insecticide through the nozzle by applying pressure.

31 of 38

AREAS OF INTEREST

Identification of the areas of interest is crucial for efficient spraying operations with UAV drones
Factors such as crop type, growth stage, weather conditions, and pest/disease infestation need to be considered
Satellite and drone imagery can be used for identification and segmentation of the areas of interest
Machine learning algorithms can be trained on labeled data to automate the process of identifying areas of interest
Effective spraying patterns can be designed based on the identified areas of interest, resulting in reduced chemical usage and improved crop health

32 of 38

PARAMETERS TO CHOOSE SPRAYING SYSTEM

Spray nozzle type: There are different types of spray nozzles, including flat fan, hollow cone, and full cone, and each type has its own advantages and disadvantages.
Spray pressure: The spray pressure can impact droplet size and coverage area.
Sprayer configuration: The sprayer configuration refers to the layout of the nozzles and their positioning on the drone. The configuration should be optimized to ensure even coverage and minimize overlap, waste, and missed areas.
Calibration: Proper calibration of the spraying system is essential to ensure accurate delivery of the crop protection product.

33 of 38

Unmanned Aerial Vehicles (UAVs) use two types of spraying techniques for pesticides.
The first type involves a boom sprayer equipped with a spraying system, while the second type has a spraying system located below the propeller rotor.
The location of the nozzle underneath the rotor influences the droplets' movement.
A pressure pump is also an essential component of the sprinkler system, which propels the insecticide through the nozzle by applying pressure.

SPRAYING PATTERN

34 of 38