Autonomous Unmanned Surface Vehicle

AIM AND OBJECTIVES

AIM

To propose a strategy to combat illegal, Unreported, and Unregulated (IUU) fishing activities and identify hazardous oil, that inflicts significant damage on the marine ecosystem using a fleet of autonomous Unmanned Surface Vehicles (USVs) capable of maritime reconnaissance and surveillance.

Objectives

Mechanical System Development

1. Design and develop the hydrodynamic hull and structure of the USV capable of operating in harsh sea conditions.

2. Design and development of the wave-powered propulsion system.

3. Design and development of the mechatronic systems, control system

Intelligence System Development

1. Data collection and data synthesis

2. Implementation of a ship detection model

3. Implementation of a distance estimation mechanism for nearby ships

4. Implementation of identification for mineral oils and other hazardous substances present on the ocean's surface

5. Model optimization for operation on edge devices, minimizing latency

6. Implementation of a real-time operational system

METHODOLOGY

Design and Development of Hull, Actuating Mechanisms, and Control Systems

Design and develop the hull, actuating mechanisms, and control systems with the required sensor array. The USV will be propelled by wave power, so the required hull and hydrofoils will be designed and manufactured. The hydrodynamics hull will be made from carbon fiber due to its low weight, and the necessary mechanisms will be implemented to control the position of the rudder for proper USV movement.

The Jetson Nano single-board computer will serve as the main controller, in conjunction with the Pixhawk flight control device. The USV will adjust its movements based on feedback from the acoustic and visual detection systems.

Ship Detection Algorithm

We are considering using the YOLOv8 object detection model for our ship detection algorithm. YOLOv8 is known for its fast inference speed and user-friendly nature. However, we are carefully evaluating all of the YOLO models to determine which one best meets the needs of our application. It is important to note that the most advanced model may not always be the best choice for a particular dataset or use case.

Regarding ship tracking, our preferred approach is to employ ByteTrack due to its swifter tracking capabilities and commendable accuracy, as established in our initial discussions. Nonetheless, we maintain an open stance by exploring alternative tracking algorithms, including DeepSORT, DeepAN, Optical Flow. This comprehensive evaluation ensures that we adopt the most suitable solution to meet our project's needs effectively.

Distance Estimation Mechanism

To accomplish this task, a comprehensive dataset needs to be collected. Additionally, we synthesize the dataset, incorporating environmental factors like wave noise and boat engine sounds to simulate real-world conditions.

Machine learning techniques can be used to extract meaningful features from sound waves. Some specific techniques that could be explored include Mel-frequency cepstral coefficients (MFCCs), spectrogram analysis, and wavelet transform analysis.

The next step is to create a regression model that can accurately estimate the distances to nearby ships. If the initial regression model does not meet performance expectations, an alternative approach is to categorize the outputs. This categorization could involve distance intervals such as "less than 100 meters", "100 to 500 meters", "500 to 1000 meters", and "greater than 1000 meters.

The final method will be selected based on the results of these approaches. The most accurate and reliable method will be used to ensure precise distance estimation for the USV's interactions with other vessels.

Identification of Hazardous Substances

To achieve this goal, we plan to use an image segmentation model to effectively segment regions of oil on the sea surface. We intend to employ the YOLOv8 segmentation model, but we are also considering other models such as DeepLab, UNet, and Mask R-CNN. These models offer different advantages in terms of precision and speed.

Model Optimization for Edge Devices

The Jetson Nano 4GB device with a battery unit has been selected as the optimal choice for building an intelligent system. This is because it is a real-time capable device that is ideal for machine learning tasks. It is also compact, energy-efficient, and has a GPU that can accelerate the execution of complex models.

To further improve efficiency, we plan to convert the models to lower precision weights. We will carefully test different quantization methods to find the best balance between model latency and accuracy. We will initially use the ONNX format, which ensures that all models are compatible with each other. This will also minimize the need for additional runtime libraries, which will simplify deployment and maintenance.

While the ONNX-based approach is a promising option, we are also exploring other ways to improve model performance. One possibility is to use TensorFlow Lite, which is compatible with TensorFlow and can also be hardware-accelerated.

The final optimization approach will be chosen based on a combination of factors, including latency reduction, accuracy preservation, and the specific requirements of the USV's edge device deployment. By carefully considering all of these factors, we can ensure that our intelligent system achieves the highest level of performance while operating on edge devices.

Real-Time Operational System

It is important to implement thread programming and load-balancing techniques in order to make sure that these models can operate in real time without causing delays.

Our strategy includes thorough unit testing for each system component independently, followed by extensive end-to-end testing. Moreover, it is essential to validate the models by conducting practical tests within the actual operational environment.