Team: 14
School: Los Alamos High
Area of Science: Engineering
Interim: All project images can be seen at: https://github.com/daniel360kim/OceanAI/tree/master/Resources/project_images
Problem Definition:
Litter in marine environments ravages ecosystems and exacts a large toll on coastal economies. (IUCN 2021). With 5.25 trillion pieces of plastic debris in the ocean (Parker 2015), understanding concentrations of ocean trash and tracking movements of debris is crucial for the efficient cleanup of the world’s seas.
Problem Solution:
The proposed solution to help evaluate the scope of marine debris consists of two key aspects. First, an autonomous underwater vehicle (AUV) to collect footage and images of the ocean, along with important metrics such as temperature and salinity. Second, a trash detection computer vision model to analyze data collected from the AUV and to build a comprehensive map of concentrations of ocean debris.
The AUV:
The AUV closely mirrors the design of an underwater glider – an unmanned vehicle that utilizes changes in buoyancy to propel itself (NOAA 2021). Underwater gliders can silently travel long distances while collecting seminal ocean metrics such as temperature and salinity. The AUV is mostly 3D printed, with other parts being widely available to reduce costs. Within the AUV, a custom circuit board has been designed to read and log sensor data and control the AUV. Due to geographical constraints, the AUV cannot be directly tested in the ocean. To overcome this limitation, the AUV will be tested using a method called hardware in the loop (HITL). HITL is a technique where real signals from a controller are “connected to a test system that simulates reality†(NI 2022). By utilizing open-source underwater data from the National Oceanic and Atmospheric Administration and running a series of underwater tests in local bodies of water, the capabilities of the glider can be evaluated, without having to travel long distances to test the AUV.
Trash Detection Model:
The trash detection model is built on a convolutional neural network that is trained using YOLOv5 with the TRASH-ICRA database. YOLOv5 functions by creating features from input images, then feeding the features through a prediction system to draw boxes around recognized objects. (Solawetz 2020). Images paired with temperature, depth, and GPS data from the AUV is inputted into the model which identifies underwater debris and pairs it with a GPS coordinate. Using this model, a comprehensive map of underwater debris concentrations can be built.
Progress to Date:
The AUV:
The AUV has been completed, with the control board and associated software all finished. The fully functional AUV is able to collect and log sensor data, take and store images, control buoyancy, and transmit data to a graphical user interface (GUI) we designed.
Trash Detection Model:
The trash detection model was trained using 1700 images of various pieces underwater debris and validated with 600 images from the same dataset. The model was able to recall 63.05% of images with 76.13% precision. So far, it can identify two classes: trash and no trash.
Expected Results:
The AUV:
Although the AUV is mostly finished, it has not undergone much testing. The AUV has been verified to be watertight, but we hope to continue underwater testing to measure its speed, runtime, and robustness. Furthermore, we hope to continue to optimize the AUV to allow it to function for longer periods of time.
Trash Detection Model:
The trash detection model will be tested with the same approach as the AUV – by feeding footage from open-sourced data and utilizing the metadata to form a comprehensive model of ocean debris. With larger datasets and better optimized training, we hope to further increase the precision of the model. As of right now, the model has two classes: not trash and trash, and we hope to add further categories to the model.
Bibliography
Marine plastic pollution. https://www.iucn.org/resources/issues-brief/marine-plastic-pollution#:~:text=Impacts%20on%20marine%20ecosystems,stomachs%20become%20filled%20with%20plastic. (accessed Jan 7, 2023).
National Oceanic and Atmospheric Administration. What is an ocean glider? https://oceanservice.noaa.gov/facts/ocean-gliders.html (accessed Jan 7, 2023).
Parker, L. Ocean trash: 5.25 trillion pieces and counting, but big questions remain. https://education.nationalgeographic.org/resource/ocean-trash-525-trillion-pieces-and-counting-big-questions-remain (accessed Jan 7, 2023).
Solawetz, J. What is Yolov5? A guide for beginners. https://blog.roboflow.com/yolov5-improvements-and-evaluation/ (accessed Jan 7, 2023).
What is hardware-in-the-loop? https://www.ni.com/en-us/solutions/transportation/hardware-in-the-loop/what-is-hardware-in-the-loop-.html (accessed Jan 7, 2023).
Team Members:
Sponsoring Teacher: NA