AI-Powered Trash Detection in Drone Imagery
Project description
In a world where pollution is becoming increasingly important, technology is essential for a sustainable solution to this problem. This AI Trash Detection project is mainly focused on using artificial intelligence to tackle waste detection on a large scale. Extensive detection models and modern drones play an important role in identifying and locating all kinds of waste on various surfaces. The reason for this project is mainly to respond to the increasing amount of litter in the Netherlands, which not only causes visual nuisance, but also causes significant damage to the environment/biodiversity. This AI Trash Detection project in collaboration with MRR can support authorities, companies, and environmental groups in cleaning up waste in public areas faster and reducing the operating costs to do so. The benefits can be seen not only in terms of efficiency and cost reduction, but also in making a contribution to a cleaner earth.
Context
MultiRotorResearch (MRR) is a company founded in 2022, specializing in providing services and products related to drone and AI technology. The company was founded by our stakeholder Sieuwe Elferink and operates in Eindhoven with a team that includes Fontys students working on innovative drone projects. These include projects like building inspections and inventory management. MRR primarily targets businesses and government agencies, with a unique selling point that caters to both experienced and beginning users, through its “Drone as a Service” model. Every year, MRR participates in offering a group assignment for the AI Advanced Semester. This year, the assignment from MRR was to work on the AI Trash Detection assignment, in which Drone Technology is combined with AI to detect and locate waste. The goal of this project was to improve the performance of the YOLOv8 model developed by the previous group, with a focus on both efficiency and accuracy (recall/precision). This was to be achieved by researching better detection models and applying hyperparameter tuning, while considering the computational resources provided by MRR. Additionally, the best-performing model was to be formulated into an API capable of communicating with MRR’s systems, enabling a more automated prediction process.
Results
At this point, we have adopted the Detectron2 model as one of the models we experimented with. This model along with the RT-DETR model were both improved through data enrichment and hyperparameter tuning. The YOLO model was dropped, as it underperformed a lot compared to the other two models used. Out of the RT-DETR and Detectron2 models, both of the newest models show similar performances, with each having some objects missed in different images. Detectron2 does have consistently higher confidence scores than RT-DETR but struggles more with smaller objects.