Automatic Waste Type Detection Using YOLO for Waste Management Efficiency

Abstract

The management of waste in Indonesia is currently suboptimal, with only 66.24% being effectively managed, leaving 33.76% unmanaged. This highlights a significant challenge in waste management, primarily due to a lack of understanding in selecting appropriate waste types. Advances in deep learning and computer vision offer promising solutions to this issue. This study employs the YOLOv8l model, a well-regarded deep learning model for object detection, to develop an automated waste type detection system integrated with trash bins. The dataset comprises 2800 images across four classes, each containing 700 images, and is split with an 80:10:5 ratio for training, validation, and testing. Evaluation on test data yields a mean Average Precision (mAP) of 96.8%, indicating robust model performance in object detection. The model's accuracy is further validated with a score of 89.98%. Real-time testing conducted at Merdeka Park, Binjai, demonstrates the system's capability to detect waste with varying confidence levels, consistently above the 0.5 threshold. The highest confidence was observed in bottle detection at 0.94, and the lowest in cans at 0.64, underscoring the system's reliability across different detection scenarios within a 30cm range.