Development of a Low-Cost Edge AI Pipeline for Waste Object Characterization and Mass Estimation
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Faculty of Sciences
Abstract
Waste characterization is an important step in modern waste management, recycling, and environmental
monitoring. The central objective of this thesis is simple: we try to estimate the
mass of waste objects using nothing more than a low-cost RGB camera. This objective is
difficult because mass is not directly visible in an image; it must be inferred from object category,
mask geometry, depth, material behavior, density, deformation, and calibration. This
thesis presents a first full version of a low-cost Edge AI pipeline for waste object characterization
and mass estimation. The system uses an RGB camera and an NVIDIA Jetson Nano
target platform to perform object segmentation, monocular depth estimation, metric feature
extraction, volume approximation, and hybrid mass prediction.
We followed a practical machine-learning engineering workflow. A segmentation
dataset was prepared by combining and adapting TACO, AquaTrash, and RealWaste. Because
several sources did not provide instance masks, the work included manual annotation,
detection-assisted annotation, Segment Anything based mask generation, and CVAT correction.
In total, 7,648 images were annotated during the wider dataset preparation effort, and
6,460 images with 9,762 instances were retained in the final segmentation training manifest.
An Ultralytics YOLO26m-seg model was fine-tuned with Optuna-assisted hyperparameter
search, ZenML pipeline organization, MLflow registration, and ONNX export. The final
evaluated model achieved mask mAP@50:95 of 0.5453 on validation and 0.5016 on test.
For mass estimation, a hybrid residual model was trained on 745 object-level samples. The
selected Extra Trees residual model reduced test MAE from 44.31 g for the physics baseline
to 23.18 g, corresponding to a 47.69% improvement.
The thesis documents the mathematical foundation, system design, training pipelines,
deployment-oriented architecture, and remaining limitations. The results show that the proposed
pipeline is technically feasible as a low-cost Edge AI approach, while also showing
that robust mass estimation requires stronger physical calibration, real measured masses, and
Jetson-specific latency validation in future work.