Non-contaminated plastic waste from healthcare facilities is largely non-infectious and rich in recoverable polymers, yet it is routinely incinerated because no automated sorting process suited to the clinical setting has been established. This paper presents the design of a sensor fusion system and a hyperspectral classification methodology for this stream; experimental results are not yet available and will be reported in a follow-up study. Near-infrared hyperspectral imaging, an RGB line-scan camera, and an inductive metal sensor are integrated on a collaborative robot platform and combined in a late-fusion pipeline for polymer identification, object detection, and metal discrimination. Two hyperspectral classification strategies are specified: reference-spectrum distance matching, which requires no labeled training data, and supervised machine learning, which depends on a labeled dataset. A conceptual comparison based on their methodological properties and on published evidence leads to their use as successive development phases, with the first strategy helping to build the training set for the second. An analytical assessment based on sensor datasheet values indicates the sensor specification required for production line speeds, which the pipeline accommodates without changes to the classification methodology. The regulatory scope of the recoverable fraction under German waste legislation is treated as a design input that determines the target material classes and the open-set rejection behavior required of the classifiers. Open design parameters and a staged validation pathway are identified.
Non-contaminated plastic waste from healthcare facilities is largely non-infectious and rich in recoverable polymers, yet it is routinely incinerated for lack of an automated sorting process suited to the clinical setting. This paper presents a sensor fusion system design and a comparative hyperspectral classification methodology for that stream. Near-infrared hyperspectral imaging, an RGB line-scan camera, and an inductive metal sensor are integrated on a collaborative robot platform, combining polymer-type identification, visual object detection, and metal discrimination in a late-fusion pipeline. Two hyperspectral classification strategies are specified and compared across labeling cost, robustness, interpretability, and stage of system maturity: reference-spectra distance matching, which needs no labeled training data, and supervised machine learning, which gains robustness as the labeled dataset grows. The two are ordered as successive development phases, the first bootstrapping the second. A throughput and spectral analysis establishes the sensor specification required for production line speeds, which the pipeline architecture accommodates without change to the classification methodology. The regulatory scope of the recoverable fraction under German waste legislation is treated as an explicit design input, determining the material classes targeted and the open-set rejection behavior required of the classifiers. Open design parameters and a validation pathway are identified.