Real-Time Waste Classification at the Edge: A Convolutional Neural Network Pipeline with Temporal Prediction Smoothing via a Moving Vote Buffer
Year: 2026
Type: conference-paper
Abstract: Effective solid waste management represents one of the most pressing environmental challenges of the twenty-first century, and the deployment of intelligent, automated classification systems offers a scalable pathway toward improved waste segregation at the source. This paper presents a real-time waste classification system that integrates a pre-trained convolutional neural network (CNN), implemented within the TensorFlow and Keras frameworks, with a live webcam-based video capture pipeline orchestrated through OpenCV. The proposed system processes individual video frames, resizes them to a standardized input resolution of 224 × 224 pixels, normalizes pixel intensity values to the [0, 1] range, and submits them to the model for inference. To mitigate the well-known problem of per-frame prediction instability caused by transient occlusions, motion blur, and environmental lighting variations, a seven-frame majority-vote smoothing buffer is introduced. This temporal aggregation mechanism selects the most frequently occurring predicted class across a rolling window of recent frames, yielding a stable, flicker-free output label suitable for live display. Experimental evaluation demonstrates that the system operates at interactive frame rates on consumer-grade hardware, correctly identifies multiple waste categories with a minimum confidence threshold of 50%, and produces visually coherent, temporally stable classification overlays. The architecture's modular design facilitates straight-forward adaptation to larger label sets and alternative backbone models, underscoring its suitability for deployment in smart-bin installations, recycling facility conveyors, and educational environmental-awareness kiosks.
Institution: Visvesvaraya Technological University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7224901190