Real-Time Object Detection and Active Learning Framework for Smart Waste Sorting from Newly Collected Data

Abstract

Campus waste sorting remains challenging because of the substantial variability in the appearance, shape, and material characteristics of waste objects, including deformed packaging, transparent plastics, paper products, and visually similar materials. This paper presents a smart waste detection and sorting framework that integrates the lightweight YOLOv12n object detector with a human-in-the-loop feedback mechanism and an active learning framework. The system is deployed on a Raspberry Pi 5 and includes a Camera Module 3 for image acquisition, a dual-servo sorting mechanism, and a 7-inch touchscreen for user interaction. Images collected during operation are ranked using detection-level entropy, after which a PCA-reduced k-means++- based diversity strategy selects informative and non-redundant samples for annotation. The detector is incrementally fine-tuned offline using the accumulated labeled dataset and subsequently redeployed to the Raspberry Pi 5. Experiments on a custom dataset of 4,223 images from 11 campus-waste categories show that the hybrid sampling strategy reaches an mAP@50 target of 80% using 1,830 labeled images, representing a 46.3% reduction in labeled training images relative to the full training split. The deployed detector operates at 4–5 FPS with a vision-pipeline latency of 198—250 ms, demonstrating the feasibility of the proposed framework for campus waste sorting.