•  
  •  
 

CIB Conferences

Abstract

In Viet Nam, recyclable municipal solid waste is often inadequately sorted and disposed of. While computer vision offers automatic sorting solutions, few studies address real-world deployment constraints on resource-limited edge devices. This research develops and evaluates an optimized edge AI framework for real-time waste classification. To better reflect practical sorting conditions, a dataset of 21,329 images was constructed, combining 15,158 internet images with 6,171 highly contaminated waste images captured in Hanoi, Viet Nam. The validation and test sets capture image variations in object shape, size, occlusion, and illumination. To enable deployment on the NVIDIA Jetson Orin Nano, the system uses YOLOv8s baseline architecture. Two independent optimization branches were applied: structural pruning reduced the model's parameters from 11.13 million to 6.28 million (43.2% reduction), while Quantization-Aware Training (QAT) compressed the baseline model into a TensorRT INT8 engine. Results show that on the test dataset, the QAT branch attains an overall mAP@50 of 0.886 and mAP@50-95 of 0.782. The pruned model, after fine-tuning, recovered to a mAP@50 of 0.830. Remarkably, the QAT INT8 engine achieved an inference speed of 34.5 FPS on the Jetson Orin Nano, representing a 69.1% improvement over the FP32 baseline. These findings illustrate the feasibility of deploying AI models on low-cost edge devices for sustainable waste management.

Keywords

Edge AI, source separation, municipal solid waste, YOLOv8, sustainable waste management

Download

Share

COinS