Vision-Based Smart Waste Segregator: An Embedded System

Abstract

Improper waste disposal is a critical environmental challenge, particularly in urban areas where manual segregation remains inefficient, unsafe, and inconsistent. This paper presents a Vision-Based Smart Waste Segregator, an automated system that classifies and sorts waste into five categories: glass, metal, paper, plastic, and organic waste. The system employs a Raspberry Pi 4B as the central processing unit, which is triggered by a Time-of-Flight (ToF) sensor upon waste detection. A Pi Camera Module captures an image of the waste item, which is classified using a novel feature-fusion deep learning model combining EfficientNetB0 and DenseNet121. The fused feature vectors from both architectures are passed through fully connected layers with a Softmax output for five-class classification. Based on the results, a dual-motor mechanism, a stepper motor for bin alignment and a servo motor for waste release, directs the waste into the appropriate bin. Bin fill-level monitoring via a second ToF sensor transmits real-time status to a web-based interface through HTTP communication. The model achieved a validation accuracy of 94-95% and was optimized using TensorFlow Lite post-training quantization for efficient edge deployment.