Designing of A Plastic Garbage Robot With Vision-Based Deep Learning Applications

Abstract

To address the issue of plastic waste, a robot using deep learning technology for visual recognition to classify plastic waste has been developed. This system includes a 3DOF robot arm, a conveyor, a camera, an electrical cabinet, and a computer. The object detection component of the system is designed using transfer learning with a pre-trained YOLOv5 model to ensure the system operates in real time. Selecting the best model by evaluating and comparing the results of models trained using labeling by bounding box and polygon methods. Then, the real-world coordinates for the origin of the robot arm are determined by utilizing matrices obtained from MATLAB through chessboard images. The computer processes the data and transmits commands to the robot arm system and conveyor, which is controlled by a PLC and 3 different Servo Drivers, for object sorting on the conveyor. The best-performing model has a Precision of 92.1% and a Recall of 87.3%, and the success rate of picking up an object is 91.5%. While the experimental results indicate complete stability in inter-device connectivity, implementing it would necessitate hardware improvements to leverage its potential.