A highly efficient garbage pick-up embedded system based on improved SSD neural network using robotic arms

With the social evolution, economic development, and continuously improved living standards, the dramatically increasing garbage produced by human beings has seriously affected our living environment. There are 3 main ways to dispose of garbage: sanitary landfill, incineration, or recycling. At present, a huge amount of labor resources is required for pre-sorting before garbage disposal, which greatly reduces efficiency, increases costs, and even leads to direct incineration without sorting. Hence, this study proposes a solution scenario of how to use object detection technology for garbage sorting. With the development of the deep learning theory, object detection technology has been widely used in all fields, thus, how to find target objects accurately and rapidly is one of the key technologies. This paper proposes a highly efficient garbage pick-up embedded system, where detection is optimized based on the Single Shot MultiBox Detector (SSD) neural network architecture and reduced model parameters. The experimental verification scenario was conducted in a dynamic environment integrating a robotic arm with a conveyor belt simulated by an electronic rotating turntable. The experimental results show that the modified model can accurately identify garbage types, with a significant speed of 27.8 FPS (Frames Per Second) on NVidia Jetson TX2, and an accuracy rate of approximately 87%.