Field Test Evaluation of an AI Model for Recycle Waste Separation running on Embedded Systems

Developing an automatic waste sorting AI model for embedded systems requires measuring performance in real-world contexts to determine if the model is acceptable. From our previous research, we selected a pre-trained model named ssd_mobilenet_v2_coco to construct and train an AI model for automatic waste sorting system running on Jetson Nano. The results of our laboratory evaluation are satisfactory. However, in real world, many environmental factors affect the accuracy of the AI model. In this research, we explored the factors that affect the accuracy of the AI model in real-word settings. The targeting factors include the weather's conditions, electricity stability, light intensity inside of the detection chamber, and the user behavior of waste littering. In the real-world setting, the AI model has an accuracy rate of 80.70%, which is decreased by 18.13% comparing to laboratory conditions.