STUDY ON AI CAMERA IMAGE TRAFFIC SURVEY INTRODUCTING OPEN DATASETS AND ADDITIONAL LEARNING

Abstract

In Japan, operational requirements and accuracy standards for AI Camera have been established as an AI observation method using CCTV cameras, aiming at constant observation of roads regardless of normal conditions or disasters. In addition, the usefulness of additional learning for AI observation is well known, and research on methods to reduce learning costs and obtain teaching data has been active. However, there is a lack of complementary research on the relationship between teacher data and accuracy requirements for AI Camera, and we believe that accuracy verification related to dataset selection is necessary for the nationwide spread of AI observations. In this study, we conducted AI observations using two open datasets that are publicly available and an original dataset created using our own semi-automated method. As a result, we clearly show the relationship between the creation and selection of teaching data and the accuracy requirements of AI Camera and discuss various issues that need to be considered.