Auto Annotation using Object Tracking with Multiple in-vehicle Cameras for Federated Learning

In recent years, significant advancements have been made in automated driving technology. To achieve fully automated driving, a vehicle must accurately acquire information regarding objects in its surroundings by gathering large amounts of data for machine learning. In many cases, data collected from each vehicle is transmitted to a central server, which is undesirable because of privacy concerns and traffic generation to retrieve the data. Federated Learning (FL) is a promising technology that addresses these concerns by enabling learning while maintaining the data with the user. However, since the local labeling accuracy for distant objects with existing models is still low, quality of the model training is affected. To address this issue, this paper proposes the use of Multiple object Tracking (MoT) with multiple cameras to improve the labeling accuracy for FL. Our method enables the application of the labeling results of nearby objects to identify distant objects by tracking time-series data from multiple cameras. The evaluation results demonstrate that the proposed method effectively labels distant objects without sending data to the central server.