RMS: Real-time Motion Segmentation over the Internet of Vehicles

In the context of autonomous driving, moving objects such as vehicles and pedestrians are of critical importance as they primarily influence the maneuvering and braking of cars. Unfortunately, due to the limited detection range of sensors, some distant and blocked objects cannot be detected, leading to slow responses when some unexpected situations occur during driving. To address this problem, a real-time motion segmentation multi-task model (RMS), running on an individual vehicle, is introduced to provide motion segmentation of moving objects within its field of view. RMS consists of a shared encoder, a multi-modal fusion module, and a dual decoder. An enhanced High Definition (HD) map constructed with the proposed RMS in line with the recommendations of the 3rd Generation Partnership Project (3GPP) Vehicle-to-everything (V2X) communication standard is produced. Extensive experiments demonstrate how RMS outperforms existing state-of-the-art motion segmentation methods in terms of multiple metrics, including mean Intersection over Union (mIoU). Additionally, Internet of Vehicles (IoV) simulation experiments show how the time required to update the map is better than the times achieved when using other methods.