Object tracking on water surfaces is an essential task for Unmanned Surface Vehicles (USVs) in the fields of coastal surveillance, environmental monitoring, and rescue operations. To achieve high accuracy and robust object tracking performance, radars and cameras are potential sensors equipped on USVs for perceiving the target's appearance, color, shape, distance and velocity in all weather conditions. This project aims to establish a deep learning model for 4D radar and camera. By analyzing the features of 4D radar point clouds and camera images, this project will perform a multi-modal feature-level fusion to track objects in real time on the water surface, which can be applied to a development kit for real-time tracking in real scenarios.