The application of ultrasound in interventional surgery faces many challenges due to its lack of clarity. While existing algorithms can process single-frame ultrasound images efficiently, they still suffer from poor accuracy and discontinuous detection. In this paper, we propose an online instance segmentation network for ultrasound-guided interventional surgery Videos. When detecting the current frame, the fusion of previous frames enhances the accuracy and continuity of the segmentation. Meanwhile, real-time 3D reconstruction of vessels and interventional instruments is also achieved through the collaboration of robotic arms. Furthermore, we construct a novel dataset for vascular interventions. It accurately labels vessels and interventional instruments in 112 ultrasound videos, making it suitable for tasks related to the detection and segmentation of vascular ultrasound images. Experiments demonstrate that the proposed network improves detection accuracy by 11.0% mAP75 compared to the state-of-the-art method.