Minimally invasive surgical robots have improved the safety and efficiency of cardiac surgery. With smaller and more flexible instruments and endoscopes, robotic systems reduce incision size and number. However, the limited operative space introduces a risk of collisions between instruments. Due to difficulties in accurate instrument localization via kinematics or sensors, this paper proposes a vision-based method for dynamic collision avoidance. Instrument positions are obtained through image segmentation, and depth maps are generated using a stereo-matching algorithm combined with a pre-trained depth estimation model, achieving an average relative error of around 7%. A KD-tree with a hierarchical strategy enables fast distance estimation. Based on instrument distance and master-side motion speed, a collision risk model is built to predict potential collisions and trigger avoidance strategies. The method is validated in a simulated environment.