Aortic valve replacement is the only definitive treatment for aortic stenosis, a highly prevalent condition in elderly population. Minimally invasive surgery brought numerous benefits to this intervention, and robotics recently provided additional improvements in terms of telemanipulation, motion scaling, and smaller incisions. Difficulties in obtaining a clear and wide field of vision is a major challenge in minimally invasive aortic valve surgery: surgeon orientates with difficulty because of lack of direct view and limited spaces. This work focuses on the development of a computer vision methodology, for a three-eyed endoscopic vision system, to ease minimally invasive instrument guidance during aortic valve surgery. Specifically, it presents an efficient image stitching method to improve spatial awareness and overcome the orientation problems which arise when cameras are decentralized with respect to the main axis of the aorta and are nonparallel oriented. The proposed approach was tested for the navigation of an innovative robotic system for minimally invasive valve surgery. Based on the specific geometry of the setup and the intrinsic parameters of the three cameras, we estimate the proper plane-induced homographic transformation that merges the views of the operatory site plane into a single stitched image. To evaluate the deviation from the image correct alignment, we performed quantitative tests by stitching a chessboard pattern. The tests showed a minimum error with respect to the image size of 0.46 ± 0.15% measured at the homography distance of 40 mm and a maximum error of 6.09 ± 0.23% at the maximum offset of 10 mm. Three experienced surgeons in aortic valve replacement by mini-sternotomy and mini-thoracotomy performed experimental tests based on the comparison of navigation and orientation capabilities in a silicone aorta with and without stitched image. The tests showed that the stitched image allows for good orientation and navigation within the aorta, and furthermore, it provides more safety while releasing the valve than driving from the three separate views. The average processing time for the stitching of three views into one image is 12.6 ms, proving that the method is not computationally expensive, thus leaving space for further real-time processing.