It is important to investigate the inside of the hull of a submerged wreck for identifying the accident circumstances and causes. Oil-tanker wrecks usually cause serious damage to natural environment. Periodical inspection is also necessary to protect ocean environment. Although these operations are normally carried out by Remotely Operated Vehicles (ROVs), they require specifically designed support vessels with heavy equipment. Thus, their mobility is inevitably restricted for immediate operation. Therefore, it is expected that the demand for the development of compact and intelligent Autonomous Underwater Vehicles (AUVs) to undertake these complicated missions will increase greatly over the next few years. Since AUV navigation into a semi-closed space is a dangerous operation, the AUV should be aware of its surroundings in real time and make careful path planning based on sensory data. Although acoustic sensors are utilized to survey underwater objects, they are seriously affected by multipath effects in a narrow closed space, and may have poor accuracy and resolution, which are not sufficient for AUV navigation. In this paper, we propose an autonomous navigation and investigation system by stereo vision for surveying closed spaces. The system was implemented in a newly constructed AUV, "Tam-Egg 2", a small size autonomous underwater vehicle, and the vehicle successfully identified the rectangular windows of a model hull placed in an experiment tank, and passed through them without touching the hull.