Underwater Stereo Vision System for Diver Rescue ROVs: Real-Time Estimation of Relative Position and Upper-Body Orientation

A diver who gets into difficulty underwater is often beyond the immediate reach of a companion, and interest is therefore growing in remotely operated vehicles (ROVs) that can stay alongside a diver and intervene when conditions deteriorate. A diver-support vehicle of this kind is controlled in a coordinate frame attached to the diver. Its controller must therefore receive the diver’s relative position, the orientation of the body, and an indication of whether the diver is in distress, continuously and from sensors the vehicle itself carries. That information is normally obtained acoustically from a transmitter worn by the diver, but a diver who needs help may or may not be carrying one, so the transmitter cannot be assumed to be available. In this study, an ROV is equipped with stereo cameras, and an onboard perception pipeline recovers the diver’s distance and lateral position, upper-body orientation, and state (normal, panic, or unconscious) from the stereo images alone, on an NVIDIA Jetson AGX Xavier at approximately 5 frames per second with an end-to-end latency of 0.20 s. A close-range fallback based on segmentation of the scuba cylinder keeps the distance estimate usable below the range at which stereo disparity becomes unstable. In test-pool evaluation at reference distances of 1.5, 2, and 3 m, the mean absolute distance error was 0.087 m over 242 frames, corresponding to about 4% of the measured distance; the geodesic orientation error against a diver-mounted attitude sensor was 12.65° over 1968 frames; and the diver-state classifier reached 99.1% accuracy on the transition from normal swimming to an unconscious posture, with a detection latency of 1.4–2.4 s. In open-water trials off Pohang at depths of 5–9 m, an onboard relative-space controller using the perception output carried out guidance, following, and rescue scenarios in 15 of 15 trials, and the perception output was sufficient for control and rescue in every scenario without a perception-related failure.