During an open-water experiment in the Caribbean Sea off the coast of Barbados, the researcher team tested the autonomous underwater vehicles (AUVs).
Credit
Photo provided by Junaed Sattar
Researchers at the University of Minnesota Twin Cities have created a first-of-its-kind artificial intelligence system that enables autonomous underwater vehicles (AUVs) to monitor a scuba diver’s physiological stress in real time by visually tracking exhaled bubbles
Published in The International Journal of Robotics Research, the study marks the first time computer vision has been successfully deployed to estimate human respiration rate (HRR) underwater.
Led by senior author Junaed Sattar and lead author Demetrious Kutzke, the breakthrough provides a non-contact, visual solution for tracking vital signs in extreme underwater environments.
Overcoming underwater sensing limits
Scuba diving carries severe physical risks, where exhaustion or hyperventilation can escalate into life-threatening emergencies. On land, monitoring breathing rate is simple, but underwater environments present severe physical constraints:
Sensor interference:
Thick neoprene wetsuits and drysuits block typical skin-contact sensors from measuring heart or breathing rates accurately.
Signal loss:
Wireless data transmission degrades rapidly underwater, preventing traditional wearables from beaming data back to monitoring systems.
By using an AUV’s built-in camera to track the timing and frequency of regulator bubble releases, the team eliminated the need for physical sensors attached to the diver.
Training AI with “fuzzy Labelling” and multimodal datasets
Because underwater video footage is frequently obscured by murkiness, floating debris, or low light, the research team developed a “fuzzy labelling” training framework.
They manually organised thousands of video frames while synchronising them with underwater audio recordings of regulator exhalations. The distinct sound profile of each breath taught the computer vision algorithm to accurately isolate and identify visual bubble plumes in challenging conditions.
To ensure the AI could perform across diverse water conditions, the team gathered visual and audio datasets from various locations, ranging from freshwater environments in Minnesota (Lake Superior and Square Lake) to warm, open-ocean conditions in the Caribbean Sea off Barbados.
Real-time safety alerts via HREyes
During field trials, the researchers implemented a visual status communication platform named HREyes. The AUV analyses visual inputs to estimate a diver’s breathing frequency, categorising their health state into three distinct brackets:
Below normal:
Fewer than 14 breaths per minute.
Normal range:
14 to 20 breaths per minute.
Above normal (Potential stress):
Greater than 20 breaths per minute.
If a diver begins hyperventilating or overexerting, the robot flags the elevated respiration rate, providing an autonomous safety buddy that can alert human divers before distress becomes critical.
Future directions for AI, diving and underwater robots
Supported in part by the U.S. Department of Defense’s SMART Scholarship and the National Science Foundation, the team plans to expand the AI’s capabilities. Future iterations will combine breathing-rate measurements with movement tracking to build holistic wellness profiles for divers during deep-sea exploration and group dive operations.