Underwater Data

Visual

Typical Images

Upward-Looking Camera

Images from the upward-looking camera are expected to show

• optionally, water surface with ripples
• optionally, human body parts performing swimming or a communication gesture
• optionally, water turbulence
• optionally, scene objects like swim lanes
• optionally, peripheral objects like walls

Forward-Looking Camera

• optionally, walls (in a pool environment)
• optionally, human body parts
• optionally, water turbulence

Processing

Background removal

The images obtained by the camera could be passed through a low-pass filter to remove the clutter caused by water surface ripples. We could also subtract the background (mostly water surface viewed from below) is to use the Stauffer-Grimson technique based on an adaptive mixture-of-Gaussians model of the background. Theoretically, we can do better than using GMM given domain knowledge: the low-frequency appearance of the water surface can be approximated as a function of the (static) lighting and the motion of the surface according to a wave model. Using a neural network we may be able to learn this function. We would still use a Gaussian mixture model for surface ripples and bubbles, though.

Swimmer detection

Strategy #1: Wait until a known, characteristic swimming pose is detected.