Sensing and perception

The active-stereo camera module

Two synchronized global-shutter sensors capture the same instant from slightly different viewpoints. The edge computer matches features between the pair, estimates disparity, and turns it into a local depth map. Global shutters reduce the skew that rolling shutters introduce when the crawler moves over rocks or the current shakes the housing. The proposed module uses a pressure-rated dome, a blue-green light ring, short-range focus, and a serviceable Ethernet/PoE connector.

Sensor class

Dual global shutter, 1–12 MP candidate range

Frame rate

30–60 fps research target

Viewport

Dome or Aquagon-style refractive correction

Link

Gigabit Ethernet / PoE concept

Working view

Recessed low-front pair, roughly 15–35° down

Why water changes the camera problem

Water absorbs red light quickly and scatters suspended silt. The useful blue-green window is roughly 400–550 nm, so the camera and light ring should be designed around that band rather than assuming daylight color will survive. A second macro camera can work at 5–50 cm for the collection tray, but focus, backscatter, and calibration targets matter more than headline megapixels. UV fluorescence and active IR structured light are experimental options—not guaranteed gold signatures or universal depth solutions.

Blue-green window

Approx. 400–550 nm

UV research band

Approx. 300–400 nm; corrosion/ions only

Active IR

Approx. 750–850 nm; test in actual turbidity

Macro range

Conceptually 5–50 cm

Water correction and computer vision

Sea-thru- or WaterGAN-inspired processing can estimate wavelength-dependent attenuation and reduce the blue-green cast before a detector sees the frame. That is a preprocessing hypothesis, not a license to manufacture detail that was never captured. A customized YOLOv8 or YOLOv10-style model could be trained on bedrock cracks, black-sand lines, coins, chains, and representative geology. The model should report calibrated confidence and an abstain state; a shiny pebble or pyrite must remain a possible false positive.

• Store raw frames beside every corrected frame.
• Train on site-specific geology and lighting, not internet stock images.
• Use confidence calibration, hard negatives, and an explicit unknown class.

Green LiDAR: a conditional detail layer

Bathymetric or subsea LiDAR uses green light near 532 nm because that wavelength can travel farther through clear shallow water than infrared. A scanning mirror and time-of-flight electronics can produce a dense local point cloud, but bubbles and turbidity quickly shorten range. Premium bathymetric systems may use dual-pulse surface/bed returns to correct water level and refraction; a small crawler would more likely use a compact experimental scanner and rely on depth/IMU calibration. Terrestrial 1064 nm IR LiDAR is not a drop-in underwater sensor.

Optical wavelength

532 nm green-light concept

Best case

Clear, shallow water and close-range structure

Failure modes

Silt, bubbles, surface glare, refraction

Role

Fine geometry after sonar finds the area

Sonar is the robust baseline

Forward-looking or multibeam sonar supplies the primary obstacle and morphology layer when optical systems lose contrast. Acoustic returns can identify the silhouette of a logjam, boulder, ledge, or undercut bank even when the camera sees only suspended particles. The fusion engine should register sonar range/angle with stereo depth, IMU attitude, depth, and track odometry, then retain the uncertainty instead of flattening every return into a crisp-looking image.

• Sonar-first route planning in turbid water.
• Optical stereo for near-field texture and sample positioning.
• Conditional LiDAR for high-detail geometry in clear water.

Pulse induction detector

The PI coil sends a high-current pulse, waits for the transmit field to collapse, and samples one or more receive windows in the decay curve. An initial laboratory sweep might explore approximately 100–1,000 pulses per second and pulse widths from tens to hundreds of microseconds. PI is attractive in mineralized or saline conditions because it does not require the same continuous ground-balance behavior as VLF, but it has minimal discrimination: a conductive iron fragment, aluminum, or steel can look like a target. Tiny sub-gram flakes remain a difficult sensitivity case.

Signal

Transient decay / conductivity anomaly

Initial sweep

100–1,000 pulses/s research range

Pulse width

Tens to hundreds of microseconds

Trade-off

Depth and stability versus discrimination

The sensor-fusion record

A target should be interesting because multiple independent observations agree: LiDAR or sonar finds a narrow trap, stereo sees a bedrock texture or black-sand line, PI records a repeatable conductive response, and the surface buoy supplies a coordinate with known offset. The resulting event stores raw sensor windows, calibration state, timestamp, depth, pose covariance, and a confidence grade. It should say ‘repeat here’ rather than ‘gold found’ until the cartridge and laboratory result support that claim.

• Raw + corrected imagery
• Sonar returns and point-cloud confidence
• PI waveform and repeated-pass comparison
• GNSS buoy position, tether offset, and uncertainty
• Cartridge ID and chain-of-custody link

Calibration and test gates

Before a river mission, the camera needs a geometric calibration target behind the dome, the LiDAR needs a range/refraction check, sonar needs a known reflector, and PI needs known conductors plus mineralized-background trials. Run the full stack in clear water, then add controlled turbidity, bubbles, moving illumination, and representative cobble. Acceptance is measured false-alert rate, localization error, dropout behavior, and safe fallback—not a single attractive demo frame.

• Bench → pressure vessel → clear-water tank → turbidity tank → permitted river
• Record temperature, depth, turbidity, and lighting with every test
• Keep an operator review path for every autonomous target