TURTLMap localises and densely maps low-texture underwater scenes in real time on a low-cost vehicle, a BlueROV2 with an NVIDIA Jetson Orin Nano. From the paper’s HTML: localisation fuses a Waterlinked DVL A50 (velocity only), an external MicroStrain IMU (the vehicle’s own IMU was too slow at about 10 Hz) and a Bar30 pressure sensor, and a downward ZED mini stereo camera supplies depth for Voxblox volumetric mapping at 5 cm voxels.
Findings
- In a tank with motion capture and a ground-truth map: 0.18 m position ATE without waves and 0.26 m with waves, against 0.48 m and 0.36 m for the best baseline (a UKF).
- Map error 0.019 m without waves and 0.040 m with waves; DVL odometry alone gives 1.733 m and 0.806 m.
- ORB-SLAM3 and SVIn2 failed in these low-texture conditions.
Limits
- Tank only, so no turbidity, currents or real seafloor.
- I read the abstract and a summary of the paper’s HTML, not the full text.
Relevance
When vision has nothing to grip, TURTLMap leans on DVL, IMU and pressure and uses the camera for the map. A small team of cheap vehicles could carry this sensor mix. See underwater perception and AQUA-SLAM.