Skip to content

Latest commit

 

History

History
40 lines (20 loc) · 2.65 KB

HardwareSpec.md

File metadata and controls

40 lines (20 loc) · 2.65 KB

Hardware

Sensor Head Specification

For our dataset, we use the following:

Example Zed Mini image stream from Tunnel Circuit:

StereolabsZedMini

  • Multiple different depth camera systems were tested before I decided that the Zed Mini was the best option for our requirements.

  • The depth estimation backbone is passive stereo matching. This seems like an odd choice of camera to send into a dark cave/mine without natural illumination. However, with our onboard active illumintation system we were able to achieve good depth estimation performance underground.

  • Since this is the primary sensor for artifact recognition, one of the primary requirements was having good quality RGB imagery, and in my comparisons with the ZED's RealSense counterparts, I believe the quality of the RGB cameras on the Zed outperform it's alternatives.

  • The Zed mini is also able to better handle dust, which was an important factor in our decision. Active illumination in a dust-filled environment brings interesting challenges with visual degradation of imagery. I am very happy with the performance of this sensor in these environments. I suspect that the choice of sensor along with the fact that this is a rolling shutter system, makes individual dust particles less prominent.

  • Lastly, the icing on the cake - the stereo camera calibration tool. Having a tool this easy to use in the field is invaluable. Hats off to the Stereolabs team for building such an intuitive calibration software. As someone who has spent the last 3 years calibrating stereo cameras, you have my respect and gratitude.

RGBImage Thermal

  • This camera appears to be somewhat of a favourite among the SubT teams. I noticed them on numerous platforms at Tunnel Circuit and it isn't surprising. I have used the Boson 320 in the 50deg and 92deg variants and they are excellent devices that require very little in the way of compute or set-up to get running.

  • For this dataset, we use the FLIR Boson 320 2.3mm 92 degree HFOV model. We use the 60Hz variant and record our data in both 8-bit AGC as well as raw 16-bit.

  • The 320 variant seemed sufficient for our task; the 640 must be much nicer but it is also significantly more expensive.

Lighting

  • There's not much to say about these lights. They're lights. They're bright lights. They're very bright lights.