Which new machine vision module would you like?

Question

Welcome everyone to JeVois. Please post as answers to this question ideas for new machine vision algorithms which you would like to see implemented on JeVois. Please include if possible:

- brief intro to the vision algorithm

- link to the source code of a library that implements key parts of the algorithm

- any rationale that it would run at a decent frame rate on the small JeVois processor

Please vote the suggestions you like up, and we will get to work on the most popular ones!

Answer 1

MonoSLAM - an algorithm that allows one to recover the trajectory in 3D space of the camera as it moves along. The algorithm computes frame-to-frame camera displacement in space through visual matching of successive frames. This can be used to estimate how the camera is moving and where it is, over time.

https://www.doc.ic.ac.uk/~ajd/Publications/davison_etal_pami2007.pdf

https://www.youtube.com/watch?v=mimAWVm-0qA

https://github.com/hanmekim/SceneLib2

https://github.com/rrg-polito/mono-slam

It worked in real time on commodity computers circa 2007 so should get a decent framerate on JeVois.

Answer 2

A module for finding landing (flat) area for drones from an aerial view. It will be interesting to use machine learning to find 'land-able' surfaces even thought simple CV algorithms can do this.

Answer 3

Note that I haven't checked to see if the existing examples might cover these suggestions. e.g. I know there's an optic flow example that might be suitable to turn into number 1)

1) Optic flow algorithm suitable for use to calculate time to impact. (Good for landing a drone)

2) Depth from disparity for successive frames for a rotating view. (Something a drone could use when hovering and turning on the spot to detect obstacles nearby)

Terry

Answer 4

Not a machine vision module, but would like to see some ROS support. Possibly rosserial?

https://wiki.ros.org/rosserial