Why are there so few published implementations of SLAM with support for heterogeneous sensors?

Heterogeneous sensors include LIDAR, monocular cameras, stereo cameras, thermal sensors from more than one manufacturer on one or multiple robots.

I am concerned that the lack of availability is not just a matter of time, effort, and politics but also of mathematical/algorithmic feasibility.

  • $\begingroup$ what do you mean by 'politics' in this context? $\endgroup$
    – FooTheBar
    Nov 21 '19 at 8:59
  • $\begingroup$ I mean preference for a particular set of sensors and intermediate representations such as maps. The SLAM implementation in Apples AR headset may not readily work on Microsofts AR headset. $\endgroup$ Nov 21 '19 at 16:04

I would somewhat disagree with the statement of lack of heterogenous systems. It is pretty common to always combine a camera sensor with an IMU. LIDAR + Camera has also been a pretty popular idea over the last few years.

In regards to lack of more exotic sensors e.g Thermal, RADAR I would say it is due to the fact that the main purpose of SLAM is to build 3D maps. The best sensors to give you 3D information is LIDAR and Cameras so the majority of focus has been on those two systems.

A couple of other reasons for lack of heterogeneous sensor systems:

  1. Cost:

    Adding more sensors, especially more exotic ones like Thermal cameras adds to the cost of your sensor platform.

  2. Computational Complexity

    More sensors means more computational cost. Not only in the backend portion of fusing the sensor information, but also your system requires now multiple frontends. This becomes especially prevalent on small robots like UAVs, where computation is a premium.

  3. Lack of benefits

    The sensors have to complement each other. A LIDAR + Stereo Pair is not a good combination, as the purpose of the Stereo Pair is made redundant with the LIDAR.

  4. Form Factor/Availability/ Ease of Use

    The sensors could be too big to put on a small vehicle, getting access to good ones could be difficult. They may require exotic connectors, or propriety software libraries. A bunch of different reasons that make it too complicated to use compared to a simple camera.

On the topic of algorithm feasibility the backends used in Sensor Fusion such as Kalman Filters and Factor Graphs are abstract enough to allow any combination sensors. So this is not an issue.

  • $\begingroup$ Thanks for an excellent, multi-perspective answer, @edwinem. Is it correct to assume that it should not be difficult to repurpose a SLAM system to use the same type of sensors but from different manufacturers? For example, if a SLAM system is designed to work with Intel Realsense, it should be straightforward to plugin some other stereo camera (RGBD) setup? $\endgroup$ Nov 22 '19 at 7:40
  • $\begingroup$ That is correct. Using a stereo slam system with a different stereo sensor should only involve you changing some of the configuration settings. Certain SLAM systems are even more configurable like ORB-SLAM2. It can allow for RGBD camera,Stereo Pair, or a Monocular Camera as its sensor. $\endgroup$
    – edwinem
    Nov 22 '19 at 19:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.