/ARM-VO

Efficient monocular visual odometry for ground vehicles on ARM processors

Primary LanguageC++

ARM-VO

Authors: Zana Zakaryaie Nejad and Ali Hosseininaveh

ARM-VO is an efficient monocular visual odometry algorithm designed for ARM processors. It uses NEON C intrinsics and multi-threading to accelerate keypoint detection and tracking. Check this video to see the performance on Raspberry Pi 3 and Odroid XU4.

Dependencies

How to build?

git clone https://github.com/zanazakaryaie/ARM-VO.git
cd ARM-VO
cmake .
make

Test on KITTI dataset

Download the odometry dataset from here. Open a terminal and type:

./ARM_VO pathToData paramsFileName

Limitations

  • ARM-VO recovers the scale if the camera height and pitch angle are provided. Thus, it is not applicable for drones or hand-held cameras.

  • The algorithm detects small-inter frame translations and pure rotations using GRIC but it doesn't decompose the estimated homography matrix. Track is lost if the camera rotates too much without translation.

Notes

  • If you get low FPS, check your power adapter. Raspberry Pi 3 runs ARM-VO at 8 frames per second (averagelly) if powered up with a 5V-2A adapter.

  • If you use ARM-VO in an academic work, please cite:
    Zakaryaie Nejad, Z. & Hosseininaveh Ahmadabadian, A. Machine Vision and Applications (2019). https://doi.org/10.1007/s00138-019-01037-5

  • ARM-VO is a part of six-wheel surveying robot project named MOOR.