Hi Taki,

If allowed, I would rather prefer using something else like visual odometery or SFM (Structure From Motion) on Odroid instead of stereo vision because

• The depth information quality from a stereo vision system deteorates rapidly with z, essentially making it a monocular vision problem.
• The point matching step of any stereo vision algorithm is very a intensive operation, unless you have a decent GPU like that on a Nvidia TX1/2. Odroid has one but it might be too small for a significant speed-up.
• Since you also want to do localization (from a map presumably), you also need to leave some computing for particle filters too, which are also a bit resource intensive. Going with stereo vision on Odroid may not leave any firepower left for global localization.

You can try Direct Sparse Odometry (https://github.com/JakobEngel/dso) or SVO2 (works best with downward facing cameras though) for getting the surrounding point cloud.

For indoor environments, if everything has to run on Odroid, my personal choice would be an RGBD camera like ASUS Xtion or Intel RealSense Camera + a visual odometry package + ROS robot_localization or some other EKF implementation for sensor fusion and scale estimation (if using VO).

Cheers!

Dear,

you need to process at least 15 frames per second to be in the real-time (RT). you need to choose the appropriate way to compute your disparity map to meet the RT requirement

Shubham Paul Than you for your response, unfortunately that's not an option for me, besides I am processing small images(320x240), and I'm using the ARM cpus on the odroid not GPU, and I'm using simple block matching with rectified stereo images. I`m targeting 30 fps for the disparity computation, so far I`m a little bit far from this, do you have an idea about a fast implementation of depth map computation on cpu ?