Pose estimation using monocular images

Hi,

I'm using the Matlab Vision Toolbox. I have to estimate nle pose of an object using a sequence of imagines. I'm using the sequence of commands illustrated in https://it.mathworks.com/help/vision/ug/monocular-visual-odometry.html ; the code can estimate correctly the relative pose from an imagine to an other. Problems coms when I try to compute also the first one. Is there a method to find the correct first Pose?

[Manfredi Calderaio]

Thanks in advance

Answer 1

[Asim]

Hello Manfredi,

Estimating the initial pose of an object in a sequence of images can be challenging because visual odometry typically relies on relative motion between frames. For the first frame, you don't have a previous frame to compare against, so the initial pose is often assumed to be the identity transformation (no movement).

To establish an initial pose, you have a few options:

1. Use Known Geometry: If you have information about the scene's geometry or certain features in the first image, you can use that information to define the initial pose.
2. Reference Object or Marker: Place a known object or marker with a predefined pose in the first frame. You can then use feature matching to estimate the camera pose relative to this object.
3. Manual Input: If it's a controlled environment, you might manually measure the pose for the first frame and input it into your system.
4. Sensor Fusion: If available, use data from other sensors (like IMUs or GPS) to estimate the initial pose.
5. Additional Frames: Consider capturing additional frames before the sequence you want to analyze, where the camera's motion or object's pose is known, to establish a baseline.

Remember that the pose in monocular visual odometry is up to scale, so your initial pose might need to account for this as well.

You may find the following research paper discussing initialization methods in monocular SLAM systems useful:

https://ieeexplore.ieee.org/document/7219438

I hope this helps.

Best Regards,

Asim Asrar