Perspective-n-point. about mit-driverless-cv-traininginfra HOT 3 CLOSED

Hello @EugenKarasev, thanks for your interest in our work.

how you found 7 points on the cone

We developed a dedicated neural network to detect the 7 keypoints on the cone: https://github.com/cv-core/MIT-Driverless-CV-TrainingInfra/tree/master/RektNet

How did we know the parameters of the calibrated camera?

We calibrated the camera in our own, manually.

How we know dimensions of the cone?

The shape of the cone is accessible through FSG's official competition manual, plus you can also measure it by yourself.

Implementation in the code

Pytorch Training: https://github.com/cv-core/MIT-Driverless-CV-TrainingInfra/tree/master/RektNet

TensorRT inference: https://github.com/cv-core/tensorrt_ros/blob/master/KeypointDetector.cpp

from mit-driverless-cv-traininginfra.

Yes, thank you.
But I am interested in exactly, how you determine the position of the cone?

from mit-driverless-cv-traininginfra.

Hi @EugenKarasev, this repo has all of the source files related to object detection. What you'd be looking for is a PnP solver for which there are multiple open-source implementations. You could try OpenCV's solvePnP or solvePnPRansac functions, alternatively this repo solves PnP using a lambda twist P3P solver which is supposed to be faster and more stable

from mit-driverless-cv-traininginfra.