After predicting the velocities (Vx, Vy, Vz) in camera sensor coordinates, we convert them to global coordinates using:
(Vx_glob, Vy_glob, Vz_glob)^T = ego_to_world * cam_to_ego * (Vx, Vy, Vz)^T
However, if the ego vehicle is turning, it has angular velocity, and the rotation matrix changes: The derivative of the rotation matrix is not 0 (https://arxiv.org/pdf/1311.6010.pdf).
If the ego vehicle is not turning, we can convert the velocity from one coordinate system to another using the relevant rotation matrix. However, when the ego vehicle is turning, we need to consider also the angular velocity in the rotation matrix derivative.
How do you propose to convert velocity vectors between different coordinates when the ego vehicle is turning?