zhiangchen / gps_vio

Odometry Fusion of PX4 GPS and Realsense T265 VIO

Image Classification

deep learning image classification

gps_vio's Issues

camera external calibration

The tf between the tracking camera and px4 controller is lacking. Also, their frame_ids are different.
Todo:

1. transform every VIO odometry to the initial px4 coordinate system before fusing it to the graph
2. change the "frame_id" value to "map"

calibration for each parameter

Now the transform calibration is divided into rotation matrix (9 params) estimation and translation matrix estimation (3 params). However, the calibration results are not always good because data samples along some directions are not sufficient, which can be a common mistake when people are not cautious about this during collecting samples. e.g. when we are calibration translation, if we only have samples from roll rotation, we can just estimate camera's x, but not y and z.

So we need to add instructions to collect samples for each parameter.

fake Odoms in gazebo

You can't use a fake odom to calibrate the translations between camera and UAV. e.g. we know gps_odom, it is not proper to have vio_odom = T * gps_odom.

The reason is that these two odometries will be geometrically identical, which means the two sensors are moving parallelly without any rotation. Thus, you can't have translation calibration.

no calibration needed

What we essentially need is to transform P_vio_vio to P_uav_uav, and we have P_uav_uav = T_vio_uav * P_vio_vio * T_uav_vio, where P_uav_uav is a linear to P_vio_vio. Thus, we don't even need to directly solve external calibration problem.

We need to estimate A from vec(P_uav_uav) = A * vec(P_vio_vio), and A is a 12x12 matrix.