I am currently writing some tools in Python to visually judge the quality of the odometry. For this I record a bag with odom topic as well as tf, lidar scans and wheel encoders.
Then I loop over the bag data, read all poses from odom and project the lidar scans to cartesian coordinates in the odom frame using the odom pose with the **closest timestamp**.
My odom is published at 100 Hz while the lidar scans are published at around 10 Hz so in worst case the time offset will be 5ms.
I then plot all lidar scans for all poses and see how well they overlap (Similar to setting decay_time of lidar scans to a very big value in Rviz)
Here is the result:
Compared to Rviz (with big decay time)
I noticed that my plot and the data displayed in Rviz (with a decay time covering all the length of the bag while doing rosbag play) looks slightly different.
Especially when zooming on known features, for example the half circle:
My plot:
Rviz:
The difference becomes very clear.
When looking in the source code of Rviz I saw that it just transforms the lidar scan to a pointcloud and then uses TF to transform it to the correct frame.
My question is how to do such a thing without having to play the bag, but only using the rosbag API, on Python? I suspect using the closest timestamp to be the source of error. Does TF interpolate to get the exact pose?
↧