Erroneous trajectory of egocentric vehicle

[jaskaran1]

Using the code in the Readme for converting imu to world coordinates, I plot the trajectory of the egocentric vehicle in 2d plane(xy) and the trajectory doesn't seem to be correct in some of the videos.

For instance, in City category for drive 93, I get this plot which doesn't look right at all

after plotting point_w array below

point_imu = np.array([0,0,0,1])
point_w = [o.T_w_imu.dot(point_imu) for o in data.oxts]

Whereas the same code for drive 15 in Road category gives the correct trajectory of the egocentric vehicle below:

Are the GPS/IMU tags noisy and are x and y in meters here?

[MouvementMondial]

I think there is a problem with the data. Take a look at this files, especially the altitude values:

.../2011_09_26/2011_09_26_drive_0093_sync/oxts_data

The readme describes the structure of the files.

[jaskaran1]

The altitude value is negative.

[leeclemnet]

It looks like the GPS/INS system didn't have a good lock on that particular sequence. Looking at the first OXTS packet in the dataset, we have

pos_accuracy=43.329367454418, 
vel_accuracy=0.34225136961012, 
navstat=4, 
numsats=2, 
posmode=0, 
velmode=0, 
orimode=0

And this seems to be representative of the entire sequence. So if the position accuracy (I assume this is a standard deviation or similar) is 20-40 meters everywhere, that would explain why your first plot doesn't make sense.

In contrast, looking at 2011_09_30_drive_0018_sync, the position accuracy is on the order of centimeters, there are significantly more satellites being used, and the pos/vel/orimode flags are different:

pos_accuracy=0.067801179930736, 
vel_accuracy=0.017029386365926, 
navstat=4, 
numsats=9, 
posmode=5, 
velmode=5, 
orimode=6

[jaskaran1]

@leeclemnet That makes a lot of sense.

I am trying to create a dataset for trajectory from all the videos in the kitti dataset. What all do I check for a particular sequence(drive) before including its trajectory information?

Do I just check the values for pos_accuracy, vel_accuracy?

[leeclemnet]

I don't have a straight answer for this because I usually just use the raw sequences corresponding to the odometry benchmark.

As far as I know, there isn't a list showing which raw sequences have good GPS and which ones don't. Excluding measurements or entire sequences with a pos_accuracy above some threshold seems like a reasonable thing to try.

I suspect posmode=0 is also a sign that something has gone wrong, but I'm not familiar with what these mode flags actually mean. You could try searching for some documentation about the OXTS packet format.