Comments (9)
First of all, thanks for your interest, @JanOpper. Yes, the IMU parameters play a very important role in the visual-inertial estimation, because those values will contribute to the covariance which is the key component for computing the state update. It would cause some critical problem if the IMU parameters are set with random numbers, as shown in your result, because this may incorrectly change the uncertainty about IMU measurement in the sense of sensor fusion with Kalman filter.
If those values are not provided with the dataset, you can instead get the reference values based on the sensor datasheet. For example, kitti uses RT3003 GNSS/INS as the inertial sensor whose datasheet (RT3000 v3) can easily be found online. Therefore, the IMU parameters can be computed yourself by referring to the 'bias stability' (n_ba, n_bg) and 'random walk' (n_a, n_g), with particular care about the unit conversion https://github.com/ethz-asl/kalibr/wiki/IMU-Noise-Model. Once you get those reference values, you can try it again to see what you can get now. Good luck.
from r-vio.
Thank you for your feedback.
I have now tried the values from https://www.oxts.com/software/navsuite/documentation/datasheets/RT3000v3_ds.pdf and hopefully converted them correctly. One example is the conversion from OXTS' Gyro "Random walk" to Kalibr's "White Noise" / "IMU.sigma_g" in the config yaml:
Those parameters yielded the following result:
Unfortunately this performs worse than with the manually set values.
I have also observed that the Mahalanobis distance check excludes most of the visual features.
I made another experiment and turned off the update step, so that only the IMU data is being integrated.
Interestingly, there are no visible turns at all. I am wondering, how much is the performance of the algorithm being affected by the IMU data rate (100dps) and the camera frame rate (10fps)?
I would also like to know: What are good start conditions for this algorithm, especially in urban driving?
from r-vio.
I have another question: To what extent does the algorithm depend on time-synchronized data? Because I noticed that EUROC is actually time synchronized, whereas Kitti is not.
from r-vio.
Currently our algorithm requires the results of both spatial and temporal calibrations i.e., the data time offset between camera and IMU is needed (as listed in the config file). So if this value is not set properly, the performance shall not be guaranteed. Especially, if the data rates of camera and IMU are too different, the influence of this value will become more significant which could lead to the result that you have shown.
from r-vio.
Both the IMU and the camera have absolute timestamps, so there should be no fixed time offset. However, the camera was coupled to the lidar and not to the IMU, i.e. the photos were taken when the lidar pointed forward. This can lead to some jitter-like fluctuations in the camera frequency.
from r-vio.
I continued working on the dataset and was able to increase the accuracy substantially.
- I lowered the threshold for the initial detection of movement to 1/10 of the value for EuroC.
- I deactivated the Sampsons error.
- I set the noise values for the IMU to 10x of the converted values from the datasheet. I think that is a reasonable way to account for any unmodeled effects.
With these changes I get the following result:
So it still jumps at some points, but much less than before.
from r-vio.
Very glad to see this result @JanOpper. The threshold for motion detection is tunable (marked in the config file) depending on the sensitivity of IMU. However, the Sampson error is used for the ransac which is also tunable, so maybe you can try to increase that value instead to allow more features to be involved. While recently I have not found time to test this dataset myself, if it is possible, could you share with me the config file and data you are using so that I can try it sometime.
from r-vio.
Interestingly I had to decrease the inlier threshold for the ransac, because there were too many features included. Regardless of this, the code ran just as well as without the Sampson error.
I also uploaded the data you requested here: https://www.dropbox.com/s/09ys3ndxuaeyc8z/Kitti_RVIO.zip?dl=0
I hope that this is a suitable way to share these large amounts of data.
And thank you again for your support.
from r-vio.
Interestingly I had to decrease the inlier threshold for the ransac, because there were too many features included. Regardless of this, the code ran just as well as without the Sampson error. I also uploaded the data you requested here: https://www.dropbox.com/s/09ys3ndxuaeyc8z/Kitti_RVIO.zip?dl=0 I hope that this is a suitable way to share these large amounts of data.
And thank you again for your support.
@JanOpper
Thank you for your contribution!
I am interesting in using R-VIO with the kitti raw dataset, and also had some problems about it. I tried your dataset and config file, it worked well. But when I changed the dataset to my kitti rosbag ( 2011_10_03_drive_0027 raw unsynced IMU data and gray left cam), the code couldn't work. I am hoping your support and you may tell how your rosbag is composed. Thank you so much!
from r-vio.
Related Issues (20)
- About the algorithm‘s harmful and unbelievable drifting when it worked in several static situations below HOT 6
- Please excuse me. Some questions about algorithmic equation deducing and a certain experiment operation details. HOT 2
- Are the urban dataset available?
- No visualization is seen in RVIZ HOT 3
- Question about the Ω(w) Matrix. HOT 1
- Question About the F matrix in code. HOT 4
- Some tips about the propagate euqation HOT 3
- Question about the Integration estimate HOT 5
- 关于初始化问题的了解 HOT 1
- run error ubuntu18.04 HOT 2
- Question about the update. HOT 2
- Question about the urban driving dataset
- PUBLISH RATE SLOW OF VIO(px4_realsense_bridge_node)
- Question about the trick R-VIO used in the real environment. HOT 2
- Question about the IMU Propagation using Robocentric formulation HOT 22
- No output
- How to save a .txt track
- How to get a dataset?
- How to install Eigen Lib?
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from r-vio.