huaizheng
commented
4 years ago 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.
JanOpper
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?
So it still jumps at some points, but much less than before.
wsh122333
Hi,
first of all thank you for sharing your code with the community! I am trying to apply the code to the kitti raw dataset, specifically a residential recording (2011_10_03_drive_0027). I am starting the estimation from a point where the car is at standstill und doing a right turn afterwards, because I was hoping the gravity and bias estimation would benefit from this. I already found out that the accuracy is highly dependent on the IMU parameters, but unfortunately I couldn't find the actual values for kitti. I tried to estimate them myself via an Allan standard deviation plot from a few seconds of standstill data from a different track, but from my understanding this is only sufficient for the white noise values, but not the bias random walk. I therefore guessed an appropriate value for the latter. I also found out that I need much higher white noise values than calculated for the turns to be recognized at all, about two orders of magnitude (used sigma_g: 0.0046, sigma_a: 0.06).
As you can see in the attached screenshot the algorithm is doing massive position corrections at each 90 degree turn. I am assuming this is because the uncertainty is reduced when the car is turning. But I am wondering why the uncertainty und deviation is so high in the first place. From what I have seen in your paper, you had much less deviation and corrections when you run the code on your urban driving test data.
I was hoping maybe you have an idea why the performance is not as good as in your paper and how I might improve this.
I will happily provide further details of the configuration or data I have used, if necessary.
Thank you very much!