When I tested with own data, reset occurs frequently. (Mono-Inertial)

[YouMinJung]

Hi. when I tested with EuRoC dataset, I get the good result. So I want to test with own data.

I collected camera images and imu data with timestamp. My dataset is like below.

Furthermore, I calibrated camera-imu with kalibr. This is the own YAML file.

However, I tested with orb-slam (mono-inertial), remapping occured very frequently. Reset occurred approximately every 4 seconds even when I walk straightly forward.

What is the problem? How to solve this problem..?? help..

[makolele12]

Are you sure your NoiseGyro value is okay? It is way to big in comparison to the other 3 parameters.

I am having the exact same behavior in Monocular-Inertial. I am using the Intel RealSense D435i camera and even when using the calibration parameters they propose, the IMU does not get initialized and the map is frequently being reinitialized.

This does not happen to me with the same camera and videos but in Monocular-only though.

[YouMinJung]

I calibrated imu sensor using imu_utils. When I record the imu data for the calibration, I shook the sensor approximately 45 minutes. Is this correct action?

[Maithilishetty]

Even I am facing frequent reset problem. Did anyone solve this?

[ynma-hanvo]

same for me. I used stereo-Inertial with my stereo cameras, parameter from kalibr result. also tracking reset every few seconds. for stereo is ok

[rancheng]

have a look here guys:

ORB-SLAM3's YAML file has the following unit: rad/s^0.5 and m/s^1.5:

you probably calibrated the noise density with kalibr, please make sure you convert it back to noise in continuous time!

Always pay attention to the details.

kalibr gyro\accel noise or random walk density can be converted back to noise or random walk by:

import numpy as np
your.gyro.noise_density / np.sqrt(1/200.0) # assume your imu frequency is 200 HZ
your.accel.noise_density / np.sqrt(1/200.0)

your.gyro.random_walk_denity * np.sqrt(1/200.0)
your.accel.random_walk_denity * np.sqrt(1/200.0)

see also: issue #280

[sandstorm666]

@rancheng I did what you said and still had frequent resets. I calibrated the parameters of the imu as follows. IMU0: Model: calibrated Update rate: 200.0 Accelerometer: Noise density: 0.0181066131887 Noise density (discrete): 0.2560661794 Random walk: 0.000479934988114 Gyroscope: Noise density: 0.00191159599129 Noise density (discrete): 0.0270340497667 Random walk: 1.34283856163e-05 This happens even though I multiply or divide by the square root.

[wwendy233]

have a look here guys:

ORB-SLAM3's YAML file has the following unit: rad/s^0.5 and m/s^1.5:

you probably calibrated the noise density with kalibr, please make sure you convert it back to noise in continuous time!

Always pay attention to the details.

kalibr gyro\accel noise or random walk density can be converted back to noise or random walk by:

import numpy as np
your.gyro.noise_density / np.sqrt(1/200.0) # assume your imu frequency is 200 HZ
your.accel.noise_density / np.sqrt(1/200.0)

your.gyro.random_walk_denity * np.sqrt(1/200.0)
your.accel.random_walk_denity * np.sqrt(1/200.0)

see also: issue #280

I don't know if I understand it wrong. Should the value in sqrt() be changed to 200 instead of 1/200? For example, when the acc_n output by kalibr is 0.004 and the frequency is 200, should it be written as 0.004/sqrt(200) in ORB-SLAM3's YAML file?

[rancheng]

@wwendy233 are you calibrating the imu with this tool: imu_utils ?

[sandstorm666]

@wwendy233 are you calibrating the imu with this tool: imu_utils ?

i solved

[mahammadirfan]

Hi @rancheng , I am using https://github.com/ori-drs/allan_variance_ros Do i need to convert the unit as well ?

[ricardo11235]

Hi @YouMinJung ， Can you tell me what type of imu you're using？Does the fps of imu affect the algorithm？my imu can only output 75fps

[LarryDong]

Hi guys, I searched for the unit and gives the conclusion:

1. Try to use https://github.com/ori-drs/allan_variance_ros but not imu_utils. Check this discussion for detials.
2. allan_variance_ros's output unit is just the same as orb-slam3, no convertion (like vins-mono, convert to discrete time) is needed. The unit is just continous form, which I think @rancheng 's comment is wrong.
3. Reset frequently comes from other issues, not just the unit.

[wqrydqklyu]

Are you sure your NoiseGyro value is okay? It is way to big in comparison to the other 3 parameters.

I am having the exact same behavior in Monocular-Inertial. I am using the Intel RealSense D435i camera and even when using the calibration parameters they propose, the IMU does not get initialized and the map is frequently being reinitialized.

This does not happen to me with the same camera and videos but in Monocular-only though.

I am using the same sensor, the Intel RealSense D435i cam, have exactly the same problem you encountered: it works in monocular-only mode but reset frequently in monocular-inertia mode. Have you solved it?