I'm in the process of tuning a robot-localization package parameters in my sensor-fusion ROS2 package. I have a dataset from various sensors mounted on a vehicle (IMU, GNSS, Lidar, Rgb Cameras, Radar, ...) without a direct odometry input from the vehicle.
The RTK GNSS I'm using has a 20Hz output, Imu publishes with a frequency of > 300Hz.
That said, I'm using a combo of ekf_node and navsat_transform_node according to tutorial where the GNSS is converted into odometry/gps relatively. I've got no map frame.
I'd like to maximize the output frequency of the package possibly up to the frequency of the IMU if this can be handled computationally (or if it even makes sense?).
I've been increasing the frequency param of the ekf_filter_node, but at around 60Hz the actual output drops down abruptly to only around 5Hz. Fiddling around with parameters that I've felt could decrease the computational time yields basically no additional FPS.
This is my config file and a relevant part of the launch file:
ekf_filter_node:
ros__parameters:
frequency: 100.0
sensor_timeout: 0.001
two_d_mode: false
reset_on_time_jump: true
print_diagnostics: false
debug: false
publish_tf: true
publish_acceleration: false
odom_frame: odom
base_link_frame: imu
world_frame: odom
odom0: odometry/gps
odom0_config: [ true, true, true,
false, false, false,
false, false, false,
false, false, false,
false, false, false ]
odom0_queue_size: 1
odom0_nodelay: false
odom0_differential: false
odom0_relative: true
imu0: imu/data
imu0_config: [ false, false, false,
true, true, true,
false, false, false,
true, true, true,
true, true, true ]
imu0_nodelay: true
imu0_differential: false
imu0_relative: false
imu0_queue_size: 1
imu0_remove_gravitational_acceleration: true
use_control: false
process_noise_covariance: [0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
initial_estimate_covariance: [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
navsat_transform_node:
ros__parameters:
frequency: 20.0
delay: 0.0
magnetic_declination_radians: 0.08726646 # For lat/long 49.2, 16.6
yaw_offset: 1.570796327 # IMU reads 0 facing magnetic north, not east
zero_altitude: false
broadcast_cartesian_transform: false
publish_filtered_gps: false
use_odometry_yaw: false
wait_for_datum: false
.
Node(
package='robot_localization',
executable='ekf_node',
name='ekf_filter_node',
output='screen',
parameters=[os.path.join(pathlib.Path().resolve(), 'params', 'ekf.yaml')],
remappings=[('imu/data', config_yaml["raw_topics"]["imu"]["imu"])]
),
Node(
package='robot_localization',
executable='navsat_transform_node',
name='navsat_transform_node',
output='screen',
parameters=[os.path.join(pathlib.Path().resolve(), 'params', 'ekf.yaml')],
remappings=[('imu', config_yaml["raw_topics"]["imu"]["imu"]),
('gps/fix', config_yaml["raw_topics"]["gnss"]["position"])]
)
I'm also adding the actual framerate and ROS2 graph as reported by the rqt
Happy to get any suggestions as to what could the bottleneck be and why the FPS drops significantly when I ask for ~60Hz and more.
