Questions about sensor fusion with Lidar - Extended Kalman Filter

Question

I am attempting to teach myself sensor fusion as I suspect I'll need to do this down the road with lidar and some other sensors. In all my research so far it sounds like a version of the extended kalman-filter is the best way to do this.

I believe I have a good intuition for how the EKF works but I am stuck on a few things. First let me describe my understanding of the process in general.

1. Starting at time $k_1$, use a motion model to predict the pose of the vehicle at time $k$
2. Use a measurement model to describe where we can expect the measurement to occur
3. Once a measurement is observed we use step 3 to decide how much of the measurement should be used to update the predicted state
4. Update the predicted state with information from step 3.
5. Repeat forever

What I am really stuck on is how do we measure the pose of the robot with a sensor like lidar or radar? Several examples I have found show the raw lidar data as if it were the measured pose of the robot, instead of the distance from a object. They use this raw data in step 4 above but this doesn't make sense to me. How can we get a better prediction of the state using the raw lidar data if we can't actually measure the state of the robot using the lidar?

Answer 1

One common way of using LiDAR to estimate the pose of a robot is by scan-matching against a previously-created occupancy grid (map), then using the match scores in a pose estimation system like AMCL. Then, that pose estimate is often fed into an EKF, along with other state measurements (wheel odometry, IMU), to "fuse" them all together to form a filtered pose estimate. If you Google terms like "AMCL" and "scan-matcher", you'll find a wealth of literature and tutorials about using LiDAR for this purpose.