position
All self-driving cars have to go through the same series of steps to navigate the world safely.
You've been learning the first step: positioning. Before vehicles can drive safely, they first use sensors and other data collected to make the best estimate of where they are.
Kalman filter
Let's review the steps required by the Kalman filter to localize a car.
1. Initial prediction
First, we make an initial prediction of the vehicle's location, and then use a probability distribution to describe our uncertainty about that prediction.
Below is a one-dimensional example. We know the vehicle is in this lane, but we don't know its exact location.
A lane and an initial uniform probability distribution.
2. Measurement update
Then, we perceive the world around the car. This step is called measurement update. We gather more information about the car's surroundings and improve our location predictions.
For example, it was measured that the vehicle was located about two grid cells before the stop sign; this measurement is not perfect, but gives us a better idea of where the car is.
Measurement update
3. Forecast (or time update)
The next step is to move. Also known as the time update or prediction step. Based on what we know about speed and current position, we need to predict where the car is going. We need to reflect movement through probability distribution drift.
In the next example, we will do a probability distribution shift to reflect that the vehicle has moved one cell to the right.
prediction step
4. Repeat
Finally, we finally formed a new estimate of the car's location! The Kalman filter simply repeats the sensing and movement (measure and predict) steps to locate the vehicle as it moves!
Steps of Kalman Filter
Tips
The beauty of the Kalman filter is that it combines inaccurate sensor measurements with inaccurate motion predictions to get a filtered position estimate that is better than all estimates from sensor readings or motion predictions alone value is better.