Range sensors (for example sonar, infrared, and lidar) are notoriously noisy. How can I characterize the noise characteristics to include these in a probabilistic localization sensor model?
This subject is covered quite nicely in the Probabilistic Robotics book by Thrun et. al. I don't have a direct reference, but there are some of his papers (such as Robust Monte Carlo Localization for Mobile Robots, pdf) essentially include the same information. Usually what is used is a mixed error model, where the probability density function consists of different parts
- A Gaussian error around the true distance reading
- A part which accounts for false positives like dynamic obstacles and so on. This is larger with smaller distances.
- A constant part which accounts for false negative readings, where the sensor gives an out of range reading.
The model needs to be fitted to your sensor and application.
Almost everybody just assumes the noise is gaussian because that way the math is relatively easy.
If you really wanted, you could experimentally determine the distribution of sensor noise, fit a model to it, and use that but it would be a lot of work for potentially no gain.