The July 2017 cohort has officially kicked off Term 2, and I am super stoked to be back working on some challenging problems in the self-driving car space. This term is all about sensors, localization, and control. Sensors meaning reading and gleaning meaningful insights from LIDAR and radar data (e.g. predicting velocity and position by combining LIDAR and radar data). Localization meaning “where is the car right now?”. And control, or determining what angle the steering wheel should be at or how fast should the car be driving?
We kicked off by diving head first into Sensors. The major algorithm that we are making use of is the Kalman Filter. Check out this low level introduction here or even go through the Udacity YouTube video here. The YouTube videos are part of Udacity’s free Artificial Intelligence for Robotics course. I found the simple idea and theory behind Kalman filters to be fairly straight forward, but the math not so much.
The basic idea is that you can predict the future state of the vehicle given a past state (e.g. past location / velocity).
The past state data will be run through a set of equations that will predict the future state (see picture below).
You can then compare this predicted state to the data you are receiving from the sensors and gather an error rate (via RMSE). This is incredibly powerful in the context of self-driving cars since you can predict where a pedestrian may be walking, which direction a biker is moving, or where / when a car will be turning.
By no means do I feel like I have a complete grasp on Kalman filters – honestly more like barely touching the tip of the ice berg in terms of understanding. But, I think it’s enough to get started on Project 1 of Term 2 (Extended Kalman Filter Project). I’ll be back with an update after I finish the first (or maybe second project) since they both utilize Kalman filters.