Learning-AI

Traffic Light Mapping and Detection

November 2020

tl;dr: Google/Waymo’s early efforts on traffic light mapping and detection.

Overall impression

The state of TFL can ONLY be perceived visually.

Maps are important. Using a prior map (that includes stop signs, speed limits, lanes), a vehicle can largely simplify its onbaord perception requirements to the estimating its position wrt the map (localization) and dealing with dynamic obstacles.

Using a map, both FP and FN are fail-safe. For FN, map indicates there should be a traffic light, and the car should take conservative actions (braking and alerting the driver). For FP (from brake taillight), the car should be braking anyway.

Key ideas

Technical details

Notes