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
- Position estimation
- at least two labels in diff images are needed, and the position estimation will be more accurate if more labels are available.
- Assuming TFLs are about 0.3 m in diameter.
- TFL Semantics
- drivers need to know which lights are relevant to their current lane and to their desired trajectory through the intersections. This can be represented as an association between a TFL and the different allowed routes through an intersection.
- Some heuristics are used to label then manually verified.
- Traffic light control:
- in the path, no red/yellow lights and at least one green, then the car is allowed to proceed. Default color to yellow.
- There are almost always multiple semantically identical TFL in an intersection, it is only necessary for the system to see one of these lights to determine the TFL state.
Technical details
- TFL Mapping has camera exposure set to a constant value. The image looks dark even during the day.
- For classification, insist on no FP green lights.
Notes
- Questions and notes on how to improve/revise the current work