From a fuzzy GPS fix to the right lane

A car drives a real San Francisco street grid. Each classic technique is layered on in turn and scored against ground truth.

GPS conditions
Drag to pan · scroll to zoom
0.0 s

Position error over time (m)

The stack, one layer at a time

StageNowRMSLateralRoadLane
  • Kalman filter fuses 1 Hz GPS with 10 Hz wheel speed and gyro, smoothing noise and coasting through the overpass, but it does not know where the roads are.
  • Map matching runs a hidden Markov model over nearby road segments, keeps only driveable sequences, and snaps the car onto the road it picks.
  • SLAM landmark map: a mapping fleet builds a map of poles and signs with graph SLAM. Localizing against it removes the GPS bias that map matching leaves along the street.
  • Lane map: a lane camera's solid and dashed markings pick the lane, and its in-lane offset places the car within it.
Loading the map…