Google Research has developed a novel unsupervised learning approach to generate accurate High-Occupancy Vehicle (HOV) lane ETAs in Google Maps. This innovation addresses the growing need for precise routing as sustainable travel modes like carpooling and electric vehicles become more prevalent. The system analyzes aggregated, anonymized traffic data to identify distinct HOV travel patterns, primarily using speed differentials between HOV and general lanes.

For instance, in Utah's Salt Lake Valley, HOV lanes average 68.18 mph versus 58.60 mph in general lanes—a 16% difference. The classification model processes trip segments by examining speed distributions and estimated lateral distance to the road center, which helps distinguish lane-specific behaviors despite GPS imprecision. The technology uses temporal clustering with weighted median approaches and soft clustering to assign probability-based labels rather than definitive ones.

Final trip classification employs a mixture of experts framework, aggregating segment-level data to determine the proportion of travel time spent in HOV lanes. This advancement enables Google Maps to offer route options that include HOV lanes with reliable arrival predictions, helping commuters save time and reduce congestion.