Why pothole evidence needs repeated drives—and lane uncertainty

Imagine a road with five lanes in one direction. One vehicle feels a sharp impact; another passes the same stretch without a noticeable response. Both recordings may be accurate. Their wheels may simply have travelled over different parts of the surface.

This is why Xinexis plans to combine multiple drives on the server. The objective is to build a fuller account of a road section: which observations recur, under what conditions, and how much usable coverage supports the interpretation. This aggregation is designed work ahead of us, not a capability already deployed in the prototype.

Match the road and direction before combining events

Nearby GPS points are not necessarily observations of the same road. Opposing traffic, parallel streets, ramps and bridges can occupy a small geographic area while representing different surfaces.

Our planned approach uses a sequence of positions, timing, travel direction and road connectivity to associate a pass with a directed road section. Newson and Krumm's map-matching research (opens in a new tab) describes using measurement noise and network structure together to interpret GPS sequences. Applying that idea here requires a suitable road network and evaluation on our own recordings.

The measured position will remain separate from any matched position. Unresolved road assignments should stay uncertain rather than quietly adding support to the nearest line on a map.

A lane count is not a lane measurement

Knowing that a road has three or five lanes does not tell us which lane a vehicle occupied. Apple defines horizontal accuracy (opens in a new tab) as a radius of location uncertainty. It is not a lane identifier.

Our initial unit of evidence will therefore be road section and direction, with lane and wheel path unknown unless separately justified. Repeating a biased GPS trace does not automatically remove the bias or reveal a precise lane.

There is also no fixed multiplier that turns lane count into the required number of drives. Traffic does not necessarily use lanes evenly, and vehicles can avoid a visible defect or miss it within the same lane. More journeys create opportunities for evidence; they do not guarantee complete surface coverage.

Count physical passes, then describe diversity

One passage over a candidate interval should contribute at most one supporting observation. Suspension ringing, multiple threshold crossings, an upload retry or a new analysis run must not create additional votes.

Repeatability and diversity are different. Several passes by one vehicle help test repeatability. Passes from different known vehicles and different days broaden the evidence. An app installation is not a verified driver, and two phones in one car do not establish two independent vehicles.

The planned review will keep these counts separate and show when independence cannot be established. That helps prevent a large file count from looking like broad confirmation.

Keep the denominator and the missing data

An event-only database can count impacts but cannot explain how often the road was observed. We also need usable movement intervals, including passes with no detected impact. Missing sensors or an ambiguous road match are excluded coverage, not evidence of a smooth surface.

Consider a purely illustrative summary: 12 supporting passes from 6 known vehicles across 4 days, within 48 usable same-direction section passes; lane unknown. The 12-to-48 ratio describes recorded responses under those collection conditions. It is not a pothole probability, and the other passes may have used different wheel paths.

Recurrence creates a review candidate

A recurring response may come from a pothole, a joint, a manhole or a speed bump. The server should first expose the supporting recordings, approximate interval, dates and uncertainty. Human-reviewed feature labels and held-out evaluation are needed before claiming a reliable cause or calibrated probability.

Recent evidence also needs its own time window. Old observations should not silently describe today's condition, and later non-detections with unknown wheel paths should not automatically mean a repair succeeded.

This is the next research step for Road Monitor: turn individual measurements into evidence that a reviewer can inspect. If you can help define useful road-maintenance review criteria, start a conversation with Xinexis.