A pattern is not a number
"Traffic on the corridor was up nine percent" is a number. It supports almost no decision on its own. What changes a decision is shape: when the peak begins, how long it holds, which movement is growing, and whether the pattern is drifting week over week or was a one-day event.
Corridor data becomes useful at the point where an agency stops asking how much traffic there was and starts asking what the traffic did.
The four questions worth instrumenting
- When does the corridor actually peak? Not the nominal rush hour — the measured one, per direction, per day of week. Peaks move with school terms, shift patterns and seasonal freight.
- What is on the road? A corridor carrying eight percent heavy goods behaves nothing like one carrying two percent, at the same total volume.
- Where do vehicles turn? Turning volume is usually the number a signal plan hinges on, and it must be counted rather than apportioned from totals.
- How long does recovery take? The time from incident clearance to normal flow tells you more about corridor resilience than the peak queue length ever will.
Counts you can defend
Analytics get used in funding submissions, public meetings and design reviews, which means somebody will eventually challenge them. Three habits make that survivable.
First, publish the confidence distribution alongside the counts, and state the floor below which detections are discarded. A count with a visible quality measure is far stronger than a bare figure.
Second, keep one dataset. When the live view and the survey report come from different pipelines they will disagree, and reconciling them after the fact is thankless. The detections that raise alerts should be the detections that get counted.
Third, audit the sensors against their own behaviour. A detection zone that is busy nearly all day with one class dominating and almost no track diversity is not a busy road — it is a zone counting something stationary. A zone with three active hours and a handful of detections is pointed at open ground.
Patterns that signal risk, not just volume
The most valuable patterns are often not about congestion at all. Repeated wrong-way or reversing detections at one interchange point at a signage or geometry problem. A queue that consistently forms at the same time in the same lane suggests a merge issue rather than demand. A steady rise in heavy vehicles on a route not designed for them is a pavement question arriving early.
None of these are visible in a monthly total. They appear when the same detections are examined by location, by class and by time of day — which is an argument for keeping the granular record rather than only the summary.
Describing a vehicle without a plate
Attribute recognition — colour, body type, and make and model — turns a witness description into a query. "Silver SUV, northbound, sometime after eleven" becomes a search rather than a request for someone to watch hours of footage.
Each attribute carries its own confidence, and they are not equal: colour and body type are considerably more reliable than make and model. Reporting them separately, rather than as a single score, is what lets an analyst judge how much weight a result deserves.
From insight to a decision someone makes
Analytics only matter if they change something. In practice the outputs that get used are narrow and repeatable: a signal retiming supported by measured turn movements; a maintenance window scheduled against the genuine overnight low; a public information posting timed to the point the corridor historically degrades; a business case supported by counts nobody can pull apart.
Start from the decision, not the dashboard. Ask which recurring judgement an agency currently makes on instinct, then instrument that. A corridor platform that answers four questions well is worth considerably more than one that produces forty charts nobody opens.