The pattern repeats in cities across the country. An agency funds adaptive signal control on a corridor, the vendor tunes it, the ribbon gets cut, and six months later the complaints are back. The engineers pull up the algorithm settings. The vendor adjusts the parameters. Nothing much changes.
The problem is almost never the algorithm. Adaptive controllers are decision engines, and they are only as good as the inputs they receive. If the detectors on that corridor are missing vehicles at night, calling phantom vehicles in the rain, or quietly dead in the pavement, the controller is optimizing for traffic that does not exist. Here is what adaptive control actually needs from detection, the four ways detection fails, and what to do about it.
A fixed-time signal needs no detection at all. A basic actuated signal needs a presence call at the stop bar. Adaptive control needs much more, continuously, on every approach:
Every one of those inputs is a place for the system to be wrong. Here is where it usually goes wrong.
Inductive loops are still the most common detector at American intersections, and they fail in ways that are easy to miss. FHWA's Traffic Detector Handbook notes that with modern electronics, the main sources of loop failure have shifted to the wire in the pavement, its connection to the lead-in cable, and installation issues. The same handbook reports that many agencies trace most of their loop maintenance problems to installation errors. Fixing them means pavement cuts, lane closures, and traffic disruption, which is why a failed loop often stays failed.
When a loop fails, the technician typically places the phase on recall so it is served every cycle. That keeps the intersection functioning. It also means the adaptive controller now sees constant demand on a phase that may have none, and allocates green to it every cycle. One dead loop can undo the savings on an entire corridor.
Video detection replaced loops at a lot of intersections because it avoids the pavement cut. It introduced a different set of problems. A Purdue evaluation for the Indiana DOT found that video detection produced significantly more false detections and missed detections than loops on most phases, and that one camera activated one to three seconds early at night, apparently from headlight reflections on the pavement. A Texas Transportation Institute study found a video system's presence and speed accuracy declined to unacceptable levels at night and in rain.
The physics have not changed. Headlight glare on wet pavement, fog, low sun angle, shadows that move across the detection zone through the afternoon, and the occasional spider web across the lens all produce the same result: a call that is early, late, or imaginary. Each one is a green that gets extended for nobody or cut short for somebody.
Plenty of intersections have working detection at the stop bar and nothing upstream. That is enough to run an actuated signal. It is not enough to run an adaptive one well. Without advance detection, the controller cannot see a platoon arriving, cannot measure how long the queue is getting, and cannot protect drivers in the dilemma zone, the stretch where a driver facing yellow has to choose between braking hard and running the red.
Research for FHWA on advance radar detection used a 400-foot detection zone and defined the dilemma zone as the range of vehicles between six and two seconds from the stop line. An adaptive system that cannot see that far upstream is making its decisions after the vehicles have already arrived.
A loop knows something metal is present. A basic camera zone knows something is moving. Neither knows whether it is a bus that needs a longer clearance, a cyclist who needs a different one, or a pedestrian in the crosswalk who has not pressed the button. Adaptive logic that cannot classify is tuning timing for a fleet of identical vehicles that does not exist, and it is blind to the road users who are most at risk when it gets the clearance wrong.
The research is consistent about what works. A University of Idaho evaluation of multiple detection systems across day, night, and adverse weather found that performance generally degraded at night and in bad weather, and that radar and hybrid systems performed with the greatest accuracy. Oregon DOT's current signal design manual makes radar the default detection type and requires detection on all vehicle phases so the agency can run signal performance measures.
Radar and video together cover each other's blind spots. Radar measures presence, speed, and range in any weather, at any hour, and reaches far enough upstream to cover the advance zone and the dilemma zone. AI video classifies what the radar sees, separating trucks from cars and picking out pedestrians and cyclists that radar alone would struggle to label. Fuse the two on the device and you get detection that is right in the conditions where loops and cameras are wrong.
That is how FusionSensor works. It fuses HD3D radar and HD video 20 times per second, classifies all road users at 98.7 percent accuracy, and processes everything on the device with TrueEdge so the call reaches the controller without a round trip to a server. One unit per approach covers the stop bar and the advance zone, delivers per-lane counts and turning movements, and detects pedestrians and cyclists without a push button, all from the existing mast arm with nothing in the pavement. We wrote more about that in FusionSensor for safer intersections.
Before you call the vendor about the algorithm, spend an afternoon on these:
If the audit turns up what it usually does, the fix is upstream of the algorithm. Our post on what cities actually save with adaptive signal timing covers why detection decides the ROI, and our checklist for specifying intersection detection covers how to write it into the next spec. Or see how we approach adaptive signal timing from the sensor up.
Subscribe to The Edge Brief for monthly field notes, deployment stories, and product updates.
Subscribe →