Automatic incident detection misses five things most often: incidents between sensors, incidents in light traffic, anything a PTZ camera sees once it has moved off its preset, events at night or in bad weather, and real alerts buried among false ones. Public evaluations by Caltrans, a multi-state DOT pooled fund and the New England Transportation Consortium all describe the same gaps.
None of these are exotic failures. They come from how AID systems are built and where they are deployed, and they show up in almost every agency review of the technology. Knowing them in advance is the difference between a pilot that proves something and one that only proves the vendor's demo.
Three numbers describe any AID system. The New England Transportation Consortium's review of incident detection algorithms (Parkany and Xie, 2005) defines them:
The same review states the catch plainly: the figures "clearly illustrate the tradeoffs between the three performance parameters." Tune for a higher detection rate and false alarms rise. Tune false alarms down and incidents get missed or detected later. A single headline accuracy figure hides which way a system has been tuned.
| Blind spot | Why it happens | What the evidence says |
|---|---|---|
| Incidents between sensors | Detection covers what a detector or camera can see, and full coverage of a network is rarely affordable | AID is "effective principally for major incidents or for incidents that occur in the immediate vicinity of a sensor" (NETC, 2005). Full coverage is "a challenging proposition… gaps will inevitably remain" (ENTERPRISE Pooled Fund, 2022) |
| Incidents in light traffic | Flow-based algorithms look for congestion, and a blockage on a quiet road may not cause any | "Most existing algorithms cannot deal with incident detection under low volume conditions very well" (NETC, 2005) |
| PTZ cameras off preset | Video analytics are calibrated to one camera position | "AID doesn't work unless the camera is in its home position" (ENTERPRISE, 2022) |
| Night, shadows and weather | Vision models lose contrast in darkness, glare, rain and snow | "lighting (darkness, glare, shadows, etc.) and weather (rain, snow) degraded AVID performance, even in modern systems" (Caltrans, 2018) |
| Real alerts among false ones | Operators stop trusting a system that alarms too often | "operator fatigue with high false alarm rates will invariably lead to AID systems being ignored" (Caltrans, 2018) |
Point detection sees only its own patch of road. Loop-based algorithms such as the California algorithm compare occupancy at two adjacent detector stations, because an incident "is likely to cause a significant increase in upstream occupancy while simultaneously reducing occupancy downstream," as the NETC review describes it. That logic works, but its speed depends on geometry: the review notes that the required distance between stations "is directly related to the time required to detect an incident."
The same review records what traffic management centers reported about the California algorithm in practice: difficult calibration, a detection rate below 50%, a high false alarm rate and long times to detection. Some centers dropped earlier algorithms altogether for those reasons. We covered how corridor-level correlation handles the gap between cameras in Automatic Incident Detection Beyond a Single Camera View.
Operators move PTZ cameras constantly, which is the point of having them. Most video AID is calibrated to one fixed view. The ENTERPRISE pooled fund study reports that with many products, detection stops until the camera returns home. MnDOT reported that detection continues off preset but with more false positives. A Caltrans pilot review traced false alarms "primarily" to environmental conditions and to "the camera being moved out of the preset position for which calibration was performed."
The practical effect is that detection is weakest exactly when an operator is looking at something, which is often when an incident is under way.
The Caltrans report, summarizing earlier research, describes a tunnel monitoring system whose false alarm rate ranged from 48% to 80% depending on the camera, caused in roughly equal parts by static shadows, snow and rain, glare and other causes. Iowa DOT told the ENTERPRISE study that its false positives "typically" occur at night with low visibility, and the same study notes that vision-based systems will not detect anything if it is completely dark.
A system evaluated on daytime footage in clear weather has only been tested on its easiest hours. Ask for results by condition.
Detection is the first stage of traffic incident management. The FHWA Freeway Management and Operations Handbook lists it ahead of verification, motorist information, response, site management, traffic management, clearance and recovery. Every later stage waits on it.
FHWA has estimated that traffic incidents cause about a quarter of all congestion, a figure that dates from the mid-2000s. Research on secondary crashes points the same way. A 2025 analysis for The Eastern Transportation Coalition found that incident duration has its strongest effect on secondary crash odds in the first ten minutes, which is the window where faster detection counts most.
Ask for detection rate, false alarm rate and time to detect, measured on your cameras, with the FAR definition stated. Then test each blind spot directly:
GoodVision Live Traffic runs on the cameras an agency already owns. When an operator moves a PTZ camera, it pauses analytics while the camera is off position and recalibrates automatically when the camera returns to its preset, with no manual reset.
It detects wrong-way driving, stopped vehicles in the travel lane and on the shoulder, and congestion as it builds and propagates along a corridor. Between cameras, it infers flow-affecting incidents from how traffic behaves at the neighboring monitoring points. That works where the points are close enough that a change at one shows up at the next. Where they sit further apart, the inference tells you a segment is affected rather than pinpointing a location, and in-view detection is what gives you a classified event with visual evidence. We size this per corridor during the pilot.
Every event is cross-checked against surrounding traffic conditions before it reaches an operator, so transient anomalies are filtered out and only verified, prioritized events surface. Events route into the agency's existing ATMS and dashboards rather than another screen. More detail is on the automatic incident detection solution page.
Want to see how detection holds up on your own cameras, including at night and with PTZ in use? Request a demo.