The ITE International & Great Lakes District Annual Meeting, taking place this July 19-22, puts thousands of traffic engineers, planners, and technology vendors in the same convention center for a few days. The exhibit hall is where data quality claims compete loudest. Every system promises accuracy. Every demo runs on clean, well-lit footage from a controlled intersection. Every pitch leads with a percentage.
Counting accuracy is the wrong thing to lead with. It matters, but it is the floor, not the differentiator. Modern AI tools count vehicles faster and at far greater scale than a person reviewing footage frame by frame, and most credible vendors land in the same accuracy range. The percentage on the slide tells you almost nothing about whether the tool will be useful on your next project.
What separates a counting tool from a data platform is what happens after the video is processed. Does the system hand you a number, or a dataset you can interrogate? Can you reshape the analysis to your exact question without re-surveying? That is where the time savings and the defensibility actually come from, and it is what the five questions below are built to surface. They apply to any vendor you speak to, including GoodVision.
Start here, then move on quickly. Accuracy percentages without methodology tell you nothing. "95% accurate" is the start of a question, not an answer. Ask: validated by whom, on what footage type, at what level of intersection complexity, and against what ground truth?
Manual classified counts remain the standard baseline in traffic engineering. A vendor who can show you a study where AI output was compared to independently reviewed manual counts, across multiple locations with mixed conditions, is making a testable claim. A vendor whose validation ran on proprietary curated footage is asking you to trust their homework.
GoodVision achieves 95%+ accuracy on turning movement counts and classifications, validated against manual counts in consultancy projects across Europe, Asia, and North America. Work with firms like A-T-R in the UK gives you externally observable proof. Get a defensible answer, then spend the rest of the conversation on the questions that actually separate one platform from another.
This is the question that exposes the real gap between tools. Counting is the visible output. The valuable part is what sits underneath it.
GoodVision extracts a full digital snapshot of the scene by generating detailed road user trajectories, geo-referenced and at millisecond granularity. Turning movement counts, saturation flows, gap acceptance, travel times, queue and occupancy data, and safety conflict measurements are all derived from that single trajectory dataset. You are not buying a count. You are buying the structured record the count was pulled from, which means the same upload answers questions you had not thought to ask when you ordered the survey.
Ask any vendor at ITE to show you what lives below the headline number. If the answer is a summary CSV, you have a counting tool. If the answer is a queryable dataset with the trajectories intact, you have something your whole team can work from.
This is the one most engineers underrate, and it is where the hours are won.
On a traditional survey, the scope is fixed the moment the count is delivered. A new movement, a different time window, a class breakdown nobody asked for upfront, any of these sends you back to the field or back to the vendor for a new order. With a trajectory-based platform, the footage is processed once and the analysis stays editable. In GoodVision you define traffic movements, zones, gates, and analytical scenarios directly in the platform, generate the outputs you need, then redefine them and generate more, with no re-processing of the original video. Process once, produce as many reports as the project requires. Filter by object class and by specific movement lanes to isolate exactly the slice you care about.
That changes how you work. A late client request becomes a five-minute reconfiguration instead of a re-survey. A reviewer's question becomes a filter, not a fee. Ask the vendor to take the dataset they just demoed and answer a question they were not prepared for, live. The platforms that can do it will show you. The ones that cannot will offer to "follow up after the show."
Standard vehicle counting at a signalized intersection is one problem. A multimodal study that needs separate cyclist volumes, pedestrian crossing counts, and a full vehicle breakdown by class is a different problem, and it is the one feeding everything downstream.
Many AI counting tools were built around motor vehicle detection and retrofitted for active travel modes. The retrofit shows in the data. Cyclists get misclassified as motorcycles at low resolution. Pedestrians crossing at speed get missed. The model was never trained for the full range of behavior at a complex urban junction, so the dataset it produces is thin in exactly the places a Vision Zero or school-zone study depends on.
GoodVision recognizes up to 40 road user classes as distinct categories, and that depth is what makes the trajectory dataset worth interrogating later. The fuller the classification, the more questions the same processed footage can answer. The full breakdown is in the GoodVision vehicle classification guide.
GDPR compliance is not only a European procurement requirement. Transport authority clients in North America and APAC increasingly ask the same questions about data residency and video ownership. A vague answer at a trade show becomes a stalled contract when your procurement team asks again.
What you need to know: Can you delete footage from the vendor's servers on demand? Where is the data physically hosted? Does the vendor use client footage to train or improve their AI model? Is a Data Processing Agreement available?
GoodVision's position is explicit: you own your data. Client footage is not used for model training. Footage can be deleted on demand. The detail is in the GoodVision data privacy overview. Ask any vendor for the equivalent documentation before you move to a proof-of-concept.
Run the five questions in order and the conversation reorganizes itself. Accuracy gets you a yes or no in two minutes. The real signal is in the next three: whether you get a dataset or a number, how much you can reshape after processing, and whether the classification is deep enough to make that dataset worth keeping. A tool that counts well but locks the output is a slower path to the same place you started. A platform that hands you an editable trajectory dataset changes what one survey is worth.
According to FHWA's Traffic Monitoring Guide, accuracy and classification consistency are foundational to any reliable traffic dataset. The standards exist. The question worth asking at ITE is whether the tool in front of you turns those standards into data you can actually work with.
Book a demo at goodvisionlive.com/request-demo/ and run your first analysis on existing footage to see how GoodVision answers each of these questions on your specific project type.
Attending ITE Annual Meeting in Detroit this year? Schedule a meeting with us: