The Future of Vision Zero Mapping: Opportunities from AI and ITS

Traditional road-safety programs depend heavily on crashes that have already happened.

Emerging transportation sensors, connected vehicle data, video analytics, near-miss detection and AI create the possibility of identifying dangerous movements before a serious collision occurs — and measuring whether a countermeasure works within weeks rather than years.

For local government, this changes road safety from periodic retrospective analysis into a potentially continuous operational process.

This matters directly to transportation and public works directors, city and county engineers, traffic engineers, GIS managers, Vision Zero and road safety staff, planners, mobility teams, city and assistant city managers and state DOT professionals. Each of these roles is asked the same question after a serious crash: where else does this risk exist, and what are we doing about it?

The background

Cities pursuing Vision Zero have mapped their crash data for years. Two limits have held that work back.

The first is the data. Crash records are infrequent and lagging. They arrive after someone is hurt, and they take three to five years to show a pattern.

The second is the people. Every map-based question needs skilled analyst time to design, run and interpret a geospatial query. Most agencies don't have enough of that time.

• Intelligent Transportation Systems (ITS). Cameras and edge computers at signalized intersections now produce leading indicators of risk around the clock. These are near-miss conflicts, rated by the injury force they could have caused.
• Agentic AI. Engineers and GIS staff can now ask questions of their safety data in plain language. The system assembles the data, builds the map or dashboard, and shows its work.

Vision Zero began in Sweden and became national policy there in 1997. It is built on the Safe System framework, and it depends on data, analysis and visualization. Many agencies have the maps but not the detailed data or the analyst capacity to answer the questions that matter.

“When you use crash data, you're taking action after an injury or death.”

The through-line: the map is where GIS staff and traffic engineers meet. What is changing is both what goes onto that map and how fast anyone can ask it a question.

Asset Mapping Intelligence distinguishes between established evidence, emerging research and industry-reported findings so readers can see not just what is being claimed, but how strongly the claim is supported.

HIGH

Government sources, standards, peer-reviewed research or strong independent evidence.

MODERATE

Credible supporting evidence with important limitations or incomplete validation.

EMERGING

Promising findings, methodologies or applications that have not yet been broadly validated.

INDUSTRY-REPORTED

Claims, examples or performance figures originating primarily from a vendor, presenter or commercial source and requiring independent confirmation.

A detailed claim-by-claim Evidence Check appears later in this publication. Throughout the text, short labels such as Agency Example Speaker Perspective Industry-Reported Independent Evidence Emerging Practice indicate where a statement comes from.

Additional Findings

• ✓Crash data has three built-in flaws. It is reactive (it arrives after harm), slow (it takes years to reveal a trend or confirm a fix) and low-resolution (a few data points per intersection, with little about root cause).
• ✓Near-miss data answers each flaw directly. It is predictive, continuous and high-resolution. Speakers cited research showing kinetic-energy-rated near misses predict five-year injury crash counts with 94% accuracy (see Evidence Check).
• ✓Time-of-day detail prevents the wrong fix. At one Missouri intersection, near misses were spread across the whole day. A 7–9 a.m. left-turn restriction would have left most risk untouched.
• ✓Signal state is essential context. Knowing whether a conflict happened on amber or red, or in the protected or permissive phase, separates a timing problem from a design problem.
• ✓AI is only as good as the data model under it. Without the controller feed, the agent knows a near miss happened but not whether the light was red.
• ✓Build a High-Injury Network that can absorb better data. LADOT's 2024 methodology uses weighted crash severity, modal networks, contextual factors and a priority score for every segment.
• ✓Crash reports hide speed. In Los Angeles, unsafe speed was listed as the primary factor in only 6% of pedestrian collisions, because officers record one primary factor. Leading indicators fill that gap.

How Near-Miss Data Changes the Safety Map

Featuring insights from: Craig Milligan, Ph.D. — Fireseeds North Infrastructure

Speaker Perspective

Craig is a consulting road safety engineer who has built hundreds of crash and safety maps, and who relies on GIS colleagues to build them. Frustrated with waiting on crash data, he developed near-miss video analytics for intersections. That technology, originally MicroTraffic, is now hosted and extended by Miovision.

Why a new data source

Canada and the US lose about 42,000 people a year to traffic crashes. Year after year, national progress is marginal, and some years go the wrong way. The argument: if the money keeps flowing into the same methods and the same data, the results won't change.

How the platform works

Industry-Reported

A 360° camera sits over the intersection. A computer in the traffic control cabinet runs computer vision to track every movement. Craig compared it to “a smartphone at the intersection”: once the hardware is in place, agencies push applications to it.

• emergency and transit priority
• detection that actuates a green light when someone is waiting
• V2X warnings
• Continuous Safety Monitoring, the near-miss application

The platform is reported as deployed in 68 countries. Near misses are rated using a kinetic-energy risk model. It combines trajectories, speeds, potential impact angle and time separation to estimate likelihood, exposure and severity. Conflicts are color-coded: green (low), yellow (medium), orange (high) and red (critical).

Three views of the data

Case — Wentzville Parkway and Great Oaks Boulevard, St. Charles County, Missouri

Emerging Practice

The site: two wide suburban arterials, a two-way center left-turn lane, protected-permissive left-turn phasing, steady traffic, and long pedestrian crossings.

Findings from 30 days of data:

• Two movement pairs dominated the conflicts: eastbound left versus westbound through, and westbound left versus eastbound through.
• Near misses were spread through the whole day, with only mild AM and PM peaks.

That second finding matters. Imagine one serious crash at 7 a.m., and local anecdote that “the morning rush is the hard time.” The common response is a 7–9 a.m. left-turn restriction. The data suggests this fix would leave roughly 80% of the problem in place. A sharp AM peak would justify a peak-hour restriction; a problem spread across the day needs a strategy that covers the whole day.

What this means for the map

• Thursday night: a pedestrian is killed by a left-turning driver.
• Friday morning: council wants to know where else this risk exists, on a map.
• 9:30 a.m.: staff send a map of exactly that issue.
• By Monday: a budget is in place and work starts.

“You never put a single geospatial query into the database to make it happen.”

Instead of one set of safety maps guiding five years of work, agencies can generate a new, precisely targeted map on any day.

How Agentic AI Changes the Way Engineers Interrogate Transportation Data

Featuring insights from: Brent Rogerson — Miovision

Industry-Reported

The layers a safety decision needs

At each intersection, Miovision One adds signal performance (ATSPMs), multimodal turning movement counts, compliance, clearance, conflicts and third-party probe data where integrated.

The problem AI addresses

After years of collecting data and integrating systems, agencies hit the same wall. Teams are small, systems are fragmented, and the analysis is still manual. The data exists; the capacity to turn it into a decision doesn't.

Mateo: three pillars

Brent described it as a co-pilot with a traffic engineer's knowledge, one that keeps working after the team goes home.

Three levels of use, demonstrated live

1. Conversational. “Which locations have the highest near-miss conflict rates over the last seven days?” Then: “Break down the top five by severity.”
2. AI Studios, a dashboard built from a prompt. A pedestrian dashboard for downtown Kitchener along King Street: volumes, compliance and conflicts, with an interactive map.
3. AI-built maps at two scales. At intersection level, a conflict cluster map; at network level, a 3D conflict heat map with time, severity and volume controls.

Adding school locations answered a common Safe Routes question: where are pedestrian near misses in relation to schools? Miovision reports work with partners, including Esri, and with agencies to wrap historical datasets in APIs so the agent can use them.

“It's only as powerful as the data you give it.”

Lessons from Los Angeles

Featuring insights from: Christopher Rider — LADOT

Agency Example

The past: the first High-Injury Network

Los Angeles adopted Vision Zero in 2015. It published its first High-Injury Network (HIN) in 2016 and updated it in 2017.

The present: the 2024 Vision Zero Safety Study

Work began in 2022 and the new methodology was adopted in 2025.

Collision weighting. All injury collisions are included, weighted using crash costs from California's Local Roadway Safety Manual:

• Contextual and roadway risk factors. These keep the network from simply following the highest-volume streets.
• The HIN itself. It is the top 5% of scoring segments, about 549 miles.
• Modal HINs. Separate networks for bicycle, pedestrian and motorcycle crashes.
• A priority score for every HIN segment and intersection. Roughly 6,000 corridors and 4,000 intersections are ranked.

The gaps: what crash data can't tell you

Speed is under-recorded. Police reports record one primary collision factor. For pedestrian collisions, that is usually a right-of-way violation. Speed may still have been involved, but it isn't counted.

Patterns without explanations. LADOT mapped KSI collisions by time of year against sunset times and found a clear spike around sunset. That tells staff when, not why or what to do.

Unknown unknowns. Sunset was easy to test because staff knew to look for it. Machine learning can surface correlations nobody thought to test.

Data sources disappear. The HIN's high-speed-roadway factor came from Wejo, which no longer exists. Pedestrian and bike activity came from StreetLight. Lane counts were left out because no good source existed.

Because LA's method treats each factor as a replaceable input, a better source can be plugged in without redesigning the network.

The future: plugging the gaps

• Volumes and speeds. The work has shifted from finding data to judging which source is most reliable.
• Behavior. Permanent near-miss monitoring, plus aggregated harsh-braking, acceleration and turning data.
• Generative AI and machine learning. Reading crash narratives once privacy can be assured, and finding patterns like the sunset effect.
• What worked. Asking an agent where past countermeasures had the most effect, what those locations had in common, and where else looks similar.

“How we don't know what we don't know.”

Asset Mapping Intelligence separates independent evidence from vendor-reported claims.

What this session did not show

• before-and-after crash reductions from near-miss-guided countermeasures
• measured accuracy rates for AI-generated answers
• installation or licensing costs

Coverage is also limited to intersections with sensors installed. These are the questions to put to vendors during a pilot.

Four practical modules for agencies moving from crash mapping toward continuous risk intelligence. For the network-level data checklist, countermeasure evaluation table and speed-management levers, see the companion Insight. Print this section.

Near-miss program — site checklist

• ☐Pick sites deliberately. Include HIN intersections, locations with recent severe crashes and a few comparison sites. Start with one to five sites, then expand.
• ☐Confirm the controller connection. Near-miss data without signal state can't separate a timing problem from a design problem.
• ☐Map crosswalks and approaches in the system. Clearance times and pedestrian conflicts depend on accurate geometry.
• ☐Establish a baseline. Collect at least 30 days before any change.
• ☐Review by movement pair, not by intersection total. The countermeasure follows the pair.
• ☐Check the hourly distribution before choosing a time-limited fix.
• ☐Check signal phase before choosing a signal fix. Red-running by one approach calls for backplates or yellow interval review. Permissive-phase conflicts call for phasing changes.
• ☐Watch the clips. Use video to confirm the diagnosis before committing budget.
• ☐Re-measure within weeks of a change. Then adjust and measure again.
• ☐Report in terms council understands. Show expected injury crashes avoided, not conflict counts.

Future-proofing your High-Injury Network (the LADOT approach)

• ✓Weight all injury crashes by cost, not KSI alone. California uses its Local Roadway Safety Manual; use your state's equivalent.
• ✓Add stacking weights for vulnerable users: youth, seniors, pedestrians, cyclists.
• ✓Treat contextual factors as replaceable inputs. Speed, activity and roadway features should each be a slot a better data source can fill later.
• ✓Build modal HINs (pedestrian, bicycle, motorcycle) alongside the combined network.
• ✓Score every segment and intersection, not just a top-N list. That lets you act when a grant, repaving project or political opportunity appears.
• ✓Report priorities by council district as well as citywide.
• ✓Log every countermeasure with location and date, so you can later ask what worked and where else it would work.

A Practical Adoption Roadmap

The June 2026 session drew a predominantly local-government and U.S.-based audience, with strong representation from public works and larger municipalities.

Geographic participation: United States 90%, Canada 5% and Europe 5%. Percentages reflect submitted poll responses; unanswered sector and municipality questions are not inferred.

Questions came from the live audience, hosted by Peter Croswell.

Terms already defined in the companion Insight are not repeated here. That includes Vision Zero, Safe System Approach, HIN, KSI, lagging and leading indicators, near miss, surrogate safety measure, VRU, CMF, SS4A and MUTCD. See the companion glossary →

Agentic AI

AI that plans and carries out multi-step tasks, such as retrieving data, analyzing it and building a chart or map, rather than only answering one question.

AI Studios

Miovision's workspace for building saved, shareable dashboards and maps from natural-language prompts.

Amber / red violation

A vehicle entering the intersection during the yellow or red interval. Where conflicts happen in the cycle points to different fixes.

API

Application Programming Interface. The connection that lets one system, such as an AI agent, read data from another.

ATSPM

Automated Traffic Signal Performance Measures. High-resolution metrics describing how well a signal serves traffic.

Biomechanical injury tolerance threshold

The level of force the human body can absorb before serious injury or death becomes likely.

Clearance time

How long a pedestrian or vehicle takes to clear a crosswalk or intersection.

Compliance

Whether road users obey the signal: red-light running, or crossing against the pedestrian signal.

Conflict configuration (movement pair)

Two specific movements that can collide, such as eastbound left versus westbound through. Each intersection has dozens of possible configurations.

Continuous Safety Monitoring

The Miovision application that detects and rates near misses at an intersection around the clock.

Contextual factors

Non-crash risk inputs in a HIN score, such as high-speed roadways, pedestrian activity, lane counts, school zones and equity areas.

Edge computing

Processing data on-site, here in the traffic cabinet, rather than sending raw video to the cloud.

Hallucination

An AI output that sounds confident but isn't supported by the underlying data.

ITS

Intelligent Transportation Systems. Sensors, communications and computing applied to transportation operations.

Kinetic-energy risk rating

A near-miss severity score based on the injury force a crash would have produced, from speeds, angle and road-user vulnerability.

Leading Pedestrian Interval (LPI)

A few seconds of walk signal before vehicles get green, so pedestrians are visible in the crosswalk first.

LLM

Large Language Model. The reasoning and language engine inside an AI agent.

Local Roadway Safety Manual (California)

Caltrans guidance for local safety analysis, including crash cost values used for severity weighting.

Mateo

Miovision's generative AI agent for traffic engineering, built into Miovision One.

Modal HIN

A High-Injury Network built from crashes involving a single mode, such as pedestrian, bicycle or motorcycle.

Primary collision factor

The single main cause recorded on a police collision report. Other factors, such as speed, may go unrecorded.

Priority score

A ranking assigned to every HIN segment or intersection so work can be ordered beyond a fixed top-N list.

Protected / permissive left turn

Protected: left turns have their own green arrow. Permissive: left turns must yield to oncoming traffic.

Signal backplate

The panel around a signal head that improves visibility. Often retroreflective borders are added to reduce red-light running.

Temporal separation

The time gap between two road users passing the same point. A small gap means a near miss.

TSM&O

Transportation Systems Management and Operations. Strategies to get more safety and reliability out of the existing network.

Turning movement count (TMC)

Counts of each movement through an intersection, by mode.

Unified data model

A single structure that connects crash, sensor, signal and context data so they can be analyzed together.

V2X

Vehicle-to-everything communication, used to send warnings between vehicles, infrastructure and road users.

Watch the presentation archive

Organizations & tools referenced

Standards & guidance referenced

Request materials

For decades, transportation safety mapping has largely answered a retrospective question: where have crashes occurred?

Emerging datasets allow agencies to ask a more useful set of questions:

• Where is risky behavior occurring now?
• Which movements are creating that risk?
• What intervention should we test?
• Did it work?

Near-miss analytics, continuous sensing and agentic AI will not eliminate the need for crash data, engineering standards or professional judgment. They can, however, give transportation professionals something they have rarely had before: a faster feedback loop between risk, intervention and outcome.

“That is the shift from crash mapping to predictive safety intelligence.”

Thank you to Miovision for sponsoring the Agentic AI and Traffic Engineering Series, and to Fireseeds North Infrastructure and LADOT for contributing their expertise.