InfrastructureIssue 084 records
Transportation
Four new items since the 2026-11-04 edition: Barcelona TMB applied AI to 4,500+ bus CCTV cameras for real-time occupancy and camera health; UMTRI received $5.66M from U.S. DOT for SAFETI intersection crash-prevention with MDOT and Ann Arbor; San José Airport’s AI hosts logged 13,600+ conversations with vendor-claimed 95% resolution; and Times of India documented AI traffic-camera false positives plus national contest rules. No fresh tolling, parking, ports or fleet evidence. Gaps remain in U.S. state DOT measured operations results.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 3 independent reporting1 academic research
- Outcomes
- 3 emerging1 cautionary
- Source freshness
- 4 new this fortnight
- Research completed
- 2026-11-06
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Roadside and enforcement AI vision needs privacy design plus human review for edge-case errors
UMTRI’s SAFETI program designs fisheye/object-class-only sensing to avoid identifying plates or faces, while Indian AI traffic cameras show guitar-as-pillion, turban-as-no-helmet and ANPR mismatch errors that forced manual validation and formal contest rules. Agencies deploying roadway vision AI should budget for demographic/edge-case testing, reviewer capacity and contest SLAs—not only detection models.
Operating questionBefore issuing AI-triggered roadway interventions or citations, what privacy constraints, false-positive tests and human-review/contest paths are mandatory?
Supporting evidenceUniversity of Michigan Transportation Research Institute (UMTRI) with Michigan DOT, City of Ann Arbor and SEMCOGBengaluru Traffic Police and other Indian traffic agencies (as reported)
Transit and airports are applying AI to existing cameras and venue hosts for passenger operations
TMB fuses AI on 4,500+ bus CCTV cameras with fare validations for real-time occupancy and camera health, while SJC’s AI wayfinding hosts logged 13,600+ conversations with strong placement effects. Both show AI layered onto assets agencies already operate, with quality loops (TMB audits/retrains) or content/placement lessons (SJC) mattering as much as the model.
Operating questionWhich existing camera, content and desk workflows can host AI first, and how will accuracy or resolution be independently audited?
Supporting evidenceTransports Metropolitans de Barcelona (TMB)San José Mineta International Airport with LiveX
Evidence ledger
Full record · every source keeps its link and limitations
Barcelona TMB uses AI on 4,500+ bus cameras for real-time passenger counts and camera health
At Tomorrow Mobility World Congress 2026, TMB said it now has real-time passenger counts per bus by applying AI to images from more than 4,500 onboard CCTV cameras covering the full bus fleet, fused with validations, surveys, occupancy alarms and video counting. The in-house bus technology team built the system with AI vendors; models are audited and continuously retrained. Outputs are intended for service regulation and passenger information, and the same stack flags failed or misaimed cameras to speed maintenance. A related bus-lane intrusion detection pilot also uses AI on cameras.
Why it matters, evidence and limitations
- Why it matters
- A fleet-wide reuse of existing transit CCTV for occupancy AI, with explicit audit/retrain loops and a dual use for camera maintenance—directly comparable to U.S. transit APC modernization.
- Evidence and measured results
- Official city press note; no published accuracy rates, latency figures or before/after reliability metrics.
- Limitations and uncertainty
- European operator; Catalan-language primary source; measured accuracy not disclosed; passenger-info use is planned rather than proven in the note.
UMTRI SAFETI wins $5.66M U.S. DOT award for AI intersection crash-prevention with MDOT and Ann Arbor
UMTRI received $5.66 million from U.S. DOT for SAFETI (Safety Actions for Everyone at Traffic Intersections), advancing open-source roadside perception (MSight) into near-miss prediction that can warn drivers and pedestrians via dashboards, in-pavement lights, phones or signal changes. U.S. DOT attributes roughly one quarter of traffic fatalities and about half of injuries to intersections. Phase 1 (~18 months) prototypes at Mcity with GM, Volkswagen and Ford; Phase 2 deploys at four Ann Arbor intersections over a 3.5-year program. Privacy design: Ann Arbor fisheye cameras cannot resolve plates or faces; the algorithm classifies object type only. Public partners include MDOT, Ann Arbor and SEMCOG.
Why it matters, evidence and limitations
- Why it matters
- A U.S. DOT-funded, DOT/city-partnered path from lab perception to signalized intersection interventions with explicit privacy design—usable as a reference architecture for state DOT safety programs.
- Evidence and measured results
- University announcement of award and planned methods; no field crash-reduction results yet.
- Limitations and uncertainty
- Pre-deployment research; Phase 2 limited to four intersections; OEM partners may shape vehicle integration; crash outcomes unknown.
San José Mineta Airport AI assistants log 13,600+ passenger conversations; vendor cites 95% resolution
Since introduction earlier in 2026 (initially for Super Bowl LX), SJC AI Hosts in Terminal A and at Gate 24 in Terminal B have handled more than 13,600 passenger conversations, resolving about 95% of questions with a 4.6/5 average rating, per vendor LiveX figures reported by Airports AI Alliance. Sessions rose 74% during the 2026 World Cup vs. the prior period. Gate 24 draws about twice the engagement of a walkway unit. Stack uses airport-controlled information and NVIDIA retrieval, with noise management for terminal audio; QR codes send routes to phones. Metrics and lessons are vendor-disclosed.
Why it matters, evidence and limitations
- Why it matters
- A U.S. public-airport passenger AI deployment with volume, resolution and placement metrics that other airport authorities can use when scoping wayfinding assistants.
- Evidence and measured results
- Trade article based on vendor disclosures; airport-released audited metrics are not shown.
- Limitations and uncertainty
- Vendor-reported KPIs; 95% resolution definition unclear; hologram/mayor features are promotional context.
Indian reporting details AI traffic-camera false positives and new national e-challan contest rules
TOI documents AI traffic-camera errors including a Bengaluru guitar mistaken for a helmetless pillion (Rs 500 challan), turban-wearing Sikh riders fined as helmetless in Udham Singh Nagar, seat-belt color confusions, Hyderabad-area totals of more than 50,000 erroneous challans in the first 10 months of 2025 (often ANPR mismatch), and double challaning. Bengaluru adopted AI enforcement in December 2022 initially without human validation, then began manually validating every AI-flagged violation after plate-reading complaints. Amended Central Motor Vehicles Rule 167 (GSR 48(E), Jan 20, 2026) gives drivers 45 days to pay or contest on a portal.
Why it matters, evidence and limitations
- Why it matters
- Cautionary evidence for U.S. DOTs and cities expanding AI camera enforcement: human review, cultural/edge-case testing, ANPR integrity and a clear contest path are operational necessities.
- Evidence and measured results
- Journalistic cases and official acknowledgments as reported; not a formal audit; Indian legal context.
- Limitations and uncertainty
- Non-U.S. legal setting; case-based reporting; Hyderabad 2025 totals are historical context within a Nov 5 article.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.