cameras that cry

wolf on the tracks

Problem statement

Fifty years of federal crossing-safety funding cut deaths at U.S. highway-rail grade crossings from 888 in 1975 to 197 in 2020, but the railroad death toll has moved rather than vanished: trespassing on the right-of-way, including suicides, now accounts for roughly 70% of all railroad-related deaths — 525 in 2020, up more than 31% over the previous decade — and FRA data indicate about 74% of non-suicide trespass casualties and 73% of suicide attempts occur within 1,000 feet of a grade crossing, because crossings are where people can get onto the tracks. Intrusion detection is a mature security technology for utility plants and banks, yet the few systems trialed on rail rights-of-way have failed on the specifics of the setting: outdoor video motion detection and passive infrared both generate false alarms (large animals were a major trigger), live high-definition video needs wireless backhaul that many track locations lack, and the deterrent effect of an installed system decayed after the second year. The unsolved problem is a detection-and-warning system that works at crossing-adjacent track — cheap enough for thousands of sites, accurate enough that police dispatchers do not learn to ignore it, and durable in its effect.

Why this matters

FRA counted 9,363 trespassing casualties in 2012–2016 with a societal cost of $43.2 billion in deaths and injuries, plus more than $56 million in train delay; each strike also stops trains and closes crossings for hours, cascading into road delays. The Section 130 program lets states spend crossing-safety funds on hazard elimination that could include detection and warning systems, but decision-makers currently have, in the problem statement's words, "limited, if not misleading, information" on which technologies work under which conditions, so the money is not spent on them. Because trespass casualties are concentrated near crossings, a system that works at that specific location type addresses the largest remaining share of rail deaths with a bounded number of installations.

What’s been tried and why it hasn’t worked

Volpe's 2007 state-of-the-art review catalogued infrastructure-based, locomotive-based, and cooperative intrusion-detection concepts (e.g., the AWARE wireless intrusion system and Nestor's Rail CrossingGuard for gate violations), predating usable computer vision. Volpe's 2012 evaluation of a video-based trespass monitoring and deterrent system on a railroad bridge in Pittsford, New York logged nearly 4,000 events (about 4.6 per day), cut trespassing 60% in the second year, but the reduction did not persist into the third — and large animals triggered many alarms that could not be masked without weakening detection. Volpe's Brunswick, Maine study with the local police (reported 2020) streamed live video from fixed cameras and found that both video motion detection and passive infrared produced false alarms, that wireless bandwidth for HD streams was a constraint at some sites, and that TV-white-space transceivers had limited range. Proprietary radar-based platform and level-crossing intrusion systems (e.g., Honeywell's) and worker-warning systems (Protran/Harsco) exist, but NCHRP staff noted the program "cannot evaluate the performance of commercial or proprietary products," so no independent comparative evaluation has been done. NCHRP funded only a $125,000 scoping study (17-122, 2024–2025) to catalogue practice and research needs; the evaluation itself remains undone.

What would unlock progress

The core technical unlock is on-device classification that separates people from animals, vehicles, and vegetation motion with a low enough false-alarm rate that alerts can go straight to a dispatcher — modern edge vision and thermal/radar fusion make this plausible where 2012-era motion detection could not — combined with event-only transmission so sites without broadband can still report. The behavioral unlock is understanding why deterrence decays: an audible warning that habituates by year three needs escalation, variation, or coupling to enforcement and to path redesign (the statement notes trespass often follows the most direct route to a destination). The adjacent solved problems are wildlife-camera trap classification and perimeter security for solar farms, both of which have driven false alarms down with edge ML on cheap hardware.

Entry points for student teams

A team could assemble a public or partner video dataset from crossing-adjacent track (or stage one on a heritage/short-line railroad with permission), train an edge classifier for person-on-ROW vs. animal/other, and quantify false-alarm and miss rates by lighting and weather against the Volpe baselines. A systems team could design a low-bandwidth, solar-powered node that sends only classified events and short clips over LTE or LoRa and cost it against the Section 130 funding constraint. A human-factors team could design and pilot-test a warning protocol that resists habituation and evaluate it with a before/after trespass count. Relevant skills: computer vision, embedded systems, human factors, and rail safety regulation.