Smarter Cars Still Aren’t Saving Pedestrians

In 2024,



7,080 pedestrians were killed and more than 71,000 injured



on American roads.



Cyclist deaths reached their highest level since at least 1980



. Overall traffic fatalities dipped below 40,000 for the first time since 2020. But nearly all of that progress benefited vehicle occupants. Pedestrians and cyclists are still dying at near-historic rates.

That gap is the story. Driver-assistance features have measurably reduced occupant fatalities over the last decade. They have not done the same for vulnerable road users, and the reason comes down to line of sight, not driver behavior. A vehicle-mounted sensor stack is constrained by the geometry of the chassis. The places where pedestrians and cyclists are most at risk are precisely the places onboard sensors are structurally weakest. Occluded intersections. Mid-block crossings. Blind corners. School zones where a child steps between parked cars. That limit applies equally to a human driver looking through a windshield, an automatic-braking system reading a forward-facing radar, and any future autonomous stack we put on a vehicle.

For most of the last decade, the entire conversation around connected vehicles, autonomous mobility, and urban robotics has been a Vehicle-to-Everything (V2X) conversation. The idea is that the vehicle talks to other vehicles, to roadside equipment, to pedestrians’ phones, and to the network. Bigger sensor suites, better models, more onboard compute, more redundancy — all centered on the vehicle itself. That framing produced real progress. It also imposed a ceiling on what vehicle-centric perception can do for pedestrians and cyclists.

The next phase of this work has a different shape. Call it Infrastructure-to-Everything, or I2X. The intersection, corridor, and surrounding infrastructure perceive and predict on behalf of whatever is moving through them. I2X is the harder half of the problem. It is also where the safety case finally closes.

V2X has a safety ceiling, and we are now hitting it

The vehicle-centric approach has been legible to investors, automakers, and regulators in a way infrastructure has not. Progress can be measured in sensor counts, model parameters, and disengagement rates. It fits neatly into a slide deck. The V2X stack has matured accordingly. Cellular V2X standards are real, roadside units are deployed in dozens of corridors, and major automotive and technology players are pouring serious capital into cooperative perception platforms.

That capital has produced genuine gains, again primarily for vehicle occupants. It has not closed the gap for vulnerable road users, and the research community is increasingly clear about why. A



recent empirical study of V2X cooperative perception systems



identifies six recurring error patterns in single-agent autonomous systems, most rooted in the same limitation: a vehicle cannot perceive beyond its line of sight. Occlusions, non-line-of-sight intersections, weather degradation, edge cases that did not appear in training data.



Separate research focused on vulnerable road user safety



reaches the same conclusion from a different angle: pedestrian and cyclist localization is the failure mode vehicle-mounted sensors are structurally bad at solving.

More LiDAR helps. More radar helps. Better models help. None of it changes the underlying geometry. A vehicle-mounted sensor will always have line-of-sight limitations, and the line of sight gets worse as urban density goes up. The places where we need autonomy to be safest are exactly the places vehicle-centric perception is structurally weakest.

I2X flips the polarity

Infrastructure-to-Everything starts from a different premise. The road, the intersection, the corridor, and the loading dock are not passive surfaces waiting to be perceived. They become active intelligence layers that perceive, interpret, and broadcast conditions outward. A vehicle approaching an occluded intersection does not need to see around the corner with its own sensors. The corner sees for it. A delivery robot working a sidewalk does not need to anticipate the pedestrian behind a parked truck. The light pole already knows the pedestrian is there.

This is the side of the problem we are building at Surge. Our deployments are LiDAR-only edge perception nodes mounted on existing urban infrastructure: light poles, signal heads, and rooftops. No cameras, no images, and no personally identifying data captured at the moment of sensing. We call the positioning “Anonymous by Physics,” because LiDAR captures movement and geometry, not faces, license plates, or identity. The output is a real-time stream of location, velocity, and trajectory. That same stream is useful to a city traffic engineer, an autonomous vehicle stack, a logistics router, and a safety researcher, all off a single sensor footprint.

Two design choices matter for the safety case. The first is that infrastructure perception is multi-tenant by default. A vehicle-mounted sensor stack is point-to-point and serves one customer at a time. A LiDAR node on a light pole serves every vehicle, every drone, and every pedestrian-safety application that needs the data, simultaneously. The economics look more like a cell tower than an oil well. The second is that corridor-level coverage matters more than intersection-level coverage. Isolated nodes are useful. Networked corridors are defensible, because pedestrian safety, autonomous vehicle training, and emergency response all depend on continuity, not snapshots.

Real-time perception is the floor. Prediction is the ceiling.

The deeper opportunity is not the real-time layer. Real-time perception solves the obvious safety cases, and that alone is valuable. The deeper unlock comes when AI models train on continuous infrastructure data over months and years instead of episodic snapshots vehicles capture.

Vehicle data is, by its nature, sparse and discontinuous. A car passes through an intersection a few times a day at most. It sees a slice. An infrastructure node watches the same intersection 24 hours a day, every day, for years. It sees the full distribution. The same place through rush hour, storms, construction, outages, events, and seasonal shifts. That is a fundamentally different kind of training data, and it produces a fundamentally different kind of model.

As that data accumulates, the system stops being reactive and becomes predictive. The gait pattern of someone about to step off the curb without looking. The deceleration profile of a vehicle that is about to run a red. The convergence geometry that precedes a near-miss between a turning bus and a bicycle in the bike lane. These are precursor signals. They are statistically observable. They do not exist in crash reports because they are not crashes. They are precursor events and occur orders of magnitude more often than crashes themselves. Crashes are statistically sparse. Near-misses are abundant. Infrastructure systems observe the precursor behaviors that crash databases never capture. A vehicle moving through an intersection will never see them at scale. Infrastructure that lives at the intersection sees them constantly.

That is the actual safety unlock. The promise of connected mobility has always been that we could intervene before a crash, not document it after. Onboard sensors plus reactive vehicle-to-everything communication get you partway there. A predictive layer trained on continuous, multi-modal, infrastructure-resident data is what gets you the rest of the way. The same logic, incidentally, applies on the energy side, where companies like HEVO are showing that fully autonomous fleets need infrastructure to deliver power as well as perception. Different domain, same conclusion: the world has to do work that the vehicle cannot do alone.

A nervous system for the urban environment

When you step back from any single deployment, what this work is really building is something cities have never actually had: a nervous system. Cities already have concrete, steel, power grids, and fiber. What they lack is a layer that senses, remembers, and predicts in real time across the physical environment.

I2X is that layer. A LiDAR node on a light pole functions like a sensory neuron. Edge compute behaves like a local reflex, fast enough to act without waiting for centralized systems. Over time, a network of nodes builds institutional memory at urban scale: how intersections behave, where near-misses occur, how flows change during storms, outages, construction, or emergencies.

The applications follow naturally. A pedestrian-safety alert in a school zone is a reflex. A traffic-signal adjustment based on observed flow is a learned response. A predictive routing recommendation for an emergency vehicle depends on both sensing and memory. Logistics, emergency management, climate resilience, and autonomous-vehicle training all become easier when the city can continuously observe and learn from its own operations. The point is not to add more cameras or dashboards. The point is to give the urban environment the capability it has always lacked: the ability to perceive, remember, and respond in real time.

Infrastructure changes the economics of autonomy

When the intelligence layer migrates from the vehicle to the infrastructure, the economics of autonomy and urban operations shift in three important ways.

First, the cost curve of onboard vehicle hardware finally has somewhere to go. Today, every autonomous vehicle is asked to carry the entire perception problem and most of the safety problem on its chassis. That is why the bill of materials for an autonomous vehicle looks the way it does. When infrastructure delivers perception over the last hundred meters and prediction on top of it, the vehicle gets lighter, cheaper, and easier to certify. The same logic applies to drones, sidewalk robots, and any other autonomous form factor waiting for its unit economics to close.

Second, the addressable market for any single infrastructure deployment expands dramatically. A LiDAR node on a light pole that serves a city’s traffic engineering team, a national logistics carrier, an autonomous shuttle operator, a safety researcher, and an insurance underwriter is a fundamentally different asset than a sensor that serves one tenant. Shared infrastructure compounds in a way point solutions do not.

Third, the financing story becomes legible to the institutional capital that has historically funded ports, towers, fiber, and utilities. Between us, we bring operating leadership in connected-infrastructure deployment and more than two decades of infrastructure project finance experience at firms including Integrated Roadways, Black & Veatch, and Diode Ventures. The pattern is familiar. Once an asset class produces multiple, contracted, long-tenor revenue streams from a single physical footprint, the cost of capital drops, the duration extends, and the buildout accelerates. That is the moment we are approaching with intelligent infrastructure. The capital has been waiting on legibility, not on the technology.

The road learns to think back

The hard fatality numbers are not going to move materially until we stop asking the vehicle to do all the work. The decade of vehicle-centric investment produced standards, deployments, and meaningful gains for vehicle occupants. It did not move the needle for the people most exposed to the consequences of vehicles getting it wrong, and the structural reason is built into the geometry of the problem.

The next chapter is infrastructure-out. Roads that perceive. Intersections that predict. Corridors that learn and intervene before crashes occur. Add the energy side later, on the same physical footprint, and you have the substrate for autonomy as a system rather than a product. More importantly, you have infrastructure cities can use for everything else they have been trying to solve for the last twenty years.

V2X taught vehicles to talk. I2X is the city learning to feel, to think back, and then to think ahead.