The Sensor Your ADAS Doesn’t Have – Why infrastructure-side AI is the missing layer in urban collision prevention

The automotive industry has achieved something remarkable over the past decade. Radar, lidar, camera fusion, and increasingly capable ADAS have made today’s vehicles dramatically safer on open roads. Highway fatality rates in countries with high new-car penetration are falling. The engineering is working.

And yet the intersection remains stubbornly lethal.

Roughly 40 to 50 percent of all serious traffic collisions globally occur at or near urban intersections. That figure has barely moved in years, despite the proliferation of automatic emergency braking, lane-keep assist, and forward collision warning systems. The reason is structural, not technical: virtually every safety system in today’s vehicle is designed around what the vehicle itself can observe. At a busy urban crossing, that is almost never the complete picture.

The occlusion problem ADAS can’t solve alone

Consider what a driver — or a vehicle’s sensor suite — actually sees when approaching a signalised intersection in a dense city. A bus stopped at the kerb blocks a cyclist in the cycle lane. A delivery van parked on the corner hides two pedestrians stepping off the pavement. A motorcycle approaches at speed from a road the forward-facing cameras don’t cover. These are not edge cases. They are the everyday geometry of urban driving.

Onboard sensors, no matter how capable, operate from a fixed vantage point at bumper height, moving with the vehicle. They cannot see around corners. They cannot detect a hazard concealed behind another road user until the two objects are on a collision course. The 2.5-second reaction window that emergency braking systems are designed around is, in these scenarios, often not enough.

V2X raises the ceiling — but connectivity has a density floor

Cellular and DSRC-based Vehicle-to-Everything communication systems address part of this gap. When both parties are connected and broadcasting, V2X dramatically extends the warning horizon. A connected pedestrian’s smartphone can alert a connected car to their presence well before either can see the other.

The challenge is adoption. A V2X network requires both sides of every potential conflict to be transmitting. In cities where smartphone penetration is high but app adoption is fragmented — and in mixed-traffic environments where two-wheelers, auto-rickshaws, and cyclists often carry no connected device at all — phone-based V2X leaves a substantial portion of road users invisible to the system.

Infrastructure as a collective perception layer

This is where the architectural conversation needs to shift. Instead of asking “how do we connect every road user,” the more tractable question may be: “how do we put a sensor at the intersection that sees everyone, regardless of what they are carrying?”

A roadside unit equipped with a 4D lidar array, multi-axis cameras, and an edge-computing module running predictive AI can observe the full geometry of an intersection — all approach vectors, all road users, all movement trajectories — and generate a warning seven seconds before a predicted conflict. That warning can be pushed to any connected receiver: a smartphone app, a vehicle’s ADAS system, a fleet telematics platform, or a city dashboard.

The output is not a reactive alert. It is a predictive one, generated from a vantage point no onboard system can replicate.

What this means for OEM architecture

Automotive engineers building next-generation ADAS stacks should be thinking now about V2I2V: vehicle-to-infrastructure-to-vehicle, the architecture in which a vehicle’s safety envelope extends to include intelligence gathered by fixed infrastructure. An RSU-equipped intersection becomes, in effect, an external sensor node for every vehicle passing through it — connected or not, ADAS-equipped or not.

Startup deep-tech company VISS-SYS is among the companies developing this infrastructure layer, combining advance lidar RSUs with AI-Augmented Kalman Filter algorithms and a zero-configuration wireless protocol designed for sub-50-millisecond warning delivery. Their PREVENT system is currently being validated in several pilots programs covering mixed urban traffic environments — the kind of dense, heterogeneous road conditions that expose the limits of vehicle-centric approaches most sharply.

The technology is not a replacement for onboard ADAS. It is the layer underneath it — the part of the safety stack that sees what the car cannot.

Intersections are not going away. Neither is the problem — unless the solution is built into the road itself.

Tsafrir Keynan is the CEO and founder of VISS-SYS, a deep-tech startup developing AI-based predictive intersection safety systems.