Camera-Only, 99.6% Accurate: Why Sensor Fusion Is the Expensive Backup Plan You No Longer Need
Rain. Fog. Night. Glare. Partial occlusion. These are the exact moments traditional detection drops off — and the moments that decide whether a signal call, a pedestrian phase, or a safety alert can actually be trusted. For years, the industry's answer has been to stack more sensors on top of the problem: add radar for redundancy, add LiDAR for precision, and hope the fused output covers for whatever any single sensor misses.
Currux Vision's latest field validation suggests that answer is no longer necessary. In difficult-to-see conditions, our AI — reading the same analog or IP video feeds agencies already have — is validated at 99.6% detection accuracy, correctly identifying vehicles, trucks, buses, motorcycles, bicycles, and pedestrians even when contrast is poor and the scene is messy.
That number changes the math on sensor fusion.
You already own the sensor. No new poles for radar, no loops cut into the pavement, no LiDAR heads to keep clean and aligned. The camera infrastructure is already there — the intelligence just needs to catch up to it.
The picture is the evidence. Every signal call, near-miss, wrong-way event, or red-light run comes with visual context, not a blob on a radar plot or an abstract point cloud. When an agency has to explain a decision — to an engineer, an auditor, or the public — video is the evidence that actually holds up.
Classification comes bundled with detection. Class, lane, speed (±2 mph), occupancy, and trajectory all come from one stream. That's a meaningful difference from a radar-plus-camera setup, where classification is bolted on after the fact from a second, separately calibrated system.
It keeps working when other sensors struggle. Glare, shadows, wet pavement, snow, and mixed lighting are exactly the conditions that confuse many radar and LiDAR setups. Vision models that track shape, texture, motion, and consistency over time are built for that mess, not undone by it.
It costs less to deploy and less to keep running. One edge box on existing cameras stands in for a multi-sensor stack — with its extra power draw, extra mounting hardware, and extra maintenance calls.
Safety analytics ride on the same feed. Near-miss detection, conflict heat maps, vulnerable-road-user (VRU) tracking, and controller actuation all run off the video stream already being captured — no second sensor layer required.
Sensor fusion sounds like insurance. In practice, it's mostly overhead. Stacking radar or LiDAR on top of video means extra hardware, extra mounts, extra power, extra calibration, and extra failure points — plus the software overhead of keeping three different data sources in agreement. Agencies pay that tax at every single approach, and still need the camera anyway for classification, evidence, and VRU detection.
At 99.6% validated camera-only accuracy in hard conditions, that complexity no longer buys a meaningful reliability gain. The incremental benefit of fusion is smaller than the cost of owning it — in dollars, in maintenance windows, and in the engineering time it takes to keep multiple sensor types calibrated and talking to each other.
Loops miss what's between the wires. Radar can lose class and context in cluttered scenes. LiDAR is expensive at every approach and still needs a camera alongside it for evidence and identification. A well-trained camera system, by contrast, handles detection, classification, measurement, and documentation in a single layer — which is precisely what shows up when you look at the underlying frames: objects that are difficult for the human eye, and harder still for legacy video systems, locked, tracked, and classified correctly.
The takeaway for agencies. Camera-only detection isn't a compromise made to save money — it's what the data actually supports once accuracy gets high enough. At 99.6% in the hard cases — the fog, the glare, the 2 a.m. low-light approach — fusion stops being insurance and becomes overhead: a second and third sensor system that still needs the camera to do the parts that matter most.
For agencies weighing a detection upgrade, the question worth asking isn't how many sensor types to stack. It's whether the sensor already mounted on every approach — the camera — has finally gotten accurate enough to stand on its own. The field validation says it has.
Currux Vision designs, builds, and supports its AI edge platform in Houston, TX, serving 300+ agencies and more than 2,000 intersections across the U.S. and Canada. To see camera-only detection validated against your own conditions, request a briefing.