A camera in a moving vehicle produces tens of frames a second. Sending that to a cloud model means paying for the upload of everything, including the overwhelming majority of frames in which nothing happens. Running the model on the camera means paying for compute once, in silicon, and uploading only what the model decided was worth sending.
The definition, once
Edge inference means the AI model runs on the device that captured the data, rather than on a server it uploads to. In a vehicle, the edge is the camera.
The practical consequence is latency and cost, in that order. A road-detection platform of this kind was built by moving inference off the cloud and onto the camera, and that move is what made a live feed possible in the first place. Not faster. Possible. A pipeline that has to upload video before it can classify it cannot produce a live view of a road network at any price.
The result, on that platform: a detection appears within a couple of seconds of the camera seeing the object, with the supporting image following within minutes. Two different clocks, because metadata is small and imagery is not.
Only we know
The confidence threshold on each object class does not gate what gets displayed. It gates what gets uploaded at all.
That is the architectural decision that people miss, and it inverts how most observers think about accuracy. The system is not uploading everything the model sees and filtering later. Thresholds are set per object class, using the precision and recall curve from the model's training to choose the operating point for that class. A class the model is weak on gets a stricter gate, and the weak detections are never transmitted.
So the model's confidence calibration is not a display feature. It is the compression algorithm. The bandwidth bill of the entire network is a function of where those thresholds sit, which is why they are set per class rather than once globally, and why they are a product decision rather than a research one.
The corollary matters for anyone evaluating this category: a vendor's stated detection accuracy and their upload economics are the same number wearing two hats.
Compute is the constraint, so the model gets smaller
A model that fits a data center does not fit a windscreen. The usual approach is to train large and distill down.
One collision-model family ships at 300 million parameters, 86 million, and 22 million. The two smaller ones are distilled from the largest. What makes this work is that most of the understanding comes from the self-supervised pre-training rather than from the distillation, which mainly adjusts where the model draws its decision boundaries. The small model is not a model that learned less.
Input shape gets tuned for the same reason. An earlier release in that family accepts between two and sixteen frames per inference, which cuts the number of tokens the model has to process and makes edge deployment more flexible. Fewer frames, fewer tokens, less compute, more hardware targets in reach.
What does not fit on the edge, and why saying so matters
Some things simply do not go. In that same collision-model family, the component that writes a natural-language explanation of a hazard is too heavy to run on the device. Only the attention-based visual output can run in real time there.
That is a useful test to apply to any vendor claim in this category. If a system advertises generated natural-language reasoning as an in-vehicle real-time feature, either the model is much smaller than the claim implies, or the reasoning is happening somewhere else, later.
The unglamorous constraints that decide deployments
Fleet buyers ask about power before they ask about accuracy, and they are right to.
On one dual-facing fleet device family, sleep draw is 11 milliamps on the newer unit and 19 milliamps on the older one, with hibernate at 1 to 2 milliamps. The newer device draws less while carrying more compute, which is what a generational improvement looks like when it is real. Storage is 128 GB embedded, with no SD card slot. The sensor package around the camera is a 6-axis inertial measurement unit, GNSS, and a Wi-Fi and Bluetooth module. Day and night imaging is handled by a mechanical infrared cut filter that physically moves between the two spectral requirements.
A moving part, in a vehicle, because the alternative is worse imagery.
Garbage does not need to reach the cloud either
Edge filtering applies before inference as well as after it. Enrichers in the data pipeline discard footage where the camera is not pointed at the road, for instance where it is facing the sky. A misaligned camera generates perfectly valid frames of nothing, and a network at scale will always contain some.
Inference runs on both consumer and fleet cameras in this network.
Limits
Figures above belong to specific products and specific releases, named in each case, and they do not transfer between them: the parameter counts are a collision model, the power figures are a fleet device, the threshold behavior is the road-detection platform. Parameter counts describe a checkpoint rather than a product line. No latency figure appears here for any named chip, because verified per-chip latency is not something this piece can stand behind.
Measured, not imagined. On the edge, the threshold is the architecture.