The category names for the two halves are different. Collision prediction usually refers to the forward-collision warning that fires before impact. Collision detection often refers to the recognition of a crash after the fact, sometimes for insurance claims, sometimes for automatic emergency response. Vendors mix the terms. A careful buyer separates them.
Why AI collision detection matters
Both halves save time and money, in different ways.
Prediction saves lives and repair budgets. The earlier a driver is warned, the more room they have to react. A collision that never happens is cheaper than one that did.
Detection saves review time and claims cost. When a crash does happen, an automatic report with video attached beats a phone call three hours later. The insurance side of the industry has been consolidating around this practice for a decade.
How the prediction side works
A prediction model has to answer one question: is a collision about to happen? It has to answer in the time budget of a warning (seconds, not minutes) and it has to answer with high enough precision that the warning does not become noise.
Nexar's BADAS ("Beyond ADAS") is a world model that watches road video and flags collision risk before impact. BADAS takes a 2-second video clip as its input. It was trained on Nexar's fleet library, which has driven more than 10 billion miles (the consolidated figure across the combined network is still being finalized), including a library of more than 70 million long-tail driving videos.
The second generation of the model was trained on 2 million clips drawn from 200,000 videos, five times the training data of the first generation. It ships in three sizes: a 300-million-parameter Large, an 86-million Flash, and a 22-million Flashlight. The smaller two are distilled from the Large model, which is what makes them small enough to run on edge hardware.
Why the architecture matters
BADAS is built on a V-JEPA style architecture that learns in latent space rather than by predicting pixels. That single choice matters because a model that predicts what to attend to is smaller and more efficient than one that predicts every pixel of the next frame; the compute budget for a road-facing edge camera cannot afford the second approach.
Most of the model's understanding comes from self-supervised pre-training; distillation mainly tunes where it draws the line.
How the detection side works
Once impact has happened, the classification problem is easier and the consequences are different. The system needs to recognize a crash reliably enough that a false negative does not lose the report, and often at a lower G-force threshold than most systems can catch.
Only we know: the model shows what caught its eye, without extra compute
Most collision detection AI produces a yes/no output. The second-generation BADAS does one extra thing that would normally require a second model: it shows what caught its eye, generating heat maps and bounding boxes straight from the model's own attention with no extra computation.
The visual explainability was an emergent property, the model was only ever trained on binary labels, never on bounding boxes.
The practical implication for a fleet safety manager or an OEM: a review after an event can show a heat map of what the model was looking at when it raised the alert, so the reviewer can tell whether it was tracking the right object. That is diagnostic information without a diagnostic model running alongside the primary one.
Where AI collision detection shows up
Three families of deployment dominate today.
Real-time driver warning. The alert has to fire on the vehicle, because a cloud round trip breaks the latency budget. Small edge-deployable models (the 86-million Flash and 22-million Flashlight variants of the second-generation BADAS) exist for this reason.
Fleet incident review. When a crash does happen, an AI collision detection system automates the "did that just happen" question so the review workflow can start immediately.
Insurance and OEM training data. The same detections used for real-time warning also produce a labeled dataset the AV and insurance sides value. Training data was mined by scoring the whole video archive with the previous BADAS version, then sending the suspicious clips to human annotators.
Limits: what AI collision detection cannot promise
Warning is not prevention. A forward-collision warning is only useful if the driver has time to act. When the closing speed is too high, no warning saves the situation. A prediction that arrives after the physical situation has been decided is a diagnostic, not a warning.
Model size caps model complexity. A frontier model does not fit on a windshield-mounted device. A prediction model deployed at the edge is a constrained model by construction; a buyer evaluating vendors is evaluating what fits in that envelope, not what fits in a data center.
Training data governs generalization. The scarce ingredient in training a prediction model is positive risk examples; negative examples are plentiful and repetitive. A vendor's ability to acquire and label enough positive examples is often the deciding factor in the model's real-world performance.
Detection has to survive false-positive economics. A collision detection system that fires on every pothole overwhelms the review queue. Tuning precision against recall is the whole product-management job, and it never finishes.
Frequently asked questions
What is AI collision detection?
AI collision detection uses computer vision to flag collision risk before impact, or to identify that a collision has occurred, or both. The prediction and post-hoc detection halves use similar signals but serve different workflows.
How does AI collision prediction work?
By training a model on a large library of driving video, so it can watch a short video clip and predict whether a collision is about to happen. BADAS, one such model, takes a 2-second clip as input and was trained on a library of more than 70 million long-tail driving videos.
Why is edge deployment important for AI collision detection?
Because the latency budget for a useful warning is seconds. A cloud round trip typically exceeds that budget. The smaller second-generation BADAS variants (22-million Flashlight and 86-million Flash) exist so the model can run on the edge inside that budget.
Can AI collision detection show why it fired?
Modern models can. The second-generation BADAS generates heat maps and bounding boxes straight from the model's own attention with no extra computation. That is the visualization that lets a reviewer confirm the model was tracking the right object.
How is training data collected for AI collision detection?
The scarce ingredient is positive risk examples; negative examples are plentiful. Training data was mined by scoring the whole video archive with the previous model version, then sending the suspicious clips to human annotators. The pipeline is why a long-tenured vendor's dataset does not commoditize easily.
How is AI collision detection different from a dashcam?
A dashcam records footage. An AI collision detection system runs a model on the footage, in the moment, to predict or recognize a collision. A modern connected dashcam can be both, if the on-device compute is enough to run the model.
Empiric Earth writes about the measured world. Numbers cited above trace to primary Nexar and Nauto product documentation.