DashStar by MOAT Safety: edge plate recognition for public alerts
DashStar is designed to read license plates on an in-vehicle edge node and check them only against official public alerts that include a vehicle or plate. Examples are AMBER, Silver and Endangered Missing alerts, stolen-vehicle alerts, and suspect alerts such as Blue and Yellow alerts in the states that issue them. Drivers never follow or engage. Today the prototype pipeline runs on one ~$600 NVIDIA Jetson Orin Nano at my desk, on recorded dashcam footage.
Status: research prototype. There is no live network yet, no real-world match has happened, and accuracy has not been measured yet. Unlike the design, the prototype keeps its plate reads and evidence clips on the device: they and the footage it has processed are my own research dataset, kept separate from the product.
- 28-33 fps offline at 1080p
- ~2.8 s match to evidence clip, offline tests
- 2,286 to 1 raw reads to one event
- 13 releases, 300+ tests
9-service DeepStream / Savant pipeline, TensorRT FP16, ZeroMQ, OpenTelemetry, Loki, Tempo, Grafana
The commitments DashStar is designed around
- Non-matches are discarded on the device within seconds.
- No location trail. Every match event carries the location and time of the sighting, because a sighting is useless without them. Nothing else ever carries a location, and drivers acknowledge this when they opt in.
- A human reviewer confirms every match before anyone is contacted. By default the driver hears only "do not follow". A driver who opts in can also see match notifications, marked unconfirmed, with the number to call, after agreeing not to handle them while driving. Drivers never follow or engage, ever.
- Training capture is opt-in only.
- Accuracy will be published, along with a quarterly transparency report.
- It matches only against official public alerts that include a vehicle or plate, never private watch-lists or ad-hoc plate requests. The full list of alert types is published only after per-state verification.
DashStar will never sell data, integrate with police networks or pay bounties.
As designed, what leaves the device for review is a cropped image plus the location and time of the sighting. The crop shows the vehicle and its plate by default, or just the plate when the vehicle crop is too large or the vehicle is not fully in frame. Other vehicles' plates and bystanders' faces in the crop are blurred on the device, and no video leaves the device. The reviewer on the MOAT Safety desk confirms each match and then calls the non-emergency number for the alert's area, or the hotline named in the alert. The location and time go to the authorities only after that confirmation. During the pilot, that reviewer is me. Vetted volunteer reviewers may assist, but they see the cropped image only, never the location or call details. Drivers give consent once at setup and can revoke it anytime. The planned DashStar Network is free.
