Active system

Doorbell Intelligence
Search the footage instead of scrubbing it.

The camera already had the useful evidence. The missing system was one that could understand the original recordings, preserve what mattered, and find it again by person, package, animal, vehicle, sound, or scene.

Activecomputer visionYOLOOpenVINOsearch
Original doorbell recordings pass through local detection, anonymous appearance grouping, and an evidence index that can be searched later.
The original recording remains the evidence. The models add an index around it.

What it is for

Questions after the alert
Visit

Who came by?

Find a person or recurring anonymous appearance, then open the exact source recording and moment.

Delivery

Was anything left?

Search package evidence that survived repeated detection checks and sits outside the person carrying it.

Scene

What happened around it?

Use visible objects, movement, transcript text, sound evidence, and semantic image matches together.

The decision

Use the camera originals

The first version watched the live stream one sampled frame at a time and rebuilt short clips around detections. The camera's SD card already held higher-resolution recordings with full frame rate and sound. Reconstructing those clips produced worse evidence.

The current system treats each completed SD recording as the source of truth. It downloads the original file locally, verifies it, persists the job in SQLite, and runs the intelligence pipeline against that recording. The camera's own app still owns immediate alerts and live view. This system answers the later question: what happened, and where is the evidence?

From recording to moment

Local processing

Detect and track

YOLO segmentation and low-frame-rate ByteTrack find people, vehicles, animals, and carried objects across each recording.

Group without naming

Person re-identification uses masked appearance crops to group repeat visitors anonymously. Weak tracks remain visits without being forced into a profile.

Keep the evidence

Events link back to the source recording, timestamp, contextual frame, clip, transcript, and sound evidence.

Search locally

Detections, transcripts, and local image embeddings make clothing, uniforms, objects, sounds, and scenes findable.

The interface shows moments

Synthetic fixture, real interface

Model output is not the unit a person wants to review. A single recording can produce several tracker fragments, overlapping detections, and multiple subjects. The interface turns that work into one moment with the complete evidence attached.

The capture uses synthetic fixture data inside the real dashboard. No camera frame, neighbor, address, private recording, or production account appears on this page.

Doorbell Intelligence dashboard populated with synthetic fixture data. It shows a calm front door, recording-level people activity, one possible package, and a recent moment.
Synthetic fixture data rendered through the production interface.

What made it reliable

Evidence before labels

A package needs repeated detections and must sit outside a person's box. A person crop must clear size and sharpness gates before it can contribute to appearance grouping. Parked vehicles can appear in the live view without becoming events. Vehicle details are kept only when the detector and an independent color check agree.

Short or poor tracks still count as visits, but they remain unprofiled. That is a better result than forcing weak evidence into a confident label.

The earlier version

System lineage

WatchPost was the first version of this idea: watch a live camera feed and keep a searchable visit log. Doorbell Intelligence keeps the useful premise but replaces reconstructed evidence with original recordings and adds anonymous appearance grouping, sound, semantic search, and stricter acceptance gates.

Back to the Lab →Read the WatchPost build →