Connect existing RTSP cameras to video AI workflows

Start with the stream, not the camera brand

Check that the camera exposes an accessible RTSP stream with H.264 or H.265 video. Codec, credentials, network reachability, and camera settings all affect compatibility. Test a representative camera before planning a fleet rollout; the presence of an RTSP URL alone is not a guarantee that every stream configuration will work.

One Kerberos Agent runs per camera on computing infrastructure near the camera. It does not replace the camera firmware. Agent gives supported cameras a consistent interface for live video, recordings, status, and controls, reducing the need to operate a different application for each camera brand.

Record locally and choose where recordings go

Agent supports motion-based and continuous recording. Keeping capture close to the camera means local recording can continue during an internet outage when the camera, local network, and storage remain available. Remote access and forwarding still depend on the connectivity of the services you configure.

Vault receives and indexes recordings in your chosen object storage, including MinIO, Storj, AWS S3, Google Cloud Storage, or Azure. You control the storage account and deployment location. Decide retention, available capacity, upload bandwidth, and access policies as part of the design rather than assuming that a cloud service must hold every recording.

Process recordings through configurable AI stages

Hub brings live streams, searchable recordings, alerts, and cases into an operator workspace. Its workflows process new recordings through configured stages such as object tracking, number plate recognition, loitering detection, scene descriptions, or forwarding to another system. Teams can also connect their own models and microservices.

This recording-driven processing path is distinct from live viewing. Choose capture settings and processing stages for the operational question you need to answer, then evaluate the outputs on representative footage. Useful detections, acceptable latency, and inference capacity depend on the model, video, infrastructure, and workflow configuration; they are not universal accuracy or real-time guarantees.

Connect the result to your operation

Operators can search recordings using supported filters and workflow markers, review relevant moments, and collect recordings into cases. Configure alert rules around the events that matter to your team instead of treating every model output as an action that must notify someone.

For an existing processing pipeline, Vault can publish new-recording events to Kafka, RabbitMQ, or SQS. Hub workflows can also integrate custom processing stages. These are integration choices: a deployment does not need every connector or every product to be useful.

Validate a small deployment before scaling

Begin with representative cameras and confirm capture, local playback, upload, storage access, and workflow results end to end. Include low light, disconnected uplinks, and the stream settings your operators actually use. Measure storage growth, processing latency, and missed or unwanted events against your own acceptance criteria.

When the pilot is ready to grow, Factory can deploy and operate one Agent per camera on your Kubernetes cluster. Use the linked product and implementation guides to choose the next component, or discuss the camera, network, storage, and inference requirements with the UUG.AI team.

Products and implementation resources

• Kerberos Agent: camera capture and local recording
• Vault: bring your own recording storage
• Hub: operator workspace and AI workflows
• Agent setup documentation
• Hub workflow documentation
• Managing camera Agents on Kubernetes