OmniVision is Nexa's video intelligence platform. It converts footage from the cameras a site already owns into an indexed, written record for search, investigation, and evidence, with inference running on hardware at the client's site.
OmniVision runs alongside existing cameras and recorders. No new cameras and no cloud upload are required.
Live RTSP and ONVIF streams, recorders, recorded files, dashcam footage, and cloud buckets feed one index.
Object detection identifies people and vehicles in each scene on the device at the site.
A vision-language model writes a plain-language caption for each scene: who, what, items carried, and location.
Operators search, ask questions, and author alerts in plain language, with the source clip attached to every result.
Detection, captioning, and search run on the device at the site. Illustrative interface.
A person in a dark jacket leaves a package by the north gate while a dark hatchback idles with its lights on.
Seven capabilities run on one index of every connected source.
Each scene carries a caption, detected objects, and a search embedding. Queries combine all three, with time expressions such as "after 9pm" applied as filters.
Written answers to questions about recorded footage, with every statement cited to its source clip and follow-up questions held in context.
Alert rules authored as plain sentences. Each rule displays its parsed object, camera, and schedule before activation.
Live view in the browser, incidents that open and collect their clips automatically, push alerts, and a written digest of the last 24 hours.
Cars, trucks, buses, and motorcycles counted once each across a drawn line, by class and direction, on live, recorded, or uploaded video.
GPS tracks read from dashcam footage and mapped with speed profiles and hard braking, sharp cornering, and speeding events. No tracker in the vehicle.
Existing recordings, USB drives, dashcam SD cards, and cloud archives are indexed and searched alongside live feeds.
Written answers built only from what the cameras recorded.
Operators ask questions in plain language and receive a short written answer. Each statement carries a citation that opens the clip it came from.
Summarise what happened at the loading bay between 6 and 9 this morning.
The bay was empty until 06:40, when a forklift moved two pallets to the doorS1. A white truck reversed into bay 2 at 07:30S2 and two people unloaded it until about 08:10S3.
Was the door left open afterwards?
Yes. The roller door is still open in the last clip at 08:52, with no one nearbyS4.
Example conversation for illustration.
Alert rules written as sentences, vehicle counts at any line, and an incident feed with a written daily digest. Illustrative interface.
Alert me when someone loiters by the back door after 11pm.
Daily digest · 07:00Quiet night. One incident at the back door at 23:14, resolved by patrol at 23:21. 47 vehicles through the north gate.
From search result to signed incident report.
Search results convert into incident reports with the relevant clips attached. Every export ships with a SHA-256 hash manifest and an HMAC-SHA256 signature, so any party can verify later that nothing was altered. An append-only audit trail records every search, view, report, and export.
Security, operations, and insurance teams query the same footage for different answers.
Offices, schools, estates, and data centres with cameras and no overnight monitoring. After-hours activity is written into the morning digest.
Who used the side entrance on Saturday?
Store operations and loss prevention from the cameras over tills and aisles: queue length, busy periods, and incident review.
When was the checkout queue longest today?
Safety and exclusion-zone rules written as sentences, overnight trespass detection, and operation over cellular or satellite links on remote sites.
Anyone in the crane zone without a hard hat this week?
Gate and dock activity counted by vehicle type and direction, and deliveries located by description, with no separate ANPR or traffic system.
How many trucks came through the north gate yesterday?
Dashcam footage located by time and place from the camera's own GPS track, with speed profiles and signed exports for insurers and courts.
Which trips had hard braking last week?
Shared areas monitored with plain-language rules for falls and wandering. Staff receive written handover summaries, and footage is processed on site.
Did anyone leave the east wing overnight?
Detection, captioning, search, and question answering run on hardware the client owns.
No third-party AI service sees the client's cameras or questions. Cameras run on an isolated network behind the device with no route to the internet, which also isolates compromised camera firmware. Nexa staff access footage only when the client grants it, and every session is recorded in the audit trail.
| Data | Standard deployment |
|---|---|
| Live and recorded video | Remains on site |
| Captions, search index, and chat | Remains on site |
| Questions to Ask OmniVision | Remains on site |
| Vehicle trip maps | Remains on site; street maps are stored on the device |
| Cloud-assisted captions | Optional, for sites without a GPU; still images only, never video |
| Push alerts | Optional; a short alert message to phone or browser |
| Cloud archive import | Optional; reads from the client's own S3, Azure, or Google Cloud bucket |
A purpose-built appliance that runs the full platform at one site.
The EdgeBox runs the portal, AI models, workers, and recordings locally. Four PoE+ ports power up to four cameras directly, a PoE+ switch extends a site to 12 or more, and the unit restarts itself when mains power returns. All three tiers share the same chassis; the GPU determines model size and caption detail. OmniVision also runs on client-owned Windows or Linux machines.
| Client hardware | Minimum | Recommended |
|---|---|---|
| RAM | 16 GB | 32 GB |
| GPU | None, or 6 GB | 12 to 16 GB NVIDIA |
| System disk | 256 GB NVMe | 500 GB NVMe |
| Recordings | About 1 to 3 GB per camera per day | |
| Operating system | Windows 11 or Ubuntu 24.04 LTS | |
Send a short description of the operation and the problem. A 30-minute scoping call follows.