Computer vision

Computer vision development — video analytics that runs on your cameras.

Appsarmy is a computer vision development company. We build on-prem video analytics and object detection that run on the cameras you already own, on hardware inside your network — so footage never leaves the site. It is the deepest applied corner of our AI development practice.

Object detectionVideo analyticsEdge AIOn-premiseRTSP
PERSON .94 PPE OK ZONE · CLEAR RTSP · 30 FPS EDGE · ON-PREM
01 /Capabilities

Computer vision development services, on your floor.

A demo that detects a hard hat in good lighting is not a system. Our computer vision development services cover the detector, the video pipeline feeding it, and the edge hardware it runs on — because on a real factory floor, all three decide whether it works.

Video analytics & object detection

YOLO-class detectors trained on your footage, ingesting RTSP streams from existing cameras. Tracking, zone logic, and dwell time on top — so the output is an event a supervisor can act on, not a raw bounding box.

Detection tuned to your site

Edge AI & on-premise deployment

On-premise AI deployment on hardware inside your network — Jetson-class devices or a GPU server, sized to your camera count. Models quantized with TensorRT or ONNX Runtime to hit the latency budget on the box you actually bought. More on the tradeoffs in on-prem vs cloud AI inference.

Inference where the cameras are

PPE detection

Hard hats, vests, gloves, and goggles — safety compliance monitored continuously instead of by spot check.

Quality inspection

Defect and anomaly detection on the line, at line speed, catching what a tiring human eye misses at hour seven.

People counting

Occupancy, queue length, and footfall by zone — the same detector, pointed at a different question.

02 /Where it runs

Your footage never leaves the site.

YOUR SITE · YOUR CAMERAS CAM 1 CAM 2 CAM 3 CAM 4 RTSP EDGE GPU detection on-site frames discarded EVENTS zone · time class · score CLOUD no footage
Cameras  →  edge GPU on site  →  events to your dashboard. The video itself never crosses the boundary.
03 /What we build

Computer vision for manufacturing, and the floors like it.

PPE & safety compliance

Hard hat, vest, glove, and goggle detection by zone, with a per-zone policy — gloves mandatory at the press, not in the corridor. Events land in a portal supervisors actually open.

Restricted-zone intrusion

Detection when a person enters a machine envelope or a marked exclusion zone — the alert that has to fire in under a second, not after the shift report.

Quality & defect inspection

Surface defects, missing components, and misalignment caught at line speed, with the borderline cases routed to a human rather than silently passed.

People counting & occupancy

Footfall, queue length, and zone occupancy for retail and logistics — counts you can trust because the line-crossing logic is tuned to your camera angle.

Vehicle & forklift movement

Forklift speed, near-miss detection, and loading-bay flow — where people and machines share a floor and the collision is the thing you are paying to prevent.

CCTV analytics retrofit

Intelligence added to the camera estate you already own. If it speaks RTSP, we can read it — no rip-and-replace to get started.

04 /How we deploy

Vision projects are won or lost at the site survey.

01

Site survey

Camera angles, resolution, lighting, and occlusion. We tell you which views will not work before you pay for a model.

02

Data

Frames pulled from your own cameras, labelled to your zone policy, with a held-out set kept back for honest evaluation.

03

Model

A YOLO-class detector fine-tuned on your footage — your lighting, your uniforms, your machines, not a public benchmark.

04

Edge sizing

Streams, frame rate, and latency budget decide the hardware. Models quantized to fit the smallest box that holds the target.

05

Deploy

Installed on site, running as a supervised service that self-recovers when a camera drops or the power blinks.

06

Tune

Thresholds tuned to the real cost of a false alarm, then retraining on the misses the first weeks surface.

05 /Technology

The vision stack we build on.

Detection & tracking
YOLOv8PyTorchOpenCVByteTrackSegmentation
Video pipeline
RTSP ingestGStreamerFFmpegFrame sampling
Edge runtime
NVIDIA JetsonTensorRTONNX RuntimeQuantizationGPU servers
Serving & portal
FastAPIDockerPostgreSQLMongoDBEvent webhooks
Operations
Supervised agentsSelf-recoveryDrift monitoringOn-prem inference
06 /Why Appsarmy

We have put vision systems on real factory floors.

01

On-prem, not lip service

Detection runs inside your network on your hardware. No footage uploaded, no per-frame cloud bill, no dependency on a link that drops at 3am.

02

Your existing cameras

If it exposes RTSP, we can use it. We would rather retrofit intelligence onto the estate you own than sell you a new one.

03

Honest about the limits

Bad angle, bad light, heavy occlusion — some views will never detect reliably. We say so at the survey instead of after the invoice.

04

Built to keep running

A supervised agent that self-recovers, monitors its own cameras, and tells you when a stream has been dead for an hour.

07 /Engagement models

Work with us the way that fits.

Site survey & feasibility

A short, fixed engagement: which camera views can support detection, what hardware it needs, and what it will cost.

Pilot on one zone

One line, one zone, one camera cluster — proven on your floor before it goes site-wide.

Site-wide rollout

Full deployment across zones, with the portal, the alerting, and the support to keep it running.

08 /FAQ

Computer vision questions.

What computer vision services do you offer?

Video analytics and object detection deployed on-premise — PPE and safety monitoring, intrusion detection, quality inspection, and people counting — plus the edge infrastructure they run on.

Do we need to replace our CCTV cameras?

Usually not. If a camera exposes an RTSP stream, we can ingest it. The site survey tells you which angles and lighting will support reliable detection.

Does our footage leave the site?

No. Inference runs on edge hardware inside your network; frames are processed on site and discarded. Only events — zone, timestamp, class, confidence — reach your dashboard.

What hardware does this need?

It depends on camera count and latency budget. A few streams run on a Jetson-class device; a large site needs a GPU server. We size it at the survey and quantize the model to fit.

How accurate is PPE detection in practice?

Camera placement and lighting matter more than the model. We measure on a held-out set from your site, tune the threshold to the cost of a false alarm, and retrain on the misses.

How does this relate to your other AI work?

It's the deepest applied corner of our AI development practice, and shares the same discipline — evaluation before training, monitoring after deployment. It also underpins the anomaly detection in our fintech work, and pairs with mobile apps when the alerts need to reach a phone.

09 /Start a project

Let's put vision on your floor.

Direct
Response
Within 24 hours,
business days
Start a project → WhatsApp