Video analytics and computer vision for business
NDDev builds video analytics for factories, warehouses and shops: the model finds people, vehicles, goods and defects in video and photos, counts objects, watches zones and sends events to your systems. We train models on frames from your own cameras. Our work includes flip chart recognition from photos for a bank, a research project under NDA.
Who it is for
Manufacturers, warehouses, retailers, logistics firms and camera-equipped sites where staff review footage, count objects or keep watch by hand.
- Trained on frames from your cameras
- Processing on a server or at the camera
- Events sent to your systems
- Accuracy measured before launch
What the model notices
We frame the task as an event the business cares about: a person enters a danger zone, a defect appears on the line, a shelf empties, a truck waits at the gate longer than it should. For each event we choose an approach: detection and counting with YOLO, segmentation, frame classification or classic OpenCV methods when they are enough.
- a description of events and response rules
- collection and labelling of frames from your cameras
- detection, segmentation and classification models
- optimisation for a server or a device at the camera
- events sent to your systems and messengers
- a dashboard with events and video clips
- a precision and recall report on held-out frames
Different cameras, lighting, seasons and times of day produce different pictures, so we fine-tune the model on frames from your cameras. We label a sample, train with PyTorch and measure precision and recall on held-out frames the model never saw in training.
Processing can run on a server or next to the camera on hardware already installed: we export the model to ONNX or TensorRT for the specific device. Video stays with you, and events and metrics go to your systems.
A video analytics project in stages
Camera and task review
We look at angles, stream quality and recordings and agree the events and metrics.
Data and labelling
We collect frames in different conditions and label them.
Model
We train and compare options and show the results on your recordings.
On-site pilot
We run it on some cameras, check events against reality and retrain the model on its errors.
Scaling up
We connect the remaining cameras and set up monitoring and a retraining routine.
What people ask about video analytics
Will our cameras work?
We check that during the review: picture detail, angle, lighting and the size of the object in the frame. If a camera gets in the way of the task, we suggest a different mounting angle or a different camera.
Where is the video processed?
On your server or on a device next to the camera. The video stays on your side, and only the events you choose leave it.
What accuracy do you guarantee?
We promise what is in our hands: the agreed process, iterations, reports and documentation. We agree target precision and recall in advance and measure them on your recordings before launch, so you see the result in numbers.
Can we start with a single task?
Yes. We usually start with one event on a few cameras: the result shows sooner and it becomes clear how to scale the system.
What does the price depend on?
A starting figure is given in the price block. The sum depends on the number of events and cameras, the labelling volume and where the model runs. Hardware and external labelling are paid for separately.
from $8,900
- Free warranty
- 80 days
Prices are starting prices for information only and are not a public offer. The final price is fixed in the specification.
Documents, video and forecasting
Related
Send us a camera frame and describe the event
A screenshot from the camera and a description of what to look for help us assess the task and give you a price.
We reply within 24 hours