Frequently asked questions

No. Tried-and-tested classical computer vision algorithms underpin our analysis, with deep learning layered on top for attribute recognition. This provides investigators with information about what was detected, which processing layer identified it, and where it occurred in the footage, not simply an unexplained confidence score.

We focus on searching for people and vehicles, as our customers are predominantly interested in mapping suspect, victim and witness movements. There are 16 search attributes in total:

• People (10): gender, age range, hair length, carrying items, hat, glasses, sleeve length, lower body clothing length, upper body colour and lower body colour.
• Vehicles (2): vehicle type and vehicle colour.
• Used in tandem (4): licence plate and OCR, face detection, direction of movement, and zone.
• If there is something else you want to search by, just ask: we do custom developments too.

Footage typical of police work: night-time, low frame rate, heavy compression, awkward camera angles, rain on the lens. Models trained largely on clean internet imagery degrade on exactly the material investigations produce. Our datasets were not scraped from the internet, and were created with privacy protection methods in place.

Kinesense's motion detection algorithm works down to 2fps, which is rare in the industry. That forms the basis for our more in-depth attribute algorithms, where resolution matters more, but its impact varies significantly by attribute type. Vehicle attributes (type and colour) are the most resilient: they can generally still be read when the vehicle is very small in frame, as little as ~12 pixels wide, since overall shape and colour remain identifiable at low resolution. Person-level attributes like hat and glasses are the most affected, since they depend on fine detail around the head that blurs first as a subject gets smaller in frame. As a rule of thumb, if a person is at least ~35 pixels wide, the full set of person attributes should perform reliably.

Install the software on the target machine first: that generates the licence request code you send to us. Turnaround is normally within 24 hours. If you know you'll need a fast turnaround, flag it to us ahead of time so we can be ready to respond immediately rather than working it into the normal queue.

Yes, Kinesense LE is compatible with most major VMS platforms. Integration is straightforward.

No. Motion detection and object classification run on CPU, on hardware most units already own. GPU is recommended and only needed for the in-depth attribute layers. That choice matters when the budget is a business case, and when the position you are deployed to has limited power. If you have a GPU, use it, and ask us for sizing figures for both.

Yes. Live triage in RECON and post-event exploitation in Kinesense LE are licensed as modules, not as separately provisioned processing estates. Ask any vendor whether a single GPU can serve both modes: the answer shapes the hardware bill.

On-premise today, with a cloud-native platform on the roadmap. Kinesense DEAM already runs in an AWS or Azure tenancy or on force hardware, with a full API.

Modularly. Start with the capability the caseload demands and add modules later, rather than buying a suite to reach one feature. We price on an annual subscription model, which means you get access to the latest version.

As our solution is sold under a subscription model, you get to tell us what you need and we take it into account. For customers who want next-level support, we offer remote technical assistance that can be instrumental in getting quick results during investigations, whether that is making an image clearer or getting video ingested with synchronised metadata.