EvolveX Gen1 watches the CCTV cameras you already have and flags loitering and fighting as they happen. It runs on a small device inside your premises, so footage never has to leave the site.
How Gen1 spots loitering and fighting, emails the alert with a snapshot, and learns from your own footage in EvolveX Studio. Camera scenes are simulated and Studio shows demo data.
1:00 · no sound · captions are in the video
Each camera gets its own zones and thresholds. A car-park entrance and a lift lobby need different definitions of “too long”, so you set them per camera rather than accepting one global rule.
A person who stays inside a marked zone longer than you allow, without moving on. Someone walking past is ignored. Someone waiting by a door for five minutes is flagged.
30 s for a lobbyFast, violent motion between people. A neural network reads short clips of each person and scores how likely it is that a fight is taking place.
p ≥ 0.70Every camera runs this sequence on its own, about ten times a second. Each step hands its result to the next, and an alert comes out the end with a snapshot attached.
Reads the camera stream (RTSP, USB or a recorded file) and scales each frame down for fast processing.
A person detector, or motion detection against a learned background, marks where people are in each frame.
Gives each person a track ID and follows them from frame to frame, keeping their path and recent appearance.
Checks every track against its zone and dwell rules, and runs the fight classifier on its recent clip.
Shows the event on the live dashboard, saves an annotated snapshot and a log entry, and can notify your system through a webhook.
Gen1 was written from first principles, from the neural-network layers up to the dashboard. That gives us full control over how it learns and where it runs.
Gen1 does not build on anyone else’s trained model. Its person detector and fight classifier are trained from scratch, on footage from the kind of scene they will watch.
Everything runs on an NVIDIA Jetson or an ordinary PC next to your recorder. Video is analysed where it is captured and does not go to a cloud service.
Beyond Python itself, the software needs only NumPy (plus ffmpeg to read streams). A small footprint is easier to audit, maintain and run on modest hardware.
Current release: version 1.24.0.
| Behaviours | Loitering (zone + dwell time), fighting (learned classifier) |
|---|---|
| Camera inputs | RTSP IP cameras, USB / V4L2 webcams, recorded video files |
| Cameras per device | Multiple. Each camera runs as its own worker with its own settings |
| Analysis rate | 10 frames per second per camera at 320 × 180 (configurable) |
| Alerts | Live web dashboard, annotated PNG evidence, JSON Lines event log, webhook |
| Alert control | Per-camera thresholds and a cooldown to stop repeat alerts |
| Hardware | NVIDIA Jetson (CUDA acceleration) or any CPU with Python 3 |
| Software | Python 3 + NumPy, with ffmpeg for decoding |
Loitering and fighting, trained and tuned for each site.
New sites add footage that the next training round learns from.
The next generation, built on what Gen1 learns in the field.
Every generation's models, results and downloads · Publications using EvolveX
We set up EvolveX on a few of your existing cameras, calibrate the zones and thresholds with your security team, and review the alerts with you. Tell us about your site and we will reply with what a pilot would involve.
We have your message and will reply by email.
Developed in collaboration with
