Video analytics software development
Find one vehicle across a city's cameras, and let a person decide.
Acasa Labs builds video analytics software for security and surveillance teams: number plate recognition, vehicle search across cameras, route reconstruction, crowd counts, and zone alerts, all reviewed by an operator. We built Acasa Sentinel for the Gujarat Police Hackathon Innovation Challenge 2026, a vendor-neutral system that runs all inference locally.
The problem
Thousands of cameras, and people scrubbing footage one by one.
Cities and campuses run thousands of cameras from different vendors, each with its own recorder and software. When something happens, tracing a vehicle means people watching footage camera by camera, often for hours. Plate reading alone does not solve it: on wide-area cameras, plates are often unreadable. Teams need software that works across vendors, finds a vehicle by more than its plate, and leaves the decision in an operator's hands.
What we build
- 01
Ingest from any camera
RTSP and HLS streams behind one interface, from any vendor, with a worker per camera under a supervisor.
- 02
Number plate recognition
Plate detection and OCR per tracked vehicle, matched against watchlists with fuzzy matching for misreads.
- 03
Search across cameras
Find a vehicle by its appearance when the plate cannot be read, and reconstruct its route across the network.
- 04
Alerts with a person in the loop
Watchlist hits, crowd counts, and zone intrusion raised in an operator console to confirm or reject, with evidence and an audit trail.
From production
What we learned building Acasa Sentinel.
Footage stays on your hardware
Acasa SentinelAll inference runs locally. No footage or number plate is sent to a third-party API.
Candidates, not verdicts
Acasa SentinelRoute reconstruction rejects physically implausible hops and says why, and watchlist hits wait in a queue for an operator to confirm or reject, with evidence crops and an audit trail.
One events table for every analytic
Acasa SentinelPlates, crowd counts, and zone intrusion all land in one events table, so a new kind of alert needs no schema change.
Measured, in a sandbox
Acasa SentinelSentinel matched 24 of 24 plates on the proof clip, raised 0 false alarms across 100,000 unrelated plates, and ran its detector at 5.7 ms per frame at 1080p on Apple Silicon. Measured on the running system against the hackathon sandbox, not a production deployment.
Honest trade-offs
A hackathon build, described as one
Acasa Sentinel is a hackathon build, run against 30 sandbox cameras, not a city-wide deployment. It shows the architecture and the engineering choices we would bring. A production system also needs site surveys, camera-by-camera tuning, load testing on your network, and a legal review of what you may record and keep. Use this page to judge how we think, and ask us how a pilot on your own cameras would work.
What we reach for
- Python
- FastAPI
- YOLOv8, ByteTrack
- ONNX plate OCR
- DINOv2
- Postgres, pgvector, PostGIS
- FFmpeg, OpenCV
- Next.js, Leaflet
Proof
Shipped, not pitched.
01Case study
Acasa Sentinel
Give it a number plate. It returns the vehicle's route across a CCTV network.
A vendor-neutral video analytics system built for the Gujarat Police Hackathon Innovation Challenge 2026. It reads number plates, re-identifies vehicles by appearance across cameras, reconstructs routes, and raises watchlist alerts for an operator to confirm.
Operator console (web) · API
Questions
Security and surveillance questions we get asked.
Does it work with our existing cameras?
Sentinel takes RTSP and HLS streams behind one interface, so it is vendor-neutral by design. We would check your recorders and stream settings in a pilot.
Is Acasa Sentinel deployed in production?
No. It was built for the Gujarat Police Hackathon Innovation Challenge 2026 and measured on a sandbox. We can build a production system on the same design, starting with a pilot.
Is footage sent to the cloud?
Not in Sentinel. All inference runs locally, and no footage or number plate goes to a third-party API.
What hardware does it need?
The same pipeline falls back from CUDA to Apple Silicon to CPU. Hardware for a real site depends on the number of cameras and the analytics you run, which a pilot measures.
Who makes the final decision on an alert?
A person. Watchlist hits and reconstructed routes are candidates for an operator to confirm or reject, with evidence crops and an audit trail.
Do you handle legal approval for surveillance?
No. Whether and how you may record, analyse, and keep footage is for you and your legal advisers. We build the access controls, retention settings, and audit logs your policy requires.
Video analytics software development
Thinking about video analytics for your cameras?
Tell us how many cameras, from which vendors, and what you need to find. We will tell you what a pilot would look like.