Hire · Freelance & consulting
Hire a freelance computer vision engineer for CCTV and video analytics
I build computer vision systems that run on real camera feeds, not just sample videos: real-time person detection and tracking on CCTV, ANPR, crowd monitoring, and the backend that turns detections into stored events and alerts. Smart India Hackathon 2023 finalist for CCTV analytics; now doing CCTV and ANPR work at Zuneko Labs.
Problems I solve with this stack
The pipeline works on sample videos but not on live RTSP feeds.
Handle the real world: dropped frames, reconnects, varying resolutions and clock drift between cameras, all part of the CCTV builds I've shipped (OpenCV capture and pre-processing, then YOLO detection).
The model runs fine on a workstation but can't keep up on the edge device.
Export to ONNX/TensorRT with the right precision (FP16/INT8 where acceptable) and benchmark on your hardware, so you know what latency to promise before deployment.
You have cameras recording, but nobody watches the footage until something goes wrong.
Real-time detection and tracking (YOLO with DeepSORT/ByteTrack) that turns frames into events (who, where, when) and sends alerts over WhatsApp or SMS.
The model is accurate on a GPU workstation but too slow on site hardware.
Export to ONNX/TensorRT, tune frame rate and resolution to the use case, and measure throughput on the device you'll actually deploy, such as NVIDIA Jetson.
Detections are noisy, so the alert channel gets ignored.
Tracking across frames plus deduplication rules, so one incident produces one alert instead of fifty.
What you get
- A detection + tracking pipeline on your RTSP/CCTV streams.
- An event store (PostgreSQL/MongoDB) and alert dispatch (WhatsApp, Twilio).
- Edge-optimized models (ONNX/TensorRT) where the hardware needs it.
- A throughput report: streams, frame rate and latency on your hardware.
- Documentation and a handover call.
How a Computer vision engagement runs
1. Review sample footage
Look at real clips from each camera: angle, lighting, resolution and what counts as an event. This shapes the model and the alert rules more than any benchmark does.
2. Run a baseline on your feeds
Pull your RTSP streams with OpenCV, run a detector over them and record what it gets right and wrong before any tuning.
3. Add tracking, dedup and alerts
Track objects across frames, deduplicate alerts, store events with camera ID and timestamp, and connect alerts to WhatsApp, SMS or webhooks.
4. Benchmark and hand over
Measure streams, frame rate and latency on the target hardware, export to ONNX/TensorRT where needed, and hand over documentation with a walkthrough call.
Frequently asked questions
Can you work with RTSP streams from IP cameras?
Yes. RTSP ingestion with reconnect and error handling is the first stage of every CCTV pipeline I've built, including the four-feed Smart India Hackathon system.
How much faster will TensorRT make our model?
It depends on the model, the precision and the device, so I benchmark on your hardware and report measured numbers instead of quoting a generic speedup. I compare FP16 and INT8 against your validation data so you choose the accuracy trade-off.
What computer vision problems do you take on?
People detection and tracking, crowd monitoring, ANPR, and turning video into stored events and alerts: the CCTV analytics I've built at Zuneko Labs and for Smart India Hackathon 2023.
Do we need new cameras or GPUs?
Often not. The pipeline reads your existing RTSP streams and the model can be optimised for the hardware you have; I benchmark first so you know before buying anything.
How do you measure whether the detector is accurate enough?
By running it on footage from your own cameras and counting missed events and false alerts against what a person sees in the same clips. Public benchmark scores do not carry over to a different camera angle or lighting.
Will it work at night or in poor light?
It depends on the camera and the lighting, which is why footage review comes first. Low-light footage may need different training data or camera settings, and I will say so after seeing samples instead of promising it up front.
Do you work with international clients?
Yes. I'm based in Pune, India, and work with teams in India and remotely worldwide. Communication is async-first with scheduled calls when needed.
What does an engagement look like?
A short call to scope the problem, a written scope (what will be built, what won't, and how we'll know it works), then small milestones you can review, with the code in your repository from day one.
More questions? Ask directly
Last updated
Need help with Computer vision?
Book a free 15-minute call below — describe the problem, I'll tell you honestly whether I'm the right person for it.
How I work with clients