Real-Time ANPR & vehicle detection

Budget: ₹12500 - ₹37500

I need an engineer who can take complete ownership of a real-time automatic number-plate recognition and traffic-analytics pipeline that ingests live IP camera streams, runs fast and accurate inference, and pushes results reliably from edge devices. The current target hardware is NVIDIA Jetson, so every design choice—from model architecture to post-processing—must respect its compute limits while still keeping total end-to-end latency under 200 ms.

The core work revolves around training, tuning, and deploying YOLO-style detectors in PyTorch (TensorFlow knowledge is welcome if it helps optimisation). You will refine the models for two challenging scenarios that matter most to our roadside installations: low-light environments and high-speed vehicle movement. Image enhancement, motion-blur compensation, and clever data-augmentation strategies are all fair game as long as they translate into measurable accuracy gains after conversion to TensorRT or ONNX.

Alongside model work, you will craft a robust video pipeline with OpenCV and Python that can pull RTSP feeds, batch or single-frame them sensibly, and expose clean JSON or MQTT outputs for the rest of our stack. Field issues do crop up, so part of the role is writing self-diagnostics and remote-update hooks to keep hundreds of Jetson boxes healthy without on-site visits.

Deliverables • Trained and optimised ANPR + vehicle detection/classification models (TensorRT engine files) • Real-time Python pipeline code with clear configuration files • Deployment scripts for Jetson (container or native) and a short setup guide • Test report demonstrating <200 ms latency and target accuracy under low-light and high-speed benchmarks

Acceptance criteria will be the successful processing of sample IP streams I provide, meeting the latency and accuracy targets, and a smooth, reproducible deployment on my Jetson dev kit. If this sounds like your kind of challenge, let’s discuss the data and milestones.