ADAS learning projects in Rust

Seven driver-assistance exercises, from lane detection to an integrated dashcam demo, implemented in Rust.

Role: Personal learning project
Organisation: Open source
Period: Started 2023
Stack: Rust, ONNX, tract, image, imageproc, ffmpeg
Links: GitHub

Dashcam frame with lane shading, object detections, a speed-limit sign and a rendered instrument cluster

Working through the systems

I used a set of ADAS course exercises to explore driver-assistance features in Rust. The workspace builds up from lane detection to sign and speed-limit recognition, object detection, traffic-light detection and an instrument cluster. A final activity runs them together on a dashcam clip.

Inference uses ONNX models through tract, image processing uses Rust libraries, and ffmpeg handles video input and output. The implementation runs without Python or OpenCV.

The integrated learning demo. Sample road footage: Udacity, MIT licence.

Handling uncertain detections

The sign classifier can be confidently wrong. The speed-limit tracker therefore waits for agreement across frames and a clear lead over competing readings before changing the displayed limit. Object and traffic-light detections have their own false positives, documented in the repository.

This is a learning project using recorded footage, separate from my production ADAS work. It does not demonstrate a validated driving system. For my experiments with vision-language planning, see EMMA-VLM.