This project was born out of a high-stakes 48-hour challenge at the Lloyd Assurances Hackathon, focusing on the bleeding edge of InsurTech and live Computer Vision. The objective was to build a comprehensive, zero-latency system capable of analyzing driver behavior to prevent accidents in real-time.
We architected a dual-stream processing engine that ingests parallel video feeds: one facing the road, and one monitoring the driver's vitals. At the core of the pipeline, we integrated YOLOv8 for high-speed external object detection (identifying pedestrians and vehicles) and MediaPipe to continuously track the driver's gaze, blink rate, and head pose.
To tie it all together, we constructed an asynchronous WebSocket-based backend in Python that fused these data streams. The system calculates a dynamic, weighted 'Driver Score' and triggers visual and auditory safety alerts the millisecond it detects microsleep, distraction, or imminent tailgating—operating flawlessly with a sub-300ms round-trip latency.
Key Takeaways
- Asynchronous WebSocket integration is paramount when fusing multiple heavy-compute ML streams (YOLO + MediaPipe).
- Building fail-safes into high-speed physical threat monitoring requires extremely efficient memory threading in Python.