abhishek.dev
All work

Case study · 2026Featured

DriveAware

Real-time driver drowsiness and distraction detection — three OpenVINO Model Zoo models chained via Async Inference, hitting ~30 FPS at under 60 ms latency on a CPU-only Intel Core i5.

Repository
  • Python
  • OpenVINO Toolkit
  • OpenVINO Model Zoo
  • OpenCV
  • NumPy
  • PERCLOS

Problem

Most published driver-drowsiness systems lean on GPUs or dedicated accelerators. The more interesting question is whether the same task runs at real-time framerates on a commodity laptop CPU. DriveAware was built to answer that, against a target of 30 FPS at under 60 ms end-to-end latency on an Intel Core i5 with no GPU.

Architecture

Webcam~30 FPSFace detectOpenVINO IR · FP16Head poseyaw / pitch / rollEye classifieropen / closedAsync Inference API(parallel)Distractionhead-pose flagsDrowsinessPERCLOS + alert

Three pre-trained OpenVINO Model Zoo models are chained through the OpenVINO Async Inference API so the CPU stays busy during the next frame's inference rather than blocking:

  1. Face detection — bounding box for every face in the frame.
  2. Head-pose estimation — yaw / pitch / roll for the primary face.
  3. Open / closed eye classification — run on each eye region.

PERCLOS (percentage of eyelid closure over a sliding window) feeds drowsiness scoring; sustained head-pose deviations flag distraction. Both states trigger an audible alert and a CSV log entry with a timestamp.

Key engineering decisions

  • OpenVINO IR over PyTorch / ONNX direct. The Intel toolkit compiles models to IR formats (FP32, FP16) tuned for the CPU's vector extensions, with meaningful speedups on commodity hardware.
  • FP16 IR for the hot path. Benchmarking on the i5 (no GPU) with FP16 hit the ~30 FPS target at under 60 ms end-to-end. FP32 was tested side-by-side as the reference.
  • Async Inference API instead of synchronous calls. Synchronous calls leave the CPU idle while inference runs. Async chaining hides that latency by starting the next frame before the current one finishes.
  • PERCLOS over instantaneous closed-eye detection. A single frame of closed eyes is a blink; a sustained closure ratio over a window is drowsiness. PERCLOS encodes the distinction directly.
  • Local CSV incident log. The whole pipeline runs offline — no network round-trip, no cloud dependency.

Stack

Python 3, OpenVINO Toolkit, OpenVINO Model Zoo (three pre-trained models), OpenCV for capture and overlay, NumPy.

Links