01 / IMPLEMENTATION & CONTRIBUTION
What the work involves
Fit the DMS model pipeline to the low-power CV181x edge SoC during the Shanghai Ben’an Intelligent internship. Tightened inference cadence and per-stage compute budgets while adapting a supplied C service, landmark backends, and model-conversion packaging.
Technical depth
Sparse SCRFD detection with per-frame landmark tracking and ROI updates; legacy 478-point and 68-point backends; callbacks and display work moved off the per-frame critical path; timestamp-window PERCLOS; per-stage latency and tracking instrumentation; ONNX to CV181x BF16 conversion.
The project family
dms_cv181x_fastpath/dms_work02 / RESULTS
What came out of it
Validated the end-to-end pipeline on real cabin video, with inference-cadence checks and per-stage compute budgeting. The internship record connects this edge pipeline to shadow-mode hard-case capture, review, and replay-based reporting.
03 / SUPPORTING EVIDENCE
Follow the source
Implementation notes, project records, and supporting artifacts.
Source context & project scope
The latest candidate-provided internship record updates the earlier repository-only hardware-validation status. It describes on-device deployment and end-to-end cabin-video validation; no numeric latency or accuracy benchmark is supplied.
20 ms remains a target, not a claimed achieved measurement.
SCRFD and PFLD are upstream models; original model training is not claimed.
Company-specific source is not mirrored on this website.
The code computes sample ratios inside a timestamp window; do not describe it as exact time-integrated PERCLOS.
Deployed and validated real-time on-device driver-state monitoring, connecting safety-relevant alerts, low-power edge inference, and replay-driven validation. Deployed and validated eight safety-relevant detector classes: eye closure, yawning, distraction, phone use, smoking, face loss, lens occlusion, and eye anomaly. Shipped a 616-test pytest suite with replay-driven regressions, inference-cadence checks, and per-detector coverage metrics alongside model development. Designed the alert state machine with hierarchical face-loss and lens-occlusion fallbacks. Speed, ignition, and gear gating suppress false alarms in non-driving states; persisted per-camera calibration profiles support fleet-grade traceability. Fit the model pipeline to the low-power CV181x edge SoC by tightening inference cadence and per-stage compute budgets. Validated end-to-end on real cabin video and built the lab CLI for shadow-mode hard-case capture, human review, and replay-based metric reporting.
CANDIDATE · 2025 – PresentHistorical architecture and bundle-status record, including hardware checks that were still open at that snapshot. Donald’s later internship update describes on-device deployment and real cabin-video validation.
SOURCE · 2026-09-17Sparse detector scheduling, tracked ROI, landmark backends and latency fields.
SOURCE · 2026-09-17Vendor SCRFD provenance, public PFLD ONNX provenance and BF16 conversion notes.
SOURCE · 2026-09-17Critical-path and API problems motivating the refactor; model availability statements are older than the main README.
SOURCE · 2026-09-17