In-Cabin Monitoring System (ICMS) & ADAS
Published:
DeltaX.ai, Seoul — 2023–2025
When a system has to make safety decisions inside a moving vehicle in real time, there's not much room for error — and even less room for a model that needs a proper GPU to run. Most of my work here was about closing that gap.
On the monitoring side, I built and deployed models for seatbelt detection, hands-on/off-wheel, drowsiness detection, child presence, and occupancy — all from IR cameras, because IR works in the dark and sidesteps a lot of privacy concerns you'd face with RGB. Getting these to run reliably on Jetson and TI boards required thinking carefully about model architecture from the start, not just quantizing something at the end and hoping for the best.
The ADAS work was a different flavor of constraint. I built a semantic segmentation model for road scenes — potholes, curbs, speed bumps, cracks — that had to be light enough for a TI embedded board while staying accurate enough to actually be useful. Lane departure and forward collision warnings ran through a monocular depth pipeline, which meant tuning something fast enough for real-time inference and stable enough that safety-relevant triggers didn't fire on noise.
Stack: PyTorch · IR/Thermal cameras · ONNX · TensorRT · TI board · NVIDIA Jetson · Hailo