Back to the scenes
Doğru Holding (DGR Project), May 2023 to Dec 2023
Face recognition with stereo liveness on Jetson Nano
98% accuracy, 30% fewer false positives, and face embeddings stored in 4 KB on-device.
98% accuracy
Problem
Face recognition had to run on a Jetson Nano, store each face compactly on the device, and reject spoofing attempts.
Approach
- Recognition on the edge. The system runs on a Jetson Nano and stores face embeddings in 4 KB on-device.
- Stereo liveness, built from scratch. Per-keypoint facial depth profiles from a stereo camera, a labelled real vs. spoofed dataset I collected, and a lightweight TensorFlow model that rejects spoofing attempts.
- Real time on constrained hardware. Performance-critical components in C++, exposed to the Python pipeline as native extensions.
Result
98% accuracy and 30% fewer false positives, with face embeddings stored in 4 KB on-device.
What I would do next
Test liveness against more spoof types, such as screens and printed masks, and report the rejection rate for each.