Hello Luxonis Support / Engineering Team,
I am developing a continuous, resource-efficient authentication system using an OAK-D Lite camera connected to a Raspberry Pi. My current DepthAI v2 pipeline utilizes face-detection-0200 combined with face-reidentification-retail-0095.
To optimize performance on the host device, I have implemented a custom tracking logic using bounding box IoU and spatial depth (Z-axis) to minimize ReID inferences, alongside an IQR-based depth analysis for anti-spoofing.
The system is deployed in a laboratory environment, which introduces a specific challenge: users frequently wear partial face coverings (PPE). This includes:
While the current setup works excellently for standard unoccluded faces, I need to optimize the recognition accuracy for these laboratory conditions. Could you please advise on the following:
Alternative Models: Are there specific pre-trained OpenVINO models (either in the public Model Zoo or internal ones) that are better suited or trained specifically for detecting and re-identifying heavily occluded faces?
Pipeline Tuning: Are there specific camera control settings (e.g., manual exposure tuning, ISP scaling) or DepthAI node configurations that could significantly improve feature extraction for the ReID network under these conditions?
Best Practices: Do you have any reference scripts or documentation tailored for facial recognition in environments where users wear PPE, specifically considering the compute constraints of a Raspberry Pi?
Thank you in advance for your technical guidance and support.
logic-ohne-zeit.bin# ============================================================
# Datei: logic.py
# ============================================================
import json
import numpy as np
Error previewing file. It may have been deleted, or the provided file ID is invalid.
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import time
import sys
import cv2
Error previewing file. It may have been deleted, or the provided file ID is invalid.
###regeln.json
{
"kamera": {
"aufloesung_breite": 960,
"aufloesung_hoehe": 540,
"bilder_pro_sekunde_fps": 15
},
"sicherheit": {
"ki_erkennung_minimum_prozent": 65,
"wartezeit_bis_abwesend_sekunden": 2.0,
"enrollment_bilder_anzahl": 15
},
"tracking": {
"erlaubte_tiefen_bewegung_mm": 180,
"spoofing_iqr_schwellenwert": 5.0,
"reverify_interval_sec": 0.5,
"track_hold_sec": 1.2,
"track_score_threshold": 1.5
},
"session": {
"datei": "/dev/shm/pi_session.json"
},
"system": {
"log_historie": "/dev/shm/status.log",
"skript_anmeldung": "/home/auth/depthai-core/auth_gui/anmeldung.py",
"skript_abmeldung": "/home/auth/depthai-core/auth_gui/abmeldung.py",
"zeige_display": false
}
}
