Smart Attendance.
Engineered by AI, Tracked in Real-Time.
Automate employee check-ins, break tracking, and shift analytics with multi-camera spatial tracking and lightning-fast FAISS vector search.
Telemetry Mini-Dashboard
A direct representation of our dynamic admin dashboard monitoring live cameras, headcount states, and check-in triggers.
AI Core Features & Speed
Powered by asynchronous multithreading, advanced vector matching, and custom tracking logic to provide instantaneous checks.
Thread-Safe Camera Capture
Asynchronous multi-threading ensures zero latency. Automatic hardware/network reconnection algorithms keep the gate and office cameras active 24/7.
L2-Normalized FAISS Search
Deep-feature embeddings matched in milliseconds using Facebook AI Similarity Search (FAISS). Precise cosine matching filters out false positives instantly.
Spatial Centroid Tracking
Distance-gated tracking tracks multiple moving targets between frames, eliminating identity flickering and keeping tracking IDs locked in crowded zones.
Bipartite Conflict Resolution
Strict one-to-one identity mapping. Demotes inferior face matches automatically so one person cannot occupy two places at once.
Smart Break State Machine
Automatically detects when an employee leaves their post, counting break timings after 30 seconds of screen absence and resuming work states automatically on return.
Auto Check-Out Guardian
Prevent stale sessions. The system automatically triggers check-out at the last-seen stamp after 3 hours of continuous invisibility.
Interactive Call-Flow Pipeline
Hover or tap on each node to witness how a single raw camera frame transforms into an immutable attendance record in milliseconds.
Multi-Threaded Real-Time Video Capture
Cameras stream raw frames. Independent non-blocking daemon threads resize incoming arrays asynchronously. This ensures zero main-loop freezing, even if camera connections flicker.
- OpenCV camera capture on separate worker thread
- Non-blocking thread-safe frame sharing buffer
- Dynamic hardware/network auto-recovery logic
def read_frames_async(self):
while self.active:
ret, frame = self.cap.read()
if not ret:
self.reconnect_stream()
continue
self.frame_queue.put(frame)512-Dimensional Face Vector Generation
The InsightFace model localizes facial landmark features inside bounding boxes. It translates facial geometry into highly descriptive 512-dimensional floating-point vectors.
- High-precision facial bounding box localization
- 512-dimension spatial normalized embeddings
- GPU accelerated tensor processing engine
def extract_embeddings(self, face_img):
faces = self.model.get(face_img)
if not faces:
return None
# L2 Normalization applied instantly
embedding = faces[0].normed_embedding
return embeddingLiveness Detection & Spoof Rejection
Before matching, the system runs a binary liveness classifier on the face crop to reject photo, video, and 3D mask attacks. Only confirmed live faces proceed to the vector search stage.
- Silent face liveness scoring via depth texture analysis
- Rejects photo, video replay & 3D mask attacks
- Configurable confidence threshold per deployment
def predict(self, face_crop):
tensor = self.preprocess(face_crop)
score = self.model(tensor)[0]
label = "Real" if score > self.threshold \
else "Spoof"
# Reject spoof before FAISS query
if label == "Spoof":
return "ACCESS_DENIED"
return label, scoreFacebook AI Similarity Search (FAISS)
The system queries the generated face vector against the registered employees database using the FAISS indexing system, searching through thousands of profiles in microseconds.
- Microsecond similarity lookup (IndexFlatIP Index)
- Cosine distance scoring for strict compliance
- Threshold demotion filtering for unknown guests
def query_vector(self, face_vector):
distances, indices = self.faiss_index.search(
np.array([face_vector]), k=1
)
score = distances[0][0]
if score > self.threshold:
return self.id_map[indices[0][0]], score
return "Unknown", scoreTemporal State & Bipartite Gating
Identities are tracked frame-to-frame. Centroid calculations and bipartite conflict resolution lock IDs to moving boxes, eliminating flicker and ghosting.
- Centroid-based spatial tracking lock-on
- Dynamic verification: 4 confirm frames / 3 doubt frames
- 1-to-1 bipartite map matching prevents duplicates
def update_tracker(self, detected_centroids):
cost_matrix = self.compute_cost(detected_centroids)
row_ind, col_ind = linear_sum_assignment(cost_matrix)
for r, c in zip(row_ind, col_ind):
self.tracks[r].centroid = detected_centroids[c]
self.tracks[r].confirmations += 1Immutable Check-In Registry
Verified check-ins are safe-written to a local thread-locked JSON database file and simultaneously mirrored/synced onto an admin Google sheet or SQL database.
- Thread-safe JSON operations using file locks
- Real-time check-in trigger hooks for HR tools
- CSV mirror logging for immediate analytics export
{
"staff_id": "028",
"name": "Ayush Singh",
"timestamp": "2026-05-18T10:14:02",
"camera": "Gate Cam",
"status": "Checked-In"
}HR Operations Analytics Panel
Observe check-in velocity, active shift metrics, and breaks patterns compiled directly from system storage files.
Hourly Peak Active Headcount
Simulated live data from headcount_log.csv
General Operations Metrics
| Metric Description | Recorded Value | Status Deviation |
|---|---|---|
| Avg. Employee Shift Duration | 8 hrs 14 mins | +2.4% (Optimal) |
| Average Cumulative Break Time | 48 mins / day | Steady |
| First-Hour On-Time Percentage | 96.4 % | +1.1% (High) |
| Database FAISS Match Queries | 142,391 | Active Real-Time |
| Total Registered Staff Indices | 142 Employees | Synchronized |
Face Enrollment Simulator
Simulate enrollment.py. Register a new user, scan their photo, and witness the FAISS vector index database update in real-time!
Register New Profile
Provide profile details to compile a new 512-dimension face embedding.
Run Spatial Scan
Simulate the facial capture algorithm localizing landmarks.
Enrollment Successful
Employee profile vectorized and indexed successfully into local DB and active mirror spreadsheet.
FAISS Index update log...