· Nº 08 — Araskova Studio · on request · TITAN v0.1.0

Titan. — edge vision with persistent identity, from Araskova Studio

Edge vision · persistent identity · on-premises

Runs next to the cameras a site already has. It tracks every person with an ID that survives blind spots (out of view through a door and back, same ID), flags weapons and enrolled faces, and keeps all footage on the premises.

Ask about Titan Measured numbers

SYS_02

The Studio · Titan

Drawn illustration of TITAN's HUD on abstract figures — tracking, threat profile, out through a door and re-acquired

· 001 — Measured

4-core CPU only

Detector latency: 27.6ms
Accuracy kept by INT8: 0.443mAP50-95
Live throughput: 23fps
Re-ID scenarios passed: 8/ 8
Re-acquisition time: 0.27s
HUD render: 1.9ms

· 002 — Capabilities

19 capabilities Verified Built, field test pending Planned

Perception[04]

●
Threat engine
YOLOv8 / YOLOv10 (NMS-free) object and weapon detection through ONNX Runtime from Rust, INT8 by default.
●
Biometric matrix
SCRFD face detection, 5-point alignment and 512-d ArcFace embeddings matched against an enrolled gallery (authorised / watchlist).
●
Context veto
A weapon class touching benign context (a knife on a dining table) raises Caution instead of Alpha.
○
Firearm-trained detector
Fine-tuned handgun / rifle / knife model from site footage plus hard negatives.

Tracking[02]

●
Multi-target tracking
ByteTrack association over decoupled Kalman filters with a 10-frame trajectory ring and projected motion vector.
●
Range and closing speed
Monocular pinhole range per person/vehicle; truncated boxes are excluded so edge-of-frame bodies don't fake an approach.

Identity[07]

●
Persistent person IDs
Every person gets a global ID (P-0007) that survives leaving view: restrooms, doorways, long occlusions.
●
Re-acquisition
Returning people are matched by face print, clothing signature and optional learned re-ID, with portal-zone priors and an ambiguity margin (never a forced merge).
●
Crossing guard
When two identified people cross and the tracker swaps them, the swap is detected after they separate and undone.
●
Threat memory
Someone who was Alpha (armed, watch-listed) when they left is Alpha again when they come back.
●
Zones and events
Portal / entrance / restricted zones; enter, exit, re-enter and zone events as JSON lines; occupancy and unique-visitor counts.
◐
Learned re-ID model
Plug in any [1,3,256,128] -> [1,D] ONNX (OSNet / FastReID) for look-alike outfits.
○
Multi-camera hand-off
Shared identity registry across cameras.

Output[02]

●
Tactical HUD
Targeting brackets, threat profiles, trajectories, range radar and identity overlays in the Anvin / Araskova ink-bone-signal palette.
●
Threat-matrix metadata
Per-frame JSON with every track, level, range, velocity, identity and timing; H.264 / RTSP video out.

Learning[02]

●
Data flywheel
Uncertain weapon detections and every Alpha onset are saved with YOLO pre-labels for review (CVAT / Label Studio).
●
Fine-tune and tune
One command trains on tactical data, exports ONNX + INT8, benchmarks accuracy vs latency and derives per-class thresholds.

Deployment[02]

◐
Live cameras
RTSP / ONVIF streams, V4L2 / MIPI devices via ffmpeg, with latest-frame (drop-stale) live semantics.
◐
GPU / NPU acceleration
TensorRT (Jetson Orin) and OpenVINO (Intel Core Ultra) execution providers behind cargo features.

· 003 — Threat engine

Accuracy vs latency · COCO128

Every exported variant is scored for accuracy and timed on the same machine. The deployed model is INT8 with the detection head kept in full precision: within a point of FP32 accuracy, about 1.5× faster.

Model	Precision	Input	mAP50-95	p50 ms
yolov8n-640	FP32 · NMS	640	0.451	41.0
yolov8n-640.int8deployed	INT8 · NMS	640	0.443	27.6
yolov8n-320	FP32 · NMS	320	0.377	13.4
yolov8n-320.int8	INT8 · NMS	320	0.374	10.8
yolov10n.e2e	FP32 · NMS-free	640	0.455	39.8

· 004 — Pipeline

1280x720

Offline throughput: 19 fps
Live throughput: 23 fps
Glass to glass p50: 97 ms
Glass to glass p95: 123 ms
Stale frames dropped: 73 / 300
Identity mixer p50: 0.07 ms

Live mode never queues: when inference falls behind, the oldest frame is dropped so the overlay is always about now.

Stage latency · p50 bar, p95 tick · ms

detector infer33.7
face detect15.0
face embed (1)14.3
hud render1.9
perception (parallel)37.7
preprocess0.9

· 005 — Persistent identity

8 / 8 scenarios

Someone goes out through a door. A few seconds later they come back: Titan gives them their old ID, and if they were flagged when they left, they're flagged when they return.

Face prints, clothing signature and the door they used are fused; a stranger leaving the same door gets a new ID, and a tracker swap when two people cross is detected and undone. Wrong-ID frames across the suite: 46 of 6830 (0.67%).

Drawn illustration — the person who leaves gets their ID back; a stranger from the same door does not

Scenario	Faces	Cue	Re-acquire	Result
stranger exits after A is back	on	face	0.27 s	Pass
stranger exits after A is back	off	appearance	0.27 s	Pass
look-alike stranger exits after A is back	on	face	0.27 s	Pass
look-alike stranger exits after A is back	off	appearance	0.27 s	Pass
look-alike stranger exits while A is inside	on	face	0.27 s	Pass
look-alike stranger exits while A is inside	off	appearance	0.27 s	Pass
stranger exits while A is inside, then crosses A	on	face	0.27 s	Pass
stranger exits while A is inside, then crosses A	off	appearance	0.27 s	Pass

· 006 — What it does

System output · drawn scenes

Crowd · system outputCrowd tracking: compact ID tags, full labels for the top targets, trajectories.

Every person, one ID

Each person gets a global ID and a trajectory the moment they appear, and keeps it.

Threat, with context

An object in hand raises the profile to Alpha. A knife on a dining table is Caution, not Alpha.

Crossing guard

When two people cross and the tracker swaps them, the swap is detected and undone.

· 007 — How these were measured

Machine: Intel(R) Xeon(R) Processor @ 2.10GHz · 4 cores
Accelerator: none (CPU)
Runtime: ONNX Runtime 1.28.0 · cpu
Build: titan 0.1.0 · 8702f29
Generated: 2026-10-01

All figures were measured on a 4-vCPU shared Intel Xeon with no GPU. Production edge hardware will differ; re-run tools/showcase.sh on the target.
The stock COCO detector has no firearm class and is weak on knives; a weapon model must be fine-tuned on tactical data (none was available here).
The re-identification suite uses composited COCO people in a synthetic corridor: it proves the logic (re-acquire, no false merge, swap correction), not field accuracy.
Sub-5 ms detection needs GPU/NPU acceleration; the TensorRT / OpenVINO paths are built but were not benchmarked here.
Face recognition is regulated (GDPR, India DPDP Act 2023, BIPA, EU AI Act): enrol only consenting people. The bundled InsightFace models are non-commercial.
The crowd frame is TITAN's HUD on a public COCO dataset image. The animations are drawn illustrations of TITAN's behaviour on abstract figures, not recordings of the system.
The deployed detector is Ultralytics YOLOv8 (AGPL-3.0). Commercial deployment needs an Ultralytics enterprise licence or a detector under a permissive licence.

· + — More from Araskova

What works today, first