Powered by OmniTrack · GPU-Accelerated

Video Safety
Intelligence

Upload any surveillance footage. YAI detects threats, tracks individuals, and delivers a full safety report — powered by YOLOv10x, ByteTrack, and AnomalyTransformer.

YOLOv10
Detection Model
6
Anomaly Types
RT
Real-Time Capable
T4
GPU Backend
Capabilities
Everything your
security needs
Object Detection
Detects people and objects with state-of-the-art accuracy using open-vocabulary CLIP classification alongside YOLOv10x.
YOLOv10x + CLIP
Multi-Person Tracking
ByteTrack with Kalman filtering maintains unique IDs across occlusions and re-entries, building complete trajectory histories.
ByteTrack + Kalman
Anomaly Detection
Transformer-based anomaly model identifies loitering, crowd surges, forbidden zone violations, speed anomalies, and suspicious trajectories.
AnomalyTransformer
Depth Estimation
Depth Anything v2 provides monocular scene depth, enabling 3D-aware threat assessment from a single camera feed.
Depth Anything v2
Real-Time Streaming
WebSocket-based streaming via FastAPI delivers frame-level results as the analysis progresses, with live confidence updates.
FastAPI + WebSocket
Safety Reports
Full structured JSON report with all alerts, tracked individual metadata, confidence scores, frame timestamps, and downloadable summary.
JSON Export
Pipeline
How OmniTrack works
01
Ingest
FastAPI Upload
Video uploaded via multipart POST, queued as a background job.
02
Detect
YOLOv10x + CLIP
Frame-by-frame detection with open-vocabulary labeling at adjustable confidence.
03
Track
ByteTrack
Kalman-filtered tracking maintains consistent IDs across frames and occlusions.
04
Analyze
AnomalyTransformer
Transformer model scores trajectories for behavioral anomalies in context.
05
Report
Structured JSON
Complete safety report with alerts, track data, depth estimates, and timestamps.

Ready to see
what's happening?

Upload your footage and get a complete AI-powered safety analysis in under a minute.

YAI Analyzer

Upload & Analyze

Drop a video file and configure your detection settings. The pipeline runs detection, tracking, and anomaly analysis automatically.

Drop video here
MP4 · MOV · AVI · MKV — up to 500 MB
Detection Settings
Confidence Threshold Detection sensitivity
0.50
0.05 0.25 0.50 0.75 0.95
Low (0.05–0.35)
Sensitive — more detections
Balanced (0.36–0.64)
Recommended for most videos
High (0.65–0.95)
Strict — fewer false positives
Uploading...
0%
Uploading video to GPU server
Running object detection — YOLOv10x
Tracking individuals — ByteTrack
Anomaly analysis — AnomalyTransformer
Building safety report
Safety Report
Alerts Detected
Tracked Individuals
Track ID Class Frames Active Avg Confidence Flagged
Download JSON Report
About YAI

The stack
behind it

What is YAI?

YAI (Video Safety Analyzer) is a web product built on top of OmniTrack — a real-time video analytics pipeline combining state-of-the-art computer vision models. Users upload a video, choose a confidence threshold, and receive a structured safety report identifying behavioral anomalies, tracked individuals, and flagged events.

Detection

YOLOv10x runs frame-by-frame inference to detect people and objects. CLIP provides open-vocabulary classification, enabling zero-shot category recognition beyond the fixed YOLO labels.

YOLOv10x CLIP PyTorch CUDA Kaggle T4

Tracking

ByteTrack paired with Kalman filtering associates detections across frames, handling occlusions and re-entries gracefully. Each person receives a persistent track ID with full trajectory history.

ByteTrack Kalman Filter Python 3.11

Anomaly Detection

AnomalyTransformer scores trajectories for unusual behavioral patterns including loitering, speed anomalies, crowd surges, forbidden zone violations, trajectory anomalies, and disappearances.

AnomalyTransformer 6 Anomaly Types

Depth Estimation

Depth Anything v2 provides monocular depth estimates per frame, giving the pipeline 3D spatial context for more accurate threat assessment without requiring stereo cameras.

Depth Anything v2 Monocular Depth

API Endpoints

GET
/health
Returns API status and model readiness
POST
/analyze
Upload video + conf_threshold → returns job_id, queues analysis pipeline
GET
/analyze/{job_id}
Poll job status → returns status + full result JSON when done
WS
/ws/{job_id}
WebSocket stream — delivers real-time frame results as they process

Infrastructure

The FastAPI backend runs on a Kaggle notebook with a free T4 GPU, tunneled via an ngrok permanent domain. The frontend is a static single-file HTML app hosted on Netlify — no build step, no framework dependencies.

FastAPI uvicorn pyngrok Kaggle T4 GPU Netlify ngrok tunnel

Open Source

OmniTrack is publicly available on GitHub. Contributions, issues, and forks are welcome.