Semantic Anomaly Stream Discovery

Overview

An unsupervised vision pipeline that ingests video streams, identifies objects, and discovers anomalies through clustering in real-time.

SASD_Visualization_Demo_compressed.mp4

SASD_Dashboard.mp4

Key Features

Unsupervised Anomaly Discovery

Leverages YoloV8, DINOv2 features, and HDBSCAN to automatically discover objects and segregate anomalies from raw video feeds, eliminating the need for extensive pretraining.

Real-Time Flow Analytics

Combines high-performance SORT tracking with custom flow analysis to calculate velocities, active counts, and movement patterns for each discovered object cluster in real-time.

Getting Started

Installation

# Clone the repository
git clone https://github.com/yourrepo/semantic-anomaly-stream-discovery.git
cd semantic-anomaly-stream-discovery

# Install dependencies
pip install -e .

Usage

# Run on a video file
python -m src.main path/to/video.mp4 --warmup 60

# Run on a YouTube stream
python -m src.main "https://youtube.com/watch?v=..." --warmup 60