Argus — Real-Time Detection & Tracking

Detection alone is not enough for real monitoring. A box around a person or car in a single frame tells you nothing about how they move or how long they stay, and plain frame-by-frame boxes flicker and swap identities in a crowd. To count and follow objects you need a stable identity on each one that survives across frames, even when the scene is dense and objects briefly overlap.

Detect people and vehicles reliably in crowded, real-world video and keep a consistent identity on each one across frames, so the footage can drive counting, flow analysis and monitoring for security, surveillance and traffic use cases.

YOLOv8 runs detection on every frame, returning class and a confidence-scored box for each person, vehicle and other object of interest. ByteTrack then associates detections across frames (matching both high and low confidence boxes) to assign and maintain a persistent track ID per object, recovering identities through brief occlusions and crowding. The result is drawn back onto the video as live overlays: class, confidence and track ID per object.

1. 1A video stream or file is decoded frame by frame
2. 2YOLOv8 detects people, vehicles and other classes, each with a confidence-scored box
3. 3ByteTrack associates detections across frames, matching high and low confidence boxes
4. 4Each object is assigned a persistent track ID, held through brief occlusion and crowding
5. 5Class, confidence and track ID are rendered back onto the frame as live overlays

• YOLOv8 object detection (Ultralytics), multi-class
• ByteTrack multi-object tracking with persistent IDs
• Confidence-aware association through occlusion and crowding
• Person and vehicle detection tuned for dense scenes
• OpenCV video decode and on-frame overlay rendering

• Detection and tracking in one real-time pipeline
• Persistent IDs per person and per vehicle across frames
• Multi-class: people, cars and other objects together
• Robust in dense day traffic and crowded night scenes
• Runs on ordinary street and CCTV footage
• Live overlays: class, confidence and track ID

A Python service decodes video frame by frame and runs YOLOv8 for detection, then passes the per-frame boxes to ByteTrack, which maintains the set of active tracks and their IDs over time. Tracked objects (class, box, confidence, ID) are drawn back onto the frame for live monitoring and can be emitted as structured events for downstream counting and analytics. The pipeline is designed to run on a single GPU box and stream continuously.

Argus is delivered as a streaming vision service: it produces annotated video with per-object class, confidence and track ID overlays, plus a structured stream of tracked objects that can feed counting, zone and flow analytics for security, surveillance and traffic dashboards.

• YOLOv8 / Ultralytics (detection)
• ByteTrack (multi-object tracking)
• OpenCV (video decode + overlay)
• Python service, streamable to RTSP / file inputs

The pipeline runs as a Python service on a GPU-equipped server, consuming live streams or recorded video and emitting annotated frames and tracked-object events. It can be deployed on-prem alongside existing camera infrastructure, which suits security and surveillance requirements.