Argus: Real-Time Detection & Tracking

The challenge

Single-frame detection does not show how objects move or how long they stay. Boxes flicker and identities swap in crowds, so stable IDs that survive occlusion are needed for counting and monitoring.

The objective

Reliably detect people and vehicles in crowded real-world video and maintain consistent identities to support counting, flow analysis and monitoring.

Our approach

YOLOv8 detects multiple classes per frame, each with a confidence-scored box. ByteTrack associates detections across frames, matching both high- and low-confidence boxes, to keep persistent track IDs through brief occlusion and crowding. Class, confidence and track ID are drawn back onto the video as live overlays and can be emitted as structured events for counting and analytics.

How it works

1. 01

   Decode a video stream or file frame by frame.

2. 02

   Detect people, vehicles and other classes with YOLOv8.

3. 03

   Associate detections across frames with ByteTrack.

4. 04

   Assign and hold a persistent ID per object.

5. 05

   Render class, confidence and ID onto each frame.

The build

Detection

YOLOv8, Ultralytics, multi-class

Tracking

ByteTrack, persistent IDs, occlusion recovery

Vision

OpenCV, video decode, overlays

Service

Python, GPU inference, RTSP and file inputs

The result