Real-Time Human & Object Detection System

Overview

A real-time detection pipeline built on YOLOv8 and OpenCV, tuned to balance accuracy and throughput for live video streams.

The problem

Off-the-shelf object detection demos often report accuracy without addressing real-time constraints — a model that’s accurate at 2 FPS isn’t usable for a live system. The goal was a pipeline that held up on both axes at once.

Approach

I built the detection pipeline around YOLOv8, with OpenCV handling video I/O and preprocessing, then iterated on model fine-tuning and preprocessing specifically to push accuracy up without sacrificing frame rate.

Architecture

1. Video capture & preprocessing via OpenCV
2. YOLOv8 inference pipeline
3. Model fine-tuning for accuracy
4. Throughput optimization for real-time playback

Results

Reached 92% mAP accuracy at 30 FPS throughput, with a further 10% accuracy gain from fine-tuning and preprocessing improvements over the initial baseline.