Multi-Object Tracking with YOLOv8 and ByteTrack

An end-to-end computer vision project for object detection, multi-object tracking, and motion visualization using YOLOv8 and ByteTrack.

This repository demonstrates how to build a clean, local inference pipeline that goes beyond object detection by adding persistent identities, motion understanding, and trajectory visualization.

---

Project Overview

This project implements a complete multi-object tracking (MOT) pipeline:

1. Fine-tune a YOLOv8 detector on a video-based dataset
2. Perform online multi-object tracking using ByteTrack
3. Visualize object trajectories and motion dynamics
4. Export structured tracking outputs for downstream analysis

The focus is on engineering clarity, reproducibility, and interpretability, rather than UI or cloud deployment.

---

Features

Object Detection

• Fine-tuned YOLOv8 (Ultralytics)
• Anchor-free, single-stage detector
• Trained on MOT17 (person class)

Multi-Object Tracking

• ByteTrack for online tracking
• Persistent object IDs across frames
• Robust to occlusions and missed detections

Motion Visualization

• Per-object trajectories with fading history
• Velocity vectors showing direction and relative speed
• Stable ID visualization across frames

Outputs

• Annotated video with boxes, IDs, trajectories, and velocity arrows
• Frame-level CSV containing tracking data

---

Design Philosophy

This project is designed to reflect real-world computer vision systems, not just model training:

• Clear separation of data, models, scripts, and outputs
• Local inference (no cloud or UI dependencies)
• Modular code that can be extended easily
• Visualizations focused on interpretability

---

Repository Structure

.
├── convert_mot17_to_yolo.py
├── data
│   ├── dataset.yaml
│   ├── images
│   └── labels
├── models
├── outputs
│   ├── tracked_trajectory_velocity.mp4
│   └── tracks.csv
├── README.md
├── requirements.txt
├── runs
│   └── detect
├── track_mot17.py
├── train_yolov8.py
├── yolo11n.pt
└── yolov8n.pt

7 directories, 10 files

Training artifacts under runs/ are generated automatically by Ultralytics and are not part of the final model deliverables.

Dataset

• MOT17 (Multiple Object Tracking Benchmark)
• Person class only
• Sequence-level train/validation split

The dataset is not included in this repository.
Please download MOT17 separately and update paths in the scripts.

Data Preparation

Convert MOT17 annotations to YOLO format:

python scripts/prepare_mot17_for_yolov8.py

This step:

• Converts MOT annotations to YOLO format
• Ensures bounding boxes are valid and normalized
• Produces a YOLOv8-compatible dataset layout

Training YOLOv8

Fine-tune a pretrained YOLOv8 model:

python scripts/train_yolov8.py

Tracking and Visualization

Run detection + tracking with motion visualization:

python scripts/track_mot17.py

This generates:

• Annotated video with:
  • Bounding boxes
  • Persistent IDs
  • Fading trajectories
  • Velocity vectors
• Frame-level CSV output with object positions

Technical Highlights

• YOLOv8 anchor-free detection
• ByteTrack motion-based association
• Trajectory history visualization
• Velocity estimation from motion
• Clean, modular Python implementation