Object Detection and Tracking

YOLOv8-based object detection, multi-object tracking, trajectory analysis, and movement visualization.

Open in Colab

Overview

This project demonstrates an end-to-end computer vision workflow for detecting and tracking objects in video or image-sequence data. It uses Ultralytics YOLOv8 with ByteTrack to assign stable object IDs across frames, generate annotated tracking output, analyze object trajectories, and visualize movement patterns with heatmaps and distribution plots.

The workflow is implemented in Object_detection.ipynb, making it easy to run locally in Jupyter or on Google Colab with GPU acceleration.

Features

• Object detection with YOLOv8 models
• Multi-object tracking with ByteTrack
• Persistent tracking IDs across frames
• Annotated output video generation
• Object trajectory collection and plotting
• Movement heatmap visualization
• Horizontal and vertical movement distribution analysis
• Basic quantitative tracking metrics, including unique object count and average trajectory length
• GPU support through PyTorch and Ultralytics

Project Structure

Object-Detection/
├── Object_detection.ipynb      # Main notebook for detection, tracking, and analysis
├── requirements.txt            # Python dependencies
├── videos/
│   └── test.mp4                # Sample input video
├── output/
│   └── tracked_output.mp4      # Sample tracked output video
├── trajectory_frame.jpg        # Example trajectory frame
├── yolov8m.pt                  # YOLOv8 medium model weights
├── yolov8n.pt                  # YOLOv8 nano model weights
├── LICENSE                     # MIT license
└── README.md

Tech Stack

Category	Tools
Detection and tracking	Ultralytics YOLOv8, ByteTrack
Computer vision	OpenCV
ML runtime	PyTorch, Torchvision
Data analysis	NumPy, Pandas
Visualization	Matplotlib
Notebook runtime	Jupyter / Google Colab

Getting Started

Prerequisites

• Python 3.10 or newer
• Jupyter Notebook or JupyterLab
• A CUDA-capable GPU is recommended for faster processing, but CPU execution is supported

Installation

Clone the repository:

git clone https://github.com/Shashwat-19/Object-Detection.git
cd Object-Detection

Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate

Install dependencies:

pip install -r requirements.txt

Launch Jupyter:

jupyter notebook Object_detection.ipynb

Usage

Run the Notebook Locally

1. Open Object_detection.ipynb.

2. Choose the model weights:

   model = YOLO("yolov8m.pt")

   Use yolov8n.pt for faster inference or yolov8m.pt for stronger detection quality.

3. Set the input source:

   VIDEO_PATH = "videos/test.mp4"

   The notebook can also process an image folder, such as a MOT dataset frame directory.

4. Run tracking:

   results = model.track(
       source=VIDEO_PATH,
       tracker="bytetrack.yaml",
       save=True,
       show=False,
       conf=0.3,
       persist=True
   )

5. Generate and inspect the tracked output video and trajectory visualizations.

Run on Google Colab

The notebook includes Colab-oriented cells for mounting Google Drive and cloning this repository. To run on Colab:

1. Open the notebook using the Colab badge above.
2. Enable GPU from Runtime > Change runtime type > GPU.
3. Mount Google Drive if your dataset is stored there.
4. Update VIDEO_PATH to point to your video or frame directory.
5. Run all notebook cells.

Outputs

The workflow can produce:

• Annotated tracking frames from YOLOv8 and ByteTrack
• MP4 tracking output, saved as tracked_output.mp4 or under output/
• Object ID logs per frame
• Trajectory history for each tracked object
• Movement heatmap
• Horizontal and vertical movement distribution plots
• Tracking metrics such as unique objects tracked and average trajectory length

Example Results

Preview	Description
output/tracked_output.mp4	Sample tracked output video
trajectory_frame.jpg	Example frame used for trajectory visualization

Configuration

Common values to tune in the notebook:

Parameter	Description	Default
VIDEO_PATH	Input video or image-frame directory	Colab Drive path / sample video
tracker	Tracking algorithm configuration	bytetrack.yaml
conf	Detection confidence threshold	0.3
persist	Reuse object IDs across frames	True
device	Inference device, such as 0 for GPU or cpu	0 in Colab cells
max_trajectories_to_show	Number of trajectories plotted	10

Notes

• The included .pt files are YOLOv8 model weights. If they are removed, Ultralytics can download supported weights automatically when requested.
• For local CPU-only runs, replace device=0 with device="cpu" or remove the device argument.
• The Colab notebook references /content/drive/MyDrive/MOT17-02-FRCNN/img1 as an example dataset path. Update it to match your own dataset location.
• Large video outputs and model weights can make the repository heavy. Consider using Git LFS for future model and media assets.

Documentation

Comprehensive documentation for this project is available on Hashnode.

At present, this README serves as the primary source of documentation.

License

This project is distributed under the MIT License.
For detailed licensing information, please refer to the LICENSE file included in this repository.

Contact

Shashwat

Machine Learning Engineer | Scalable AI Systems

🔹 ML systems: (CV, NLP) + data pipelines
🔹 End-to-end: training → deployment
🔹 Backend & Cloud: Python, Flask, Node.js, Docker, AWS
🔹 Projects: Traffic AI, Video Summarizer, AI Assistants

---

Open Source | Tech Innovation

Building robust applications and leveraging cloud technologies for high-performance solutions.

---

Find me here:

Feel free to connect for tech collaborations, open-source contributions, or brainstorming innovative solutions!