Skip to content

Latest commit

 

History

History

real-time-daily-activity-recognizer

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
experiment_logs
experiment_logs
 
 
figures
figures
 
 
notebook
notebook
 
 
.gitignore
.gitignore
 
 
README.md
README.md
 
 
requirements.txt
requirements.txt
 
 

README.md

🎥 Real-Time Daily Activity Recognizer

"Your webcam knows what you're doing — let's teach it how."

This project builds a real-time human activity recognition system powered by computer vision and deep learning.
The system uses a webcam and a transfer-learned CNN model to classify everyday activities such as brushing teeth, reading, or walking.

🎯 Objectives

  • Build a live activity recognizer using webcam input and a CNN-based classifier
  • Collect a custom dataset of daily activities using Selenium-based image crawling
  • Train an image classifier using transfer learning (EfficientNetB0, MobileNetV2, etc.)
  • Use OpenCV to perform real-time inference with webcam video feed

📁 Dataset

  • Source: Google Images (via Selenium automation)

  • Classes:

    • brushing_teeth
    • drinking
    • eating
    • typing
    • sleeping
    • reading
    • washing_face
    • walking

  • Images per class: Approx. 350~3000 images each (customizable)

  • Availability:
    The dataset is excluded from the repository to avoid copyright violations.

⚠️ Images are collected solely for educational purposes and stored locally.
You can build your own dataset using notebook/activity-crawling.ipynb.

🛠️ Techniques

  • Convolutional Neural Networks (CNN)
  • Transfer Learning with EfficientNetB0 / MobileNetV2
  • Strong image augmentation using ImageDataGenerator
  • Real-time classification via OpenCV
  • Metrics visualization (accuracy, loss curves)
  • Evaluation with confusion matrix and prediction overlay

📁 Directory Structure

real-time-daily-activity-recognizer/
├── .gitignore                # Excludes checkpoints, temp files, dataset
├── experiment_logs/          # Experiment logs in Markdown
├── figures/                  # Accuracy/loss plots, confusion matrix images
├── images/                   # Crawled images organized by class (not included in repo)
├── notebook/                 # Jupyter Notebooks (training, crawling, inference)
├── README.md                 # Project description and instructions
└── requirements.txt          # Python dependencies

🚀 How to Run

1. **Install Dependencies**  
   pip install -r requirements.txt  

2. **Crawl Your Own Dataset (Optional)**  
   Edit and run: `notebook/activity-crawling.ipynb`  

3. **Train the Model**  
   Run notebook: `notebook/01_train_model.ipynb`  
   Check experiment logs in `experiment_logs/`  

4. **Run Inference (Webcam)**  
   Use: `notebook/03_realtime_inference.ipynb`  
   Or run a standalone script using `cv2.VideoCapture`  
   Press `q` to close the webcam window

🚫 Limitations

  • Dataset is relatively small and crawled from the web (quality and labeling errors may exist)
  • Limited to 8 predefined activity classes
  • Real-time inference may be sensitive to lighting, background, and camera quality
  • Not optimized for deployment on resource-constrained devices

🔭 Next Steps

  • Expand dataset with more balanced and diverse activity classes
  • Fine-tune additional architectures (EfficientNetV2, Vision Transformers)
  • Integrate real-time data augmentation and smoothing for more stable predictions
  • Deploy as a lightweight application (Streamlit / mobile demo)

👤 Author

Maintained by hojjang98
📅 Last updated: September 2025