Action Classifier using CNN and Transfer Learning

Overview

This project implements an action classifier using Convolutional Neural Networks (CNN) enhanced by transfer learning with the VGG16 architecture. It classifies human actions into two categories: Running and Walking. The model is developed using Python, TensorFlow, and Keras, with hyperparameter tuning performed using Keras Tuner.

Key Features

• Dataset Management: Efficient splitting of the dataset into training, validation, and test sets using train_test_split.
• Data Augmentation: Implemented using ImageDataGenerator to increase robustness and reduce overfitting.
• Transfer Learning: Leveraged the pre-trained VGG16 model to extract powerful feature representations.
• Hyperparameter Optimization: Utilized Keras Tuner's RandomSearch for optimal parameter selection.
• Video and Image Prediction: Capable of predicting actions from both images and video streams using OpenCV.

Findings from Training

• Best Validation Accuracy: Achieved 83.59%.
• Optimal Model Parameters:
  • Convolution Blocks: 1
  • Filters: 96
  • Kernel Size: 3x3
  • Dense Units: 128
  • Dropout Rate: 0.2
  • Optimizer: RMSProp

Performance on Test Data

• Accuracy: 86.04%
• Loss: 0.4691

The results demonstrate the effectiveness of using transfer learning and systematic hyperparameter tuning for image-based action classification tasks.

Technologies Used

• Python
• TensorFlow & Keras
• OpenCV
• Keras Tuner
• Google Colab

Usage

• Clone the repository.
• Install required dependencies (tensorflow, keras-tuner, opencv-python, etc.).
• Organize your dataset following the provided structure.
• Train and evaluate your model using the provided scripts.

Future Improvements

• Extend the dataset with additional action classes.
• Explore more complex models and deeper CNN architectures.
• Integrate real-time video stream prediction for live-action classification.