Action Recognition in Videos

This script identifies the start and end of a round in boxing video footage.

Key Features:

• Video Processing: Load a video and process it frame by frame.

• Action Recognition: Utilizes the R(2+1)D-18 model, a 3D ResNet model pretrained on the Kinetics-400 dataset.

• Frame Transformation: Applies image transformations (resize, center crop, and normalization) using the albumentations library.

• Output Video: Generates an output video with frames labeled with the detected action.

• 

Requirements:

• Python 3.x
• torch
• torchvision
• cv2 (OpenCV)
• numpy
• tqdm
• albumentations

Usage:

1. Setup:

   pip install torch torchvision opencv-python-headless numpy tqdm albumentations

Directory Structure:

• input/: This directory should contain the video files you want to process.
• outputs/: The processed videos will be saved in this directory with the action label superimposed on the frames.

2. Run the Script:
   • By default, the script processes the video "input/video.mp4".

   python action_recognition.py
   `
   

3. Output:
   • The processed video will be saved in the "outputs" directory with the action label superimposed on the frames.

Customization:

• To process a different video, change the input_video variable in the main function to the desired video path.

Future Work:

• Enhance the script to provide timestamps indicating intervals where no boxing activity is detected for a specified duration.