Rsharan622244/human-action-recognition-lrcn

Human Action Recognition using LRCN

Overview

This project is a deep learning-based Human Action Recognition system that classifies actions from video sequences using an LRCN (Long-term Recurrent Convolutional Network) architecture.

The application supports both:

• —Video file uploads
• —YouTube video URLs

and predicts the action being performed in the video.

The model was trained on four action classes and achieved 87% classification accuracy.

Features

• —Upload local video files for prediction
• —Predict actions directly from YouTube videos
• —Real-time video playback inside the application
• —Confidence scores for all classes
• —Streamlit-based interactive web interface
• —CNN + LSTM architecture for spatiotemporal feature learning

Action Classes

The model is trained to classify the following activities:

• —PullUps
• —TaiChi
• —Skijet
• —Billiards

Model Architecture

The project uses an LRCN (Long-term Recurrent Convolutional Network) architecture:

Video Frames → CNN Feature Extraction → Temporal Sequence Modeling (LSTM) → Softmax Classification

Architecture

• —TimeDistributed Conv2D (16 filters)

• —TimeDistributed MaxPooling2D

• —TimeDistributed Dropout

• —TimeDistributed Conv2D (32 filters)

• —TimeDistributed MaxPooling2D

• —TimeDistributed Dropout

• —TimeDistributed Conv2D (64 filters)

• —TimeDistributed MaxPooling2D

• —TimeDistributed Dropout

• —TimeDistributed Conv2D (64 filters)

• —TimeDistributed MaxPooling2D

• —TimeDistributed GlobalAveragePooling2D

• —LSTM (32 Units)

• —Dropout (0.5)

• —Dense (4, Softmax)

Dataset Preprocessing

Each video is processed as follows:

• —Extract 20 frames per video
• —Resize frames to 64 × 64 pixels
• —Normalize pixel values by dividing by 255
• —Create temporal sequences for LSTM processing

Input Shape:

(20, 64, 64, 3)

Performance

Project Structure

Human-Action-Recognition/

├── app.py
├── predict.py
├── LRCN_model_4class_87acc.h5
├── requirements.txt
└── README.md

Installation

Clone the repository:

git clone <repository-url>
cd Human-Action-Recognition

Create a virtual environment:

python -m venv venv

Activate environment:

Windows

venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

Running the Application

Start the Streamlit server:

streamlit run app.py

Open:

http://localhost:8501

Usage

Option 1: Upload Video

1. 1.Upload a supported video file.
2. 2.Wait for preprocessing and inference.
3. 3.View the predicted action and confidence scores.

Option 2: YouTube URL

1. 1.Paste a YouTube video URL.
2. 2.Download and process the video.
3. 3.View the prediction results.

Technologies Used

• —Python
• —TensorFlow / Keras
• —OpenCV
• —NumPy
• —Streamlit
• —yt-dlp

Future Improvements

• —Support for additional action classes
• —Real-time webcam prediction
• —Batch video processing
• —GPU inference optimization
• —Action localization within videos

Author

Rachuri Sharan

AI / Machine Learning Projects Human Action Recognition using LRCN