CUDA Real-time Video Processor

A GPU-accelerated application for real-time video processing, enhancement, and analysis using CUDA.

Overview

This capstone project demonstrates the power of GPU computing for real-time video processing. The application leverages CUDA to perform computationally intensive operations on video frames in parallel, achieving significant performance improvements compared to CPU-based processing.

Key features:

• Multiple GPU-accelerated image filters (blur, sharpen, edge detection, emboss, sepia, etc.)
• Real-time video processing from files or camera input
• Advanced visual effects like thermal vision and night vision simulation
• Motion detection between consecutive frames
• Optical flow visualization
• Simple object detection
• Batch processing for higher throughput
• Performance benchmarking tools

Requirements

• NVIDIA CUDA Toolkit (11.0+)
• OpenCV 4.x
• C++17 compatible compiler
• CMake 3.10+

Installation

Option 1: Building from Source

1. Clone this repository:

   git clone <repository-url>
   cd GPU-Specialisation-Capstone

2. Create a build directory and compile:

   ./build.sh

   Or manually:

   mkdir -p build
   cd build
   cmake ..
   make -j$(nproc)

Option 2: Using Docker

A Dockerfile is provided for containerized development and execution:

docker build -t cuda-video-processor .
docker run --gpus all -it --rm cuda-video-processor

Usage

The application supports various command-line arguments:

./video_processor [options]

Options

• --input <source>: Input source (video file path or camera index)
• --output <filename>: Output video file (optional)
• --filter <filter_type>: Filter to apply (default: none)
• --transform <transform>: Transform to apply (default: none)
• --intensity <value>: Filter intensity (0.0-1.0, default: 0.5)
• --detect-motion: Enable motion detection
• --optical-flow: Enable optical flow visualization
• --detect-objects: Enable simple object detection
• --benchmark: Run performance benchmark
• --batch-size <size>: Batch processing size (default: 1)
• --help: Display help message

Available Filters

• none: No filter
• blur: Gaussian blur
• sharpen: Sharpen effect
• edge_detect: Edge detection
• emboss: Emboss effect
• sepia: Sepia tone
• grayscale: Grayscale conversion
• negative: Color inversion
• cartoon: Cartoon effect
• sketch: Sketch effect
• night_vision: Night vision effect
• thermal: Thermal vision effect

Available Transformations

• none: No transformation
• rotate_90: Rotate image by 90 degrees
• rotate_180: Rotate image by 180 degrees
• rotate_270: Rotate image by 270 degrees
• flip_h: Flip horizontally
• flip_v: Flip vertically

Examples

Process a video file with a blur filter:

./video_processor --input input_video.mp4 --output output_video.mp4 --filter blur --intensity 0.7

Use camera input with thermal vision effect:

./video_processor --input 0 --filter thermal --intensity 0.8

Enable motion detection with edge detection filter:

./video_processor --input input_video.mp4 --filter edge_detect --detect-motion

Run performance benchmark:

./video_processor --input input_video.mp4 --benchmark

How it Works

CUDA-Accelerated Processing Pipeline

1. Frame Acquisition: Frames are captured from video files or camera.
2. GPU Transfer: Frame data is transferred to the GPU memory.
3. Parallel Processing: Each pixel is processed in parallel using thousands of CUDA threads.
4. Filter Application: Mathematical operations are applied to transform pixel values.
5. Result Transfer: Processed frames are transferred back to CPU memory.
6. Display/Storage: Results are displayed in real-time and/or saved to disk.

CUDA Kernels

The project implements various CUDA kernels for different image processing tasks:

• Convolution Kernel: Applies convolution filters (blur, sharpen, edge detection, etc.)
• Color Transformation Kernels: Applies specialized color effects (sepia, grayscale, etc.)
• Motion Detection Kernel: Computes differences between consecutive frames
• Special Effect Kernels: Implements complex effects like thermal vision and night vision

Performance

The GPU acceleration provides significant performance improvements over CPU-based implementations:

• Processing HD video (1920x1080) in real-time at 30+ FPS
• 10-20x speedup compared to equivalent CPU implementation
• Efficient batch processing for higher throughput

Actual performance depends on GPU specifications, filter complexity, and frame resolution.

Future Improvements

• Advanced object detection and tracking using CUDA-accelerated ML models
• Deep learning-based video enhancement
• Support for multiple GPU devices
• Hardware-accelerated video encoding/decoding
• More complex visual effects and transformations

License

MIT License