Efficient parallel implementation of active appearance model fitting algorithm on GPU.

Area of Science:

• Computer Vision
• Parallel Computing
• Image Processing

Background:

• Active Appearance Models (AAM) are powerful for object detection and tracking.
• High-dimensional texture representation in AAM leads to computationally intensive processes, hindering real-time applications.
• Modern Graphics Processing Units (GPUs) offer parallel processing capabilities to address computational bottlenecks.

Purpose of the Study:

• To develop an efficient parallel implementation of the AAM fitting algorithm for GPUs.
• To overcome the computational challenges associated with high-dimensional texture data in AAM.
• To achieve real-time performance for AAM-based tracking systems.

Main Methods:

• Proposed a fine-grained parallelism strategy, distributing AAM texture data to thousands of GPU threads.
• Implemented the parallel AAM fitting algorithm using CUDA on an Nvidia GTX 650 GPU (Kepler architecture).
• Evaluated performance using sixteen face AAM models with varying texture dimensionality.

Main Results:

• The parallel AAM fitting algorithm demonstrates significant speed improvements.
• Achieved real-time performance for video processing, even with very high-dimensional textures.
• The fine-grained parallelism effectively leverages the GPU architecture for enhanced computation.

Conclusions:

• The proposed GPU-based parallel AAM fitting algorithm successfully addresses computational limitations.
• Enables real-time object detection and tracking applications previously constrained by processing time.
• Highlights the potential of GPU computing for accelerating complex computer vision algorithms.