GPU-Accelerated Particle Filtering for 3D Model-Based Visual Tracking

Abstract

<p>Model-based approaches to 3D object tracking and pose estimation that employ a particle filter are effective and robust , but computational complexity limits their efficacy in real-time scenarios. This thesis describes a novel framework for acceleration of particle filtering approaches to 3D model-based, markerless visual tracking in monocular video using a graphics processing unit (GPU). Specifically, NVIDIA compute unified device architecture (CUDA) and Direct3D are employed to harness the single-instruction multiple- thread (SHvIT) programming model used by the GPU's massively parallel streaming multiprocessors (SMs) for simulation (3D model rendering) and evaluation (segmentation, feature extraction, and weight calculation) of hundreds of particles at high speeds. The proposed framework addresses the computational intensity that is intrinsic to all particle filter approaches, including those with modifications and extensions that strive to reduce the number of required particles while maintaining tracking quality.</p> <p>The sampling importance resampling (SIR) particle filter and its utility in 3D model-based tracking is reviewed and a detailed overview of relevant GPU-programming concepts is presented. The proposed framework is formulated as a series of interconnected steps and the functionality and implementation of each is described in det ail. Rigid and articulated tracking examples are presented in the context of human-computer interact ion (HCI) and augmented reality (AR) applications, with a focus on bare hand tracking. Performance and tracking quality results demonstrate markerless, model-based visual t racking on consumer-grade hardware with pixel-level accuracy up to 95 percent at 30+ frames per second. The framework accelerates particle evaluation up to 25 times over a comparable CP -only implementation. providing an increased particle count while maintaining real-time frame rates.</p>