Head and Body Orientation Estimation Using Convolutional Random Projection Forests.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Estimating human pose from low-resolution images is challenging due to difficulties in extracting facial features and body parts.
• Existing methods struggle with noise, occlusion, and motion blur in low-resolution imagery.

Purpose of the Study:

• To propose an efficient and accurate algorithm for estimating head pose and body orientation from low-resolution images.
• To address the limitations of current methods in handling degraded image quality.

Main Methods:

• Developed a Convolutional Random Projection Forest (CRPforest) algorithm.
• Utilized a Convolutional Random Projection Network (CRPnet) at each forest node to map images to a high-dimensional feature space.
• Employed a rich filter bank designed for sparse responses, enabling efficient computation via compressive sensing.
• Used sparse random projection matrices to retain essential information with fewer computations.

Main Results:

• The CRPforest algorithm demonstrates high accuracy, achieving over 98% on the HIIT dataset.
• The method is computationally efficient, requiring minimal convolutions and processing images without a GPU.
• Achieved a slight accuracy trade-off (less than 2%) for significant speed improvements.

Conclusions:

• The proposed CRPforest algorithm offers a favorable performance compared to state-of-the-art methods for pose estimation in challenging low-resolution images.
• The algorithm is robust to noise, occlusion, and motion blur.
• The method provides an efficient solution for real-time applications where GPU resources are limited.