MARCOnI-ConvNet-Based MARker-Less Motion Capture in Outdoor and Indoor Scenes

Area of Science:

• Computer Vision
• Human Motion Analysis
• Machine Learning

Background:

• Marker-less motion capture struggles with low camera counts and general outdoor scenes.
• Existing methods often fail to reliably track articulated human body motion in complex environments.

Purpose of the Study:

• To develop an accurate marker-less motion capture method for articulated skeleton motion.
• To enable tracking of multiple subjects in general indoor and outdoor scenes using as few as two cameras.

Main Methods:

• Combines discriminative Convolutional Network (ConvNet)-based joint detection with model-based generative motion tracking.
• Utilizes a unified pose optimization energy incorporating unary potentials from ConvNet and appearance-based similarity.
• Employs iterative local optimization for efficient pose computation with analytic derivatives.

Main Results:

• Achieves state-of-the-art accuracy and temporal stability in articulated joint angle tracking with minimal cameras.
• Demonstrates superior performance over related work on various indoor and outdoor datasets.
• Introduces MPI-MARCOnI, a new, extensive dataset for marker-less motion capture evaluation.

Conclusions:

• The proposed method offers efficient and accurate marker-less motion capture for articulated human skeletons.
• The approach is suitable for general scenes, low camera counts, and parallel hardware implementation.
• The new dataset facilitates further research and evaluation in the field.