Human pose modelling and body tracking from monocular video sequences

Abstract

This paper proposes a computer vision-based approach to automatically detect human body parts and estimate the human body poses efficiently from a monocular video sequence. Human body parts detection is performed using colour, contours and silhouettes cues. It determines the 2D spatial locations of joints of a human body without any special markers on the body. The input image is segmented using silhouette extraction function to obtain a silhouette human figure. A parametric skin distribution modelling method is then utilized to detect the face and limbs of a person. Hue-Saturation-Value (HSV) colour space is chosen for this application. Radon transform is used to get more accurate orientation of the upper arms when they are inside the body perimeter. Various physical and motion constraints regarding the human body is then used to construct the upper body configuration. Our algorithm can estimate poses for the person wearing short sleeve and long sleeve shirt. It could estimate the human poses even under illumination changes, self-occlusion occurrence, and distance variations. Then, seventeen body poses are considered for classification. Eight features are extracted from each pose. Feed-forward neural network is used to classify the defined human body poses.