Automated Body Parts Estimation and Detection using Salient Maps and Gaussian Matrix Model
Ayesha Arif, Ahmad Jalal
Abstract
Estimation and detection different human body Portion from different scenes of videos and images is an important step for most model based systems. Human body-portion detection from a single still image estimate the layout of human body portion, the position of body portion (head, torso, arms, and legs), size and orientation within the scene to recognize the action. For the foreground segmentation technique; we have used salient-object detection via Structured Matrix Decomposition (SMD) and skin-tone detection. After extraction of silhouette; body-portion estimation is applied by using Gaussian Mixture Model (GMM). Five basic parts are determined by using classical expectation maximization (EM) algorithm. The minimum number of twelve ellipsoids represented on the image with the centroid of each ellipse. The estimated distance between the centroids of ellipses are compared. The experimental results over dataset as PAMI09_release has accuracies of 86.2%, respectively. Our proposed pose descriptors outperform other state-of-the-art body portion detection model.