Ghost: A Human Body Part Labeling System Using Silhouettes Ismail Haritaoglu, David Harwood and Larry S. Davis Computer Vision Laboratory, University of Maryland, USA hismail,harwood,lsd@umiacs.umd.edu 指導教授 張元翔 報告人員 陳昱辰
Introduction Ghost is a real time system for estimating human body posture and detecting body parts in monochromatic imagery. we consider four main postures: standing, sitting, crawling-bending, and laying down
Algorithm A: The body posture which yields the highest similarity measure is taken as the estimated posture. B: A recursive convex-hull algorithm is applied to find possible body part locations on the silhouette boundary.
Algorithm C: The location of the head is predicted using the major axis of the silhouette D: Map the remaining hull vertices to the body parts using a topological-order preserved distance transform calculation.
Detection of Human Body Parts Segmented from the background by a four stage process: thresholding, noise cleaning, morphological filtering and object detection.
2D body modeling using silhouettes 6 primary body parts (head, hands, feet, and torso) 10 secondary parts( elbows, knees, shoulders, armpits, hip, and upper back)
Examples of the order of the body parts on the silhouette boundary
Estimation of the human body posture We observed that four different main postures have large differences in the order of body parts. A body posture is represented by the normalized horizontal and vertical projection histograms
Vertical and horizontal The vertical and horizontal normalized projections of standing, crawling-bending and laying down postures used in the body posture estimation
Front left-side right-side
Detection of the convex and concave hulls on silhouettes We need to find those points on the silhouette in order to find the set of locations which will be labeled as body parts.
Prediction of the head location The head is a stable body part compared to the others, and its location can be easily predicted.
Prediction of the feet and hands and torso locations Relative path distances are used to check if the mapping is consistent with the order.
Conclusion In current implementation, shadows create problems in locating body parts which are too close to the ground. Better results could be obtained if shadows are removed from the silhouette