Automatic Feature Detection in 3D Human Body Scans

Abstract

Human body scanners generate meshes, consisting of over 100,000 points and triangles, defining a human shape model. The underlying anthropometric landmarks are not scanned, but necessary for many applications. In the CAESAR database these anthropometric landmarks have been premarked by attaching small markers to the human body. The positions of these anthropometric landmarks have been extracted semi-automatically and are available as part of the CAESAR data. Attaching markers to humans is time consuming and is therefore often omitted in other surveys. Hence, there is a need for methods for fully automated extraction of landmarks from human body scans. We investigated three fully automatic detection methods for landmark extraction. The first method uses a function that is fitted onto the region around the marker of interest. The second method analyzes the curvature in the area around the marker. The third and last method is template matching, which has been successfully applied in other fields. The sellion and the four malleoli were selected for evaluation. The latter landmarks are important to locate the ankle joint. For the malleoli the template matching method was best (mean deviation 15 ± 3 mm) followed by curvature calculation (15 ± 3 mm) and function fitting (28 ± 10 mm). For all methods the deviation for the sellion was considerable larger. Template matching proves to be the most consistent and furthermore has the potential of being refined.

A Survey of 3D Face Recognition Methods

“…Suikerbuik [11] proposed to use Gaussian curvatures to find 5 landmarks in a 3D model. He could find the correct landmark point with a maximal error of 4 mm .…”

Automatic measurement of scanned human body in fixed posture

“…Suikerbuik [8] investigated three methods for landmarks extraction. The first uses a function fitted on to the region around the marker of interest.…”

Virtual Human Imaging

“…Some algorithms only work within small bounding boxes that do not deliver an acceptable performance. For example, if a feature detection algorithm takes one hour to process, then it is not useful for a security screening system [31,32]. In this project, we want to develop a model that is invariant to poses and coordinates.…”

“…From a computer vision point of view, detecting features from 3D body scan data is nontrivial because human bodies are diverse. The technical methodology of function fitting has been used for extracting special landmarks, such as ankle joints, from 3D body scan data [31,32], similar to the method for extracting landmarks on terrain [21,22]. Curvature calculation is also introduced from other fields such as the sequence dependent curvature structure of DNA [19,20].…”