An Optimized Method for 3D Body Scanning Applications Based on KinectFusion

Abstract

KinectFusion is a powerful method for 3D reconstruction of indoor scenes. It uses a Kinect camera and tracks camera motion in real-time by applying ICP method on successive captured depth frames. Then it merges depth frames according their positions into a 3D model. Unfortunately the model accuracy is not sufficient for body scanner applications because the sensor depth noise affects the camera motion tracking and deforms the reconstructed model. In this paper we introduce a modification of the KinectFusion method for specific 3D body scanning applications. Our idea is based on the fact that, most body scanners are designed so that the camera trajectory is a fixed circle in the 3D space. Therefore each camera position can be determined as a rotation angle around a fixed axis (rotation axis) passing through a fixed point (rotation center). Because the rotation axis and the rotation center are always fixed, they can be estimated offline while filtering out depth noise through averaging many depth frames. The rotation angle can be also precisely measured by equipping the scanner motor with an angle sensor.