Application of a Robust and Efficient ICP Algorithm for Fitting a Deformable 3D Human Torso Model to Noisy Data

We investigate the use of an iterative closest point (ICP) algorithm in the alignment of a point distribution model (PDM) of 3D human female torsos to sample female torso data. An approximate k-d tree procedure for efficient ICP is tested to assess whether it improves the speed of the alignment process. The use of different error norms, namely L₂ and L₁, are compared to ascertain if either offers an advantage in terms of convergence and in the quality of the final fit when the sample data is clean, noisy or has some data missing. It is found that the performance of the ICP algorithm used is improved in both speed of convergence and accuracy of fit through the combined use of an approximate and exact k-d tree search procedure and with the minimisation of the L₁ norm even when up to 50% of the data is noisy or up to 25% is missing. We demonstrate the use of this algorithm in providing, via a fitted torso PDM, smooth surfaces for noisy torso data and valid data points for torsos with missing data.