Abstract

A variety of practical algorithms can be used to extract a set of feature points associated with objects containing in image data and an important task is to detect shape classes of interest in such point cloud data. This problem is challenging because the data is typically noisy, cluttered, and unordered. In addition, once a shape is detected in a point cloud, we want to reconstruct the shape from the cloud.


Project Description


Challenges

What makes the problem of shape detection in point clouds difficult? Here are some of the issues:
1. Unknown Pose and Scale: A shape can be present in arbitrary pose and scale as shown in the top row of Figure 1 and one does not know these prior information.
2. Shape Variability: Even in a same shape class, there is also shape variability.
3. Noise and Clutter: Furthermore, there are observation noise and clutter in real point clouds as shown in the bottom row of Figure 1.


Fundamental Issue

The comparison of unlabeled point patterns is a difficult problem and most recent algorithms suffer difficulties with multi-modality, including Rangarajan, Chui and Bookstein (1997); Walker (1999); Taylor, Mardia and Kent (2003); Kent, Mardia and Taylor (2004); Green and Mardia (2006); Dryden, Hirst and Melville (2007). The key is to find global solutions for unknown variables, especially the pose and scale. We would like to build a statistical model and find global solutions for unknown variables efficiently.

Our Approach

We present a fully statistical framework for detecting pre-determined shape classes in point clouds. An important goal is to provide likelihood, and thus a confidence, of find a shape class in a given data. We develop a model-based approach. The data is modeled using Poisson process on the object’s boundary, corrupted by an additive noise and a clutter process. Using analytical likelihood functions dictated by the model, we develop a generalized likelihood ratio test for detecting a shape class. This ratio test is based on optimizing pose, scale, and shape variability associated with hypothesized objects. To guarantee the global optima, we have used grid search on some variables, i.e., scale and noise variance.

Related Publications

J. Su, Z. Zhu, A. Srivastava and F. Huffer. Detection of shapes in 2D point clouds generated from images. ICPR, pp. 2640-2643, August 2010.