Shape estimation and object recognition using spatial probability distributions

Abstract

This thesis describes an algorithm for performing object recognition and shape estimation from sparse sensor data. The algorithm is based on a spatial likelihood map which estimates the probability density for the surface of the object in space. The spatial likelihood map is calculated using an iterative, finite element approach based on a local probabilistic model for the object's surface. This algorithm is particularly useful for problems involving tactile sensor data. An object classification algorithm using the spatial likelihood map was developed and implemented using simulated tactile data. The implementation for the tactile problem was in two dimensions for the sake of clarity and computational speed, and is easily generalized to three dimensions. The spatial likelihood map is also useful for multi-sensor data fusion problems. This is illustrated with an application drawn from the study of mobile robots.