Self-location is a critical component of unmanned ground vehicle operation. While GPS has met this need for many applications, there are others, indoors and outdoors, where GPS is not available or reliable. This paper discusses the investigation of a method to self-locate using an onboard 3D range sensor, when a 3D map of the environment is available. We have examined the performance in terms of what is needed for data coverage and terrain types. Localization is accomplished through surface fitting. Surface fitting is an important component of Simultaneous Localization and Mapping (SLAM) for 3D terrains. Three environments were chosen for testing, including a hilly desert scene, an indoor corridor scene, and a forest scene. A simulated 3D range sensor was used to scan models corresponding to these three environments. Various parameters were modified to identify strong and weak points in the scheme.

The need to register data is abundant in applications such as: world modeling, part inspection and manufacturing, object recognition, pose estimation, robotic navigation, and reverse engineering. Registration occurs by aligning the regions that are common to multiple images. The largest difficulty in performing this registration is dealing with outliers and local minima while remaining efficient. A commonly used technique, iterative closest point, is efficient but is unable to deal with outliers or avoid local minima. Another commonly used optimization algorithm, simulated annealing, is effective at dealing with local minima but is very slow. Therefore, the algorithm developed in this paper is a hybrid algorithm that combines the speed of iterative closest point with the robustness of simulated annealing. Additionally, a robust error function is incorporated to deal with outliers. This algorithm is incorporated into a complete modeling system that inputs two sets of range data, registers the sets, and outputs a composite model.

A three-dimensional world model is crucial for many robotic tasks. Modeling techniques tend to be either fully manual or autonomous. Manual methods are extremely time consuming but also highly accurate and flexible. Autonomous techniques are fast but inflexible and, with real-world data, often inaccurate. The method presented in this paper combines the two, yielding a highly efficient, flexible, and accurate mapping tool. The segmentation and modeling algorithms that compose the method are specifically designed for industrial environments, and are described in detail. A mapping system based on these algorithms has been designed. It enables a human supervisor to quickly construct a fully defined world model from unfiltered and unsegmented real-world range imagery. Examples of how industrial scenes are modeled with the mapping system are provided.

The ability to successfully use and interact with a computerized world model is dependent on the ability to create an accurate world model. The goal of this project was to develop a prototype system to remotely deploy sensors into a workspace, collect surface information, and rapidly build an accurate world model of that workspace. A key consideration was that the workspace areas are typically hazardous environments, where it is difficult or impossible for humans to enter. Therefore, the system needed to be fully remote, with no external connections. To accomplish this goal, an electric, mobile platform with battery power sufficient for both the platform and sensor electronics was procured and 3D range sensors were deployed on the platform to capture surface data within the workspace. A radio Ethernet connection was used to provide communications to the vehicle and all on-board electronics. Video from on-board cameras was also transmitted to the base station and used to teleoperate the vehicle. Range data generated by the on-board 3D sensors was transformed into surface maps, or models. Registering the sensor location to a consistent reference frame as the platform moved through the workspace allowed construction of a detailed 3D world model of the extended workspace.

World modeling is defined as the process of creating a numerical geometric model of a real world environment or workspace. This model is often used in robotics to plan robot motions which perform a task while avoiding obstacles. In many applications where the world model does not exist ahead of time, structured lighting, laser range finders, and even acoustical sensors have been used to create three dimensional maps of the environment. These maps consist of thousands of range points which are difficult to handle and interpret. This paper presents a least squares technique for fitting range data to planar and quadric surfaces, including cylinders and ellipsoids. Once fit to these primitive surfaces, the amount of data associated with a surface is greatly reduced up to three orders of magnitude, thus allowing for more rapid handling and analysis of world data.