Building three-dimensional visual maps of interior space with a new hierarchical sensor-fusion architecture

Hyukseong Kwon, Purdue University

Abstract

It is now generally recognized that sensor-fusion is a desirable approach to the accurate construction of environment maps by a sensor-equipped mobile robot. Typically, range data collected with a range sensor is combined with the reflectance data obtained from one or more cameras mounted on the robot. Sensor fusion for map building is made challenging by the need to build maps hierarchically. That is, the low level structures extracted from the sensed data collected at any single position of the robot must be merged into higher level structures when information is combined from the different positions of the robot. Many of the previous approaches to sensor fusion for accurate map building confine data fusion to the lowest-level data abstractions in their processing architectures. What that implies is that only the fused data is made available to the processing steps that are designed to deal with larger data abstractions. This makes it impossible to correct the data collected by a sensor by bringing to bear on it any top-down constraints. One example of such top-down constraints would be the continuity of, say, the floor boundary edges in a hallway across multiple positions of the robot. Prematurely early fusion of the data from the different sensors also makes it difficult to remove the outliers that can be rejected or corrected by using a more reliable sensor to become a source of continuity constraints on the data generated by a less reliable sensor. Our proposed new hierarchical approach to map building eliminates these difficulties. Our approach simultaneously fuses together and keeps separate the data abstractions extracted from the outputs of the different sensors. This allows the different sensors to interact at different abstraction levels, the result being a more accurate final global map. This dissertation includes an experimental proof of the proposed hierarchical architecture. We will show the robot constructing high-quality 3D visual maps of a hallway system at a fairly rapid rate.

Degree

Ph.D.

Advisors

Kak, Purdue University.

Subject Area

Robotics

COinS