next up previous contents
Next: Sensor Inversion Up: Previous Work Previous: Kalman Filtering

Feature-Based Methods

Localisation methods which operate on the basis of geometric reasoning rely on the reliable extraction and recognition of features from sensor data. Image-based methods extract features based on edge formations, such as corners or straight lines [64, 33, 54, 31], or perform segmentation on the basis of intensity or colour [59], while sonar-based methods attempt to link sonar points into lines and structures [36, 40]. Feature extraction and correspondence is plagued by the inherent noise of almost all sensors, which often leads to instability in the extracted features. For most feature-based methods, the choice of which features to employ is often sensor dependent and constrained to a particular application domain. For example, Sugihara, Krotkov, Yagi et al. and Kosaka and Kak all present methods which depend on the reliable extraction of vertical lines in an image. Vertical lines are chosen under the assumption that if the camera pose is constrained such that it is vertically aligned at all times, vertical structures in the environment will always appear as vertical lines in the image [33, 54, 62, 63, 31]. These assumptions break down easily if a camera is poorly mounted, the terrain is rough, or there is a paucity of fixed vertical structures in the environment, such as in outdoor scenes.

Feature-based methods are concerned primarily with optimising feature correspondence, and are susceptible to local minima in the functional to be optimised, especially when employed with large-scale maps. Furthermore, these methods often rely on an accurate a priori map which is usually obtained from architectural drawings, or by manual measurement, which can fail to account for the presence of furnishings such as desks or chairs, or the issues of the dynamics of human and robot interaction with the environment.

A popular alternative to extracting naturally occurring features from sensor data is to employ artificial landmarks, that is, features which are not natural to a particular environment, but which are inserted, affixed, or otherwise deployed on the basis that they can be more robustly detected and extracted by a sensor. Artificial landmarks benefit from the ability to easily extract parameters based on a priori knowledge of landmark geometry, or through explicit labelling, such as bar codes or ultrasonic beacons [37, 22]. The use of artificial landmarks can greatly simplify the problem of position estimation but there are significant drawbacks in the facts that they require prior (and often human) intervention, and can impose other costly or impractical requirements on the environment.


next up previous contents
Next: Sensor Inversion Up: Previous Work Previous: Kalman Filtering

Robert Sim
Tue Jul 21 10:30:54 EDT 1998