Feature selection for position estimation using an omnidirectional camera

a r t i c l e i n f o This paper considers visual feature selection to implement position estimation using an omnidirectional camera. The localization is based on a maximum likelihood estimation (MLE) with a map from optimally selected visual features using Gaussian process (GP) regression. In particular, the collection of selected features over a surveillance region is modeled by a multivariate GP with unknown hyperparameters. The hyperparameters are identified through the learning process by an MLE, which are used for prediction in an empirical Bayes fashion. To select features, we apply a backward sequential elimination technique in order to improve the quality of the position estimation with compressed features for efficient localization. The excellent results of the proposed algorithm are illustrated by the experimental studies with different visual features under both indoor and outdoor real-world scenarios. Minimizing levels of location uncertainties in sensor networks or robotic sensors is important for regression problems, e.g., prediction of environmental fields [1,2]. Localization of a robot relative to its environment using vision information (i.e., appearance-based localization) has received extensive attention over the past few decades from the robotic and computer vision communities [3–5]. Vision-based robot positioning may involve two steps. The first step involves learning some properties of vision data (features) with respect to the spatial position where observation is made, so-called mapping. The second step is to find the best match for the new spatial position corresponding to the newly observed features, so-called matching. The mapping from these visual features to the domain of the associated spatial position is highly nonlinear and sensitive to the type of selected features. In most cases, it is very difficult to derive the map analytically. The features shall vary as much as possible over the spatial domain while varying as small as possible for a given position over the disturbance. For example, they should be insensitive to changes in illumination and partial obstruction. Motivated by the aforementioned situations, we consider the problem of selecting features from the original feature set in order to improve the localization performance of a robot. The central assumption when using a feature selection technique is that the original feature set contains many redundant or irrelevant features. To facilitate further discussion, let us consider a configuration where the input vector X is the robot position and the output feature vector Y is the collection of extracted features from the vision …